Predicting behavior with artificial intelligence has the potential to produce disastrous policy errors. To find the most effective levers for change, health and development initiatives need to learn how to apply causal models, which provide a deeper explanation for why individuals behave in certain ways. The Surgo Foundation has made this article freely accessible.
The majority of commonly used artificial intelligence (AI) is devoted to behavior prediction. It makes an effort to predict what you will buy, click on a mouse, or do next at work. However, there may be issues with these methods when analyzing data for development and health initiatives. Without understanding the underlying causes of behavior, it would be easy for us to support unfair and ineffective policies and make bad choices.
Healthcare systems can now anticipate which individuals are most likely to require specialized medical care, thanks to AI, for instance. About 200 million Americans are enrolled in risk-prediction programs that use data on projected future costs to the healthcare system to determine whether patients would benefit from additional medical care now. It makes use of predictive machine learning, a subset of self-adaptive algorithms whose accuracy increases with the addition of fresh data. However, as health researcher Ziad Obermeyer and his colleagues demonstrated in a recent Science magazine article, this specific method had an unexpected side effect: patients of color who were more likely than white patients to have chronic illnesses were not identified as requiring more care.
What went incorrectly? Using information from insurance claims, the system forecasted individuals’ future medical requirements based on past medical expenses. However, the creators of the algorithm had overlooked the fact that, for reasons unrelated to their level of illness, such as lack of insurance, insufficient access to healthcare, or other issues, black Americans often spend less on healthcare than white Americans with comparable conditions.
The predictive algorithm’s use of healthcare expenditures as a stand-in for sickness resulted in appropriate recommendations for white patients (fewer health issues meant lower healthcare spending). However, this practice perpetuated racial prejudices in the treatment of black patients. After notifying the manufacturer, the researchers conducted experiments using their own data, which verified the issue and helped them eliminate bias from the algorithm.
This tale highlights one of the dangers associated with some forms of AI. No matter how advanced, users of prediction algorithms may make the mistake of confusing correlation with causation, or that just because event X comes before occurrence Y, therefore X must be the reason behind Y. To determine the relationship between an occurrence and an outcome, a predictive model is helpful. The statement reads, “We can predict that Y will occur when we observe X.”
However, this is not the same as demonstrating that X causes Y to occur. Higher rates of illness (X) were accurately linked with higher health-care costs (Y) for white patients in the context of the health-care algorithm. Healthcare costs were a reliable indicator of future disease and medical demands since X caused Y. However, higher rates of disease did not always translate into higher expenses for Black patients, and the algorithm was unable to anticipate their future medical needs. A correlation existed, but not causality.
This is significant since AI is being used more and more to assist address urgent health and development issues. In fields as diverse as health care, justice, and agriculture, depending exclusively on AI predictive models runs the danger of disastrous outcomes when correlations are misconstrued for causality. As a result, decision-makers must take into account causal AI, another AI technique that may be used to precisely pinpoint cause and effect linkages.
The benefit of utilizing causal AI algorithms to pose hypothetical questions is not limited to its ability to identify the underlying causes of outcomes; it also allows for the modeling of interventions that could alter those outcomes. For instance, how much might we anticipate an improvement in student arithmetic test scores if a particular training program is put in place to increase teacher competency? Long-term experiments in the field can be avoided by simulating scenarios to assess and compare the possible impact of an intervention (or set of interventions) on a result. This saves time and money.
When implemented and utilized properly, predictive AI algorithms undoubtedly have a significant role to play. Precision agriculture is a prime example. It employs predictive AI to analyze data from sensors and satellite pictures to assist farmers in predicting crop yields, identifying various plant species, and detecting weeds and diseases. However, predicting something does not mean that you know why it will happen. It’s one thing to predict that a farmer will have a reduced crop output this year; it’s another to know why, and to act accordingly.
A further difficulty with relying solely on predictive models is the basic ignorance of the underlying assumptions that lead them to generate specific forecasts in the first place. Precision agriculture uses a predictive AI called deep learning, which has an issue. The way that human brain cells are arranged (in “layers”) and interact with one another—taking in signals from cells in one layer, altering them, and then sending the modified signals to cells in another layer—served as the model for deep learning.
Deep learning can map variables to outcomes with far more complex relationships between them than is possible with commonly used methods for outcome prediction, such as regression, a traditional statistical technique that maps the relationships between variables to the predicted outcome with a single best mathematical formula. Deep learning algorithms may learn input-output correlations considerably more complicated than a single mathematical formula and use them to predict outcomes by merging numerous layers between the input variables and outputs.
But because the connections and bridges between variables are “black boxed,” it is difficult for users—and for the algorithm developers—to understand how the variables relate to one another and to the final result. This implies that it is frequently impossible to determine which input features were employed by deep learning models to generate their predictions.
When it comes to the course of people’s lives, like in the US criminal justice system, this opacity is intolerable. One in 111 adult Americans, or 2.3 million, were incarcerated in 2016; the federal and state governments bore the heavy financial burden of housing them. To cut jail expenses by reducing the number of convicts without increasing crime, courts around the country have used “recidivism scores.” The probability that a person convicted of a crime would commit another one is estimated using a prediction algorithm, which yields a single number known as the recidivism score.
The score could theoretically enable a judge to concentrate on incarcerating those who have a higher likelihood of committing new crimes and even help to eliminate any potential bias in sentencing. However, because recidivism scores are derived from risk-assessment instruments that identify statistical correlations rather than causal relationships, they are fundamentally flawed.
Low income, for instance, is not the cause of crime; it is merely connected with it. However, those who originate from low-income families can be given a high recidivism score out of the blue, which makes them more likely to receive a prison sentence. Understanding the root causes of crime rather than just its correlations is essential to improving the criminal justice system.
A deeper examination of causal AI will reveal how it can crack open the “black box” that houses the functioning of solely predictive AI models. Beyond correlation, causal AI can draw attention to the exact connections between causes and consequences.
Controlled Randomized Trials
The domains of development and health are not new to the significance of testing causation. A simple method is to randomly assign individuals to one population group (the treatment group) and carry out an intervention there while carrying out no intervention in a group that is otherwise identical (the control group). It is feasible to identify the intervention’s impact by contrasting the outcomes of the two groups. This is referred to as A/B testing in marketing research and as a randomized controlled trial in clinical investigations.
The Nobel Prize in Economics was given to development economists Michael Kremer, Abhijit Banerjee, and Esther Duflo in 2019 for their pioneering work in applying randomized controlled trials to pinpoint the underlying causes of development problems and devise remedies. Some long-held beliefs regarding causality have been refuted by these trials.
For instance, several observational studies have found links between low vitamin D levels and higher risks of cancer, diabetes, heart disease, and hypertension. However, vitamin D supplementation does not lower the risk of these disorders, as shown by randomized controlled trials, which also failed to find a causal relationship between vitamin D supplementation and health outcomes.
Nevertheless, randomized controlled trials have drawbacks. To guarantee that the results aren’t skewed or impacted by accidental, outlier features like age, sex, health status, or educational attainment, large groups of people are needed. Due to this, these trials are typically very costly (in the millions of dollars) and labor-intensive (they can take years to complete).
Furthermore, despite the complexity of health and social outcomes and their multitude of underlying causes, randomized controlled trials are limited in their ability to evaluate the impact of a single or small number of packaged interventions at a time. Lastly, they are only able to forecast whether an intervention will have an impact on a typical treatment group member rather than a particular person.
Herein lies the role of causal AI. Together with the capacity to reveal the underlying complexity, it provides new avenues for the quicker and more effective testing of causation in both individuals and population groupings. By utilizing pre-existing data, it enables researchers and program designers to replicate an intervention and deduce causality.
Two Methods for Determining Causality
The potential outcomes framework and causal graph models are two well-established methods for implementing causal AI. Using actual data, these methods allow one to test the impact of a proposed intervention. The strong underlying algorithms that are employed to identify the causal patterns in huge data sets are what distinguish them as AI. However, the quantity of plausible reasons that they can examine varies.
Consider the following hypothetical situation to better understand the two approaches, how they operate, and how they differ: Researchers sought to find out if an antismoking advertising campaign made individuals want to stop, but since the commercials were released nationwide, there was no control group. All they had was a data set that revealed whether or not the people saw the advertisements, whether or not they quit smoking, and details about their demographics and other health-related habits. Causal AI offers methods to infer causation even in the absence of a control group.
The potential outcomes framework was introduced by statisticians Paul Rosenbaum and Donald Rubin in 1983. It compares an individual’s actual outcome—quit smoking—with an implied “potential outcome” of the same individual if they had not been exposed to the cause of interest, which is the antismoking advertisement. Of course, the problem is that, for an individual who was exposed to the ad, there is no data on nonexposure results.
Therefore, the AI algorithms find one individual in the data set who was not exposed to the advertisement but who is identical in other important aspects (such as age, race, and education) for every person who was exposed to the advertisement. Stated differently, an artificial control group is designed to resemble a randomized controlled trial by reverse engineering. The prospective outcomes framework can evaluate the impact of a single prespecified intervention at a time—in this example, did the advertising campaign influence the person’s decision to stop smoking—despite being able to address the issue of lacking a control group?
In contrast, causal graph models are capable of doing more than just determining the cause-and-effect link between a single pair of variables. They can be applied as exploratory instruments to map all possible causal pathways leading to a desired outcome and illustrate the relationships between various factors. An analysis of our antismoking campaign using a causal graph would reveal that while some people stopped smoking immediately after seeing the advertisement in a drugstore, others had to purchase nicotine patches before they could give up.
Multiple causal graph models exist. The structural equation model is a popular technique where researchers identify potential variables and how they might interact, and the model then evaluates the data to see whether the interactions occur. The entire network of interactions between various variables must be established using known information, even if this model is capable of testing a large number of these associations in the data.
This model’s drawback is that it only examines the relationships between the factors that are hypothesized; if the variables that produce the effect are not included in the list, they will not be compared to the other possibilities.
The term “causal Bayesian network,” which honors English statistician Thomas Bayes from the 18th century, was first used in the 1980s by computer scientist and philosopher Judea Pearl. With this approach, the relationships between each variable in a data collection are estimated. The outcome is an understandable visual map that indicates which variables affect one another and how much. This methodology has the advantage of being a true discovery method because these interactions do not need to be stated before the test, unlike in a structural equation model.
The potential of causal Bayesian networks is intriguing for several reasons, even though they need a large amount of data to capture the universe of relevant variables. It makes it possible to simultaneously identify many causal linkages through data-driven discovery.
A causal Bayesian network may demonstrate how advertising and the availability of various quit-smoking aids influenced people’s behavior in the anti-smoking ad campaign, or it could demonstrate the influence of people’s own goals. Crucially, in contrast to the opaque nature of predictive AI, the causal AI approach makes evident to researchers, program administrators, and policymakers the connections between the factors (ad exposure, nicotine patch availability), and the outcome (stopped smoking).
It is also feasible to simulate numerous potential interventions at once using causal graphic models. What if, for instance, distinct anti-smoking advertisements catered to distinct age demographics or integrated a broad campaign with outreach by fellow educators?
They also make it possible to counteract any potential drawbacks of a purely data-driven strategy by incorporating expert knowledge. For example, experts can help decide which variables to include in the model, apply conditions to the model to increase accuracy, and assist in deciphering results that defy logic.
Useful Implementation
Causal AI is a rapidly developing field. Researchers are utilizing it in a wide range of sectors, including health and climate change, as its potential becomes more evident.
Climate change | To determine whether and how humans contribute to climate change, as well as what influences people’s perceptions about it, causal AI techniques have been applied to the problem. To answer this question, British scientists employed a causal AI method known as counterfactual event attribution in the potential outcomes framework to ascertain whether greenhouse gas emissions from human activity were a contributing factor in the 2003 European heatwave, which was estimated to have killed over 70,000 people.
The researchers replicated summer temperatures across Europe in 2003, both with and without human interference, using historical data, solar data, information on volcanic eruptions, and atmospheric data on greenhouse gases, aerosols, and ozone. They discovered that when air travel and electricity production were taken out of the model, the likelihood of the heatwave happening was significantly higher.
One of the first studies to relate human activity to an extreme weather event, it was published in 2004 and offered compelling evidence to support the idea of lowering the greenhouse gases produced by such activity. The Intergovernmental Panel on Climate Change of the United Nations has cited the research.
The elements that cause people’s opinions about climate change to become more divisive have also been discovered via causal AI. Researchers employed Bayesian networks to model how various individuals responded to a variety of climate change messaging after surveying participants from the US and Australia. They discovered that Americans who actively mistrusted climate scientists updated their beliefs in opposition to the facts they were given when faced with consensus information about climate change in an online survey. A new method for estimating the interdependent links between worldviews, scientific beliefs, and scientific confidence was made possible by this causal paradigm.
Such insights are crucial in influencing the public’s sense of the urgency of taking action against climate change. These findings offer a framework for creating interventional messaging that considers potential participant reactions to information in light of their histories and beliefs.
Childhood diarrhea | Where previous methods have failed, Causal AI presents chances to treat common and complicated health issues. Diarrhea in children is one instance. For children under five, this disease is the second leading cause of death worldwide. Although many variables contribute to diarrhea, it is very difficult to identify the biological and structural causes of the illness. This complicates the process of creating successful interventions.
Data collected from a nationwide survey including over 110,000 people from over 15,000 households were used in a study conducted in Pakistan. Economic, social, environmental, and household factors were all included in the poll. The researchers discovered 12 home characteristics that were strongly related with diarrhea when they used multivariate regression, a conventional statistical technique.
They were difficult to comprehend, though; for instance, one variable was the number of rooms in the residence. In contrast, a causal Bayesian network analysis of the same data set revealed a network map that identified three factors that were directly associated with children’s diarrheal disease: the use of dry-pit latrines as opposed to toilets that were connected to drains; reliance on a water source other than piped, river, or stream water; and absence of formal trash collection. These discoveries may result in successful strategies to lower diarrheal illness in children if they are implemented into society or national policy.
Maternal and neonatal mortality rates: In many low-income nations, the death rate for mothers and their babies continues to be resolutely high. Women must give birth in medical facilities to ensure the survival and welfare of both the mother and the child.
The Indian government has rapidly improved the rate of institutional delivery through a national incentive scheme that pays families to deliver their babies at facilities (300 Indian rupees [about $4] for the hospital delivery itself, and an additional 300 Indian rupees if the mother has also used antenatal care). But in a lot of Indian states, this growth has leveled off at roughly 80%.
At Surgo Foundation, we investigated the reasons behind women’s non-choice of institutional delivery as well as the kinds of extra interventions required to persuade them to do so. Our research has employed a range of methods, such as causal AI, to determine the reasons behind the decision made by some families to give birth at home. We measured a wide range of possible drivers of institutional delivery in the more than 230 million-person state of Uttar Pradesh using a series of large-scale quantitative surveys. The factors influencing this behavior were then found using a causal Bayesian network, and the most promising targets for a public health intervention were determined.
While delivering at a medical facility was associated with a wide range of factors, causal AI pinpointed the root causes. Contrary to popular assumption, and much to our surprise, the mother’s accessibility to transportation was more important than her closeness to a medical facility. This implied that rather than constructing more healthcare facilities closer to homes, the government should address transportation-related difficulties.
Surprisingly, views regarding hospital cleanliness, staff skills, and staff biases were not nearly as important as beliefs regarding whether hospital deliveries were safer than home deliveries. A delivery plan and the mother’s knowledge of financial incentives both raised the chance of an institutional delivery, confirming the effectiveness of the government’s incentive program.
The results of this study are presently being utilized to model theoretical situations and test an intervention wherein frontline health workers assist mothers in Uttar Pradesh in creating comprehensive prenatal plans, including where the mother will give birth, how she will get to the hospital, and how she will cover any additional expenses.
Seven Pointers for Scaling
Businesses and governments are embracing AI to streamline operations, find solutions to issues, and increase productivity. It is similarly crucial for those involved in development and health issues to research and apply causal AI more widely. It provides a path ahead with clear benefits above just predictive AI. For example, predictive algorithms can determine if a mammography reading indicates breast cancer or not. They can also provide strong and frequently accurate information.
However, causal AI can assist by determining the complex network of underlying reasons behind a behavior or event and offering crucial insights that predictive models are unable to deliver. This can result in interventions that are more successful and produce desired results. Furthermore, causal AI reduces the possibility of biases similar to those previously mentioned by enabling researchers to verify the rationale behind the model.
The era of causal AI is upon us, as indicated by three convergent variables. First, developments in AI are drawing attention to the wide range of uses for causal techniques. Additionally, when models are improved, expanded, and used in new contexts, more is discovered about the benefits and drawbacks of these models. Secondly, access to large-scale data sets is increasing.
Similar to a 4K ultra-high definition television that has more pixels per square inch of screen real estate than an older standard definition TV, more data improves the clarity and accuracy of predictions and increases trust in the conclusions drawn from causal networks. Last but not least, the fields of health and development are emphasizing precision policy more and more. This involves developing treatments with the best outcomes in order to allocate scarce resources where they can work the hardest. Causal AI is in a perfect position to tackle this task.
It will take effort to successfully implement these strategies. The following list of seven suggestions can help make causal AI more widely adopted and utilized.
Utilize data more effectively and raise their caliber. Over the past ten years, funds have been allocated to several extensive data collection initiatives. These data sets, though, are frequently underutilized and have more potential for mining to yield additional insights. Even while data is growing, there are still issues. Quality varies and data sets are sometimes fragmented.
Another difficulty in linking disparate data sets is when information is recorded at the individual, regional, or national levels in one data set and at the individual, national, or regional level in another. Creating standard indicators that are used in all national data collection initiatives would aid in optimizing data sets after they are connected.
Get more thorough data collection. To successfully use causal AI, one must have a thorough grasp of every aspect that could impact behavior, including individual beliefs, motives, biases, and influencers in addition to structural factors like laws and policies. The causative variables that underpin actions or events may be overlooked during data collection if there are too many preconceived notions about what should be collected. This could result in incorrect causal relationships being constructed.
Create open-source, scalable, high-performing tools for using causal AI algorithms. Because proprietary algorithm platforms are expensive, the health and development sectors often cannot afford them. Because more individuals may review the source codes and offer criticism, open-sourcing eventually results in software that is more accessible, free, and of higher quality. While there are some open-source algorithms (like unlearn), their accuracy and speed still need to be increased.
It is important for practitioners who are not causal AI experts to know how to use this method in their field. The Surgo Foundation is creating resources to help enterprises that are new to causal AI avoid process mistakes and minimize access barriers. An open-source program that determines which algorithms are optimal to employ on a given data set and if it can be applied to Bayesian networks is one example. To facilitate the transition of causal AI from theoretical study to real-world application, Surgo is also creating a workflow guide.
Combine humans with artificial intelligence. Development issues cannot be resolved by a data-driven strategy alone. It is imperative to incorporate expert knowledge at every stage of the process to ensure accurate interpretation of causal networks by researchers and software developers.
By incorporating limitations that are based on a real-world understanding of how systems operate and determining if well-known confounding variables are absent from the data, experts can enhance the performance of causal AI. Additionally, as causal AI becomes more widely used, ethicists and policy experts will be crucial in ensuring that the method stays clear of the bias and inaccuracy traps that have occasionally plagued the employment of predictive AI models.
Enhance methods for assessing algorithmic performance. Researchers in computer science are working to enhance the general robustness and accuracy of causal AI algorithms. Comparing outcomes with established causal linkages is a common method of assessing the accuracy of causal models. But if a model cannot be validated by established causal links, what should a researcher do? (After all, the initial purpose of executing causal AI is frequently to find those correlations.)
Moreover, what happens when a causal AI model produces conclusions that disagree with the body of current expert knowledge? Making fictional data sets that resemble real data sets but have predefined causal linkages between variables would be one way to solve the problem. Researchers can determine expected performance on a real data set with similar features by assessing a causal AI model’s performance on an artificial data set.
Prove the usefulness of causal AI in the field of development. The above-mentioned examples are strong, but they are rare in number. Increased public knowledge of the ongoing research will facilitate the adoption of causal techniques. The Surgo Foundation is utilizing causal AI to determine which interventions should be scaled up to enhance student learning, how to maximize the effectiveness of frontline health professionals, and how to increase the uptake of contemporary family planning techniques. As the foundation develops, we hope to evaluate the use of causal AI in fields like climate change and agriculture.
Increase the important stakeholders’ awareness and expertise. For people outside the field, the concept of causal AI is still quite novel. For policymakers, funders, program administrators, monitoring and evaluation specialists, and experts in the various industries where causal AI could find application, to comprehend these techniques, at least conceptually, work needs to be done to convey their potential.
The Next Reasonably Appropriate Action
People consider and consider patterns that recur to make sense of the world. We have gone a long way from making up stories to explain the weather to employing mathematical modeling and meticulous data collection to forecast the next hurricane’s route or the amount of rainfall. However, the boundaries of what we can see and the tools at our disposal for data analysis keep getting in the way.
The next natural development is causal AI, which has been made possible by recent advances in technology and the growing amount of data available. Its ability to pinpoint the exact causative factors that directly result in specific behaviors or outcomes, as well as its ability to evaluate various strategies for altering those behaviors or outcomes, give it an advantage over several other social science fields and even predictive AI.
This advantage helps practitioners and academics concentrate on the optimal combination of interventions for tackling some of the most pressing problems of our time, such as health care and climate change. Programs can accomplish more with fewer resources and save time by using better causal inferences. Programs can also avoid errors that occur when people—or the machines or software that they create—ignore important contexts or fall into the trap of mistaking correlation for causation by combining causal AI with human expertise.
In the end, understanding the “why” behind complicated issues enables us to discern the proper course of action to take to reach our goals. An ounce of causal AI might still prove to be worth a pound of prediction.