fair trade case study geography a level

Ethical Concerns

Local sourcing - getting food, resources locally (from an area close by)

Sustainability - meeting the needs for now, without compromising the needs for the future.

food miles - the distance from crop to plate that a food takes.

Transition town - towns that aim to reduce their carbon emissions and increase independence.

Fair trade- ensuring that a fair price is paid for goods that are produced so that producers and workers get a fairer share of the money and are better protected (rights. Tends to support small scale, democratic organisations in LICs.

Ethical shopping A deliberate choice of buying goods for ethical reasons considering the ethical and social costs of purchasing them.

Like most HICs, the UK is living well beyond its ‘environmental means’. To supply resources for every country at the UK’s current level of consumption it would take 3.1 Earths. At the current rates of growth and global consumption, researchers believe that we would need two planets by 2050.

What can be done?

Responding locally – transition towns

Some local groups and non-governmental organisations (NGOs) promote local sourcing of goods to increase sustainability (relocalisation).

Totnes in Devon (pop 8000) was the world’s first ‘Transition Town’. Now a global network exists using the internet and social media to spread the idea of ‘transition’. By 2016 transition has become a movement of up to 50 countries attempting to reduce their carbon footprints and increase their independence.

Transition towns promote:

• Reducing consumption by repairing or reusing items

• Reducing waste, pollution and environmental damage

• Meeting local needs through local production where possible (farmers market)

Evaluate Ethical Consumption

Ethical Consumerism:

The UK’s retail sector is becoming increasingly aware of ethical issues associated with shopping. M&S now only sells Fairtrade teas and coffees, plus naturally died fabrics (to reduce carbon emissions). All supermarket chains now display ethical shopping credentials for those who want to buy ethical.

Local produce such as meat and milk have named suppliers on the product. This shows a shift back towards farmers markets e.g. Lidl adverts proving where they source their produce (Mussels, Scottish Beef, and Turkeys).

The issues with ethical shopping:

· Buying organic destroys more forests – less use of fertilisers and pesticides mean that more land is needed to produce the same amount.

· Fairtrade does raise farmers’ incomes but it also increases potential overproduction – causing prices to fall, which leaves farmers no better off.

· Growing cash crops even under Fairtrade can mean some farmers end up not growing enough food to feed themselves and their families.

· Buying local food minimises ‘food miles’ and helps the local economy – but most consumers still use cars to go shopping and as a result more energy is used.

Example: Bristol Pound

In 2012 Bristol introduced the ‘Bristol Pound’ (a community currency) to encourage people to spend locally rather than chain stores. However strategies like this also threaten global economic growth because they reduce the demand for items overseas.

Every £10 spent in local businesses is actually worth £23 for the local economy – through what economist call the ‘Multiplier Effect’ (Employees and suppliers are paid). In contrast to £10 in a chain supermarket is worth only £13 locally because it will return to its head office (e.g. Aldi and Lidl are German).

Some services are coordinated centrally (e.g. transport) so it’s hard to influence them. It’s also been argued that doing transition in a large city e.g. London can be difficult.

Due to TNC exploitation of poorer countries many fair trade companies and campaigners began to appear. They want simple things such as fair pay, better working conditions and improvements in infrastructures. The Fairtrade mark is a guarantee that the product is ethically produced and that a fair price has been paid to by the producers.

In 2015 Ethical Consumer gave coffee retailers a rank of out 20 for using ethical products. The marks received were not generally high.

Degradation

Waste and recycling:

In 2012 the UK generated 200 million tonnes of waste! (That’s around 32 fully loaded container ships). We need to do something about how we dispose of our waste. The UK has a long way to go in order to improve it waste management. Many efficient European countries recycle more than half of their waste!

Actions for the Future

The circular economy is an approach to sustainable development calling for careful management of materials. The ultimate idea is to ‘design out’ waste all together.

Could this concept be used into the future to create a more sustainable world?

welcome to the geography portal!

Become a member and discover where geography can take you.

• Global trade
• Investigating ...

Investigating fairtrade

This lesson highlights the positive impact that buying fairtrade products has on communities in other countries

In this lesson pupils investigate the fairtrade approach to global trade. Pupils learn the geographical terms ‘more developed’ (developed) and ‘less developed’ (developing) countries and how these relate to fairtrade. Pupils learn about the positive impact that people buying fairtrade has on communities of farmers and manufacturers in less developed countries, for example through better working conditions and a fair working wage. In the main activity pupils create a poster ‘Why Pay More?’, outlining the benefits of fairtrade and link different types of fairtrade products such as flowers and jewellery to their source location on a map.

Ideally, this lesson should take place after a Mathematics data handling lesson in the morning, when pupils create a bar chart of the price of a range of fairtrade and non-fairtrade items.

Pupils use the Price Data and Bar chart resource provided (table only, bar chart is only used if pupils do not complete the linked Maths lesson) which displays images of a range of popular items (chocolate, brown sugar, black tea, mango smoothie, cotton wool, gold jewellery, beauty products, cotton, coffee, flowers, Balasport footballs) and their fair and non-fairtrade retail prices.

They record the data on the fair and non-fairtrade price of these products in table format with clear column labels and title before creating a bar chart on graph paper. Pupils conclude that fairtrade products do cost more to buy. Explain that in the afternoon geography lesson we will learn the reasons why pay more for fairtrade.

Linked Maths Lesson

Mathematics: National Curriculum KS2 Ma 4 (Data Handling)  Pupils to select and use handling data skills to discover trends in global trade and identify data needed to solve word problems related to trade.  Ma 4 (2a) Solve problems involving data; b) interpret tables, lists and charts involving data on global trade, construct tables and charts; d) Use the terms ‘mode’ and ‘range’ to describe data sets; e) recognise whether data is discrete or continuous; f) draw conclusions from statistics and graphs) If the pupils do not carry out the linked maths lesson, they use the table and bar chart included in the resources Price Data and Bar chart resource and Price comparison resources to support lesson four.

Key questions

What is fairtrade?

Do fairtrade products cost more to produce and purchase than non-fairtrade products?

Why might fairtrade products cost the consumer more?

Why should we pay more for fairtrade products? What is the benefit?

Additional resources needed

Additional links.

Go to the Balasport website to find out about fairtrade football

Go to the HK jewellery website to find out about fairtrade gold

Go to YouTube to watch a video about roses from Ecuador Fairtrade Association

Recap that trade is global (imports, exports, and global supply chain).

Explain there are huge benefits to global trade, however it needs to be done in a way that benefits the workers in the early stages of the supply chain (farmers, miners etc).

Introduce the terms ‘less developed’ and ‘more developed’ countries on the Investigating Faritrade PPT and link this to stages of the supply chain (primary, secondary, tertiary).  Often the primary stage is in less economically developed countries and the tertiary in the more economically developed).  Show the table of the wealth of different continents on Investigating Faritrade PPT and discuss.

Main Activity

Define the ‘fairtrade’ approach to global trade: “Trade between companies in developed countries and producers in developing countries in which fair prices are paid to the producers”.  Go to YouTube website to play an introductory video on Fairtrade

Ask a pupil to volunteer to read the Fairtrade Foundation statement on the Investigating Faritrade PPT. Go through the benefits to fairtrade on the following slide and discuss as a whole class or in small groups. 

Examine the pie chart of fairtrade products by volume and ask pupils whether they can remember any other products you can buy fairtrade from the homework research task set last week.

Go to the Fairtrade foundation website to see an interactive world map of Fairtrade producers on the Interactive Whiteboard.

Revisit the morning findings in Mathematics and bar charts created (if did optional maths lesson), otherwise look at example bar chart on Investigating Fairtrade PPT/ downloadable resource Price Data and Bar Chart) showing difference in price of fair and non-fairtrade items. Conclude fairtrade items are more expensive to buy. Encourage the pupils to think about why this may be (Fairtrade minimum price, Fairtrade premium).

Pupils create a poster using the template provided Why Pay More? (Enlarge to A3).

They write the reasons why people should pay more for fairtrade products and the positive impact of buying fairtrade products on people in developing countries.  They illustrate their poster with pictures of different fairtrade products (five from the price comparison table- football, chocolate, gold ring, roses, face cream) linking their source location the correct location on the world map.

Success criteria:

Pupils show five products and link these to five places they have located using their atlas.

They list at least three reasons why consumers should pay more for fairtrade products.

Pose the question: which product has the biggest price jump when it is fairtrade?

Pupils present their posters to the class and explain content as a persuasive argument for choosing fairtrade products. 

Pupils offer feedback to one other using the teacher’s preferred peer review technique e.g. two stars and a wish.

Discussion questions: What are the positive impacts on the communities when we purchase fairtrade items? When we are purchasing items in our shops should we try to understand why an item may be cheaper or more expensive? (E.g. consider those who have made the items and their working conditions).

If time permits, teacher can also follow the links below so pupils learn more about specific fairtrade products:

Go to the Balasport website to find out about fairtrade football   Go to the HK jewellery website to find out about fairtrade gold

Ask pupils to bring in a fairtrade product each for a fairtrade tea party next week.

File name Files

Global Trade Lesson 5 Lesson Plan

Global Trade Lesson 5 Lesson Plan (1)

Global Trade Lesson 5 Investigating Fair Trade

Global Trade Lesson 5 Price Data and Pie Chart

Global Trade Lesson 5 Price Data and Pie Chart (1)

Global Trade Lesson 5 Poster Template 'Why Pay More?'

Global Trade Lesson 5 Poster Template 'Why Pay More?' (1)

Global Trade Lesson 5 Price Comparison

This resource has been developed as part of the Rediscovering London's Geography project, funded by the GLA through the London Schools Excellence Fund. It seeks to improve the quality of teaching and learning of geography in London’s schools, in addition to encouraging more pupils to study geography

Back To Top

Case Study: Fair Trade

• Fair Trade is an organisation that sets standards fro trade with LEDC’s
• It guarantees a fair price for the farmer
• This pays for the product and investment in local community development projects
• In return the farmer must farm in an environmentally friendly way and treat their workers fairly too!
• More than 900 products are labelled fair trade
• Retail trade is growing by 40% a year!
• In Costa Rica coffee growers have formed co-operatives (groups ) and now grow and trade their own coffee

Skip to content

Get Revising

Join get revising, already a member.

Fair Trade and Free Trade

• Globalisation
• Created by: FloraD
• Created on: 01-06-16 16:04

Importance of Trade

• Countries with strong economies have greater influence over international trade
• Many TNCs or MNCs strike hard bargains with the producers of their primary products.
• General Motors, Wal-Mart and Exxon Mobil all have revenues greater than the GDPs of the 48 least developed countries
• A country should aim to have trade surplus - the opposite is trade deficit
• Free Trade assumes that there are no barriers to trade between countries
• A balance is created between how much the  producer wants and how much the consumer is prepared to pay
• If a commodity is scarce, the producer has the greatest influence
• If there is a glut, the consumer has greatest influence
• Focuses on trading with poor and marginalized groups
• Pays fair prices that cover production costs and provides living wage
• Provides credit for producers
• A premium can be paid that will provide funds for social development in local communities
• Fair treatment of all workers + safe workplace conditions
• Development of long term trading relationships

Kenya - Disadvantaged

Kenya is classified by the World Bank as a LOW INCOME COUNTRY

• The cost of Kenya's imports is rising more quickly than the value of Kenya's exports
• Most of the exports are primary products: tea, coffee and fresh flowers
• Tourism is a growing industry, but in susceptible to political instability
• Farmers in countries like the EU and USA recieve subsidies - can sell goods at cheap prices
• Countries whose farmers are not given subsidies cannot compete with these subsidised producers

Traidcraft in Kenya

Kenya is ranked 134 out of 177 in HDI - 60% of the population live on less than $2 a day.

AIMS OF TRAIDCRAFT

• Help local businessed develop and enhance their capacities
• Promoteethical trading policies = increased productivity & reduced poverty
• improve skills of businesses

KENYA PLANS

• Craft productts from Bombolulu Workshops for the Handicapped in Mombassa
• Machakos District Co-operative Union - 7000 artisans
• Tea from a range of producers on the FLO register
• Business Linkages Project - small & medium businesses create links in supply chain
• Coastal farmers and wood carver co-operatives to gain access to new markets
• Fair Trade Tourism - promoting people-to-people tours

Role of WTO

The WTO has 162 members (2016). It remits to reduce tariffs and other barriers to trade. 

Significant progress has been made; the average rate of import tax has gone down from 40% in the 1940s to 4% today.

• CANCUN WTO MEETING 2003
• Discussing on agricultural and on-agricultural goods, as well as services
• If talks were successful, it was said that global income could be raised by $500 billion a year and over 60% of teh inc. would occur in poorer countries
• The real discussion was centred around LEDCs opening their markets compeltely to free trade, whilst the MEDCs cut subsidies for farmers
• Meeting failed as no one was willing to compromise

There is a growing recognition that it is in everyone's interest to have a system of trade that allows LEDCs and LDCs to develop their industries and compete with MEDCs.

Market realities

Comparisons

• The EU is home to many dairy farmers who are subsidised on average $2 for every cow
• In Burkina Faso, a cotton famrer with a small plot of landwithout fertiliser or irrigation etc. has an average annual income of just $250
• The USA is the world's biggest cotton exporterand gave its cotton farmers $4.7 billion in subsidies in 2005
• This figure is larger than the amount of aid given to sub-Saharn Africa in the same year

Import Levies

• In order to develop their industries, LEDCs may want to move away from producing and exporting raw amterials, and instead export manufactured goods
• This is already difficult as they do not have the same level of knowledge or technologies as MEDCs
• Whats more, richer countries charge higher import levies on manufactured goods than on raw materials

TRIPS (Trade-Related Aspects of Intellectual Prope

• TRIPS is a policy that was agreedon by the WTO
• TRIPS states that the inventors of medical treatments have monopoly rights over their inventiions for 20 years
• During the AIDS epidemic, the standard drug used in MEDCs was too expensive for the LEDCs, so South Africa purchased generic drugs from India and Brazil - this was in breach of the TRIPS agreeement
• LEDCs are unable to treat their patients as they are unable to create their own drugs (limited knowledge). 

What is needed to make trade fairer?

LEDCs need help in developing their trade. This comes in the form of...

• Protection of their fledgling industries from cheaper imports
• Rich countries sharing technical expertise and knowledge
• The dismissal of policies which support the protection of such informations
• In terms of the WTO, poorer countries often lack the expertise of the teams of specialist advisers that support trade ministers from richer countries
• All the G7 countries meeting their target of giving 0.7% of GDP in aid

No comments have yet been made

Similar Geography resources:

Fair Trade Vs Free Trade 0.0 / 5

Bananas 0.0 / 5

3.2.1.3 international trade, Global features and trends 0.0 / 5

International Trade Agreements 0.0 / 5

approaches to reducing development Gap 0.0 / 5

Trade or Aid 3.0 / 5 based on 1 rating

AQA Geography-International trade and access to markets 0.0 / 5

GLOBAL INTERDEPENDENCE- TRADE FLOWS AND TRADING PATTERNS 0.0 / 5

Global Systems and Global Governance 4 5.0 / 5 based on 1 rating Teacher recommended

Key players in the development gap 0.0 / 5

Live revision! Join us for our free exam revision livestreams Watch now →

Reference Library

Collections

• See what's new
• All Resources
• Student Resources
• Assessment Resources
• Teaching Resources
• CPD Courses
• Livestreams

Study notes, videos, interactive activities and more!

Economics news, insights and enrichment

Currated collections of free resources

Browse resources by topic

• All Economics Resources

Resource Selections

Currated lists of resources

Study Notes

Fair Trade and Development

Last updated 9 Apr 2023

• Share on Facebook
• Share on Twitter
• Share by Email

The Fair Trade movement covers over 650 producer organisations in more than 60 countries

One of the driving forces behind Fair Trade was a desire to correct for market failures in industries for many primary sector commodities. These included the effects of monopsony power among transnational food processors and food manufacturers that often led to farmers in some of the world's poorest countries receiving an inequitably low and unsustainable price for their products.

Sales of fair-trade labels have risen from an estimated Euro 830m in 2004 to Euro 4.9bn in 2011

The Fair Trade Foundation web site explains fair trade as follows:

"Fairtrade is about better prices, decent working conditions, local sustainability, and fair terms of trade for farmers and workers in the developing world. By requiring companies to pay sustainable prices (which must never fall lower than the market price), Fairtrade addresses the injustices of conventional trade, which traditionally discriminates against the poorest, weakest producers. It enables them to improve their position and have more control over their lives."

Key aims of Fair Trade

• Guarantee a higher / premium price to certified producers
• Achieve greater price stability for growers
• Improve production standards . A grower will be able to receive a Fair Trade licence if it can improve working conditions , better pay and guarantees of environmental sustainability
• A premium price can be offered - for direct investment in improving businesses and communities

Criticisms of Fair Trade

• Impact on non-participating farmers: Some claim that by encouraging consumers to buy their products from Fairtrade sources, this cuts demand and revenues for farmers in poorer nations not covered by the Fairtrade label thereby worsening the risk of extreme poverty
• Who captures the gains (producer surplus) from Fair-Trade coffee ? There is some evidence that a large part of the premium price goes to processors and distributors rather than the farmers themselves.
• The fundamental causes of poverty are not addressed by Fairtrade. Greater investment needs to be made in raising farm productivity , reducing vulnerability to climate change, and reaching multi-lateral trade agreements between countries to reduce import tariffs and improve access for poor countries into the markets of rich advanced nations. Other investment might be better targeted at encouraging farmers to establish producer co-operatives of their own and create their own branded products selling direct to consumers.
• Some free market think-tanks believe that the fair trade movement has resulted for example in excess production of coffee, which has driven down world coffee prices.
• Many producers certified to use the fair trade label come from richer more diversified advanced and middle income countries rather than poorer one
• Producer surplus
• Terms of Trade

You might also like

Australia lags in Commitment to Development Index

7th February 2015

UK & the European Union

Topic Videos

Causes and Effects of China's Growth Slowdown - Views from Economists

27th November 2015

Why Can’t Economists Sell Free Trade?

28th May 2017

EU agrees biggest free trade deal with Japan

10th December 2017

African Free Trade Agreement (Chain of Analysis)

Practice Exam Questions

Welfare Loss - Applying The Concept in Exams

Our subjects.

• › Criminology
• › Economics
• › Geography
• › Health & Social Care
• › Psychology
• › Sociology
• › Teaching & learning resources
• › Student revision workshops
• › Online student courses
• › CPD for teachers
• › Livestreams
• › Teaching jobs

Boston House, 214 High Street, Boston Spa, West Yorkshire, LS23 6AD Tel: 01937 848885

• › Contact us
• › Terms of use
• › Privacy & cookies

© 2002-2024 Tutor2u Limited. Company Reg no: 04489574. VAT reg no 816865400.

Jack Williams Geography

Case Study: The Banana Trade

The banana market is dominated by one species of banana, the Cavendish banana. There are other types of bananas that do not fit the specified description of one and so are excluded from trade. (e.g red bananas)

Quick Facts

Facts about bananas

• Staple food (commonly eaten) for over 400 million people
• One of the 5 most eaten fruits worldwide
• 90 calories per 100g
• 2013 16.5 million tons exported. (Latin America and the Caribbean)
• Grown mainly in India, Philippines
• 30kg of active ingredients (pesticides, fungicides, insecticides, herbicides) added per hectare of bananas.
• Mass production in developing countries is bad for the environments
• WTO will support free trade at all cost
• TNCs have a large influence
• Supermarket price wars decide on the price
• Power and control has moved to supermarkets

Negatives of banana farms

• Deforestation to build (greenfield)
• Chemicals added to farms to increase yield can cause eutrophication
• Kill biomass decreasing biodiversity
• Decreased soil fertility due to contaminants

Positives of banana farms

Four main TNCs

TNCs used to own 80% of  the banana trade. This has changed as TNCs have released direct control of farms and as of 2002 TNCs only control 60% of the trade. They are, however, large influences in the labour standards. Recently retailers have been gaining power and due to the large number of up-coming banana companies exporting. This allows for retailers to demand low prices from suppliers as they can threaten to stop imports from that company.

• ACP  – Africa, the Caribbean and Pacific
• ‘ dollar producers’  (so called due to the heavy investment from US TNCs) – Central American Republics, mainly Ecuador. and Colombia.
• EU (4.5 million tonnes)
• USA (4.5 million tonnes)

Banana Trade War

• Lasted 20 years (1992 – 2012)
• Geneva Banana Agreement
• Between US and EU
• EU negotiated trade agreements with 71 African, Caribbean and Pacific former European colonies ( Lomé Convention ) . They were given special and differential treatment (SDT) when supplying to EU to allow them to develop without European aid.
• The deal was extended to include Cameroon, Dominican Republic, Belize, Ivory Coast, Jamaica, Ghana, Suriname and the Windward Isles.
• The idea of the convention was to protect the smaller, family run business to develop as around 75% of crop supplied to the EU were from large mechanised South American TNCs plantations.
• In 1997 the WTO ruled against the EU
• EU proposals did not satisfy TNCs and were put under pressure and so imposed a number of WTO sanctions on various EU products.
• A compromise was reached in 2009 between the EU and 11 Latin American countries. The Geneva Banana Agreement was ratified in 2012

Race to the Bottom

Supermarkets are paying low prices to suppliers and so TNCs are looking for cheaper places to produce bananas. They are settling in West Africa where there is weaker legislation and low labour costs. This has lead to some very bad social impacts, such as the long, hot hours for workers.

Fair trade bananas help smaller-scale producers in the Caribbean. Fair trade gives consumers organic products that are at a slightly higher price and help developing small suppliers.

Further Reading

http://www.fao.org/economic/est/est-commodities/bananas/en/

Summary: Global exports of bananas in 2017 were expected to reach 18.1 million tonnes which is good as in 2015 export fell to 16.7 million.

https://www.express.co.uk/news/world/1003246/Brexit-news-Africa-banana-trade-wiped-out-fear-Latin-America-European-Union-UK

Summary:  UK import 20% of all bananas that are imported to the EU. Brexit threatens food quotas and suppliers that rely on UK imports

Share this:

One thought on “ case study: the banana trade ”.

So cool. do you need any more or not atm?

Leave a comment Cancel reply

• Already have a WordPress.com account? Log in now.
• Subscribe Subscribed
• Copy shortlink
• Report this content
• View post in Reader
• Manage subscriptions
• Collapse this bar

• Media centre
• Resources library
• Campaign materials

Fairtrade case studies

Choose Fairtrade and you’ll be joining a strong and growing Fairtrade market. Over 400 businesses already choose us and work with us in many different ways.

Whether you are interested in sustainable sourcing on Fairtrade terms, tackling supply chain challenges with a Fairtrade Programme, or are simply intrigued to see how the FAIRTRADE Mark might benefit your marketing, read our case studies from businesses and brands below.

• Jump down to sharing expertise and innovative ways of working with Fairtrade
• Jump down to Fairtrade partnerships for long-term impact

Sourcing with Fairtrade

Greggs: building back better with fairtrade commitments.

Why Greggs are continuing to grow their Fairtrade range despite today’s tough climate.

Despite the challenges faced by the out-of-home sector as a result of the pandemic, Greggs announced in 2021 that they would be using Fairtrade cocoa across their supply chain. Fairtrade cocoa will be in all chocolate products sold through their shops by the end of 2023.

We caught up with them to find out more about why they’re continuing to grow their Fairtrade range despite today’s tough climate.

Cru Kafe: Fighting climate change with Fairtrade

The standout reason CRU Kafe chooses Fairtrade is to fight back against climate change.

Hear from Katie Colvin, Head of Marketing and Communications at CRU Kafe, about how Fairtrade equips the coffee farmers they work with, to deal with climate-related diseases like La Roya. What are the kinds of support Fairtrade coffee farmers in CRU Kafe’s supply chain choose? ‘Certain tools or certain new seed varieties to overcome climate change and get the best from their crop.’

Fyffes: Empowering communities with Fairtrade

Empowering communities is the best reason Fyffes can think of to explain why they are proud to work with Fairtrade.

The combination of the Fairtrade Minimum Price, a safety net that farmers receive when the market price drops, and the Fairtrade Premium, a fund that the community can choose to spend wherever it needs, is unbeatable for farmers to take control of their lives. ‘We work in some of the poorest communities in the world and are proud to see the changes this creates.’ John Hopkins, Procurement Director, Fyffes Group Limited.

Dip & Doze: Secure livelihoods and sustainable incomes

Dip & Doze feel that not knowing where your next pay is coming from, or even how much it might be, can be really stressful and unsustainable for families and communities. They explain why Fairtrade helps them support farmers with more financial security.

Why do businesses choose Fairtrade? Jennie Blake, Head of Brand & Content at Dip & Doze answers ‘with Fairtrade we can support secure livelihoods and sustainable incomes.’ They know that everyone deserves a good night’s sleep, without worrying about having enough money to put food on the table, for school books or for healthcare. We couldn’t have put it better ourselves, Jennie.

Grumpy Mule: Farmers invest in quality

The quality of Grumpy Mule’s coffee is down to the ability of their coffee farmers to invest in their business.

We asked Dave Jameson, Coffee Programme Manager from Grumpy Mule why do they choose Fairtrade coffee? Here’s Dave’s response, including how Fairtrade coffee farmers are using the Fairtrade Premium to invest in sorting machines to produce the best, ripest coffee beans. ‘I know that Fairtrade coffee farmers will invest their Premium in improving the coffee that they grow for us.’

Cafédirect: Direct trade and Fairtrade

Cafédirect was set up as a mission-led business to deliver impact for smallholder farmers worldwide, and working alongside Fairtrade they have created a brilliant model incorporating both direct trade, and Fairtrade.

Cafédirect has a holistic, grower led approach, that is facilitated through direct relationships with growers, and the framework of Standards and pricing that Fairtrade provides.

LEON Fairtrade coffee: Protecting the rainforest and regenerating land in one beautiful blend

LEON buys their 100% Arabica beans from Puro, who are trailblazers in sustainable coffee production. For years, Puro has helped LEON source the best beans from Peru and Honduras , picked at high altitude to give them natural sweetness, then roast them to bring out their full flavours.

The beans are organic, Fairtrade certified, and support the World Land Trust. Puro’s efforts to protect biodiversity have earned them the honour of having not one, but three new rainforest species named after them. 

Sharing expertise

Tate & lyle: fighting child labour with fairtrade.

Tate & Lyle know any labour issues in their supply chain can be tackled in a timely, responsible and ethical way.

Tate & Lyle Sugars buys raw sugar from Belize, where sugar supports the livelihoods of more than 40,000 people, and their communities. Farmers in Belize are dealing with the worldwide Covid-19 pandemic, the climate crisis, and since 2015 they have been addressing the social, cultural and economic factors that drive child labour. And Tate & Lyle Sugars have been by their side.

A programme partnership to create more resilient flower supply chains

This case study demonstrates how partnering with Fairtrade can generate industry-wide innovation.

The project aimed to provide vital support to Kenyan flower farms and workers to ensure they could operate safely and bridge income gaps during the pandemic. It also aimed to carry out research and hold forums to lead discussions on how to create a more sustainable future for the floriculture sector. These two approaches were highly complementary since the pandemic had highlighted systemic issues within these supply chains.

Fairtrade Partnerships

Divine: empowering farmers and consumers.

Divine is driven by its social mission to empower farmers and consumers through creating a supply chain that shares value more equitably; a mission aligned closely to Fairtrade.

Sophi Tranchell, CEO of Divine Chocolate ‘I hope that we have demonstrated that there really is a different way of doing business, putting the cocoa farmer at the heart of it.’

Clipper: Fairtrade means education, health and an empowered future

Clipper believe that small things like a cup of tea can make a big difference. That’s why they’re proudly organic and have supported Fairtrade for the last 25 years.

Clipper don’t just make tea, they make every cup count. And being FAIR is a vital part of that. FAIR stands for Fairness and education for the workers’ children. For Aid and healthcare for all. For Influencing the future and for Rights and empowerment for the workers. And the added bonus for Clipper? Campaigning for act-TEA-vism! The tastiest way to change the world.

Read more about the Clipper Fairtrade campaign

Waitrose & Partners and Fairtrade: Partnering on sourcing and producer voice

Waitrose & Partners are a long-standing partner of Fairtrade, helping to provide stability to producers in the global south since 1994 when they were one of the first retailers to stock Fairtrade products.

This case study demonstrates the huge impact of their sourcing commitments, which benefit Fairtrade farmers and workers, and how we’ve worked together to amplify their voices.

Get in touch

This page is currently under maintenance.

We apologise for any inconvenience caused.

To provide the best experiences, we use technologies like cookies to store and/or access device information. Consenting to them allow us to process data such as browsing behaviour on this site.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 28 June 2024

The limits of fair medical imaging AI in real-world generalization

• Yuzhe Yang   ORCID: orcid.org/0000-0002-7634-8295 1   na1 ,
• Haoran Zhang 1   na1 ,
• Judy W. Gichoya   ORCID: orcid.org/0000-0002-1097-316X 2 ,
• Dina Katabi 1 &
• Marzyeh Ghassemi   ORCID: orcid.org/0000-0001-6349-7251 1 , 3  

Nature Medicine ( 2024 ) Cite this article

1644 Accesses

126 Altmetric

Metrics details

• Radiography

As artificial intelligence (AI) rapidly approaches human-level performance in medical imaging, it is crucial that it does not exacerbate or propagate healthcare disparities. Previous research established AI’s capacity to infer demographic data from chest X-rays, leading to a key concern: do models using demographic shortcuts have unfair predictions across subpopulations? In this study, we conducted a thorough investigation into the extent to which medical AI uses demographic encodings, focusing on potential fairness discrepancies within both in-distribution training sets and external test sets. Our analysis covers three key medical imaging disciplines—radiology, dermatology and ophthalmology—and incorporates data from six global chest X-ray datasets. We confirm that medical imaging AI leverages demographic shortcuts in disease classification. Although correcting shortcuts algorithmically effectively addresses fairness gaps to create ‘locally optimal’ models within the original data distribution, this optimality is not true in new test settings. Surprisingly, we found that models with less encoding of demographic attributes are often most ‘globally optimal’, exhibiting better fairness during model evaluation in new test environments. Our work establishes best practices for medical imaging models that maintain their performance and fairness in deployments beyond their initial training contexts, underscoring critical considerations for AI clinical deployments across populations and sites.

Similar content being viewed by others

Generative models improve fairness of medical classifiers under distribution shifts

AI in health and medicine

Physical imaging parameter variation drives domain shift

As artificial intelligence (AI) models are increasingly deployed in real-world clinical settings 1 , 2 , it is crucial to evaluate not only model performance but also potential biases toward specific demographic groups 3 , 4 . Although deep learning has achieved human-level performance in numerous medical imaging tasks 5 , 6 , existing literature indicates a tendency for these models to manifest existing biases in the data, causing performance disparities between protected subgroups 7 , 8 , 9 , 10 , 11 . For instance, chest X-ray (CXR) classifiers trained to predict the presence of disease systematically underdiagnose Black patients 12 , potentially leading to delays in care. To ensure the responsible and equitable deployment of such models, it is essential to understand the source of such biases and, where feasible, take actions to correct them 13 , 14 .

Recent studies have unveiled the surprising ability of deep models to predict demographic information, such as self-reported race 15 , sex and age 16 , from medical images, achieving performance far beyond that of radiologists. These insights raise the concern of disease prediction models leveraging demographic features as heuristic ‘shortcuts’ 17 , 18 —correlations that are present in the data but have no real clinical basis 18 , for instance deep models using the hospital as a shortcut for disease prediction 19 , 20 .

In this work, we investigated four questions. First, we consider whether disease classification models also use demographic information as shortcuts and whether such demographic shortcuts result in biased predictions. Second, we evaluate the extent to which state-of-the-art methods can remove such shortcuts and create ‘locally optimal’ models that are also fair. Third, we consider real-world clinical deployment settings where shortcuts may not be valid in the out-of-distribution (OOD) data, to dissect the interplay between algorithmic fairness and shortcuts when data shift. Finally, we explore which algorithms and model selection criteria can lead to ‘globally optimal’ models that maintain fairness when deployed in an OOD setting.

We performed a systematic investigation into how medical AI leverages demographic shortcuts through these questions, with an emphasis on fairness disparities across both in-distribution (ID) training and external test sets. Our primary focus is on CXR prediction models, with further validation in dermatology (Extended Data Fig. 1 ) and ophthalmology (Extended Data Fig. 2 ). Our X-ray analysis draws upon six extensive, international radiology datasets: MIMIC-CXR 21 , CheXpert 22 , NIH 23 , SIIM 24 , PadChest 25 and VinDr 26 . We explored fairness within both individual and intersectional subgroups spanning race, sex and age 12 . Our assessment uncovers compelling new insights into how medical AI encodes demographics and the impact that this has on various fairness considerations, especially when models are applied outside their training context during real-world domain shifts, with actionable insights on what models to select for fairness under distribution shift.

Datasets and model training

We used six publicly available CXR datasets, as described in Table 1 . We focused on four binary classification tasks that have been shown to have disparate performance between protected groups 7 , 27 : ‘No Finding’, ‘Effusion’, ‘Pneumothorax’ and ‘Cardiomegaly’. The detailed prevalence rates of the diseases for each demographic subgroup are shown in Extended Data Table 1 .

We also examined medical AI applications in dermatology and ophthalmology. Specifically, we used the ISIC dataset 28 with ‘No Finding’ as the task for dermatological imaging (Extended Data Fig. 1a ) and the ODIR dataset 29 with ‘Retinopathy’ as the task for ophthalmology images (Extended Data Fig. 2a ).

To evaluate fairness, we examined the class-conditioned error rate that is likely to lead to worse patient outcomes for a screening model. For ‘No Finding’, a false positive indicates falsely predicting that a patient is healthy when they are ill, which could lead to delays in treatment 12 ; we, therefore, evaluated the differences in false-positive rate (FPR) between demographic groups. For all other diseases, we evaluated the false-negative rate (FNR) for the same reason. Equality in these metrics is equivalent to equality of opportunity 30 . We choose to study fairness through the notion of equalized odds, as it has been widely used in previous work in the CXR and fairness literature 7 , 12 . In addition, shortcut learning using a particular demographic attribute leads to differences in class-conditioned error rates (that is, FPR and FNR gaps) across attributes 31 , 32 , and so studying these gaps allows us to glean insight into the severity of shortcut learning. Finally, FPR and FNR (as enforced to be equal by equalized odds) are meaningful metrics in the clinical setting, as they correspond to error rates of decision-making at the individual level 12 .

To understand and quantify the types and degrees of distribution shifts in our study, we examined whether there are significant statistical differences in distributions between demographic groups in the ID settings as well as across different datasets in the OOD settings. Specifically, we analyzed prevalence shifts P(Y|A) and representation shifts P(X|A) across different subgroups for ID scenarios and added label shifts P(Y) and covariate shifts P(X) for OOD scenarios ( Methods ). Our analyses indicate that all the distributions that we examined show statistically significant shifts, affecting most demographic groups in the ID context (Extended Data Table 2 ) and across various sites in the OOD context (Extended Data Table 3 ). We note that our analysis does not presuppose specific types of distribution shifts; instead, we simulated real-world deployment conditions where any of these shifts might occur, aiming for results that are generalizable to complex, real-world scenarios.

We trained a grid of deep convolutional neural networks 33 on MIMIC-CXR (radiology), CheXpert (radiology), ODIR (ophthalmology) and ISIC (dermatology), varying the classification task. Our approach follows previous work that achieves state-of-the-art performance in these tasks 8 , 12 using empirical risk minimization (ERM) 34 . We also evaluated algorithms designed to remove spurious correlations or increase model fairness during training. We categorized these algorithms into those that (1) reweight samples based on their group to combat underrepresentation (ReSample 35 and GroupDRO 36 ); (2) adversarially remove group information from model representations (DANN 37 and CDANN 38 ); and (3) more generically attempt to improve model generalization—that is, exponential moving average (MA 39 ). In total, our analysis encompassed a total of 3,456 models trained on MIMIC-CXR, corresponding to the cartesian product of four tasks, four demographic attributes, six algorithms, 12 hyperparameter settings and three random seeds. We summarized our experimental pipeline in Fig. 1 .

a , We trained a grid of deep learning models on medical images from a variety of modalities on several clinical tasks. We applied a variety of state-of-the-art algorithms to mitigate shortcuts, for up to four demographic attributes (where available). b , We evaluated each model ID (that is, on the same dataset where it is trained), along the axis of performance, fairness, amount of demographic encoded and calibration. c , We evaluated the performance and fairness of CXR classification models on OOD domains. To mimic a realistic deployment setting where OOD samples are not observed, we chose the ‘best’ model based on several ID selection criteria.

Algorithmic encoding of attributes leads to fairness gaps

We separately trained deep learning models for our four distinct CXR prediction tasks (‘No Finding’, ‘Cardiomegaly’, ‘Effusion’ and ‘Pneumothorax’) as well as ‘Retinopathy’ in ophthalmology and ‘No Finding’ in dermatology. Each model consists of a feature extractor followed by a disease prediction head. We then employed a transfer learning approach, wherein we kept the weights of the feature extractor frozen and retrained the model to predict sensitive attributes (for example, race). This allowed us to assess the amount of attribute-related information present in the features learned by each model as measured by the area under the receiver operating characteristic curve (AUROC) for attribute prediction ( Methods ). Previous work 15 , 40 demonstrated that deep models trained for disease classification encode demographic attributes, and such encoding could lead to algorithmic bias 41 . We extend the investigation to a broader array of datasets, attributes and imaging modalities. As Fig. 2a,c,e confirms, the penultimate layer of different disease models contains significant information about four demographic attributes (age, race, sex and the intersection of sex and race), and that is consistent across different tasks and medical imaging modalities.

a , The AUROC of demographic attribute prediction from frozen representations for the best ERM model. We trained ERM models on MIMIC-CXR to predict four different binary tasks. ERM representations encode demographic attributes to a high degree. b , The fairness gap, as defined by the FPR gap for ‘No Finding’, and the FNR gap for all other tasks for the best ERM model. ERM models exhibit high fairness gaps, especially between age groups. c , The AUROC of demographic attribute prediction from frozen representations for the best ERM model on the ODIR dataset (ophthalmology), following the same experimental setup. d , The fairness gap for the best ERM model on the ODIR dataset (ophthalmology). e , The AUROC of demographic attribute prediction from frozen representations for the best ERM model on the ISIC dataset (dermatology), following the same experimental setup. f , The fairness gap for the best ERM model on the ISIC dataset (dermatology). a – f , Each bar and its error bar indicate the mean and standard deviation across three independent runs. g , The correlation between attribute prediction performance and fairness for all learned models. We excluded models with suboptimal performance—that is, with an overall validation AUROC below 0.7. The attribute prediction AUROC shows a high correlation with the fairness gap (‘No Finding’, age: R  = 0.82, P  = 4.7 × 10 −8 ; ‘No Finding’, sex and race: R  = 0.81, P  = 8.4 × 10 −9 ; ‘Cardiomegaly’, age: R  = 0.81, P  = 1.9 × 10 −7 ; ‘Effusion’, race: R  = 0.71, P  = 6.4 × 10 −6 ; ‘Pneumothorax’, sex: R  = 0.59, P  = 2.3 × 10 −3 ; all using two-sided t -test). The center line and the shadow denote the mean and 95% CI, respectively.

We then assessed the fairness of these models across demographic subgroups as defined by equal opportunity 30 —that is, discrepancies in the model’s FNR or FPR for demographic attributes. We focused on underdiagnosis 12 —that is, discrepancies in FPR for ‘No Finding’ and discrepancies in FNR for other diseases. For each demographic attribute, we identified two key subgroups with sufficient sample sizes: age groups ‘80–100’ ( n  = 8,063) and ‘18–40’ ( n  = 7,319); race groups ‘White’ ( n  = 32,732) and ‘Black’ ( n  = 8,279); sex groups ‘female’ ( n  = 25,782) and ‘male’ ( n  = 27,794); and sex and race groups ‘White male’ ( n  = 18,032) and ‘Black female’ ( n  = 5,027). In all tasks, we observed that the models displayed biased performance within the four demographic attributes, as evidenced by the FNR disparities (Fig. 2b ). The observed gaps can be as large as 30% for age. The same results hold for the other two imaging modalities (Fig. 2d,f ). Similar results for overdiagnosis (FNR of ‘No Finding’ and FPR for disease prediction) can be found in Extended Data Fig. 3 .

We further investigated the degree to which demographic attribute encoding ‘shortcuts’ may impact model fairness. When models use demographic variables as shortcuts, previous work showed that they can exhibit gaps in subgroup FPR and FNR 31 , 40 . We note that a model encoding demographic information does not necessarily imply a fairness violation, as the model may not necessarily use this information for its prediction. For each task and attribute combination, we trained different models with varying hyperparameters ( Methods ). We focused on the correlation between the degree of encoding of different attributes and the fairness gaps as assessed by underdiagnosis. Figure 2g shows that a stronger encoding of demographic information is significantly correlated with stronger model unfairness (‘No Finding’, age: R  = 0.82, P  = 4.7 × 10 −8 ; ‘No Finding’, sex and race: R  = 0.81, P  = 8.4 × 10 −9 ; ‘Cardiomegaly’, age: R  = 0.81, P  = 1.9 × 10 −7 ; ‘Effusion’, race: R  = 0.71, P  = 6.4 × 10 −6 ; ‘Pneumothorax’, sex: R  = 0.59, P  = 2.3 × 10 −3 ; all using two-sided t -test). Such consistent observations indicate that models using demographic encodings as heuristic shortcuts also have larger fairness disparities, as measured by discrepancies in FPR and FNR.

Mitigating shortcuts creates locally optimal models

We performed model evaluations first in the ID setting, where ERM models trained and tested on data from the same source performed well. We compared ERM to state-of-the-art robustness methods that were designed to effectively address fairness gaps while maintaining overall performance. As shown in Fig. 3a , ERM models exhibited large fairness gaps across age groups when predicting ‘Cardiomegaly’ (that is, models centered in the top right corner, FNR gap of 20% between groups ‘80–100’ and ‘18–40’). By applying data rebalancing methods to address prevalence shifts during training (for example, ReSample), we observed reduced fairness gaps in certain contexts. By applying debiasing robustness methods that correct demographic shortcuts, such as GroupDRO and DANN, the resulting models were able to close the FNR gap while achieving similar AUROCs (for example, the bottom right corner). Our results hold when using the worst group AUROC as the performance metric (Extended Data Fig. 4 ) and across different combinations of diseases and attributes (Fig. 3b and Extended Data Fig. 4 ).

a , Tradeoff between the fairness gap and overall AUROC for all trained models, for ‘Cardiomegaly’ prediction using ‘age’ as the attribute. We plotted the Pareto front—the best achievable fairness gap with a minimum constraint on the performance. b , Tradeoff between the fairness gap and overall AUROC for all trained models, with more disease prediction tasks and attributes. c , Tradeoff between the fairness gap and the overall AUROC on the ODIR dataset (ophthalmology). d , Tradeoff between the fairness gap and the overall AUROC on the ISIC dataset (dermatology).

To demonstrate the value of model debiasing, we further plotted the set of ‘locally optimal models’—those on the Pareto front 42 that balance the performance–fairness tradeoff most optimally on ID data (Fig. 3a ). Those models that lie on this front are ‘locally optimal’, as they have the smallest fairness gap that can be achieved for a fixed performance constraint (for example, AUROC > 0.8). In the ID setting, we found several existing algorithms that consistently achieve high ID fairness without losing notable overall performance for disease prediction (Fig. 3a,b and Extended Data Fig. 4 ).

Similar to our observations in radiology, we identified fairness gaps within subgroups based on age and sex in dermatology and ophthalmology, respectively (Fig. 2d,f ). We further verified the Pareto front for both attributes, where similar observations hold that algorithms for fixing demographic shortcuts could improve ID fairness while incurring minimal detriments to performance (Fig. 3c,d ). The steepness of the Pareto front suggests that small sacrifices in performance could yield substantial gains in fairness.

Locally optimal models exhibit tradeoffs in other metrics

We examined how locally optimal models that balance fairness and AUROC impact other metrics, as previous work showed that it is a theoretical impossibility to balance fairness measured by probabilistic equalized odds and calibration by group 43 . We found that optimizing fairness alone leads to worse results for other clinically meaningful metrics in some cases, indicating an inherent tradeoff between fairness and other metrics. First, for the ‘No Finding’ prediction task, enforcing fair predictions across groups results in worse expected calibration error gap (ECE Gap; Extended Data Fig. 5a ) between groups. Across different demographic attributes, we found a consistent statistically significant negative correlation between ECE Gap and Fairness Gap (age: R  = −0.85, P  = 7.5 × 10 −42 ; race: R  = −0.64, P  = 6.1 × 10 −15 ; sex: R  = −0.73, P  = 4.4 × 10 −28 ; sex and race: R  = −0.45, P  = 1.9 × 10 −8 ; all using two-sided t -test).

We explored the relationship between fairness and other metrics, including average precision and average F1 score. For ‘No Finding’ prediction, fairer models lead to both worse average precision and F1 score (Extended Data Fig. 5a ). The same trend holds across different diseases—for example, for ‘Effusion’ (Extended Data Fig. 5b ). These findings stress that these models, although being locally optimal, exhibit worse results on other important and clinically relevant performance metrics. This uncovers the limitation of blindly optimizing fairness, emphasizing the necessity for more comprehensive evaluations to ensure the reliability of medical AI models.

Local fairness does not transfer under distribution shift

When deploying AI models in real settings, it is crucial to ensure that models can generalize to data from unseen institutions or environments. We directly tested all trained models in the OOD setting, where we report results on external test datasets that are unseen during model development. Figure 4 illustrates that the correlation between ID and OOD performance is high across different settings, which was observed in previous work 44 . However, we found that there was no consistent correlation between ID and OOD fairness. For example, Fig. 4b shows an instance where the correlation between ID fairness and OOD fairness is strongly positive (‘Effusion’ with ‘age’ as the attribute; R  = 0.98, P  = 3.0 × 10 −36 , two-sided t -test), whereas Fig. 4c shows an instance where the correlation between these metrics is actually significantly negative (‘Pneumothorax’ with ‘sex and race’ as the attribute; R  = −0.50, P  = 4.4 × 10 −3 , two-sided t -test). Across 16 combinations of task and attribute, we found that five such settings exhibited this negative correlation, and three additional settings exhibited only a weak ( R  < 0.5) positive correlation (see Extended Data Fig. 6a,b for additional correlation plots). Thus, improving ID fairness may not lead to improvements in OOD fairness, highlighting the complex interplay between fairness and distribution shift 45 , 46 .

a , We plotted the Pearson correlation coefficient of ID versus OOD performance versus the Pearson correlation coefficient of ID versus OOD fairness. Here, each point was derived from a grid of models trained on a particular combination of task and attribute. We found that there was a high correlation between ID and OOD performance in all cases, but the correlation between ID and OOD fairness was tenuous. b , One particular point where fairness transfers between ID and OOD datasets (‘Effusion’ with ‘age’ as the attribute; R  = 0.98, P  = 3.0 × 10 − 36 , two-sided t -test). The center line and the shadow denote the mean and 95% CI, respectively. c , One particular point where fairness does not transfer between ID and OOD datasets (‘Pneumothorax’ with ‘sex and race’ as the attribute; R  = −0.50, P  = 4.4 × 10 − 3 , two-sided t -test). The center line and the shadow denote the mean and 95% CI, respectively. d , The ID Pareto front for ‘Cardiomegaly’ prediction using ‘race’ as the attribute. e , The transformation of the ID Pareto front to the OOD Pareto front, for ‘Cardiomegaly’ prediction using ‘race’ as the attribute. Models that are Pareto optimal ID often do not maintain Pareto optimality OOD.

In addition, we investigated whether models achieving ID Pareto optimality between fairness and performance will maintain in OOD settings. As shown for ‘Cardiomegaly’ prediction using race as the attribute, models originally on the Pareto front ID (Fig. 4d ) do not guarantee to maintain Pareto optimality when deployed in a different OOD setting (Fig. 4e ). We show additional examples of this phenomenon in Extended Data Fig. 6c .

Dissecting model fairness under distribution shift

To disentangle the OOD fairness gap, we present a way to decompose model fairness under distribution shift. Specifically, we decompose and attribute the change in fairness between ID and OOD to be the difference in performance change for each of the groups—that is, the change in fairness is determined by how differently the distribution shift affects each group ( Methods ).

In Extended Data Fig. 7 , we show examples of transferring a trained model from ID setting to OOD setting. For example, Extended Data Fig. 7d illustrates an ERM model trained to predict ‘No Finding’ on CheXpert (ID) and transferred to MIMIC-CXR (OOD) while evaluating fairness across sex. We found that the model was fair with respect to the FPR gap in the ID setting (−0.1% gap, not significant) but had a significant FPR gap when deployed in the OOD setting (3.2%), with females being underdiagnosed at a higher rate (Extended Data Fig. 7e ). We then segmented this FPR gap by sex and found that females experienced an increase in FPR of 3.9%, whereas males experienced an increase in FPR of 0.8% (Extended Data Fig. 7f ). In other words, the model becomes worse for both groups in an OOD setting but to a much larger extent for female patients. This decomposition suggests that mitigation strategies that reduce the impact of the distribution shift on females could be effective in reducing the OOD fairness gap in this instance.

We further extended this study to a larger set of tasks and protected attributes (Extended Data Fig. 7 ). Across all settings, the disparate impact of distribution shift on each group was a significant component, indicating that mitigating the impact of distribution shift is as important as mitigating ID fairness, if the goal is to achieve a fair model OOD.

Globally optimal model selection for OOD fairness

Figure 4 shows that selecting a model based on ID fairness may not lead to a model with optimal OOD fairness. Here, we examined alternate model selection criteria that may lead to better OOD fairness, when we have access only to ID data. Our goal is to find ‘globally optimal’ models that maintain their performance and fairness in new domains. First, we subsetted our selection only to models that had satisfactory ID overall performance (defined as those with overall validation AUROC no less than 5% of the best ERM model). This set of models also had satisfactory OOD performance (Supplementary Fig. 1 ).

Next, we proposed eight candidate model selection criteria (Fig. 5a ), corresponding to selecting the model from this set that minimizes or maximizes some ID metric. We evaluated the selected model by its OOD fairness across five external datasets, each containing up to four attributes and up to four tasks, corresponding to a total of 42 settings. We compared the OOD fairness of the selected model to the OOD fairness of an ‘oracle’, which observes samples from the OOD dataset and directly chooses the model with the smallest OOD fairness gap. For each setting, we computed the increase in fairness gap of each selection criteria relative to the oracle. In Fig. 5a , we report the mean across the 42 settings as well as the 95% confidence interval (CI) computed from 1,000 bootstrap iterations. We found that, surprisingly, selecting the model with the minimum ID fairness gap may not be optimal. Instead, two other criteria based on selecting models where the embedding contains the least attribute information lead to a lower average OOD fairness gap. For instance, we observed a significantly lower increase in OOD fairness gap by selecting models with the ‘Minimum Attribute Prediction Accuracy’ as compared to ‘Minimum Fairness Gap’ ( P  = 9.60 × 10 −94 , one-tailed Wilcoxon rank-sum test). The result echoes our finding in Fig. 2 that the encoding of demographic attributes is positively correlated with ID fairness.

a , We varied the ID model selection criteria and compared the selected model against the oracle that chooses the model that is most fair OOD. We plotted the increase in OOD fairness gap of the selected model over the oracle, averaged across 42 combinations of OOD dataset, task and attribute. We used non-parametric bootstrap sampling ( n  = 1,000) to define the bootstrap distribution for the metric. We found that selection criteria based on choosing models with minimum attribute encoding achieve better OOD fairness than naively selecting based on ID fairness or other aggregate performance metrics (‘Minimum Attribute Prediction Accuracy’ versus ‘Minimum Fairness Gap’: P  = 9.60 × 10 −94 , one-tailed Wilcoxon rank-sum test; ‘Minimum Attribute Prediction AUROC’ versus ‘Minimum Fairness Gap’: P  = 1.95 × 10 −12 , one-tailed Wilcoxon rank-sum test). b , We selected the model for each algorithm with the minimum ID fairness gap. We evaluated its OOD fairness against the oracle on the same 42 settings. We found that removing demographic encoding (that is, DANN) leads to the best OOD fairness (‘DANN’ versus ‘ERM’: P  = 1.86 × 10 −117 , one-tailed Wilcoxon rank-sum test). On each box, the central line indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to 1.5 times the interquartile range. Points beyond the whiskers are plotted individually using the ‘+’ symbol.

Finally, we studied the fairness of each algorithm in the OOD setting. We maintained the performance cutoff described above and selected the model for each algorithm with the lowest ID fairness gap. In Fig. 5b , we report the mean increase in OOD fairness gap relative to the oracle across the same 42 settings. We found that methods that remove demographic information from embeddings (specifically, DANN) lead to the lowest average OOD fairness gap (‘DANN’ versus ‘ERM’: P  = 1.86 × 10 −117 , one-tailed Wilcoxon rank-sum test). Our findings demonstrate that evaluating and removing demographic information encoded by the model ID may be the key to ‘globally optimal’ models that transfer both performance and fairness to external domains.

We demonstrated the interplays between the demographic encoding of attributes as ‘shortcuts’ in medical imaging AI models and how they change under distribution shifts. Notably, we validated our findings across global-scale datasets in radiology (Table 1 ) and across multiple medical imaging modalities (Extended Data Figs. 1 and 2 ). The results show that algorithmic encoding of protected attributes leads to unfairness (Fig. 2 ) and mitigating shortcuts can reduce ID fairness gaps and maintain performance (Fig. 3 ). However, our results also show that there exists an inherent tradeoff for clinically meaningful metrics beyond fairness (Extended Data Fig. 5 ), and such fairness does not transfer under distribution shift (Fig. 4 ). We provide initial strategies to dissect and explain the model fairness under distribution shifts (Extended Data Fig. 7 ). Our results further reveal actionable algorithm and model selection strategies for OOD fairness (Fig. 5 ).

Our results have multiple implications. First, they offer a cautionary tale on the efficacy and consequences of eliminating demographic shortcuts in disease classification models. On the one hand, removing shortcuts addresses ID fairness, which is a crucial consideration in fair clinical decision-making 12 . On the other hand, the resulting tradeoffs with other metrics and non-transferability to OOD settings raises the question about the long-term utility in removing such shortcuts. This is particularly complex in the healthcare setting, where the relationship between demographics and the disease or outcome label is complex 47 , variables can be mislabeled 48 and distribution shifts between domains are difficult to quantify 1 .

Second, we frame demographic features as potential ‘shortcuts’, which should not be used by the model to make disease predictions. However, some demographic variables could be a direct causal factor in some diseases (for example, sex as a causal factor of breast cancer). In these cases, it would not be desirable to remove all demographic reliance but, instead, match the reliance of the model on the demographic attribute to its true causal effect 49 . In the tasks that we examined here, demographic variables, such as race, may have an indirect effect on disease (for example, through socioeconomic status) 50 , which may vary across geographic location or even time period 51 . Whether demographic variables should serve as proxies for these causal factors is a decision that should rest with the model deployers 14 , 47 , 52 , 53 .

Third, we present a preliminary decomposition for diagnosing OOD model fairness changes, by expressing it as a function of the ID fairness gap and the performance change of each group. We found that the disparate impact of distribution shift on per-group performance is a major contributor to lack of fairness in OOD settings. Our work suggests that, for practitioners trying to achieve fairness in models deployed in a different domain, mitigating ID fairness is at least as important as mitigating the impact of distribution shift for particular groups. However, building models robust to arbitrary domain shifts is, in general, a challenging task 54 , 55 . Having some knowledge or data about how the distributions may shift, or even the ability to actively collect data for particular groups, may be necessary 56 . Developing methods and deriving theoretical characterizations of fairness under distribution shift is an active area of research 45 , 46 .

Fourth, the US Food & Drug Administration (FDA), as the primary regulatory body for medical technologies, does not require external validation of clinical AI models, relying instead on the assessment by the product creator 57 . Our findings underscore the necessity for regular evaluation of model performance under distribution shift 58 , 59 , challenging the popular opinion of a single fair model across different settings 60 . This questions the effectiveness of developer assurances on model fairness at the time of testing and highlights the need for regulatory bodies to consider real-world performance monitoring, including fairness degradation 61 . Finally, when a model is deployed in any clinical environment, both its overall and per-group performance, as well as associated clinical outcomes, should be continuously monitored 62 .

Finally, although we imply that smaller ‘fairness gaps’ are better, enforcing these group fairness definitions can lead to worse utility and performance for all groups 43 , 63 , 64 , 65 , 66 , and other fairness definitions may be better suited to the clinical setting 8 , 67 . We note that these invariant notions of fairness could have drawbacks 66 , as equalized odds are incompatible with calibration by group (Extended Data Fig. 5 ), and enforcing equalized odds often lead to the ‘leveling down’ effect in overall performance 63 , 64 . We present the Pareto curve showing the tradeoff between fairness and accuracy, allowing the practitioner to select a model that best fits their deployment scenario. In general, we encourage practitioners to choose a fairness definition that is best suited to their use case and carefully consider the performance–equality tradeoff. The impact of minimizing algorithmic bias on real-world health disparities, the ultimate objective, is complex 68 , and there is no guarantee that deploying a fair model will lead to equitable outcomes. In addition, although we constructed several models for clinical risk prediction, we do not advocate for deployment of these models in real-world clinical settings without practitioners carefully testing models on their data and taking other considerations into account (for example, privacy, regulation and interpretability) 1 , 3 .

Datasets and pre-processing

The datasets used in this study are summarized in Extended Data Table 1 . Unless otherwise stated, we trained models on MIMIC-CXR 21 and evaluated on an OOD dataset created by merging CheXpert 22 , NIH 23 , SIIM 24 , PadChest 25 and VinDr 26 . We included all images (both frontal and lateral) and split each dataset into 70% train, 15% validation and 15% test sets. Note that only MIMIC-CXR and CheXpert have patient race information available, and we extracted race (and other attributes) following established protocols 69 . For MIMIC-CXR, demographic information was obtained by merging with MIMIC-IV 70 . For CheXpert, separate race labels were obtained from the Stanford Center for Artificial Intelligence in Medicine & Imaging ( https://aimi.stanford.edu/ ) website. Where applicable, we dropped patients with missing values for any attribute.

For all datasets, we excluded samples where the corresponding patient has missing age or sex. For ODIR and ISIC, we dropped samples from patients younger than 18 years and older than 80 years due to small sample sizes (that is, smaller than 3% of the total dataset).

Owing to computational constraints, we mainly chose four prediction tasks for CXRs (that is, ‘No Finding’, ‘Effusion’, ‘Cardiomegaly’ and ‘Pneumothorax’). We selected these tasks for several reasons: (1) diversity in presentation: ‘Effusion’, ‘Cardiomegaly’ and ‘Pneumothorax’ each present distinctively and occur in different locations on a CXR, allowing for a comprehensive evaluation across varied pathologies and underlying causes; (2) prevalence in clinical and research settings: these labels are not only common in clinical practice but also frequently studied in prior academic work 7 , 12 , 63 and used in commercial diagnostic systems 71 ; and (3) performance and fairness considerations: these labels are among those with both the highest diagnostic accuracy and substantial fairness gaps on MIMIC-CXR, making them particularly relevant for exploring the relationship between model performance and fairness 7 , 12 .

We scaled all images to 224 × 224 for input to the model. We applied the following image augmentations during training only: random flipping of the images along the horizontal axis, random rotation of up to 10° and a crop of a random size (70–100%) and a random aspect ratio (3/4 to 4/3).

Evaluation methods

To evaluate the performance of disease classification in medical imaging, we used the following metrics: AUROC, TPR, TNR and ECE.

The TPR and TNR are calculated as (FN, false negative; FP, false positive; TP, true positive; TN, true negative):

When reporting the sensitivity and specificity, we followed previous work 12 , 72 in selecting the threshold that maximizes the F1 score. This threshold optimization procedure is conducted separately for each dataset, task, algorithm and attribute combination. We followed standard procedures to calculate the 95% CI for sensitivity and specificity.

We also reported AUC, which is the area under the corresponding ROC curves showing an aggregate measure of detection performance. Finally, we report the expected calibration error (ECE) 73 , which we computed using the netcal library 74 .

Assessing the fairness of machine learning models

To assess the fairness of machine learning models, we evaluated the metrics described above for each demographic group as well as the difference in the value of the metric between groups. Equality of TPR and TNR between demographic groups is known in the algorithmic fairness literature as equal odds 75 . As the models that we studied in this work are likely to be used as screening or triage tools, the cost of an FP may be different from the cost of an FN. In particular, for ‘No Finding’ prediction, FPs (corresponding to underdiagnosis 12 ) would be more costly than FNs, and so we focused on the FPR (or TNR) for this task. For all remaining disease prediction tasks, we focused on the FNR (or TPR) for the same reason. Equality in one of the class-conditioned error rates is an instance of equal opportunity 30 .

Finally, we also examined the per-group ECE and ECE gap between groups. Note that zero ECE for both groups (that is, calibration per group) implies the fairness definition known as sufficiency of the risk score 75 . We emphasize that differences in calibration between groups is a significant source of disparity, as consistent under-estimation or over-estimation of risk for a particular group could lead to under-treatment or over-treatment for that group at a fixed operating threshold relative to the true risk 76 .

Quantifying the distribution shifts

We examined and quantified the types and degrees of distribution shifts in both ID and OOD settings in this study. Inspired by previous work 46 , 77 , we performed a series of hypothesis tests to determine if there were significant statistical differences in distributions between demographic groups and across different pairs of datasets. All P values were adjusted for multiple testing using Bonferroni correction 78 .

We studied the following distribution shifts in the ID setting:

Prevalence shift: P(Y|A)

For binary outcomes Y across groups, we calculated the total variational distance between the probability distributions of Y conditioned on different groups and used a two-sample binomial proportion test, where the null hypothesis corresponds to P(Y|A =  a 1 ) = P(Y|A =  a 2 ):

Representation shift: P(X|A)

When comparing the distribution of the input images X, we first encoded them into representations derived from a frozen foundation model f that is trained in a self-supervised manner on diverse CXR datasets 79 , 80 . We then used the mean maximum discrepancy (MMD) distance and a permutation-based hypothesis test following ref. 81 to test if demographic groups differed statistically in their distribution of representations:

OOD setting

We studied the following distribution shifts in the OOD setting (the null hypothesis is P ID (·) = P OOD (·)):

Label shift: P(Y)

We calculated the total variational distance between the probability distributions of binary outcomes Y across ID and OOD datasets using a two-sample binomial proportion test:

Prevalence shift: P(Y|A = a)

We similarly evaluated the distance between the distributions of Y conditioned on specific demographic subgroups (A) between ID and OOD datasets:

Covariate shift: P(X)

We again encoded X into representations derived from a frozen foundation model f and then used the MMD distance and a permutation-based hypothesis test 81 to examine if ID and OOD datasets differed statistically in their distribution of representations:

Representation shift: P(X|A = a)

Similarly, we calculated the MMD distance conditioned on subgroup A to evaluate shifts in the representation space:

We provide additional results on quantifying various distribution shifts in both ID and OOD settings in the Supplementary Information (Supplementary Tables 1 – 3 ).

Training details

We trained DenseNet-121 (ref. 33 ) models on each task, initializing with ImageNet 82 pre-trained weights. We evaluated six algorithms: ERM 34 , ReSample 35 , GroupDRO 36 , DANN 37 , CDANN 38 and MA 39 .

For each combination of task, algorithm and demographic attribute, we conducted a random hyperparameter search 83 with 12 runs. During training, for a particular attribute, we evaluated the validation set worst-group validation AUROC every 1,000 steps and early stopped if this metric had not improved for five evaluations. We tuned the learning rate and weight decay for all algorithms and also tuned algorithm-specific hyperparameters as mentioned in the original works. We selected the hyperparameter setting that maximized the worst-attribute validation AUROC. CIs were computed as the standard deviation across three different random seeds for each hyperparameter setting.

We also explored a multi-label training setup, where models were trained simultaneously on 14 binary labels available in MIMIC-CXR 7 . We followed the same experimental protocol as outlined in the main paper, including hyperparameter tuning and model selection. Our findings in the multi-label setup mirrored those seen in the binary task setup (Supplementary Figs. 2 and 3 ).

To obtain the level of demographic encoding within representations (Fig. 2 ), we first computed representations using a trained disease prediction model. We froze these representations and trained a multi-class multi-nomial logistic regression model to predict the demographic group using the training set using the scikit-learn library 84 . We varied the L2 regularization strength between 10 −5 and 10 and selected the model with the best macro-averaged AUROC on the validation set. We report the macro-averaged AUROC on the test set.

Decomposing OOD fairness

Here we present a first approach toward decomposing the fairness gap in an OOD environment as a function of the ID fairness gap and the impact that the distribution shift has in each group. In particular, let D src and D tar be the source and target datasets, respectively. Let g ∈ G be a particular group from a set of groups. Let L f ( g , D ) be an evaluation metric for a model f , which is decomposable over individual samples—that is, L f ( g , D ) =  \({L}_{f}(g,D)={\sum }_{(x,y,g{\prime} )\in D;g{\prime} =g}l(f(x),y)\) . Examples of such metrics are the accuracy, TPR or TNR. Then, we can decompose:

The left-hand term is the fairness gap in the OOD environment, and the three terms on the right are (1) the fairness gap in the ID data, (2) the impact of the distribution shift on g 2 and (3) the impact of the distribution shift on g 1 . We note that, to achieve a low fairness gap in the OOD environment, it is important not only to minimize the ID fairness gap (term 1) but also to minimize the difference in how the distribution shift impacts each group (term 2 − term 3).

Evaluation with different medical imaging modalities

In addition to radiology, we also examined medical AI applications in dermatology and ophthalmology to corroborate our findings. Specifically, Extended Data Fig. 1 shows the results for dermatological imaging. We used the ISIC dataset 28 , which contains 32,259 images sourced from multiple international sites. We focused on the ‘No Finding’ task, taking into account ‘sex’ and ‘age’ as the sensitive demographic attributes (Extended Data Fig. 1a ). Similar to our observations in radiology, we identified fairness gaps within subgroups based on age and sex (Extended Data Fig. 1b ), although these disparities were less significant than those observed in CXR assessments (for example, fairness gaps smaller than 2%). This was further confirmed by the Pareto front plot, where most models, including ERM, could achieve a good performance–fairness tradeoff (Extended Data Fig. 1c ).

We extended our analysis to ophthalmology images, specifically focusing on retinopathy detection, using the ODIR dataset 29 with 6,800 images (Extended Data Fig. 2 ). The task that we considered was ‘Retinopathy’, with ‘sex’ and ‘age’ being used as demographic attributes (Extended Data Fig. 2a ). Notably, significant subgroup fairness gaps were observed in age (43% FNR gap between groups ‘60–80’ and ‘18–40’). In contrast, the fairness gap based on sex was less significant, with a 3% FNR difference between ‘female’ and ‘male’ subgroups. We further verified the Pareto front for both attributes, where similar observations hold that algorithms for fixing demographic shortcuts could improve ID fairness while incurring minimal detriments to performance (measured in AUROC).

Analysis on underdiagnosis versus overdiagnosis

In evaluating fairness metrics, our primary study centered on underdiagnosis, specifically the disparities in FPR for ‘No Finding’ and discrepancies in FNR for other conditions. However, an alternative approach involves focusing on overdiagnosis, defined as variances in FNR for ‘No Finding’ and differences in FPR for other diseases. We present findings between their relationship in Extended Data Fig. 3 . An analysis spanning two datasets (MIMIC and CheXpert) and various tasks revealed a consistent pattern: larger gaps in underdiagnosis tend to correspond with more significant overdiagnosis discrepancies. Nonetheless, certain task and attribute combinations exhibited more complex trends, indicating a necessity for deeper exploration and informed decision-making regarding the most appropriate fairness metrics for critical disease evaluations in practical medical settings.

Direct prediction of demographic attributes

We provide analyses for demographic encoding of attribute information. In the main paper, we analyzed the predictiveness of attributes (for example, age, sex and race) based on the embeddings from a disease classification model. The distinct predictiveness between attributes in these domains could be attributed to the intrinsic characteristics of the datasets or the nature of the conditions being studied. To delve deeper, we conducted an experiment training an ERM model to predict these attributes directly using the dermatology dataset (ISIC), and we show the results in Supplementary Table 4 . We observed that certain attributes are indeed less predictive compared to others (that is, age versus sex), suggesting that age may be inherently more challenging to encode within the studied dermatology dataset. Furthermore, the results reveal variations in the predictiveness of age across different subgroups (for example, age groups ‘18–40’ and ‘60–80’ exhibit higher AUROC than the ‘40–60’ group).

Analysis using multi-label models

Prior works 7 studied CXR classification in the multi-label setting, where a model contains an encoder, followed by an individual linear classification head for each of the downstream tasks. We followed the setup 7 to study the following 14 binary labels in MIMIC-CXR: Atelectasis, Cardiomegaly, Consolidation, Edema, Enlarged Cardiomediastinum, Fracture, Lung Lesion, Airspace Opacity, No Finding, Effusion, Pleural Other, Pneumonia, Pneumothorax and Support Devices.

We adapted the following methods to the multi-label setting, (1) ERM, (2) DANN and (3) CDANN, and we followed the same experimental protocol in the main paper in terms of hyperparameter tuning. Note that GroupDRO and ReSample are challenging to adapt to the multi-label setting, as the number of groups is exponential in the number of tasks. For each combination of algorithm and attribute, we selected the multi-label model that maximizes the worst-attribute AUROC, averaged across the 14 tasks.

First, we examined the level of demographic encoding present in the embeddings of the best multi-task ERM model and found that it also encodes a variety of demographic information, similar to the single-label case (Supplementary Fig. 2a ). We further showed the fairness gaps of this best multi-label ERM model and observed that a variety of fairness gaps exist and are statistically significant across all tasks (Supplementary Fig. 2b ). In addition, we plotted the correlation between the fairness gap and the attribute prediction AUROC, across all trained multi-label models. We found a strong and statistically significant positive correlation among all combinations of task and attribute, similar to the single-label case (Supplementary Fig. 2c ).

We also present Pareto plots showing the tradeoff between the fairness gap and overall AUROC across all models, for each combination of task and attribute (Supplementary Fig. 3 ). Overall, we found that the Pareto fronts for the multi-label models demonstrate similar behavior as the single-task models—where the multi-label ERM exhibits the best overall AUROC but has high fairness gap. In addition, with debiasing methods such as multi-label DANN and CDANN, we are able to achieve fair models with minimal loss in overall AUROC.

Test set rebalancing for prevalence shift

We investigated whether eliminating the prevalence shift in the test set would address the observed fairness gaps. Following prior work 40 , we balanced the test set for multiple attributes—age and race—ensuring that demographic proportions (for example, ‘White’ aged ‘20–40’ versus ‘Black’ aged ‘60–80’) and disease prevalence are uniform across all attribute combinations. This approach aims to eliminate prevalence shifts within the test set. Our findings in Supplementary Fig. 4 suggest that, although test set rebalancing can reduce fairness gaps for certain task and attribute combinations (for example, ‘No Finding’ for ‘race’ and ‘Cardiomegaly’ for ‘age’, as compared to Fig. 2b ), there exists significant gaps even after rebalancing the test set, indicating that fairness gaps are influenced by multiple shifts beyond just prevalence shifts.

Statistical analysis

Correlation.

To calculate the correlations between variables, we used Pearson correlation coefficients and their associated P value (two-sided t -test, α = 0.05). 95% CI for the Pearson correlation coefficient was calculated.

Increase in OOD fairness gap

One-tailed Wilcoxon rank-sum test (α = 0.05) was used to assess the increase in OOD fairness gap compared to oracle models.

We used non-parametric bootstrap sampling to generate CIs: random samples of size n (equal to the size of the original dataset) were repeatedly sampled 1,000 times from the original dataset with replacement. We then estimated the increase in OOD fairness gap compared to oracle using each bootstrap sample (α = 0.05).

All statistical analysis was performed with Python version 3.9 (Python Software Foundation).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

All datasets used in this study are publicly available. The MIMIC-CXR ( https://www.physionet.org/content/mimic-cxr-jpg/2.1.0/ ) and VinDr-CXR ( https://physionet.org/content/vindr-cxr/1.0.0/ ) datasets are available from PhysioNet after completion of a data use agreement and a credentialing procedure. The CheXpert dataset ( https://stanfordmlgroup.github.io/competitions/chexpert/ ), along with associated race labels, is available from the Stanford Center for Artificial Intelligence in Medicine & Imaging website ( https://aimi.stanford.edu/ ). The ChestX-ray14 (National Institutes of Health) dataset ( https://nihcc.app.box.com/v/ChestXray-NIHCC ) is available to download from the National Institutes of Health Clinical Center. The PadChest dataset ( https://academictorrents.com/details/96ebb4f92b85929eadfb16761f310a6d04105797 ) can be downloaded from the Medical Imaging Databank of the Valencia Region. The SIIM-ACR Pneumothorax Segmentation dataset ( https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation ) can be downloaded from its Kaggle contest page. The ISIC 2020 dataset ( https://challenge.isic-archive.com/data/#2020 ) can be downloaded from the SIIM-ISIC Melanoma Classification Challenge page. The ODIR dataset ( https://odir2019.grand-challenge.org/dataset/ ) can be obtained from the ODIR 2019 challenge hosted by Grand Challenges.

Code availability

Code that supports the findings of this study is publicly available with an open-source license at https://github.com/YyzHarry/shortcut-ood-fairness .

Zhang, A., Xing, L., Zou, J. & Wu, J. C. Shifting machine learning for healthcare from development to deployment and from models to data. Nat. Biomed. Eng. 6 , 1330–1345 (2022).

Article   PubMed   Google Scholar  

Sendak, M. P. et al. A path for translation of machine learning products into healthcare delivery. EMJ Innov . https://doi.org/10.33590/emjinnov/19-00172 (2020).

Wiens, J. et al. Do no harm: a roadmap for responsible machine learning for health care. Nat. Med. 25 , 1337–1340 (2019).

Article   CAS   PubMed   Google Scholar  

Ahmad, M. A., Patel, A., Eckert, C., Kumar, V. & Teredesai, A. Fairness in machine learning for healthcare. In Proc. 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining 3529–3530 (Association for Computing Machinery, 2020).

McKinney, S. M. et al. International evaluation of an AI system for breast cancer screening. Nature 577 , 89–94 (2020).

Burlina, P. et al. Utility of deep learning methods for referability classification of age-related macular degeneration. JAMA Ophthalmol. 136 , 1305–1307 (2018).

Article   PubMed   PubMed Central   Google Scholar  

Seyyed-Kalantari, L., Liu, G., McDermott, M., Chen, I. Y. & Ghassemi, M. CheXclusion: Fairness Gaps in Deep Chest X-ray Classifiers . Pacific Symposium on Biocomputing (World Scientific Publishing Company, 2020); https://psb.stanford.edu/psb-online/proceedings/psb21/seyyed-kalantari.pdf

Zong, Y., Yang, Y. & Hospedales, T. MEDFAIR: benchmarking fairness for medical imaging. In Proc. 11th International Conference on Learning Representations (ICLR, 2023); https://openreview.net/forum?id=6ve2CkeQe5S

Kinyanjui, N. M. et al. Fairness of classifiers across skin tones in dermatology. In Medical Image Computing and Computer Assisted Intervention — MICCAI 2020 Vol. 12266 (eds Martel, A. L. et al.) 320–329 (Springer, 2020); https://doi.org/10.1007/978-3-030-59725-2_31

Lin, M. et al. Improving model fairness in image-based computer-aided diagnosis. Nat. Commun. 14 , 6261 (2023).

Article   CAS   PubMed   PubMed Central   Google Scholar  

Weng, N., Bigdeli, S., Petersen, E. & Feragen, A. Are sex-based physiological differences the cause of gender bias for chest X-ray diagnosis? In Clinical Image-Based Procedures, Fairness of AI in Medical Imaging, and Ethical and Philosophical Issues in Medical Imaging (CLIP 2023, EPIMI 2023, FAIMI 2023) Vol. 14242 (eds Wesarg, S. et al.) 142–152 (Springer, 2023); https://doi.org/10.1007/978-3-031-45249-9_14

Seyyed-Kalantari, L., Zhang, H., McDermott, M. B., Chen, I. Y. & Ghassemi, M. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat. Med. 27 , 2176–2182 (2021).

Adamson, A. S. & Smith, A. Machine learning and health care disparities in dermatology. JAMA Dermatol. 154 , 1247–1248 (2018).

McCradden, M. D., Joshi, S., Mazwi, M. & Anderson, J. A. Ethical limitations of algorithmic fairness solutions in health care machine learning. Lancet Digit. Health 2 , e221–e223 (2020).

Gichoya, J. W. et al. AI recognition of patient race in medical imaging: a modelling study. Lancet Digit. Health 4 , e406–e414 (2022).

Adleberg, J. et al. Predicting patient demographics from chest radiographs with deep learning. J. Am. Coll. Radiol. 19 , 1151–1161 (2022).

Geirhos, R. et al. Shortcut learning in deep neural networks. Nat. Mach. Intell. 2 , 665–673 (2020).

Article   Google Scholar  

Banerjee, I. et al. ‘Shortcuts’ causing bias in radiology artificial intelligence: causes, evaluation, and mitigation. J. Am. Coll. Radiol. 20 , 842–851 (2023).

Zech, J. R. et al. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLoS Med. 15 , e1002683 (2018).

DeGrave, A. J., Janizek, J. D. & Lee, S.-I. AI for radiographic COVID-19 detection selects shortcuts over signal. Nat. Mach. Intell. 3 , 610–619 (2021).

Johnson, A. E. et al. MIMIC-CXR-JPG, a large publicly available database of labeled chest radiographs. Preprint at https://arxiv.org/abs/1901.07042 (2019).

Irvin, J. et al. CheXpert: a large chest radiograph dataset with uncertainty labels and expert comparison. In Proc. of the AAAI Conference on Artificial Intelligence Vol. 33, 590–597 (Association for Computing Machinery, 2019); https://doi.org/10.1609/aaai.v33i01.3301590

Wang, X. et al. ChestX-ray8: hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition 2097–2106 (IEEE, 2017); https://openaccess.thecvf.com/content_cvpr_2017/papers/Wang_ChestX-ray8_Hospital-Scale_Chest_CVPR_2017_paper.pdf

Zawacki, A. et al. SIIM-ACR pneumothorax segmentation. kaggle https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/ (2019).

Bustos, A., Pertusa, A., Salinas, J.-M. & De La Iglesia-Vaya, M. PadChest: a large chest x-ray image dataset with multi-label annotated reports. Med. Image Anal. 66 , 101797 (2020).

Nguyen, H. Q. et al. VinDr-CXR: an open dataset of chest X-rays with radiologist’s annotations. Sci. Data 9 , 429 (2022).

Larrazabal, A. J., Nieto, N., Peterson, V., Milone, D. H. & Ferrante, E. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proc. Natl Acad. Sci. USA 117 , 12592–12594 (2020).

Rotemberg, V. et al. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Sci. Data 8 , 34 (2021).

Ocular disease recognition. kaggle https://www.kaggle.com/datasets/andrewmvd/ocular-disease-recognition-odir5k (accessed 5 September 2023).

Hardt, M., Price, E. & Srebro, N. Equality of opportunity in supervised learning. In Proc. 30th Conference on Neural Information Processing Systems (NIPS, 2016); https://proceedings.neurips.cc/paper_files/paper/2016/file/9d2682367c3935defcb1f9e247a97c0d-Paper.pdf

Brown, A. et al. Detecting shortcut learning for fair medical AI using shortcut testing. Nat. Commun. 14 , 4314 (2023).

Makar, M. et al . Causally motivated shortcut removal using auxiliary labels. In Proc. 25th International Conference on Artificial Intelligence and Statistics Vol. 151, 739–766 (PMLR, 2022); https://proceedings.mlr.press/v151/makar22a/makar22a.pdf

Huang, G., Liu, Z., Van Der Maaten, L. & Weinberger, K. Q. Densely connected convolutional networks. In Proc. of the 2017 IEEE Conference on Computer Vision and Pattern Recognition 4700–4708 (IEEE, 2017); https://doi.ieeecomputersociety.org/10.1109/CVPR.2017.243

Vapnik, V. Principles of risk minimization for learning theory. In Advances in Neural Information Processing Systems 4 (NeurIPS, 1991); https://proceedings.neurips.cc/paper_files/paper/1991/file/ff4d5fbbafdf976cfdc032e3bde78de5-Paper.pdf

Idrissi, B. Y., Arjovsky, M., Pezeshki, M. & Lopez-Paz, D. Simple data balancing achieves competitive worst-group-accuracy. In Proc. 1st Conference on Causal Learning and Reasoning 336–351 (PMLR, 2022); https://proceedings.mlr.press/v177/idrissi22a.html

Sagawa, S., Koh, P. W., Hashimoto, T. B. & Liang, P. Distributionally robust neural networks for group shifts: on the importance of regularization for worst-case generalization. In Proc. of the International Conference on Learning Representations (ICLR, 2020); https://openreview.net/pdf?id=ryxGuJrFvS

Ganin, Y. et al. Domain-adversarial training of neural networks. J. Mach. Learn. Res. 17 , 2096–2030 (2016).

Google Scholar  

Li, Y. et al. Deep domain generalization via conditional invariant adversarial networks. In Proc. of the European Conference on Computer Vision (ECCV) 624–639 (ECCV, 2018); https://openaccess.thecvf.com/content_ECCV_2018/papers/Ya_Li_Deep_Domain_Generalization_ECCV_2018_paper.pdf

Polyak, B. T. & Juditsky, A. B. Acceleration of stochastic approximation by averaging. SIAM J. Control Optim. 30 , 838–855 (1992).

Glocker, B., Jones, C., Bernhardt, M. & Winzeck, S. Algorithmic encoding of protected characteristics in chest X-ray disease detection models. EBioMedicine 89 , 104467 (2023).

Jones, C., Roschewitz, M. & Glocker, B. The role of subgroup separability in group-fair medical image classification. In Medical Image Computing and Computer Assisted Intervention — MICCAI 2023 Vol. 14222 (eds Greenspan, H. et al.) 179–188 (Springer, 2023); https://doi.org/10.1007/978-3-031-43898-1_18

Wei, S. & Niethammer, M. The fairness-accuracy Pareto front. Stat. Anal. Data Min. 15 , 287–302 (2022).

Kleinberg, J., Mullainathan, S. & Raghavan, M. Inherent tradeoffs in the fair determination of risk scores. In Proc. 8th Innovations in Theoretical Computer Science Conference (ITCS 2017) Vol. 67 (ed. Papadimitriou, C. H.) 1–23 (2017); https://doi.org/10.4230/LIPIcs.ITCS.2017

Miller, J. P. et al. Accuracy on the line: on the strong correlation between out-of-distribution and in-distribution generalization. In Proc. 38th International Conference on Machine Learning 7721–7735 (PMLR, 2021); https://proceedings.mlr.press/v139/miller21b/miller21b.pdf

An, B., Che, Z., Ding, M. & Huang, F. Transferring fairness under distribution shifts via fair consistency regularization. In Proc. 36th Conference on Neural Information Processing Systems (NeurIPS, 2022); https://proceedings.neurips.cc/paper_files/paper/2022/file/d1dbaabf454a479ca86309e66592c7f6-Paper-Conference.pdf

Schrouff, J. et al. Diagnosing failures of fairness transfer across distribution shift in real-world medical settings. In Proc. 36th Conference on Neural Information Processing Systems (NeurIPS, 2022); https://proceedings.neurips.cc/paper_files/paper/2022/file/7a969c30dc7e74d4e891c8ffb217cf79-Paper-Conference.pdf

Vyas, D. A., Eisenstein, L. G. & Jones, D. S. Hidden in plain sight—reconsidering the use of race correction in clinical algorithms. N. Engl. J. Med. 383 , 874–882 (2020).

Jain, S. et al. VisualCheXbert: addressing the discrepancy between radiology report labels and image labels. In Proc. of the Conference on Health, Inference, and Learning 105–115 (Association for Computing Machinery, 2021); https://doi.org/10.1145/3450439.3451862

Kumar, A., Deshpande, A. & Sharma, A. Causal effect regularization: automated detection and removal of spurious attributes. In Proc. 37th Conference on Neural Information Processing Systems https://openreview.net/pdf?id=V5Oh7Aqfft (NeurIPS, 2023).

Basu, A. Use of race in clinical algorithms. Sci. Adv. 9 , eadd2704 (2023).

Chandra, A. & Skinner, J. Geography and Racial Health Disparities (National Bureau of Economic Research, 2003); https://doi.org/10.3386/w9513

Suriyakumar, V. M., Ghassemi, M. & Ustun, B. When personalization harms: reconsidering the use of group attributes in prediction. In Proc. 40th International Conference on Machine Learning 33209–33228 (PMLR, 2023); https://proceedings.mlr.press/v202/suriyakumar23a.html

Manski, C. F., Mullahy, J. & Venkataramani, A. S. Using measures of race to make clinical predictions: decision making, patient health, and fairness. Proc. Natl Acad. Sci. USA 120 , e2303370120 (2023).

Gulrajani, I. & Lopez-Paz, D. In search of lost domain generalization. In Proc. of the International Conference on Learning Representations (ICLR, 2021); https://openreview.net/pdf?id=lQdXeXDoWtI

Zhang, H. et al. An empirical framework for domain generalization in clinical settings. In Proc. of the Conference on Health, Inference, and Learning 279–290 (Association for Computing Machinery, 2021); https://doi.org/10.1145/3450439.3451878

Branchaud-Charron, F., Atighehchian, P., Rodríguez, P., Abuhamad, G. & Lacoste, A. Can active learning preemptively mitigate fairness issues? Preprint at https://arxiv.org/abs/2104.06879 (2021).

Artificial intelligence and machine learning (AI/ML)-enabled medical devices. US Food & Drug Administration https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices (2022).

Koh, P. W. et al. WILDS: a benchmark of in-the-wild distribution shifts. In Proc. 38th International Conference on Machine Learning 5637–5664 (PMLR, 2021); https://cs.stanford.edu/people/jure/pubs/wilds-icml21.pdf

Yang, Y., Zhang, H., Katabi, D. & Ghassemi, M. Change is hard: a closer look at subpopulation shift. In Proc. 40th International Conference on Machine Learning 39584–39622 (Association for Computing Machinery, 2023); https://proceedings.mlr.press/v202/yang23s/yang23s.pdf

Mitchell, M. et al. Model cards for model reporting. In Proc. of the Conference on Fairness, Accountability, and Transparency 220–229 (Association for Computing Machinery, 2019); https://doi.org/10.1145/3287560.3287596

Joint statement on enforcement efforts against discrimination and bias in automated systems. Federal Trade Commission https://www.ftc.gov/legal-library/browse/cases-proceedings/public-statements/joint-statement-enforcement-efforts-against-discrimination-bias-automated-systems (2023).

Gallifant, J. et al. Disparity dashboards: an evaluation of the literature and framework for health equity improvement. Lancet Digit. Health 5 , e831–e839 (2023).

Zhang, H. et al. Improving the fairness of chest x-ray classifiers. In Proc. of the Conference on Health, Inference, and Learning Vo. 174, 204–233 (PMLR, 2022); https://proceedings.mlr.press/v174/zhang22a/zhang22a.pdf

Zietlow, D. et al. Leveling down in computer vision: Pareto inefficiencies in fair deep classifiers. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 10410–10421 (IEEE, 2022); https://openaccess.thecvf.com/content/CVPR2022/papers/Zietlow_Leveling_Down_in_Computer_Vision_Pareto_Inefficiencies_in_Fair_Deep_CVPR_2022_paper.pdf

Petersen, E., Holm, S., Ganz, M. & Feragen, A. The path toward equal performance in medical machine learning. Patterns 4 , 100790 (2023).

Petersen, E., Ferrante, E., Ganz, M. & Feragen, A. Are demographically invariant models and representations in medical imaging fair? Preprint at https://arxiv.org/abs/2305.01397 (2023).

Martinez, N., Bertran, M. & Sapiro G. Minimax Pareto fairness: a multi objective perspective. In Proc. of the International Conference on Machine Learning Vol. 119, 6755–6764 (PMLR, 2020); https://proceedings.mlr.press/v119/martinez20a.html

Paulus, J. K. & Kent, D. M. Predictably unequal: understanding and addressing concerns that algorithmic clinical prediction may increase health disparities. npj Digit. Med. 3 , 99 (2020).

Movva, R. et al. Coarse race data conceals disparities in clinical risk score performance. In Proc. of the 8th Machine Learning for Healthcare Conference . https://proceedings.mlr.press/v219/movva23a.html 443-472 (PMLR, 2023).

Johnson, A. E. W. et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci. Data 10 , 1 (2023).

Fanni, S. C. et al. Artificial intelligence-based software with CE mark for chest X-ray interpretation: opportunities and challenges. Diagnostics (Basel) 13 , 2020 (2023).

Lipton, Z. C., Elkan, C. & Naryanaswamy, B. Optimal thresholding of classifiers to maximize F1 measure. Mach. Learn. Knowl. Discov. Databases https://doi.org/10.1007/978-3-662-44851-9_15 (2014).

Guo, C., Pleiss, G., Sun, Y. & Weinberger, K. Q. On calibration of modern neural networks. In Proc. of the 34th International Conference on Machine Learning 1321–1330 (PMLR, 2017); https://proceedings.mlr.press/v70/guo17a/guo17a.pdf

Kuppers, F., Kronenberger, J., Shantia, A. & Haselhoff A. Multivariate confidence calibration for object detection. In Proc. of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) 1322–1330 (IEEE, 2020); https://doi.org/10.1109/CVPRW50498.2020.00171

Barocas, S., Hardt, M. & Narayanan, A. Fairness and Machine Learning: Limitations and Opportunities (MIT Press, 2019).

Pfohl, S. et al. Net benefit, calibration, threshold selection, and training objectives for algorithmic fairness in healthcare. In Proc. of the 2022 ACM Conference on Fairness, Accountability, and Transparency 1039–1052 (Association for Computing Machinery, 2022); https://doi.org/10.1145/3531146.3533166

Bernhardt, M., Jones, C. & Glocker, B. Potential sources of dataset bias complicate investigation of underdiagnosis by machine learning algorithms. Nat. Med. 28 , 1157–1158 (2022).

Aickin, M. & Gensler, H. Adjusting for multiple testing when reporting research results: the Bonferroni vs Holm methods. Am. J. Public Health 5 , 726–728 (1996).

Wang, Z., Wu, Z., Agarwal, D. & Sun, J. MedCLIP: contrastive learning from unpaired medical images and text. In Proc. of the 2022 Conference on Empirical Methods in Natural Language Processing (eds Goldberg, Y. et al.) 3876–3887 (Association for Computational Linguistics, 2022); https://doi.org/10.18653/v1/2022.emnlp-main.256

Tiu, E. et al. Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning. Nat. Biomed. Eng. 6 , 1399–1406 (2022).

Rabanser, S., Günnemann, S. & Lipton, Z. Failing loudly: an empirical study of methods for detecting dataset shift. In Proc. 33rd Conference on Neural Information Processing Systems (NeurIPS, 2019); https://proceedings.neurips.cc/paper_files/paper/2019/file/846c260d715e5b854ffad5f70a516c88-Paper.pdf

Deng, J. et al. Imagenet: a large-scale hierarchical image database. In Proc. of the 2009 IEEE Conference on Computer Vision and Pattern Recognition 248–255 (IEEE, 2009); https://doi.org/10.1109/CVPR.2009.5206848

Bergstra, J. & Bengio, Y. Random search for hyper-parameter optimization. J. Mach. Learn. Res. 13 , 281–305 (2012).

Pedregosa, F. et al. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12 , 2825–2830 (2011).

Download references

Acknowledgements

H.Z. was supported, in part, by a Google Research Scholar Award. J.W.G. is a 2022 Robert Wood Johnson Foundation Harold Amos Medical Faculty Development Program awardee and declares support from Radiological Society of North America health disparities grant number EIHD2204, the Lacuna Fund (no. 67), the Gordon and Betty Moore Foundation, a National Institutes of Health (National Institute of Biomedical Imaging and Bioengineering) Medical Imaging and Data Resource Center grant under contracts 75N92020C00008 and 75N92020C00021 and National Heart, Lung, and Blood Institute award number R01HL167811. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Author information

These authors contributed equally: Yuzhe Yang, Haoran Zhang.

Authors and Affiliations

Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, USA

Yuzhe Yang, Haoran Zhang, Dina Katabi & Marzyeh Ghassemi

Department of Radiology, Emory University School of Medicine, Atlanta, GA, USA

Judy W. Gichoya

Institute for Medical Engineering & Science, Massachusetts Institute of Technology, Cambridge, MA, USA

Marzyeh Ghassemi

You can also search for this author in PubMed   Google Scholar

Contributions

Y.Y., H.Z. and M.G. conceived and designed the study. Y.Y. and H.Z. performed data collection, processing and experimental analysis. Y.Y., H.Z., J.W.G., D.K. and M.G. interpreted experimental results and provided feedback on the study. Y.Y., H.Z., J.W.G. and M.G. wrote the original manuscript. M.G. supervised the research. All authors reviewed and approved the manuscript.

Corresponding author

Correspondence to Yuzhe Yang .

Ethics declarations

Competing interests.

D.K. is a co-founder of Emerald Innovations, Inc. and serves on the scientific advisory board of Janssen and on the data and analytics advisory board of Amgen. The other authors declare no competing interests.

Peer review

Peer review information.

Nature Medicine thanks Melanie Roschewitz and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Lorenzo Righetto, in collaboration with the Nature Medicine team.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended data fig. 1 evaluation results for models on the isic dataset for no finding prediction..

a , Dataset statistics. b , Subgroup fairness gaps of the ERM model as defined by FPR. Each subgroup contains at least 100 samples for analysis (age: subgroup ‘60–80’ vs ‘18–40’; sex: subgroup ‘female’ vs ‘male’). We find that disparities in FPR are small and statistically insignificant in the case of age. Each bar and its error bar indicate the mean and standard deviation across 3 independent runs. c , Trade-off between the fairness gap and overall AUROC for all trained models, evaluated against sensitive attribute age and sex, respectively. We find that most models, including ERM, achieve a good fairness-performance tradeoff.

Extended Data Fig. 2 Evaluation results for models on the ODIR dataset for Retinopathy prediction.

a , Dataset statistics. b , Subgroup fairness gaps of the ERM model as defined by FNR. Each subgroup contains at least 100 samples for analysis (age: subgroup ‘60–80’ vs ‘18–40’; sex: subgroup ‘female’ vs ‘male’). We find a significant FNR gap between age groups. Each bar and its error bar indicate the mean and standard deviation across 3 independent runs. c , Trade-off between the fairness gap and overall AUROC for all trained models, evaluated against sensitive attribute age and sex, respectively. We find, similar to the chest X-ray setting, that algorithms for fixing demographic shortcuts could improve in-distribution fairness while incurring minimal detriments to performance.

Extended Data Fig. 3 Trade-offs between the FPR gap and the FNR gap for each task and attribute, for models trained on MIMIC-CXR or CheXpert and evaluated on the same dataset for (a) No Finding and (b) Effusion prediction.

We evaluate these metrics across age (‘80–100’ vs ‘18–40’), sex (‘female’ vs ‘male’), race (‘White’ vs. ‘Black’), and the intersection of sex and race (‘White male’ vs. ‘Black female’). We find for the most part, there is a positive correlation, indicating that fairer models achieve fairness with respect to both FPR and FNR (that is, equal odds). All p values are calculated using two-sided t-test. The center line and the shadow denote the mean and 95% CI, respectively.

Extended Data Fig. 4 Algorithms for removing demographic shortcuts mitigate in-distribution fairness gaps and maintain performance.

a, b , Trade-off between the fairness gap and overall AUROC for all trained models. c, d , Trade-off between the fairness gap and the worst-group AUROC for all trained models. Each plot represents a specific disease prediction task (for example, Cardiomegaly) with a specific attribute (for example, age). In each case, we plot the Pareto front, the best achievable fairness gap with a minimum constraint on the performance.

Extended Data Fig. 5 Trade-offs between the fairness gap and the expected calibration error (ECE) gap, average F1, and average precision, for models trained and evaluated on MIMIC-CXR for (a) No Finding and (b) Effusion prediction.

We evaluate these metrics across age (‘80–100’ vs ‘18–40’), sex (‘female’ vs ‘male’), race (‘White’ vs. ‘Black’), and the intersection of sex and race (‘White male’ vs. ‘Black female’). We find that fairer models tend to exhibit larger ECE gaps, worse average F1, and worse average precision, indicating the undesirable tradeoff between fairness with other performance and calibration metrics. All p values are calculated using two-sided t-test. The center line and the shadow denote the mean and 95% CI, respectively.

Extended Data Fig. 6 The transfer of performance (overall AUROC) and fairness between the ID (MIMIC-CXR) and OOD datasets.

a , The OOD performance versus the ID performance for each task and attribute combination. b , The OOD fairness versus the ID fairness for each task and attribute combination. c , The Pareto front between fairness and performance for ID and OOD, finding that models that are Pareto optimal ID often do not maintain Pareto optimality OOD. All p values are calculated using two-sided t-test. The center line and the shadow denote the mean and 95% CI, respectively.

Extended Data Fig. 7 Examining the behavior of specific ERM models transferred from ID setting to OOD setting.

a, d, g, j , The receiver operating characteristic (ROC) curves for each group and each dataset, marking the operating point of the model. b, e, h, k , This shift in the operating point results in a change in the FPR and FNR values of subgroups on the OOD dataset. Each bar and its error bar indicate the mean and standard deviation across 3 independent runs. c, f, i, l , We decompose the OOD fairness gap as a function of the ID fairness gap, and the change in FPR for each of the groups. Each bar and its error bar indicate the mean and standard deviation across 3 independent runs.

Supplementary information

Supplementary tables 1–4, figs. 1–4 and note 1., reporting summary, rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Yang, Y., Zhang, H., Gichoya, J.W. et al. The limits of fair medical imaging AI in real-world generalization. Nat Med (2024). https://doi.org/10.1038/s41591-024-03113-4

Download citation

Received : 08 December 2023

Accepted : 05 June 2024

Published : 28 June 2024

DOI : https://doi.org/10.1038/s41591-024-03113-4

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

• Environment
• Science & Technology
• Business & Industry
• Health & Public Welfare
• Topics (CFR Indexing Terms)
• Public Inspection
• Presidential Documents
• Document Search
• Advanced Document Search
• Public Inspection Search
• Reader Aids Home
• Office of the Federal Register Announcements
• Using FederalRegister.Gov
• Understanding the Federal Register
• Recent Site Updates
• Federal Register & CFR Statistics
• Videos & Tutorials
• Developer Resources
• Government Policy and OFR Procedures
• Congressional Review
• My Clipboard
• My Comments
• My Subscriptions
• Sign In / Sign Up
• Site Feedback
• Search the Federal Register

The Federal Register

The daily journal of the united states government.

• Legal Status

This site displays a prototype of a “Web 2.0” version of the daily Federal Register. It is not an official legal edition of the Federal Register, and does not replace the official print version or the official electronic version on GPO’s govinfo.gov.

The documents posted on this site are XML renditions of published Federal Register documents. Each document posted on the site includes a link to the corresponding official PDF file on govinfo.gov. This prototype edition of the daily Federal Register on FederalRegister.gov will remain an unofficial informational resource until the Administrative Committee of the Federal Register (ACFR) issues a regulation granting it official legal status. For complete information about, and access to, our official publications and services, go to About the Federal Register on NARA's archives.gov.

The OFR/GPO partnership is committed to presenting accurate and reliable regulatory information on FederalRegister.gov with the objective of establishing the XML-based Federal Register as an ACFR-sanctioned publication in the future. While every effort has been made to ensure that the material on FederalRegister.gov is accurately displayed, consistent with the official SGML-based PDF version on govinfo.gov, those relying on it for legal research should verify their results against an official edition of the Federal Register. Until the ACFR grants it official status, the XML rendition of the daily Federal Register on FederalRegister.gov does not provide legal notice to the public or judicial notice to the courts.

Fair Lending Report of the Consumer Financial Protection Bureau

A Notice by the Consumer Financial Protection Bureau on 07/02/2024

Document Details

Information about this document as published in the Federal Register .

Document Statistics

Enhanced content.

Relevant information about this document from Regulations.gov provides additional context. This information is not part of the official Federal Register document.

Published Document

This document has been published in the Federal Register . Use the PDF linked in the document sidebar for the official electronic format.

Enhanced Content - Table of Contents

This table of contents is a navigational tool, processed from the headings within the legal text of Federal Register documents. This repetition of headings to form internal navigation links has no substantive legal effect.

FOR FURTHER INFORMATION CONTACT:

Supplementary information:, 1. fair lending enforcement and supervision, 1.1. risk-based prioritization, 1.2. fair lending enforcement, 1.2.1. ecoa-related public enforcement actions, 1.2.2. hmda-related public enforcement actions, freedom mortgage, bank of america, 1.2.3. ecoa referrals to department of justice, 1.2.4. implementing enforcement orders, 1.3. fair lending supervision, 2. rulemaking and guidance, 2.1. rulemaking, 2.1.1. small business lending data collection rulemaking, 2.1.2. automated valuation models rulemaking, 2.2. guidance, 2.2.1. proposed interagency guidance on reconsiderations of value for residential real estate valuations, 2.2.2. consumer financial protection circular 2023-03: adverse action notification requirements and the proper use of the cfpb's sample forms provided in regulation b, 2.2.3. coverage of franchise financing under ecoa, including the small business lending rule, 2.2.4. supervisory highlights, 2.2.5. hmda guidance and resources, 3. stakeholder engagement, 3.1. promoting and broadcasting the fair lending and access to credit mission, 3.1.1. cfpb blog posts, press releases, and other communications, 3.1.2. cfpb engagements with stakeholders, 3.2. data and reports, 3.2.1. state community reinvestment act: summary of state laws, 3.2.2. banking and credit access in the southern region of the united states, 3.2.3. consumer finances in rural areas of the southern region, 3.2.4. availability of 2022 hmda data, 3.2.5. report on the home mortgage disclosure act rule voluntary review, 3.2.6. data point: 2022 mortgage market activity and trends, 4. interagency engagement, 4.1. special purpose credit program interagency roundtable, 4.2. joint statements, 4.3. appraisal bias, 5. amicus program and other litigation, 5.1. amicus briefs and statements of interest, 5.2. litigation, 6. interagency reporting on ecoa and hmda, 6.1. reporting on ecoa enforcement, 6.1.1. public enforcement actions, 6.1.2. cfpb enforcement actions, colony ridge, 6.1.3. interagency enforcement actions, 6.1.4. number of institutions cited for ecoa/regulation b violations, 6.1.5. violations cited during ecoa examinations, 6.1.6. referrals to the department of justice, 6.2. reporting on hmda, 7. looking forward & focus on digital discrimination, appendix a: hmda resources, appendix b: defined terms, signing authority, enhanced content - submit public comment.

• This feature is not available for this document.

Enhanced Content - Read Public Comments

Enhanced content - sharing.

• Email this document to a friend

Enhanced Content - Document Print View

• Print this document

Enhanced Content - Document Tools

These tools are designed to help you understand the official document better and aid in comparing the online edition to the print edition.

These markup elements allow the user to see how the document follows the Document Drafting Handbook that agencies use to create their documents. These can be useful for better understanding how a document is structured but are not part of the published document itself.

Enhanced Content - Developer Tools

This document is available in the following developer friendly formats:.

• JSON: Normalized attributes and metadata
• XML: Original full text XML
• MODS: Government Publishing Office metadata

More information and documentation can be found in our developer tools pages .

Official Content

• View printed version (PDF)

This PDF is the current document as it appeared on Public Inspection on 07/01/2024 at 8:45 am. It was viewed 0 times while on Public Inspection.

If you are using public inspection listings for legal research, you should verify the contents of the documents against a final, official edition of the Federal Register. Only official editions of the Federal Register provide legal notice of publication to the public and judicial notice to the courts under 44 U.S.C. 1503 & 1507 . Learn more here .

Consumer Financial Protection Bureau.

Fair Lending Report of the Consumer Financial Protection Bureau.

The Consumer Financial Protection Bureau (CFPB) is issuing its eleventh Fair Lending Report of the Consumer Financial Protection Bureau (Fair Lending Report) to Congress. The CFPB is committed to ensuring fair, equitable, and nondiscriminatory access to credit for both individuals and communities. This report describes our fair lending activities in supervision and enforcement; guidance and rulemaking; interagency coordination; and outreach and education for calendar year 2023.

The CFPB released the 2023 Fair Lending Report on its website on June 26, 2024.

Susan Grutza, Senior Policy Counsel, Fair Lending, at 1-855-411-2372. If you require this document in an alternative electronic format, please contact [email protected] .

Because Congress charged the Consumer Financial Protection Bureau (CFPB) with the responsibility of overseeing many lenders and products, the CFPB has long used a risk-based approach to prioritizing supervisory examinations and enforcement activity. This approach helps ensure that the CFPB focuses on areas that present substantial risk of credit discrimination for consumers and small businesses. [ 1 ]

As part of the prioritization process, the CFPB identifies emerging developments and trends by monitoring key consumer financial markets. If this field and market intelligence identifies Start Printed Page 54787 fair lending risks in a particular market, that information is used to determine the type and extent of assets applied to address those risks.

The prioritization process incorporates a number of additional factors, including tips and leads from industry whistleblowers, advocacy groups, and government agencies; supervisory and enforcement history; consumer complaints; and results from analysis of Home Mortgage Disclosure Act (HMDA) and other data.

As a result of its annual risk-based prioritization process, in 2023 the CFPB focused much of its fair lending supervision efforts on: mortgage origination (including redlining, property valuation bias, and HMDA and Regulation C compliance); credit card marketing and the use of alternative data in digital marketing; and on the use of automated systems and models, sometimes marketed as artificial intelligence (AI) and machine learning models, in credit card originations.

As in previous years, the CFPB's 2023 mortgage origination work continued to focus on redlining (intentional discrimination against applicants and prospective applicants living or seeking credit in minority neighborhoods, including by discouragement). The CFPB's mortgage work also included assessing potential discrimination in mortgage underwriting and pricing processes, including assessing whether there were disparities in application, underwriting, and pricing processes, and whether there were weaknesses in fair lending-related compliance management systems. The CFPB's mortgage origination work also included reviewing residential property appraisal service providers to identify risks that may arise due to potential discrimination or bias as well as HMDA data integrity and validation reviews.

The CFPB's credit card work included assessing credit card lenders' digital marketing practices relating to credit cards, as well as credit card lenders' use of alternative data in that marketing. The CFPB's credit card work also included evaluation of automated systems and models, sometimes marketed as artificial intelligence and machine learning models, used by credit card lenders in credit card originations, as well as assessing whether there were disparities in application, underwriting, and pricing processes, and whether the institutions searched for less discriminatory alternatives to the models used.

Across multiple markets, the CFPB continued to assess whether lenders complied with the adverse action notice requirements of the Equal Credit Opportunity Act (ECOA) and Regulation B and evaluated whether lenders maintain policies and procedures that unlawfully exclude property on the basis of geography in underwriting decisions, unlawfully exclude certain types of income, and treat criminal history in an unlawful manner.

Congress authorized the CFPB to bring actions to enforce the requirements of eighteen enumerated statutes, including ECOA, HMDA, and the Consumer Financial Protection Act of 2010 (CFPA), which prohibits unfair, deceptive, and abusive acts or practices. The CFPB is able to engage in research, conduct investigations, file administrative complaints, hold hearings, and adjudicate claims through the CFPB's administrative enforcement process. The CFPB also uses its independent litigation authority to file cases in Federal court alleging violations of fair lending laws under the CFPB's jurisdiction. Like other Federal regulators, the CFPB is required to refer matters to the Department of Justice (DOJ) when it has reason to believe that a creditor has engaged in a pattern or practice of lending discrimination. [ 2 ]

In 2023, the CFPB announced two ECOA-related public enforcement actions, relating to discrimination on the basis of race and national origin, one against Citibank N.A. (Citibank) and the other against Colony Ridge Development, LLC, and Colony Ridge BV, LLC, and affiliate mortgage company Colony Ridge Land, LLC (collectively, the Colony Ridge defendants). For more information on these ECOA-related enforcement actions, see section 6.1.2 of this report.

HMDA, its implementing Regulation C, and Regulation B require mortgage lenders to report certain information about loan applications and originations to the CFPB and other Federal regulators. HMDA data are the most comprehensive source of publicly available information on the U.S. mortgage market. Both the public and regulators can use this information to monitor whether financial institutions are serving the housing needs of their communities, as well as to identify possible discriminatory lending patterns.

In 2023, the CFPB announced public enforcement actions against two repeat offenders for reporting false, erroneous, or incorrect HMDA data: Freedom Mortgage Corporation (Freedom Mortgage) and Bank of America, N.A.

The CFPB will continue to monitor the rate at which mortgage lenders fail to collect and report applicants' demographic information. The rate of nonreporting of demographic information has been increasing since 2019, potentially compromising the ability of the CFPB and other financial regulators, enforcement agencies, academics, other mortgage lenders, and civil rights and consumer advocates, to detect and remedy redlining, discouragement, and other forms of discrimination in the mortgage market. The CFPB's evaluations will include assessments of lenders' demographic reporting practices and HMDA compliance systems to ensure they are monitoring for inaccurate or incomplete demographic information reporting and complying with HMDA.

On October 10, 2023, the CFPB filed a lawsuit against Freedom Mortgage, a residential mortgage loan originator and servicer, alleging that it submitted legally-required mortgage loan data that were riddled with errors. [ 3 ] In 2020, Freedom Mortgage reported HMDA data on over 700,000 applications and originated nearly 400,000 HMDA-reportable loans worth almost $100 billion, making it the third largest mortgage lender in the United States by origination volume. Freedom Mortgage is a repeat offender: at the time the CFPB filed its complaint, Freedom was already under a CFPB Consent Order related to previous HMDA violations. In 2019, the CFPB issued an order against Freedom finding that it intentionally misreported certain HMDA data fields from at least 2014 to 2017. [ 4 ] In the CFPB's lawsuit, the CFPB alleges that the mortgage loan data for 2020 that Freedom Mortgage submitted contained widespread errors across multiple data fields, in violation of HMDA and Regulation C. The CFPB's complaint further alleges that by reporting inaccurate HMDA mortgage loan data for 2020, Freedom Mortgage also violated the 2019 order and the CFPA. The CFPB seeks appropriate injunctive relief and a civil money penalty.

On November 28, 2023, the CFPB issued an order against Bank of America for routinely submitting falsified HMDA data. [ 5 ] The CFPB found that between 2016 and late 2020, hundreds of Bank of America's loan officers failed to ask applicants for their race, ethnicity, and sex, as required by law, and instead falsely recorded that the applicants chose not to provide this information, in violation of HMDA, Regulation C, and the CFPA. The CFPB's order requires Bank of America to pay a $12 million civil money penalty and to develop policies and procedures to ensure compliance with HMDA and Regulation C, including recording and auditing phone applications to make sure that HMDA data are accurately collected and recorded.

The CFPB must refer to DOJ any matter when it has reason to believe that a creditor has engaged in a pattern or practice of lending discrimination in violation of ECOA. [ 6 ] The CFPB may refer other potential ECOA violations to DOJ as well. [ 7 ] In 2023, the CFPB referred 18 matters to DOJ pursuant to 15 U.S.C. 1691e(g) . More information on these referrals can be found in section 6.1.6 of this report.

When an enforcement action is resolved through a public enforcement order, the CFPB (together with other government entities, when relevant) takes steps to ensure that the respondent or defendant complies with the requirements of the order. Depending on the specific requirements of individual public enforcement orders, the CFPB may take steps to ensure that borrowers who are eligible for compensation receive remuneration and that the defendant has complied with the injunctive provisions of the order, including implementing a comprehensive fair lending compliance management system.

The CFPB's Supervision program assesses compliance with Federal consumer financial protection laws and regulations at banks and nonbanks over which the CFPB has supervisory authority. As a result of the CFPB's efforts to fulfill its fair lending mission during 2023, the CFPB initiated 28 fair lending examinations or targeted reviews.

In 2023, two of the most frequently identified fair lending issues in supervisory communications related to the granting of pricing exceptions and HMDA violations.

In 2023, the CFPB issued several fair lending-related Matters Requiring Attention and entered Memoranda of Understanding directing entities to take corrective actions that the CFPB will monitor through follow-up supervisory actions. In these communications, the CFPB directed mortgage lenders to correct violations relating to redlining, including by institutions providing consumer remediation designed to spur lending in redlined areas. The CFPB also directed lenders to enhance their fair lending compliance management systems in several ways, including by directing institutions to, when testing and approving credit scoring models, document the specific business needs the models serve, as well as document specific standards for assessing whether a model serves each stated business need. Further, the CFPB also directed the institutions to test credit scoring models for prohibited basis disparities and to require documentation of considerations the institutions will give to how to assess those disparities against the stated business needs. To ensure compliance with ECOA and Regulation B, institutions were directed to develop a process for the consideration of a range of less discriminatory models. Additionally, institutions were directed to test and validate the methodologies used to identify principal reasons in adverse action notices required under ECOA and Regulation B. Finally, institutions were directed to implement policies, procedures, and controls designed to effectively manage HMDA compliance, including regarding integrity of data collection.

During 2023, informed by the Director's priority to address risks of consumer harm from advanced and emerging technologies in consumer finance, the CFPB continued to increase its technical capacity and analyses to ensure that the use of this technology does not pose risks to consumers or violate Federal consumer financial law.

During 2023, the CFPB issued a final rule on small business lending data collection and issued a notice of proposed rulemaking on automated valuation models (AVMs).

The CFPB publishes an agenda of its planned rulemaking activity biannually, which is available at: https://www.consumerfinance.gov/​rules-policy/​regulatory-agenda .

In section 1071 of the Dodd-Frank Act, Congress directed the CFPB to adopt regulations governing the collection of small business lending data. [ 8 ] Section 1071 amended ECOA to require financial institutions to compile, maintain, and submit to the CFPB certain data on applications for credit for women-owned, minority-owned, and small businesses.

Congress enacted section 1071 for the purpose of facilitating enforcement of fair lending laws and enabling communities, governmental entities, and creditors to identify business and community development needs and opportunities for women-owned, minority-owned, and small businesses.

On March 30, 2023, the CFPB issued a final rule amending Regulation B to implement changes to ECOA made by section 1071 of the Dodd-Frank Act. [ 9 ] Consistent with section 1071, covered financial institutions are required to collect and report to the CFPB data on applications for credit for small businesses, including those that are owned by women or minorities. The rule also addresses the CFPB's approach to privacy interests and the publication of section 1071 data; shielding certain demographic data from underwriters and other persons; recordkeeping requirements; enforcement provisions; and the rule's effective and compliance dates.

In light of court orders in ongoing litigation, the CFPB has announced plans to extend the compliance dates in the small business lending rule. [ 10 ] More information about pending litigation is contained in section 5 of this report.

On June 1, 2023, the CFPB, along with its interagency partners, the Board of Governors of the Federal Reserve System (FRB), Office of the Comptroller of the Currency (OCC), Federal Deposit Insurance Corporation (FDIC), National Start Printed Page 54789 Credit Union Administration (NCUA), and Federal Housing Finance Agency (FHFA) (collectively, the Agencies) requested public comment on a proposed rule designed to ensure the credibility and integrity of models used in real estate valuations. [ 11 ] In particular, the proposed rule would implement quality control standards for AVMs used by mortgage originators and secondary market issuers in valuing real estate collateral securing mortgage loans. AVMs are used as part of the real estate valuation process, driven in part by advances in database and modeling technology and the availability of larger property datasets. While advances in AVM technology and data availability have the potential to contribute to lower costs and reduce loan cycle times, it is important that institutions using AVMs take appropriate steps to ensure the credibility and integrity of their valuations. It is also important that the AVMs that institutions are using adhere to quality control standards designed to comply with applicable nondiscrimination laws.

The proposed standards are designed to ensure a high level of confidence in the estimates produced by AVMs; help protect against the manipulation of data; seek to avoid conflicts of interest; require random sample testing and reviews; and promote compliance with applicable nondiscrimination laws.

The comment period for the proposed rule closed on August 21, 2023.

The CFPB issues guidance to its various stakeholders in many forms, including Consumer Financial Protection Circulars (Circulars), advisory opinions, interpretive rules, statements, bulletins, publications such as Supervisory Highlights.

On June 8, 2023, the CFPB, along with FRB, FDIC, NCUA, and OCC requested public comment on proposed guidance addressing reconsiderations of value (ROV) for residential real estate transactions. [ 12 ] ROVs are requests from a financial institution to an appraiser or other preparer of a valuation report to reassess the value of residential real estate. A ROV may be warranted if a consumer provides information to a financial institution about potential deficiencies or other information that may affect the estimated value. The proposed guidance advises on policies that financial institutions may implement to allow consumers to provide financial institutions with information that may not have been considered during an appraisal, or if deficiencies are identified in the original appraisal.

The comment period for the proposed guidance closed on September 19, 2023.

On September 19, 2023, the CFPB released a circular pertaining to certain legal requirements that lenders must adhere to, including when using artificial intelligence and other complex models. [ 13 ] The circular describes how, under ECOA and Regulation B, lenders must make available to an applicant a statement of specific and accurate reasons when taking adverse action against the applicant and cannot simply use the CFPB sample adverse action forms and checklists if they do not reflect the actual reason for the denial of credit or other adverse action. This requirement is especially important with the growth of advanced algorithms and personal consumer data in credit underwriting. The legal requirement to explain the reasons for adverse actions helps improve consumers' chances for future credit and protect consumers from illegal discrimination and serve an educational role, allowing consumers to understand the reasons for a creditor's action and take steps to improve their credit status or rectify mistakes made by creditors.

On June 5, 2023, the CFPB published a document affirming the extent to which ECOA and Regulation B apply with respect to franchisees seeking credit to finance their businesses. [ 14 ] Franchising is a significant portion of the small business ecosystem, and franchisees generally obtain credit either directly from the franchisor or from third party finance companies, which could be independent of the franchisor or brokered by or affiliated with the franchisor. These financing arrangements are likely “credit” and “business credit” under ECOA and Regulation B.

The CFPB's Supervisory Highlights reports provide general information about the CFPB's supervisory activities at banks and nonbanks without identifying specific entities. These reports communicate the CFPB's key examination findings and operational changes to the CFPB's supervision program. In 2023, the CFPB published three issues of Supervisory Highlights. [ 15 ]

The CFPB released the 30th edition of Supervisory Highlights on July 26, 2023, which covered examinations completed between July 1, 2022, and March 31, 2023. [ 16 ] This report included findings of ECOA and Regulation B violations in several areas, including pricing discrimination and discriminatory lending restrictions. Specifically, examiners found that mortgage lenders violated ECOA and Regulation B by discriminating in the incidence of granting pricing exceptions for competitive offers across a range of ECOA-protected characteristics, including race, national origin, sex, and age.

Additionally, this edition detailed examiners' findings on certain lending restrictions, including how lenders handled the treatment of applicants' criminal records. The use of criminal history in credit decisioning may create a heightened risk of violating ECOA and Regulation B. In this review, examiners uncovered risky policies and procedures relating to the use of criminal history information at several institutions in several areas of credit, including mortgage origination, auto lending, and credit cards, but most notably within small business lending.

Further, examiners identified institutions improperly treating income derived from public assistance. In some instances, lenders imposed stricter standards on income derived from public assistance programs, while in Start Printed Page 54790 other instances, institutions excluded income derived from certain public assistance programs.

In 2023, the CFPB issued two other editions of Supervisory Highlights, which pertained specifically to junk fees.

All issues of Supervisory Highlights are available at: https://www.consumerfinance.gov/​compliance/​supervisory-highlights/​ .

Given the importance of accurate HMDA data, including to the CFPB's fair lending mission and for transparency in the mortgage market, the CFPB maintains a comprehensive suite of resources on its public website to help filers fulfill their reporting requirements under HMDA and Regulation C and to allow others to evaluate and study mortgage lending. A complete accounting of the CFPB's materials for HMDA data users and filers can be found in Appendix A of this report.

The CFPB engages with external stakeholders, including Tribal governments, consumer advocates, civil rights organizations, industry, academia, and other government agencies. This engagement comes in varied forms, including disseminating the CFPB's work and policy priorities through blogs, press releases, or speeches, as well as reaching out directly to advocates and consumers through website updates and social media. The CFPB also regularly issues research and reports analyzing data and market conditions. To further an all-of-government approach to fair lending enforcement, the CFPB also participates in several interagency groups.

The CFPB regularly uses blog posts, statements, press releases, guides, brochures, social media, media interviews, and other tools to timely and effectively communicate with stakeholders.

In 2023, the CFPB published numerous blog posts relating to fair lending topics, including: the joint letter sent to The Appraisal Foundation, urging it to revise its draft ethics rule;  [ 17 ] the CFPB's Statement of Interest filed in Connolly & Mott v. Lanham et al. and the CFPB's commitment to ensuring fair and accurate appraisals;  [ 18 ] the CFPB's Statement of Interest filed in Roberson v. Health Career Institute LLC;   [ 19 ] an interagency proposed rulemaking on AVMs;  [ 20 ] a blog explaining how chatbots, including those supported by large language models and those marketed as AI can fail to provide adequate customer service;  [ 21 ] the CFPB's Amicus brief in Saint-Jean v. Emigrant Mortgage Company;   [ 22 ] the publication of the 2022 Annual Fair Lending Report to Congress;  [ 23 ] the CFPB's initiative to better understand the financial experiences of immigrants in the United States;  [ 24 ] and the Appraisal Subcommittee's November 1 public hearing to discuss the challenges and solutions to preventing bias in the home appraisal process. [ 25 ]

The CFPB also issued several press releases relating to fair lending topics, including announcements regarding: the availability of the 2022 HMDA modified loan application register data;  [ 26 ] the finalization of the small business lending rule, [ 27 ] the issuance of a joint statement confirming that automated systems and advanced technology is not an excuse for law-breaking behavior;  [ 28 ] the publication of the proposed AVM rule and request for public comment;  [ 29 ] an issue spotlight on AI chatbots in banking;  [ 30 ] the publication of two new reports on the financial opportunities and challenges facing Southern communities;  [ 31 ] the availability of 2022 HMDA data;  [ 32 ] a roundtable on special purpose credit programs (SPCPs);  [ 33 ] the issuance of Consumer Financial Protection Circular 2023-03, Adverse action notification requirements and the proper use of the CFPB's sample forms provided in Regulation B;  [ 34 ] the Freedom Mortgage enforcement action for reporting allegedly erroneous data under HMDA;  [ 35 ] the issuance of the CFPB and DOJ's joint statement reminding financial institutions that all credit applicants are protected from Start Printed Page 54791 discrimination on the basis of race, national origin, race, and other characteristics covered by ECOA, regardless of their immigration status;  [ 36 ] the publication of a new analysis on State Community Reinvestment Act (CRA) laws, highlighting how states ensure financial institutions' lending, services, and investment activities meet the credit needs of their communities;  [ 37 ] the Citibank enforcement action;  [ 38 ] the Bank of America enforcement action;  [ 39 ] and the Colony Ridge enforcement action. [ 40 ]

The CFPB often engages directly with external stakeholders to inform the CFPB's policy developments and message the CFPB's priorities and recent work. In 2023, CFPB staff participated in 69 stakeholder engagements related to fair lending and access to credit issues. Through speeches, presentations, podcasts, roundtables, webinars, and other smaller discussions on fair lending topics, the CFPB strives to keep abreast of economic and market realities that impact the lives of individuals, small businesses, and communities the CFPB is charged with protecting.

Throughout 2023, numerous engagements centered around the use of advanced technologies including their use in discriminatory targeting, consumer surveillance, and digital redlining; redlining; discrimination on the basis of receipt of public assistance income; false and erroneous HMDA data reporting; student lending; and credit reporting.

On November 2, 2023, the CFPB published a new analysis of state-specific versions of CRA laws, highlighting how States ensure financial institutions' lending, services, and investment activities meet the credit needs of their communities. Many States adopted laws similar to the Federal CRA in the decades following the 1977 passage of the landmark Federal anti-redlining law. The report examined the laws of seven States (Connecticut, Illinois, Massachusetts, New York, Rhode Island, Washington, West Virginia) and the District of Columbia, and found that data collected by Federal agencies, such as HMDA, are often used for State CRA compliance and other oversight purposes. [ 41 ]

On June 21, 2023, the CFPB published a data spotlight, Banking and Credit Access in the Southern Region of the U.S. [ 42 ] Spanning the States of Alabama, Arkansas, Georgia, Louisiana, Mississippi, North Carolina, South Carolina, and Tennessee, this report seeks to identify gaps as well as opportunities to increase financial access in the region, particularly through branch presence and bank account access, and capital access such as mortgage lending and small business lending. The analysis looks at trends by State, the region as a whole, and differences between rural and non-rural areas. Utilizing HMDA data, the analysis also identified differences for mortgage originations and denials by race and ethnicity in both rural and non-rural communities.

On June 21, 2023, the CFPB published a data spotlight, Consumer Finances in Rural Areas of the Southern Region. [ 43 ] This report is the second in a series profiling the finances of consumers in rural communities. Nearly 48 million people live in the southern region examined in this report, which includes Alabama, Arkansas, Georgia, Louisiana, Mississippi, North Carolina, South Carolina, and Tennessee. Intended to provide a starting point in better understanding the financial lives of consumers in rural areas of the southern United States, this report takes a broad survey of consumer financial profiles, including credit scores, financial distress, medical debt, and other debt categories and compares profiles of consumers in the rural South to those in other geographies. Among other things, the report examines originations for auto loans by credit score and majority-minority census tracts, by State and for the region as a whole.

On March 20, 2023, the CFPB announced the initial availability of the 2022 HMDA modified loan application register data on the Federal Financial Institutions Examination Council's (FFIEC) HMDA Platform for approximately 4,394 HMDA filers. [ 44 ] These published data contain loan-level information filed by financial institutions, modified to protect consumer privacy. [ 45 ]

On June 29, 2023, the FFIEC announced the availability of static “Snapshot” HMDA data, a static dataset of 2022 mortgage lending transactions at 4,460 financial institutions reported under HMDA as of May 1, 2023. [ 46 ] These data include a total of 48 data points providing information about the applicants, the property securing the loan or proposed to secure the loan in the case of non-originated applications, the transaction, and identifiers.

On March 3, 2023, the CFPB published a report containing the findings of the CFPB's voluntary review of the CFPB's final HMDA rule (issued in October 2015) and related Start Printed Page 54792 amendments (collectively, the HMDA Rule). [ 47 ] The report analyzed, among other key issues, how changes in reporting thresholds and other amendments affected HMDA data coverage and the available data on the supply over time of open-ended lines of credit and closed-end mortgage loans; how new or revised HMDA data points have contributed to predicting underwriting and pricing outcomes; and how revised and expanded reporting of race and ethnicity helped provide additional data on subpopulation groups in the residential mortgage market.

On September 27, 2023, the CFPB released its annual report on residential mortgage lending activity and trends for 2022. [ 48 ] The report shows that in 2022, mortgage applications and originations declined markedly from the prior year, while rates, fees, discount points, and other costs increased. Overall affordability declined significantly, with borrowers spending more of their income on mortgage payments and lenders more often denying applications for insufficient income. Most refinances during the reported period were cash-out refinances, and, in a reversal of recent trends, the median credit score of refinance borrowers declined below the median credit score of purchase borrowers. As in years past, independent lenders continued to dominate home mortgage lending, with the exception of home equity lines of credit.

The CFPB regularly coordinates with other Tribal, Federal, State, county, municipal, and international government entities; policymakers; and the organizations that represent them regarding current and emerging fair lending risks. Through numerous interagency organizations and taskforces, the CFPB coordinated its 2023 fair lending regulatory, supervisory, and enforcement activities to promote consistent, efficient, and effective enforcement of Federal fair lending laws.

The CFPB, along with the Department of Housing and Urban Development (HUD), Federal Trade Commission (FTC), FDIC, FRB, NCUA, OCC, DOJ, and FHFA, constitute the Interagency Task Force on Fair Lending. This Task Force meets regularly to discuss fair lending enforcement efforts, share current methods of conducting supervisory and enforcement fair lending activities, and coordinate fair lending policies. In 2023, the NCUA was the Chair of this Task Force.

Through the FFIEC, the CFPB has robust engagements with other partner agencies that focus on fair lending issues. For example, throughout the reporting period, the CFPB has continued to chair the HMDA and CRA Data Collection Subcommittee, a subcommittee of the FFIEC Task Force on Consumer Compliance. This subcommittee oversees FFIEC projects and programs involving HMDA data collection and dissemination, the preparation of the annual FFIEC budget for processing services, and the development and implementation of other related HMDA processing projects as directed by this Task Force.

Together with DOJ, HUD, and FTC, the CFPB also participates in the Interagency Working Group on Fair Lending Enforcement, a standing working group of Federal agencies that meets regularly to discuss issues relating to fair lending enforcement. The agencies use these meetings to also discuss fair lending developments and trends, methodologies for evaluating fair lending risks and violations, and coordination of fair lending enforcement efforts.

As required by section 1022 of the Dodd-Frank Act, the CFPB also consults with other agencies as part of its rulemaking process. [ 49 ] For example, in 2023, while developing its small business lending data collection final rule, the CFPB consulted or offered to consult with FRB, FDIC, NCUA, OCC, HUD, DOJ, FTC, the Department of Agriculture, the Department of the Treasury, the Economic Development Administration, the Farm Credit Administration (FCA), the Financial Crimes Enforcement Network, and the Small Business Administration (SBA) including, among other things, on consistency with any prudential, market, or systemic objectives administered by such agencies.

In addition to the established interagency organizations, CFPB personnel meet regularly with personnel from other agencies, including with DOJ, HUD, FTC, FHFA, State Attorneys General, and the prudential regulators to coordinate and discuss the CFPB's fair lending work.

On September 12, 2023, the CFPB, along with HUD, OCC, and FHFA hosted a roundtable discussion on SPCPs. [ 50 ] In addition to remarks by the respective leaders of the participating agencies, the event included a roundtable discussion with representatives from community groups and trade organizations that are focused on the opportunities and benefits of SPCPs. The event was open to the public via livestream.

On April 25, 2023, the CFPB, along with DOJ, Equal Employment Opportunity Commission, and FTC issued a joint statement committing to enforcement efforts against discrimination and bias in automated systems. [ 51 ] In the statement, all four agencies resolved to vigorously enforce their collective authorities and to monitor the development and use of automated systems, including those sometimes marketed as AI.

On October 12, 2023, the CFPB along with DOJ, issued a joint statement reminding financial institutions that, while ECOA and Regulation B do not expressly prohibit consideration of immigration status, they prohibit creditors from using immigration status to discriminate on the basis of national origin, race, or any other characteristic covered by ECOA. [ 52 ]

Appraisal bias is a key fair lending priority of the CFPB. Throughout 2023, the CFPB has been very active with its interagency partners to advance work to combat appraisal bias through the FFIEC Appraisal Subcommittee (ASC), correspondence, court briefs, proposed guidance, and work of the Property Start Printed Page 54793 Appraisal and Valuation Equity Task Force.

The ASC comprises designees from the CFPB and certain other Federal agencies, including FDIC, HUD, FRB, OCC, NCUA, and FHFA, and is tasked with providing Federal oversight of State appraiser and appraisal management company regulatory programs, as well as a monitoring and reviewing framework for The Appraisal Foundation, the private, nongovernmental organization that sets appraisal standards. CFPB Deputy Director Zixta Martinez currently serves as the Chairperson of the ASC. Through the ASC, the CFPB addresses topics including discriminatory bias in home appraisals.

The ASC held its first-ever hearing about appraisal bias on January 24, 2023. The hearing served to raise awareness of the issue of appraisal bias by focusing on its scope and impact, and to provide information on the role of the ASC in the appraisal regulatory system. On May 19, 2023, the ASC held its second public hearing, which explored the appraisal regulatory system and focused on appraisal standards, appraiser qualification criteria and barriers to entry into the profession, appraisal practice, and State regulation. The ASC held its third public hearing on November 1, 2023, which discussed how a residential appraisal is developed and reviewed, the ROV process for residential real estate valuations, and the development of rural appraisals. These hearings were the first three in a series of four planned hearings relating to appraisal bias.

On February 14, 2023, senior officials from the CFPB, FDIC, HUD, NCUA, FRB, DOJ, OCC, and FHFA submitted a joint letter to The Appraisal Foundation. The letter urged The Appraisal Foundation to revise its draft Ethics Rule for appraisers to include a detailed statement of Federal prohibitions against discrimination that exist under the FHA and ECOA. The agencies expressed concern that some appraisers may be unaware of these prohibitions and, of particular concern, that the draft Ethics Rule emphasized that “[a]n appraiser must not engage in unethical discrimination,” implying that appraisers may engage in “ethical” discrimination, a concept foreign to current law and practice.

On March 13, 2023, the CFPB filed a joint statement of interest with DOJ in Connolly & Mott v. Lanham et al., explaining the application of the FHA and ECOA to lenders relying on discriminatory home appraisals. For more information on this statement of interest, see section 5.1 of this report.

On June 1, 2023, the CFPB, in conjunction with the FRB, FDIC, FHFA, NCUA, and OCC, proposed a rule regarding quality control standards for AVMs. For more information on this rulemaking, see section 2.1.2 of this report.

On June 8, 2023, the CFPB, in conjunction with the FRB, FDIC, NCUA, and OCC, requested public comment on proposed guidance addressing ROV for residential real estate transactions. The proposed guidance would advise on policies that financial institutions may implement to allow consumers to provide financial institutions with information that may not have been considered during an appraisal or if deficiencies are identified in the original appraisal. For more information on this proposed guidance, see section 2.2.1 of this report.

In 2023, the CFPB also continued to engage with other agencies on issues of bias in home appraisals through the Interagency Task Force on Property Appraisal and Valuation Equity. More information on this Task Force is available at https://pave.hud.gov .

The CFPB files amicus, or “friend-of-the-court,” briefs in significant court cases concerning Federal consumer financial protection laws, including cases involving ECOA. These briefs provide courts with the CFPB's views and help ensure that consumer financial protection statutes are correctly and consistently interpreted. In 2023, the CFPB filed two fair lending related amicus briefs and a statement of interest.

On June 23, 2023, the CFPB filed an amicus brief in Saint-Jean et al. v. Emigrant Mortgage Co. & Emigrant Bank in support of Plaintiffs who won a jury verdict against Emigrant Mortgage Company and Emigrant Bank (Emigrant) for violating ECOA. [ 53 ] The jury found that Emigrant had for years targeted Black and Latino borrowers and neighborhoods in New York City with predatory mortgage loans and practices. The CFPB's brief addresses three issues raised on appeal to explain why the jury verdict should be affirmed: (1) the timeliness of Plaintiff's claims under the doctrine of equitable tolling, (2) the propriety of the district court's jury instructions under ECOA, and (3) the public policy goals undermined by enforcing a waiver of claims in a loan modification agreement. [ 54 ]

On April 14, 2023, the CFPB filed an amicus brief in Roberson v. Health Career Institute LLC. [ 55 ] In the brief, the CFPB explained that discriminatory targeting violates ECOA. In particular, the CFPB's brief explains that ECOA's prohibition on discrimination applies to “any aspect of a credit transaction,” meaning it covers every aspect of a borrower's dealings with a creditor, not just the specific terms of a loan—like the interest rate or fees. The CFPB's brief also explains that in order to survive a motion to dismiss under ECOA, plaintiffs need only plead facts that plausibly allege discrimination, rather than the elements of a prima facia case, which is not a pleading requirement but rather an evidentiary standard. [ 56 ]

On March 13, 2023, the CFPB and DOJ filed a joint statement of interest in Connolly & Mott v. Lanham et al. explaining that relying on discriminatory home appraisals can violate ECOA. [ 57 ] The law is clear that mortgage lenders cannot take race, sex, or any other prohibited bases into account when evaluating the creditworthiness of an applicant. As such, lenders cannot rely on a discriminatory appraisal if they knew, or should have known, that the appraisal was discriminatory. The statement of interest also explains that, to survive a motion to dismiss under ECOA, plaintiffs need only plead facts that plausibly allege discrimination, rather than establish a prima facie case, which is not a pleading requirement but rather an evidentiary standard. In the statement of interest, the Department of Justice also addresses how the FHA applies to discriminatory appraisals. [ 58 ]

More information regarding the CFPB's amicus program is available on the CFPB's website. [ 59 ]

In September 2022, the CFPB was sued in the U.S. District Court for the Start Printed Page 54794 Eastern District of Texas by the U.S. Chamber of Commerce et al., challenging an update to the UDAAP section of the CFPB's examination manual. The updated manual clarified that discriminatory conduct may violate the CFPA's prohibition on unfair practices and provided guidance to examiners on how discriminatory conduct should be examined to determine whether it violates the unfairness prohibition. The court granted plaintiffs' motion for summary judgment, vacated the manual update, and permanently enjoined the CFPB from engaging in any examination, supervision, or enforcement action against any member of the plaintiff associations based on the CFPB's interpretation of its unfairness authority set forth in the updated manual. The CFPB filed a notice of appeal in November 2023, and the appeal was stayed by the Fifth Circuit pending the Supreme Court's resolution of CFPB v. CFSA. [ 60 ]

On March 30, 2023, the CFPB issued its final rule on small business lending under ECOA, as required by section 1071 of the Dodd-Frank Act. [ 61 ] On April 26, 2023, the Texas Bankers Association and Rio Bank sued the CFPB in the U.S. District Court for the Southern District of Texas challenging the validity of the final rule. The court entered a preliminary injunction enjoining the CFPB from enforcing or implementing the rule against plaintiffs (including the American Bankers Association, who had joined as a plaintiff via an amended complaint filed on May 14, 2023) and their members, and stayed the compliance dates for plaintiffs and their members pending a decision in CFPB v. CFSA. [ 62 ] On October 26, 2023, the court extended that order to apply to all covered entities following the intervention of other plaintiffs seeking to join the lawsuit. Separately, on August 11, 2023, the Kentucky Bankers Association and several Kentucky banks sued to challenge the rule in the U.S. District Court for the Eastern District of Kentucky. The court preliminarily enjoined the CFPB from enforcing the rule pending a decision in CFPB v. CFSA. [ 63 ] A third lawsuit was filed on December 26, 2023, in the U.S. District Court for the Southern District of Florida by the Revenue Based Finance Coalition, a trade association representing merchant cash advance providers.

The CFPB is statutorily required to file a report to Congress annually describing the administration of its functions under ECOA, summarizing public enforcement actions taken by other agencies with administrative enforcement responsibilities under ECOA, and providing an assessment of the extent to which compliance with ECOA has been achieved. [ 64 ] In addition, the CFPB's annual HMDA reporting requirement calls for the CFPB, in consultation with HUD, to report annually on the utility of HMDA's requirement that covered lenders itemize certain mortgage loan data. [ 65 ] The information below provides the required reporting.

The enforcement and compliance efforts and assessments made by the eleven agencies assigned enforcement authority under section 704 of ECOA are discussed in this section, as reported by the agencies.

In 2023, of the Federal agencies with ECOA enforcement authority, the CFPB, FDIC, and FTC brought a total of four fair lending enforcement actions. Information on the DOJ's fair lending Start Printed Page 54796 program and fair lending related public enforcement actions can be found at: https://www.justice.gov/​crt/​fair-lending-program-0 .

In 2023, the CFPB brought two fair lending enforcement actions: Citibank and Colony Ridge.

On November 8, 2023, the CFPB ordered Citibank, N.A. to pay $25.9 million in fines and consumer redress for intentionally and illegally discriminating against credit card applicants the bank identified as Armenian American. [ 69 ] From at least 2015 through 2021, Citibank discriminated against retail services credit card applicants with surnames that Citibank employees associated with consumers of Armenian national origin, targeting applicants with surnames ending in “-ian” and “-yan” as well as applicants in or around Glendale, California. Nicknamed “Little Armenia,” Glendale is home to approximately 15 percent of the Armenian American population in the U.S. When Citibank identified credit card applicants as potentially being of Armenian national origin, the bank applied more stringent criteria to these applications, including denying them outright and requiring additional information or placing a block on the account.

Further, Citibank supervisors conspired to hide the discrimination by instructing employees not to discuss the discriminatory practices in writing or on recorded phone lines. Citibank employees also lied about the bases of denial, providing false reasons to denied applicants.

On December 20, 2023, the CFPB, together with DOJ, filed a complaint against the Texas-based Colony Ridge defendants. [ 70 ] The lawsuit alleges Colony Ridge sells unsuspecting families flood-prone land without water, sewer, or electrical infrastructure, and that the company sets borrowers up to fail with loans they cannot afford. As alleged in the complaint, roughly one in four Colony Ridge loans ends in foreclosure, after which the company repurchases the properties and sells them to new borrowers. As alleged in the complaint, Colony Ridge targets Hispanic borrowers. In particular, Colony Ridge advertises almost exclusively in Spanish, often on TikTok or other social media platforms, often featuring national flags and regional music from Latin America. In their marketing, Colony Ridge promised consumers the American dream of home ownership with its own seller financing: an easy-to-obtain loan product that requires no credit check and only a small deposit. The complaint alleges that foreclosure and property deed records from September 2019 through September 2022 show that Colony Ridge initiated foreclosures on at least 30 percent of seller-financed lots within just three years of the purchase date, with most loan failures occurring even sooner. Records also confirm that Colony Ridge accounted for more than 92 percent of all foreclosures recorded in Liberty County, Texas between 2017 and 2022.

In the complaint, the CFPB and DOJ allege that defendants violated ECOA by targeting consumers of Hispanic origin with a predatory loan product. The CFPB separately alleges that the Colony Ridge defendants violated the CFPA by making deceptive representations to consumers; that Colony Ridge Development and Colony Ridge BV violated the Interstate Land Sales Full Disclosure Act (ILSA) by making untrue statements, omitting material facts, failing to provide required accurate translations, and failing to report and disclose required information; and that defendants violated the CFPA by virtue of their violations of ECOA and ILSA, respectively. DOJ further alleges defendants' conduct violated the FHA.

The joint complaint seeks, among other things, injunctions against defendants to prevent future violations of Federal consumer financial laws, redress to consumers, damages, and the imposition of civil money penalties.

In 2023, the FTC, along with the State of Wisconsin, brought an enforcement action in Federal court against Rhinelander, a Wisconsin auto dealer group, its current and former owners, and general manager Daniel Towne, alleging, among other things, that defendants violated ECOA and Regulation B by discriminating against American Indian consumers by charging them higher financing costs and fees. [ 71 ] Among other things, the settlement with Rhinelander's current owners and Defendant Towne requires the company to establish a comprehensive fair lending program that will, among other components, allow consumers to seek outside financing for a purchase, and cap the additional interest markup Rhinelander can charge consumers, as well as require the current owners and Defendant Towne to pay $1 million to refund affected consumers. [ 72 ] The former owners, Rhinelander Auto Center, Inc. and Rhinelander Motor Company, agreed to a separate settlement, that requires the companies to permanently wind down the businesses and pay $100,000 to refund affected consumers. [ 73 ]

On March 8, 2023, the FDIC issued a public consent order for Cross River bank under section 3(q) of the Federal Deposit Insurance Act (Act), 12 U.S.C. 1813(q) . The FDIC determined that Cross River bank engaged in unsafe or unsound banking practices related to its compliance with applicable fair lending laws and regulations by failing to establish and maintain internal controls, information systems, and prudent credit underwriting practices in conformance with the Safety and Soundness Standards contained in appendix A of 12 CFR part 364 , or the violations of ECOA, 15 U.S.C. 1691 , et seq., as implemented by Regulation B, 12 CFR part 1002 , and the Truth in Lending Act, 15 U.S.C. 1601 , et seq., as implemented by Regulation Z, 12 CFR part 1026 .

In 2023, the agencies and the CFPB collectively reported citing 189 institutions with violations of ECOA and/or Regulation B.

Among institutions examined for compliance with ECOA and Regulation B, the FFIEC agencies reported that the most frequently cited violations were as follows: Start Printed Page 54797

Table 3—Regulation B Violations Cited by FFIEC Agencies, 2023

Regulation B violations: 2023FFIEC Agencies reporting , : Discrimination —Discrimination on a prohibited basis in a credit transaction; improperly requiring the signature of the applicant's spouse or other personNCUA, CFPB. , , : Inquiring about protected class —Inquiring about the race, color, religion, national origin, or sex of an applicant or any other person in connection with a credit transaction, except as permitted in sec. 1002.5(b)(1) and (b)(2), or sec.1002.8 in the case of a special purpose credit program; requesting any information concerning an applicant's spouse or former spouse, except as permitted in sec. 1002.5(c)(2); requesting the marital status of a person applying for individual, unsecured credit, except as permitted in sec. 1002.5(d)(1) (for credit other than individual, unsecured, a creditor may inquire about the applicant's marital status, but must only use the terms “married,” “unmarried,” and “separated”); inquiring as to whether income stated in an application is derived from alimony, child support, or separate maintenance payments, except as permitted in sec.1002.5(d)(2); or requesting information about birth control practices, intentions concerning the bearing or rearing of children, or capability to bear children, except as permitted in sec. 1002.5(d)(3)FDIC, OCC. , : Specific rules concerning use of information —Improperly evaluating age, receipt of public assistance in a credit transaction.CFPB. , , ; ; : Adverse Action —Failure to provide notice to the applicant 30 days after receiving a completed application concerning the creditor's approval of, counteroffer to, or adverse action on the application; failure to provide appropriate notice to the applicant 30 days after taking adverse action on an incomplete application; failure to provide sufficient information in an adverse action notification, including the specific reasons for the action takenOCC, NCUA, FRB, FDIC, CFPB. , : Information for Monitoring Purposes —Failure to obtain information for monitoring purposes; failure to request information on an application pertaining to the applicant's ethnicity or raceOCC. , , , : Appraisals and Valuations —Failure to provide appraisals and other valuationsOCC, NCUA.

Among institutions examined for compliance with ECOA and Regulation B, the Non-FFIEC agencies reported that the most frequently cited violations were as follows:

TABLE 4: Regulation B Violations Cited by Non-FFIEC Agencies Enforcing ECOA, 2023

Regulation B violations: 2023Non-FFIEC agencies reporting , : Adverse Action —Failure to provide notice to the applicant 30 days after receiving a completed application concerning the creditor's approval of, counteroffer to, or adverse action on the application; failure to provide sufficient information in an adverse action notification, including the specific reasons for the action takenFCA. : Failure to request and collect information for monitoring purposes —Failure to obtain information for monitoring purposesFCA.

The AMS, SEC, and the SBA reported that they received no complaints based on ECOA or Regulation B in 2023. The FTC is an enforcement agency and does not conduct compliance examinations.

The agencies assigned enforcement authority under section 704 of ECOA must refer a matter to DOJ when there is reason to believe that a creditor has engaged in a pattern or practice of lending discrimination in violation of ECOA. [ 85 ] They also may refer other potential ECOA violations to DOJ. [ 86 ] In 2023, 5 agencies (FDIC, NCUA, FRB, OCC and CFPB) collectively made 33 such referrals to DOJ involving discrimination in violation of ECOA. This is an increase of 175 percent in such referrals since 2020 (12 referrals). A brief description of those matters follows.

In 2023, the CFPB referred 18 fair lending matters to DOJ. The referrals included: discrimination on the basis of race and national origin in mortgage lending (redlining); discrimination in underwriting on the basis of receipt of public assistance income; predatory targeting on the basis of race and national origin; discrimination in pricing exceptions on the basis of race, national origin, sex, and age; and discrimination in credit cards on the basis of national origin and race.

In 2023, the FDIC referred seven fair lending matters to DOJ. The referrals included: one matter involving discrimination on the basis of race in mortgage lending (redlining); three matters for discrimination on the basis of race and national origin in mortgage lending (redlining); one matter involving discrimination in underwriting in commercial loans on the basis of race, color, national origin, and religion; one matter involving discrimination in auto loan pricing on the basis of sex or gender; and one matter for discrimination in auto loan pricing on the basis of race and national origin.

NCUA referred six ECOA matters to DOJ which involved discrimination Start Printed Page 54798 based on age and discrimination based on marital status.

The OCC made one referral to DOJ for a matter that involved discrimination on the basis of race, color, or national origin in mortgage lending (redlining).

The FRB referred one fair lending matter to DOJ. The matter involved discrimination on the basis of marital status in agricultural and commercial lending.

The CFPB's annual HMDA reporting requirement calls for the CFPB, in consultation with HUD, to report annually on the utility of HMDA's requirement that covered lenders itemize loan data in order to disclose the number and dollar amount of certain mortgage loans and applications, grouped according to various characteristics. [ 87 ] The CFPB, in consultation with HUD, finds that itemization and tabulation of these data furthers the purposes of HMDA.

The CFPB has made clear that the same laws and regulations apply to all technologies, regardless of the complexity or novelty of the technology deployed by institutions, including when it comes to combatting unlawful discrimination or explaining how certain credit decisions are made. ECOA is a powerful means to address unlawful digital discrimination in any aspect of a credit transaction. In 2023, the CFPB continued to combat digital discrimination through enforcement matters, [ 88 ] supervisory matters, [ 89 ] rulemaking, [ 90 ] guidance, [ 91 ] and using an all-of-government interagency approach. [ 92 ]

Looking forward, the CFPB will continue to enforce the law to root out unlawful discrimination, including when discrimination may be disguised by other processes within credit transactions. This includes actions that financial institutions take around the selection and procurement of data for use in advanced technological methods. Data brokers sell myriad types of personal data and sensitive information about consumers, some of which may directly implicate protected bases under ECOA. These data, alone or in combination with other data, may create proxies for, or have a disparate impact on, any of the ECOA prohibited bases. Creditors subject to ECOA and Regulation B may violate these laws if they use these data to engage in discriminatory targeting, steering, redlining, or in other ways that create unlawful discrimination.

The same holds true for fraud screens purported to facilitate compliance with other consumer protection and banking laws. While fraud detection compliance regimes may serve important purposes, institutions that are subject to ECOA and Regulation B may not use fraud screens and associated policies and procedures as an excuse to violate or circumvent fair lending laws. [ 93 ]

Further, as the CFPB continues to monitor markets and institutions for fair lending compliance, the CFPB will also continue to review the fair lending testing regimes of financial institutions. Robust fair lending testing of models should include regular testing for disparate treatment and disparate impact, including searches for and implementation of less discriminatory alternatives using manual or automated techniques. CFPB exam teams will continue to explore the use of open-source automated debiasing methodologies to produce potential alternative models to the insitutions' credit scoring models.

In 2024 and beyond, the CFPB will continue to combat digital discrimination and also continue to take steps to be a leader when it comes to building the Federal government's capabilities to address these types of transformative technologies.

As stated in section 2.2.5, the CFPB maintains a comprehensive suite of resources pertaining to the reporting and use of HMDA data, in addition to the annual HMDA filing guides released annually by the FFIEC. These resources include: Executive Summaries of HMDA rule changes;  [ 94 ] Small Entity Compliance Guide;  [ 95 ] Institutional and Transactional Coverage Charts;  [ 96 ] Reportable HMDA Data Chart;  [ 97 ] sample data collection form;  [ 98 ] FAQs;  [ 99 ] a Beginners Guide to Accessing and Using HMDA Data;  [ 100 ] and downloadable webinars, [ 101 ] which provide an overview of the HMDA rule. In June of 2023, the CFPB published a summary of the 2022 data on mortgage lending. [ 102 ] The CFPB also provides on its website an interactive version of Regulation C that is easier to access and navigate than the printed version of Regulation C. [ 103 ]

Together with the Federal Financial Institutions Examination Council Start Printed Page 54799 (FFIEC), [ 104 ] the CFPB also routinely updates its HMDA resources throughout the year to ensure HMDA reporters have the most up-to-date information. For example, in November 2023, the CFPB released the 2024 Filing Instructions Guide, [ 105 ] an online interactive Filing Instructions Guide, [ 106 ] and the 2023 Supplemental Guide for Quarterly Filers. [ 107 ] Together with the FFIEC, in March of 2023, the CFPB also published the 2023 edition of the HMDA Getting it Right Guide. [ 108 ] The CFPB also works with the FFIEC to publish data submission resources for HMDA filers and vendors on its Resources for HMDA Filers website, https://ffiec.cfpb.gov .

In addition, HMDA reporters can ask questions about HMDA and Regulation C, including how to submit HMDA data, by emailing the CFPB's HMDA Help at [email protected] . The CFPB also offers financial institutions, service providers, and others informal staff guidance on specific questions about the statutes and rules the CFPB implements, including ECOA and Regulation B and HMDA and Regulation C, through its Regulation Inquiries platform at www.reginquiries.consumerfinance.gov .

TermDefinitionAMSAgricultural Marketing Service of the U.S. Department of Agriculture.ASCFFIEC's Appraisal Subcommittee.AVMAutomated Valuation Models.CFPAConsumer Financial Protection Act of 2010.CFPBConsumer Financial Protection Bureau.CRACommunity Reinvestment Act.Dodd-Frank ActDodd-Frank Wall Street Reform and Consumer Protection Act.DOJU.S. Department of Justice.DOTU.S. Department of Transportation.ECOAEqual Credit Opportunity Act.FCAFarm Credit Administration.FDICFederal Deposit Insurance Corporation.FHAFair Housing Act.FHFAFederal Housing Finance Agency.Federal Reserve Board or FRBBoard of Governors of the Federal Reserve System.FFIECFederal Financial Institutions Examination Council—the FFIEC member agencies are the Board of Governors of the Federal Reserve System ( ), the Federal Deposit Insurance Corporation ( ), the National Credit Union Administration ( ), the Office of the Comptroller of the Currency ( ), and the Consumer Financial Protection Bureau (CFPB). The State Liaison Committee was added to FFIEC in 2006 as a voting member.FTCFederal Trade Commission.HMDAHome Mortgage Disclosure Act.HUDU.S. Department of Housing and Urban Development.ILSAInterstate Land Sales Full Disclosure Act.NCUANational Credit Union Administration.OCCOffice of the Comptroller of the Currency.ROVReconsideration of Value.SBASmall Business Administration.SECSecurities and Exchange Commission.SPCPSpecial Purpose Credit Program.UDAAPUnfair, Deceptive, or Abusive Acts or Practices.USDAU.S. Department of Agriculture.

The Director of the Bureau, Rohit Chopra, having reviewed and approved this document, is delegating the authority to electronically sign this document to Laura Galban, a Bureau Federal Register Liaison, for purposes of publication in the Federal Register .

Laura Galban,

Federal Register Liaison, Consumer Financial Protection Bureau.

1.   See Risk-Based Approach to Examinations, Supervisory Highlights Summer 2013 at 23, https://files.consumerfinance.gov/​f/​201308_​cfpb_​supervisory-highlights_​august.pdf , for additional information regarding the CFPB's risk-based approach in prioritizing supervisory examinations.

2.   See 15 U.S.C. 1691e(g) .

3.   See https://www.consumerfinance.gov/​enforcement/​actions/​freedom-mortgage-corporation-hmda-2023/​ .

4.   See https://files.consumerfinance.gov/​f/​documents/​cfpb_​freedom-mortgage-corporation_​consent-order_​2019-05.pdf .

5.   See https://www.consumerfinance.gov/​enforcement/​actions/​bank-of-america-na-hmda-data-2023/​ .

6.   15 U.S.C. 1691e(g) .

7.   Id.

8.   15 U.S.C. 1691c-2 .

9.  CFPB, Small Business Lending under the Equal Credit Opportunity Act (Regulation B) (Mar. 30, 2023), https://www.consumerfinance.gov/​rules-policy/​final-rules/​small-business-lending-under-the-equal-credit-opportunity-act-regulation-b/​ .

10.  More information is available at: https://www.consumerfinance.gov/​1071-rule/​ , a page compiling key materials related to the CFPB's small business rulemaking, including information on the interim final rule to extend compliance deadlines.

11.  CFPB, OCC, FHFA, FRB, FDIC, NCUA. Quality Control Standards for Automated Valuation Models (June 1, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​automated-valuation-models_​proposed-rule-request-for-comment_​2023-06.pdf .

12.  CFPB, OCC, FRB, FDIC, NCUA, Interagency Guidance on Reconsideration of Value of Residential Real Estate Valuations (June 8, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​interagency-guidance-reconsiderations-of-value-of-residential-real-estate_​2023-06.pdf .

13.  CFPB, Consumer Financial Protection Circular 2023-03 Adverse action notification requirements and the proper use of the CFPB's sample forms provided in Regulation B (Sept. 19. 2023), https://www.consumerfinance.gov/​compliance/​circulars/​circular-2023-03-adverse-action-notification-requirements-and-the-proper-use-of-the-cfpbs-sample-forms-provided-in-regulation-b/​ .

14.  CFPB, Coverage of Franchise Financing Under the Equal Credit Opportunity Act, Including the Small Business Lending Rule (May 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​coverage-of-franchise-financing_​2023-05.pdf .

15.  CFPB Issue 29, Junk Fees Special Edition, Winter 2023; Issue 30, Summer 2023; Issue 31, Junk Fees Update Special Edition Fall 2023.

16.  CFPB, Issue 30, Summer 2023 (July 31, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​supervisory-highlights_​issue-30_​2023-07.pdf .

17.  Patrice Alexander Ficklin and Tim Lambert, Appraisal standards must include federal prohibitions against discrimination (Feb. 14, 2023), https://www.consumerfinance.gov/​about-us/​blog/​appraisal-standards-must-include-federal-prohibitions-against-discrimination .

18.  Seth Frotman, Zixta Q. Martinez, and Jon Seward, Protecting homeowners from discriminatory home appraisals, (Mar. 13, 2023), https://www.consumerfinance.gov/​about-us/​blog/​protecting-homeowners-from-discriminatory-home-appraisals/​ .

19.  Seth Frotman, Protecting people from discriminatory targeting (Apr. 14, 2023), https://www.consumerfinance.gov/​about-us/​blog/​protecting-people-from-discriminatory-targeting/​ .

20.  Rohit Chopra, Algorithms, artificial intelligence, and fairness in home appraisals (June 1, 2023), https://www.consumerfinance.gov/​about-us/​blog/​algorithms-artificial-intelligence-fairness-in-home-appraisals/​ .

21.  Eric Halperin and Lorelei Salas, The CFPB has entered the chat (June 7, 2023), https://www.consumerfinance.gov/​about-us/​blog/​cfpb-has-entered-the-chat/​ .

22.  Seth Frotman, Protecting consumers' right to challenge discrimination (June 26, 2023), https://www.consumerfinance.gov/​about-us/​blog/​protecting-consumers-right-to-challenge-discrimination/​ .

23.  Patrice Alexander Ficklin, The CFPB's 2022 fair lending annual report to congress (June 29, 2023), https://www.consumerfinance.gov/​about-us/​blog/​the-cfpbs-2022-fair-lending-annual-report-to-congress/​ .

24.  Sonia Lin, Protecting immigrant access to fair credit opportunities, (Oct. 12, 2023), https://www.consumerfinance.gov/​about-us/​blog/​protecting-immigrant-access-to-fair-credit-opportunities/​ .

25.  CFPB, Next public hearing on appraisal bias: November 1 (Oct. 23, 2023), https://www.consumerfinance.gov/​about-us/​blog/​next-public-hearing-on-appraisal-bias-november-1/​ .

26.  CFPB, 2022 HMDA Data on Mortgage Lending Now Available (Mar. 20, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​2022-hmda-data-on-mortgage-lending-now-available/​ .

27.  CFPB, CFPB Finalizes Rule to Create a New Data Set on Small Business Lending in America (Mar. 30, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-finalizes-rule-to-create-a-new-data-set-on-small-business-lending-in- america/.

28.  CFPB, CFPB and Federal Partners Confirm Automated Systems and Advanced Technology Not an Excuse for Lawbreaking Behavior (Apr. 25, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-federal-partners-confirm-automated-systems-advanced-technology-not-an-excuse-for-lawbreaking-behavior/​ .

29.  CFPB, Agencies Request Comment on Quality Control Standards for Automated Valuation Models Proposed Rule (June 1, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​agencies-request-comment-on-quality-control-standards-for-automated-valuation-models-proposed-rule .

30.  CFPB, CFPB Issue Spotlight Analyzes “Artificial Intelligence” Chatbots in Banking (June 6, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-issue-spotlight-analyzes-artificial-intelligence-chatbots-in-banking/​ .

31.  CFPB, CFPB Releases Reports on Banking Access and Consumer Finance in Southern States (June 20, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-releases-reports-on-banking-access-and-consumer-finance-in-southern-states/​ .

32.  CFPB, FFIEC Announces Availability of 2022 Data on Mortgage Lending (June 29, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​ffiec-announces-availability-of-2022-data-on-mortgage-lending/​ .

33.  CFPB, Agencies to Host Roundtable on Special Purpose Credit Programs (Aug. 24, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​agencies-to-host-roundtable-on-special-purpose-credit-programs/​ .

34.  CFPB, CFPB Issues Guidance on Credit Denials by Lenders Using Artificial Intelligence (Sept. 19, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-issues-guidance-on-credit-denials-by-lenders-using-artificial-intelligence/​ .

35.  CFPB, CFPB Sues Repeat Offender Freedom Mortgage Corporation for Providing False Information to Federal Regulators (Oct. 10, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-sues-repeat-offender-freedom-mortgage-corporation-for-providing-false-information-to-federal-regulators/​ .

36.  CFPB, CFPB and Justice Department Issue Joint Statement Cautioning that Financial Institutions May Not Use Immigration Status to Illegally Discriminate Against Credit Applicants (Oct. 12, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-and-justice-department-issue-joint-statement-cautioning-that-financial-institutions-may-not-use-immigration-status-to-illegally-discriminate-against-credit-applicants/​ .

37.  CFPB, CFPB Issues New Report on State Community Reinvestment Laws (Nov. 2, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-issues-new-report-on-state-community-reinvestment-laws/​ .

38.  CFPB, CFPB Orders Citi to Pay $25.9 Million for Intentional, Illegal Discrimination Against Armenian Americans (Nov. 8, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-orders-citi-to-pay-25-9-million-for-intentional-illegal-discrimination-against-armenian- americans/.

39.  CFPB, CFPB Orders Bank of America to Pay $12 Million for Reporting False Mortgage Data (Nov. 28, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-orders-bank-of-america-to-pay-12-million-for-reporting-false-mortgage-data/​ .

40.  CFPB, C FPB and Justice Department Sue Developer and Lender Colony Ridge for Bait-and-Switch Land Sales and Predatory Financing (Dec. 20, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​cfpb-and-doj-sue-developer-and-lender-colony-ridge-for-bait-and-switch-land-sales-and-predatory-financing/​ .

41.  CFPB, State Community Reinvestment Act: Summary of State Laws (Nov. 2, 2023), https://www.consumerfinance.gov/​data-research/​research-reports/​state-community-reinvestment-acts-summary-of-state-laws/​ .

42.  CFPB, Banking and Credit Access in the Southern Region of the U.S. (June 21, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​ocp-data-spotlight_​banking-and-credit-access_​2023-06.pdf .

43.  CFPB, Consumer Finances in Rural Areas of the Southern Region (June 21, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​or-data-point_​consumer-finances-in-rural-south_​2023-06.pdf .

44.  CFPB, 2021 HMDA Data on Mortgage Lending Now Available (Mar. 20, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​2022-hmda-data-on-mortgage-lending-now-available/​ .

45.  Additional activity has occurred since the close of this reporting period. On March 26, 2024, the CFPB announced the availability of the HMDA modified loan application data for 2023, available at https://ffiec.cfpb.gov/​data-publication/​modified-lar/​2023 .

46.  CFPB, FFIEC Announces Availability of 2022 Data on Mortgage Lending (June 29, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​ffiec-announces-availability-of-2022-data-on-mortgage-lending/​ .

47.  CFPB, Report on the Home Mortgage Disclosure Act Rule Voluntary Review (Mar. 3, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​hmda-voluntary-review_​2023-03.pdf .

48.  CFPB, Data Point: 2022 Mortgage Market Activity and Trends (Sept. 27, 2022), https://www.consumerfinance.gov/​data-research/​research-reports/​data-point-2022-mortgage-market-activity-trends/​ .

49.   12 U.S.C. 5512 .

50.  CFPB, Agencies to Host Roundtable on Special Purpose Credit Programs, (Aug. 24, 2023), https://www.consumerfinance.gov/​about-us/​newsroom/​agencies-to-host-roundtable-on-special-purpose-credit-programs/​ .

51.  CFPB; Dept. of Justice Civil Rights Div.; Equal Opportunity Comm'n; Federal Trade Comm'n; Joint Statement on Enforcement Efforts Against Discrimination and Bias in Automated Systems, (Apr. 25, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​joint-statement-enforcement-against-discrimination-bias-automated-systems_​2023-04.pdf .

52.  CFPB; Dept. of Justice Civil Rights Div. Joint Statement on Fair Lending and Credit Opportunities for Noncitizen Borrowers under the Equal Credit Opportunity Act (Oct. 12, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb-joint-statement-on-fair-lending-and-credit-opportunities-for-noncitizen-b_​jA2oRDf.pdf .

53.  Brief for CFPB as Amici Curiae Supporting Plaintiff-Appellees, Saint-Jean v. Emigrant Mortg. Co., 50 F. Supp. 3d 300 (E.D.N.Y. 2014) (No. 22-3094).

54.   See https://files.consumerfinance.gov/​f/​documents/​cfpb_​saint-jean-et-al-v-emigrant-mortgage-coemigrant-bank_​2023-06.pdf .

55.  Statement of Interest of the CFPB in Support of Plaintiffs, Roberson et al v. Health Career Institute LLC, et al. (S.D.Fla. 2023) (No. 9:22CV81883).

56.   See https://www.consumerfinance.gov/​compliance/​amicus/​briefs/​roberson-v-health-career-institute-llc/​ .

57.  Statement of Interest for the United States, Connolly et al. v. Lanham et al., 685 F.Supp.3d 312 (No. 1:22CV02048).

58.   See https://www.consumerfinance.gov/​compliance/​amicus/​briefs/​connolly-mott-v-lanham-et-al/​ .

59.   See generally https://www.consumerfinance.gov/​policy-compliance/​amicus/​ .

60.  Additional activity has occurred since the close of this reporting period. On May 16, 2024, the Supreme Court issued a decision in CFPB v. CFSA. See CFPB v. Cmty. Fin. Servs. Ass'n of Am., Ltd., 601 U.S. 416 (2024).

61.   See https://www.consumerfinance.gov/​rules-policy/​final-rules/​small-business-lending-under-the-equal-credit-opportunity-act-regulation-b/​ .

62.  Additional activity has occurred since the close of this reporting period. See n.65, supra.

63.  Additional activity has occurred since the close of this reporting period. See n.65, supra.

64.   15 U.S.C. 1691f .

65.   12 U.S.C. 2807 .

66.  Collectively, the Board of Governors of the Federal Reserve System (FRB), the Federal Deposit Insurance Corporation (FDIC), the National Credit Union Administration (NCUA), the Office of the Comptroller of the Currency (OCC), and the Consumer Financial Protection Bureau (Bureau) comprise the Federal Financial Institutions Examination Council (FFIEC). The State Liaison Committee was added to FFIEC in 2006 as a voting member. Federal Financial Institutions Examination Council, http://www.ffiec.gov (last visited Mar. 30, 2021).>

67.  The Grain Inspection, Packers and Stockyards Administration (GIPSA) was eliminated as a stand-alone agency within USDA in 2017. The functions previously performed by GIPSA have been incorporated into the Agricultural Marketing Service (AMS), and ECOA reporting comes from the Packers and Stockyards Division, Fair Trade Practices Program, AMS.

68.   15 U.S.C. 1691c .

69.   See https://www.consumerfinance.gov/​enforcement/​actions/​citibank-n-a/​ .

70.   See https://www.consumerfinance.gov/​enforcement/​actions/​colony-ridge/​ .

71.   FTC v. Rhinelander Auto Ctr., Inc., No. 23-cv-737 (W.D. Wis., filed Oct. 24, 2023), available at https://www.ftc.gov/​system/​files/​ftc_​gov/​pdf/​1-ComplaintbyFTC-WIagainstRhinelander.pdf .

72.   FTC v. Rhinelander Auto Ctr., Inc., No. 23-cv-737 (W.D. Wis. Nov. 6, 2023) (stipulated order for permanent injunction, monetary judgment, and other relief as to Defendants Rhinelander Auto Group LLC, Rhinelander Import Group LLC, and Daniel Towne), https://www.ftc.gov/​system/​files/​ftc_​gov/​pdf/​18-ConsentJudgmentEnteredastoRAGRMGandTowne.pdf .

73.   FTC v. Rhinelander Auto Ctr., Inc., No. 23cv737 (W.D. Wis. Nov. 6, 2023) (stipulated order for permanent injunction, monetary judgment, and other relief as to Defendants Rhinelander Auto Center, Inc., and Rhinelander Motor Company), https://www.ftc.gov/​system/​files/​ftc_​gov/​pdf/​17-ConsentJudgmentEnteredastoRACandRMC.pdf .

74.   12 CFR 1002.4(a) .

75.   12 CFR 1002.4 , 1002.7(d)(1) .

76.   12 CFR 1002.5(b)-(d) .

77.   12 CFR 1002.5(b) .

78.   12 CFR 1002.9(a)(2) ; 1002.9(a)(1)(i) ; 1002.9(b)(2) .

79.   12 CFR 1002.9(a)(1) ; 1002.9(a)(2) ; 1002.9(b)(2) .

80.   12 CFR 1002.9(a)(1)(i) ; 1002.9(b)(2) .

81.   12 CFR 1002.9(a)(1) ; (a)(2) ; (b)(2) .

82.   12 CFR 1002.9(a)(1) , (2) ; 1002.9(b) ; 1002.9(c) .

83.   12 CFR 1002.14(a)(1) ; 1002.14(a)(2) .

84.   12 CFR 1002.14(a)(2) .

85.   15 U.S.C. 1691e(g) .

86.   Id.

87.   12 U.S.C. 2807 .

88.   See https://www.consumerfinance.gov/​enforcement/​actions/​colony-ridge/​ .

89.   See section 1.3, supra.

90.  CFPB, OCC, FHFA, FRB, FDIC, NCUA, Quality Control Standards for Automated Valuation Models (June 1, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​automated-valuation-models_​proposed-rule-request-for-comment_​2023-06.pdf .

91.  CFPB, Consumer Financial Protection Circular 2023-03 Adverse action notification requirements and the proper use of the CFPB's sample forms provided in Regulation B (Sept. 19. 2023), https://www.consumerfinance.gov/​compliance/​circulars/​circular-2023-03-adverse-action-notification-requirements-and-the-proper-use-of-the-cfpbs-sample-forms-provided-in-regulation-b/​ .

92.  CFPB, OCC, FRB, FDIC, NCUA, Interagency Guidance on Reconsideration of Value of Residential Real Estate Valuations (June 8, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​interagency-guidance-reconsiderations-of-value-of-residential-real-estate_​2023-06.pdf ; CFPB; Dept. of Justice Civil Rights Div.; Equal Opportunity Comm'n; Federal Trade Comm'n; Joint Statement on Enforcement Efforts Against Discrimination and Bias in Automated Systems (Apr. 25, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​joint-statement-enforcement-against-discrimination-bias-automated-systems_​2023-04.pdf .

93.  Discrimination on a prohibited basis can violate ECOA and Regulation B when it occurs in any aspect of a credit transaction, including when it occurs through practices that entities may characterize as related to fraud detection. See, e.g., https://www.consumerfinance.gov/​enforcement/​actions/​citibank-n-a/​ .

94.  CFPB, Executive Summary of the 2020 Home Mortgage Disclosure Act (Regulation C) Final Rule (Apr. 16, 2020), https://files.consumerfinance.gov/​f/​documents/​cfpb_​rule-executive-summary_​hmda-2020.pdf ; https://www.consumerfinance.gov/​about-us/​blog/​changes-to-hmda-closed-end-loan-reporting-threshold/​ . Summaries for different reporting years are available at: https://www.consumerfinance.gov/​compliance/​compliance-resources/​mortgage-resources/​hmda-reporting-requirements/​ .

95.  CFPB, Home Mortgage Disclosure (Regulation C) Small Entity Compliance Guide (Feb. 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​hmda_​small-entity-compliance-guide_​2023-02.pdf .

96.  CFPB, HMDA Institutional Coverage Chart, https://files.consumerfinance.gov/​f/​documents/​cfpb_​hmda-institutional-coverage_​2023.pdf ; CFPB, HMDA Transactional Coverage Chart, https://files.consumerfinance.gov/​f/​documents/​cfpb_​hmda-transactional-coverage_​2023.pdf .

97.  CFPB, Reportable HMDA Data: A Regulatory and Reporting Overview Reference Chart for HMDA Data Collected in 2023 (Feb. 9, 2023), https://files.consumerfinance.gov/​f/​documents/​cfpb_​reportable-hmda-data_​regulatory-and-reporting-overview-reference-chart_​2023-02.pdf .

98.  CFPB, Sample Data Collection Form, https://files.consumerfinance.gov/​f/​documents/​201708_​cfpb_​hmda-sample-data-collection-form.pdf .

99.  CFPB, Home Mortgage Disclosure Act FAQs, https://www.consumerfinance.gov/​compliance/​compliance-resources/​mortgage-resources/​hmda-reporting-requirements/​home-mortgage-disclosure-act-faqs/​ .

100.  CFPB, A Beginner's Guide to Accessing and Using Home Mortgage Disclosure Act Data (June 13, 2022), https://files.consumerfinance.gov/​f/​documents/​cfpb_​beginners-guide-accessing-using-hmda-data_​guide_​2022-06.pdf .

101.  CFPB, HMDA Webinars, https://www.consumerfinance.gov/​compliance/​compliance-resources/​mortgage-resources/​hmda-reporting-requirements/​webinars/​ .

102.  CFPB, Summary of 2022 Data on Mortgage Lending (June 29, 2023), https://www.consumerfinance.gov/​data-research/​hmda/​summary-of-2022-data-on-mortgage-lending/​ .

103.   Interactive Bureau Regulations, Regulation C, https://www.consumerfinance.gov/​rules-policy/​regulations/​1003/​ .

104.  Collectively, the Board of Governors of the Federal Reserve System (FRB), the Federal Deposit Insurance Corporation (FDIC), the National Credit Union Administration (NCUA), the Office of the Comptroller of the Currency (OCC), and the CFPB comprise the Federal Financial Institutions Examination Council (FFIEC). The State Liaison Committee was added to FFIEC in 2006 as a voting member. Federal Fin. Instit. Examination Council, http://www.ffiec.gov (last visited June 5, 2024).

105.  CFPB, Filing instructions guide for HMDA data collected in 2024 (Nov. 2023), https://s3.amazonaws.com/​cfpb-hmda-public/​prod/​help/​2024-hmda-fig.pdf .

106.  2023 FIG (Filing Instructions Guide), https://ffiec.cfpb.gov/​documentation/​fig/​2023/​overview .

107.  CFPB, Supplemental Guide for Quarterly Filers for 2024 (Aug. 2023), https://s3.amazonaws.com/​cfpb-hmda-public/​prod/​help/​supplemental-guide-for-quarterly-filers-for-2024.pdf .

108.  Federal Fin. Instit. Examination Council, A Guide to HMDA Reporting, Getting it Right! (Mar. 23, 2023), https://www.ffiec.gov/​hmda/​pdf/​2023Guide.pdf .

BILLING CODE 4810-AM-P

[ FR Doc. 2024-14533 Filed 7-1-24; 8:45 am]

• Executive Orders

Reader Aids

Information.

• About This Site
• Accessibility
• No Fear Act
• Continuity Information

Trading Relationships ( AQA A Level Geography )

Revision note.

Geography Content Creator

Trading Relationships

• Trading relationships between HDE , EME and LDE countries tend to follow a similar pattern.
• They specialised in producing high-tech products that require money and expertise
• They are wealthier so people have more disposable income to spend on goods
• They are more likely to have trade agreements facilitating trade
• They have better infrastructure to make trade quicker and easier
• They have lower labour costs making it cheaper to manufacture products – this attracts FDI
• They are experiencing rapid econ omic growth which is creating demand for more products as incomes rise
• They often have large and growing populations which are creating new consumer markets
• These factors help to explain why China is now the world’s largest exporter of goods and second-largest importer of goods
• They are less likely to be well-connected with infrastructure to manufacture and transport goods
• They have lower GDP meaning they lack the capital to invest in infrastructure and their consumer markets are smaller due to lower disposable income
• They are more likely to be suffering political instability which could deter FDI
• Although they are starting to trade more, growth has been much slower than for EME countries
• LDE countries rely mostly on the export of primary commodities whereas HDE and EME countries rely more on the export of secondary commodities

Examples of the most common goods traded between countries

It is important to recognise that the trading relationship between LDE , EME and HDE countries makes it difficult for LDE countries to achieve significant economic growth due to lack of access to markets and restrictions that prevent them from producing more high-value secondary commodities e.g. EU places higher tariffs on imports of roasted nuts compared to imports of raw nuts which can make it difficult for LDE countries to access the market for processed goods

Impacts of Trade in Metals

• There has been a dramatic increase in demand for metals, driven mostly by EME countries such as China which now dominates consumption of iron ore, copper, steel and other metals
• This has led to a shift in trading relationships, with EME countries becoming more influential and able to manipulate trade to their advantage
• This has caused a global shift in metal extraction from HDE countries to EME and LDE countries, particularly African countries such as Zambia
• It has also caused metal production to decline in HDE countries
• When China’s economic growth slowed, Chinese steel companies had a huge surplus of steel which they looked to export by selling at low prices – this is known as ‘dumping’
• It contributed to a further decline in steel industries in UK, Europe and USA

Trade in Copper

• China is the world’s largest importer of copper, accounting for over 40% of global imports (three times more than Japan, in second place)
• This gives China significant bargaining power and causes global copper prices to fluctuate as demand in China rises and falls
• China imports copper from all over the world, though its largest sources are Chile, Peru and Mexico
• China has also invested heavily in African mining in countries such as Zambia to help meet demand
• This is leading to the development of new trading routes and relationships with exports from LDE countries increasing

You've read 0 of your 0 free revision notes

Get unlimited access.

to absolutely everything:

• Downloadable PDFs
• Unlimited Revision Notes
• Topic Questions
• Past Papers
• Model Answers
• Videos (Maths and Science)

Join the 100,000 + Students that ❤️ Save My Exams

the (exam) results speak for themselves:

Did this page help you?

Author: Rhiannon Molyneux

Rhiannon graduated from Oxford University with a BA in Geography before training as a teacher. She is enthusiastic about her subject and enjoys supporting students to reach their full potential. She has now been teaching for over 15 years, more recently specialising at A level. Rhiannon has many years of experience working as an examiner for GCSE, IGCSE and A level Geography, so she knows how to help students achieve exam success.