Causal Machine Learning: How AI Moves from Correlation to Causation

Key takeaways

Causal machine learning identifies underlying cause-and-effect relationships between variables, helping researchers uncover why certain events occur. Here are some important things to know:

• Researchers predict the global causal artificial intelligence (AI) market will reach $757.74 billion by 2033 [1]. 

• Common causal machine learning techniques include targeted maximum likelihood estimation (TMLE), augmented inverse probability weighting (AIPW), and double/debiased machine learning (DML).

• You can use causal machine learning to assess public policy impact, personalize patient care, and emulate randomized controlled trials from observational data.

Discover how mastering causal machine learning methods can help you build models that produce reliable outcomes and inform smarter decisions. If you’re ready to start learning, enroll in the IBM Machine Learning with Python & Scikit-learn Professional Certificate. You’ll have the opportunity to compare and contrast different machine learning algorithms and master up-to-date machine learning skills.

What is causal inference in machine learning?

Causal inference in machine learning aims to discover how changes in one factor (variable) directly lead to changes in another. The global causal artificial intelligence (AI) market is expected to reach $757.74 billion by 2033, up from $40.55 billion in 2024, reflecting a growing demand for systems that can go beyond prediction to explain why outcomes occur [1]. Increasingly, professionals are using machine learning techniques to streamline this process, analyzing nonlinear, complex data to assess whether variables are conditionally dependent on one another. This allows them to go beyond correlation to automate causal discovery. 

Causation versus correlation is a distinction that’s important for making real-world decisions. For example, a predictive model might show that patients who take a certain medication tend to recover more quickly than those who don’t, demonstrating a correlation. However, this doesn’t inherently mean that taking a certain medication caused a quicker recovery. 

For instance, maybe patients exposed to the medication were at a healthier baseline than unexposed patients. Perhaps exposed patients had better access to health care, which influenced their recovery rate. A patient’s age and their respective family histories could also play a factor. Causal inference in machine learning uses careful model design and a series of assumptions to uncover causal effects, helping to identify cause-and-effect relationships with reasonable certainty. 

What does causal mean in AI? 

Causal artificial intelligence (AI) refers to understanding why things happen within an AI analysis by discovering the underlying cause-and-effect relationships. This allows models to consider the effects and interactions of more variables, and to take hypothetical scenarios into account, even when those records aren’t directly reflected in the data. These models allow professionals to simulate interventions and infer causality even without access to randomized controlled trial data. 

Types of causal machine learning 

You can use several specialized machine learning methods to estimate cause-and-effect relationships while minimizing bias and uncertainty. Consider three of the most widely used approaches below.

Targeted maximum likelihood estimation (TMLE)

You can use TMLE to model both the exposure and outcome with machine learning methods, providing better model flexibility by allowing specifications to dynamically adapt to the data. 

First, the algorithm uses your machine learning model to estimate the outcome, given the treatment and covariates. This first model produces initial predictions of the expected outcome for each individual under two scenarios: exposure and no exposure. You can interpret these predictions as the estimated counterfactual outcomes, which are the outcomes that theoretically would have happened if the individual had the opposite exposure to what they actually did. 

The second step of the model estimates the propensity score, which is the probability of receiving treatment given the covariates. The algorithm computes a weight from this information, known as a “targeting” step, which it then uses to update the outcome estimates in order to reduce bias. This step makes TMLE “doubly robust,” which means the estimate is accurate even if one of the two models (outcome or treatment) is misspecified. 

Augmented inverse probability weighting (AIPW)

AIPW is another type of doubly robust estimation that uses a combination of outcome regression and propensity score weighting to model the average treatment effect (ATE). Similar to other double robust methods, this method only requires you to specify either the outcome or the propensity score model correctly, helping to reduce bias due to misspecification. However, unlike TMLE, AIPW doesn’t have a targeting step that iteratively updates predictions.

The model starts by calculating the propensity score model and then the outcome model, which is the predicted outcome for both treated and untreated individuals. Following this, the AIPW algorithm weights each observation by the inverse of treatment probability (the propensity score) to create a pseudopopulation that mimics a randomized controlled trial. This method is “target trial emulation,” which essentially aims to turn observational or non-random data into a data set that you can analyze in a similar way to a randomized controlled trial.

Double/debiased machine learning (DML)

Double/debiased machine learning (DML) is a causal inference framework that estimates causal parameters, such as ATE or regression coefficients, without bias from regularization. Regularization bias arises from situations where you use traditional machine learning algorithms, which are optimized for prediction accuracy rather than causal estimation, and then use their outputs in causal formulas, which distorts the resulting estimates.

DML helps to address this issue through constructing an orthogonal (debiased) score function by estimating two sets of models: the main outcome model and an auxiliary model for the treatment or confounding structure. It then combines these models to produce an estimator that you can use for statistical inference. To prevent overfitting, DML uses cross-fitting, which is where it uses one set of data for training and the other to estimate the causal effect. Due to its ability to analyze complex and high-dimensional data, researchers often opt for DML when they have complex data sets with images, a high number of variables, and textual data.

What is causal machine learning used for?

Causal machine learning estimates treatment effects, meaning the goal is to uncover how a specific intervention, policy, or action influences an outcome. You can use causal machine learning with both experimental data, such as randomized controlled trials, and observational data, such as clinical registries, electronic health records, and medical databases. Using this data, you can estimate both individualized treatment effects and conditional ATEs.

For example, in health care, you can use these models to predict how patients respond to different treatments to improve the safety and personalization of care. You could use causal machine learning to identify for whom a treatment would be most effective, or predict how someone would respond under different treatment regimens to make informed decisions for patient care. 

You can use similar methods in economics to evaluate interventions and measure heterogeneous effects across populations. For example, you might apply machine learning methods to evaluate how factors such as occupation, education level, and income interact to influence the causal relationship between policy interventions and the decision to work. This helps you make more informed decisions when shaping policies and predicting their societal impacts, supporting evidence-based policymaking.

Who uses causal machine learning?

Professionals in a variety of disciplines, from data science to health care, use causal machine learning to understand why things happen. For example:

• As a data scientist or machine learning engineer, you might build these models directly, focusing on algorithm development or deployment. 

• As an economist or policy designer, you might use causal machine learning to assess the impact of social programs and economic interventions. 

• As an epidemiologist or researcher, you might use causal structures to hypothesize relationships between your variables and estimate statistical parameters. 

Learn more: 4 Careers in Designing Machine Learning Systems 

What is the difference between causal ML and traditional ML?

Traditional machine learning (ML) algorithms focus on predicting outcomes, while causal machine learning answers “what if” questions. For example, you might use traditional machine learning models to predict whether a patient is high risk for cancer recurrence, while causal machine learning might assess how the risk of recurrence might change in response to a new medication.

Pros and cons of causal inference in machine learning

When deciding whether to use causal inference machine learning methods, understanding the strengths and limitations can help you make an informed decision. Consider the following key advantages and challenges. 

Advantages: 

• Handles many covariates effectively

• Captures nonlinear relationships and interaction effects between variables

• Detects effect heterogeneity, supporting personalized or targeted decision-making

• Able to use deep learning and neural networks for causal analysis on complex data

• Allows for causal inference in contexts where experimentation is infeasible 

Disadvantages:

• Largely black box in nature, making it more difficult to assess reliability

• May require significant computational power

• Dependent on correct model specifications, which require domain expertise

• Data limitations may limit the applicability of causal models in certain contexts

Steps to learn how to use machine learning for causal inference

To start learning causal machine learning, a good first step is to become comfortable with the basics of causal inference. This includes understanding the foundations of randomized controlled trial design, counterfactual reasoning, and how to use observational studies within causal inference frameworks.

Once you’ve learned causal inference basics, you can explore underlying techniques for machine learning and AI in causal contexts. This involves experimenting with machine learning algorithms, causal discovery algorithms, deep learning, and reinforcement learning.

Start building machine learning skills with our free resources

If you’re interested in learning more about machine learning, start exploring machine learning topics and trends by subscribing to our LinkedIn newsletter, Career Chat. You can also check out the following resources to keep learning:

• Learn from experts: 6 Questions With a Google AI Research Director 

• Watch on YouTube: The #1 Machine Learning Course for Complete Beginners | Start Your AI Journey 

• Explore career resources: Machine Learning Career Paths: Explore Roles & Specializations

Whether you want to develop a new skill, get comfortable with an in-demand technology, or advance your abilities, keep growing with a Coursera Plus subscription. You’ll get access to over 10,000 flexible courses.

Build job-ready skills with Coursera Plus

Start 7-day free trial

• 
• 
• 
• 

Start 7-day free trial