1 Foundations of causal inference
Causal inference examines whether and how a change in one variable brings about a change in another. It aims to move beyond simple association and toward defensible statements about cause and effect. Because real-world data often reflect complex systems with many interacting influences, causal analysis relies on explicit assumptions, carefully defined questions, and study designs that support interpretation.
1.1 Causation versus association
Association describes whether two variables vary together, while causation implies that altering one variable would change the other. Two quantities may be correlated for reasons that have nothing to do with a direct causal link, such as shared background factors or coincidence. Causal inference therefore asks not only whether variables are related, but whether the relation persists under an intervention or hypothetical change.
1.2 Counterfactual reasoning
Counterfactual reasoning compares what actually happened with what would have happened under an alternative condition. For example, one asks what outcome a person would have experienced if exposed versus unexposed, even though only one of those possibilities can be observed. This framework is central because it provides a clear way to define causal effects, despite the fact that the unobserved alternative must be inferred indirectly.
1.3 Causal questions and estimands
A causal question must be translated into a precise estimand, or target quantity, such as an average treatment effect, a direct effect, or a risk difference. Clear definition of the estimand determines which data, assumptions, and methods are appropriate. In practice, causal work often begins by specifying the intervention of interest, the population to which the result should apply, and the contrast being evaluated.
1.4 Assumptions in causal analysis
Causal conclusions depend on assumptions that are not fully testable from data alone. These may include correct specification of the causal structure, adequate measurement of relevant variables, and the absence of hidden confounding after adjustment. The strength of a causal claim is therefore tied to how plausible these assumptions are in the context of the study.
2 Historical development
Causal thinking has roots in philosophy, natural science, and statistics. Over time, the field developed from informal reasoning about cause and effect into formal mathematical frameworks. Modern causal inference combines graphical, counterfactual, and model-based traditions, each contributing tools for identifying and estimating causal effects.
2.1 Early philosophical roots
Early discussions of causation appear in classical philosophy, where scholars examined necessity, sequence, and explanation. Later thinkers distinguished between mere succession of events and genuine causal dependence. These ideas shaped the eventual scientific view that causes should be understood through interventions, mechanisms, and consistent patterns of change.
2.2 Statistical development
Statistical approaches to causality grew from agricultural experiments, epidemiologic studies, and econometric analysis. As researchers confronted observational data, methods were developed to reduce bias from nonrandom assignment. The challenge of making causal claims from incomplete and imperfect data pushed the field toward formal definitions and stronger design principles.
2.3 Modern causal frameworks
Modern causal inference is often organized around three influential frameworks: potential outcomes, structural causal models, and directed acyclic graphs. These approaches differ in language and emphasis, but they share a goal of clarifying how causal claims can be stated and evaluated. Together, they form the conceptual backbone of much contemporary work.
2.3.1 Potential outcomes approach
The potential outcomes approach defines causal effects by comparing outcomes under different hypothetical treatments. Each unit has a set of potential responses, only one of which is observed in practice. This framework is especially useful for defining average effects, treatment contrasts, and the role of randomized assignment.
2.3.2 Structural causal models
Structural causal models represent relationships through equations that describe how variables generate one another. Interventions are modeled by replacing or fixing parts of the system, allowing analysts to trace how changes propagate through the network. This approach is well suited to reasoning about mechanisms, mediation, and complex dependency structures.
2.3.3 Directed acyclic graphs
Directed acyclic graphs, or DAGs, use arrows to represent causal relations between variables. They help researchers visualize assumptions, identify confounding paths, and determine which variables should or should not be adjusted for. Their main value lies in making causal assumptions explicit and easier to scrutinize.
3 Core concepts
Causal inference uses a set of core ideas that recur across applications and methods. These concepts describe the roles of variables, the sources of bias, and the ways causal effects may operate differently across settings. Understanding them is essential for building a valid analysis.
3.1 Treatment, exposure, and outcome
A treatment or exposure is the factor whose causal impact is being studied, while the outcome is the result of interest. In medicine, the exposure might be a drug; in economics, a policy change; and in social science, an educational intervention. Careful definition of these terms is necessary because causal conclusions depend on exactly what intervention is being considered.
3.2 Confounding
Confounding occurs when an outside factor influences both the exposure and the outcome, creating a misleading association. If such a factor is not properly accounted for, the observed relationship may overstate, understate, or even reverse the true causal effect. Controlling confounding is one of the central tasks of causal analysis.
3.3 Mediation
Mediation refers to situations in which an exposure affects an outcome partly through an intermediate variable. The mediator lies on the pathway from cause to effect and helps explain how the causal influence operates. Studying mediation can reveal mechanisms, though it also introduces additional assumptions and analytic complexity.
3.4 Moderation and effect heterogeneity
Moderation occurs when the size or direction of a causal effect varies across levels of another variable. This variation is often called effect heterogeneity. Identifying moderators helps explain why an intervention may work better for some groups or in some contexts than in others.
3.5 Selection bias
Selection bias arises when the individuals included in an analysis are not representative of the target population in a way that is related to exposure and outcome. This can happen through nonresponse, attrition, conditioning on a collider, or other forms of selective inclusion. The result is a distorted estimate of the causal effect.
3.6 Measurement error
Measurement error occurs when the recorded value of a variable differs from its true value. Error in exposure, outcome, or confounder measurement can weaken or distort causal estimates. The consequences vary by setting, but inaccurate measurement often reduces credibility and may introduce systematic bias.
4 Research designs
Research design strongly shapes what causal conclusions are possible. Some designs, such as randomized trials, directly support causal interpretation, while others rely on assumptions and comparison strategies to approximate experimental conditions. The choice of design depends on ethical constraints, feasibility, and the nature of the causal question.
4.1 Randomized controlled experiments
Randomized controlled experiments assign exposure by chance, which helps balance observed and unobserved factors across groups. This design is often considered the strongest basis for causal inference because it reduces confounding at the point of assignment. When properly implemented, it provides a clear comparison between treated and untreated conditions.
4.2 Natural experiments
Natural experiments exploit situations in which assignment is shaped by external events or rules rather than individual choice. Examples include policy thresholds, random shocks, or administrative processes that approximate randomization. Although not truly randomized, these settings can yield credible causal evidence when the assignment mechanism is well understood.
4.3 Observational studies
Observational studies examine data in which the researcher does not control exposure assignment. Because treatment is often related to participant characteristics, causal interpretation is more difficult than in experiments. Such studies remain important, however, because they are often the only feasible way to study real-world effects at scale.
4.4 Quasi-experimental designs
Quasi-experimental designs use structured features of data or policy environments to approximate experimental comparison. They are especially valuable when randomization is impossible or unethical. Their validity depends on assumptions about timing, threshold behavior, or the stability of trends.
4.4.1 Difference-in-differences
Difference-in-differences compares changes over time between a group affected by an intervention and a group that is not. By focusing on differences in change rather than raw levels, the method attempts to remove stable between-group differences. Its credibility rests on the assumption that, absent the intervention, the groups would have followed parallel trends.
4.4.2 Regression discontinuity design
Regression discontinuity design uses a cutoff in an assignment rule, comparing observations just above and below the threshold. Units near the cutoff are often similar, making the local comparison informative about causal impact. The method can be highly persuasive when the running variable and threshold are measured accurately and the cutoff is not manipulated.
4.4.3 Instrumental variables
Instrumental variables estimation uses an external factor that influences exposure but affects the outcome only through that exposure. A valid instrument can help recover causal effects even when exposure is confounded. The challenge is finding a variable that satisfies the required exclusion and relevance conditions.
4.4.4 Matching methods
Matching methods pair treated and untreated observations with similar characteristics to improve comparability. The goal is to reduce imbalance on measured covariates before estimating the effect. Matching can be helpful, though it cannot eliminate bias from unmeasured confounders.
5 Identification strategies
Identification refers to the process of determining whether a causal effect can be expressed from observed data and assumptions. Before estimating an effect, analysts must know whether the causal quantity is recoverable in principle. Identification strategies formalize the conditions under which this is possible.
5.1 Backdoor criterion
The backdoor criterion specifies sets of variables that block spurious paths between exposure and outcome. Adjusting for such variables can remove confounding if the causal graph is correctly specified. It offers a principled rule for deciding which covariates to control for.
5.2 Frontdoor criterion
The frontdoor criterion identifies causal effects through an intermediate variable when direct confounding adjustment is not sufficient. If certain pathways are observed and relevant assumptions hold, the effect can be recovered even when exposure and outcome share unmeasured confounders. This is a more specialized but conceptually important strategy.
5.3 Ignorability and exchangeability
Ignorability means that, conditional on measured covariates, exposure assignment is independent of potential outcomes. Exchangeability is a closely related idea stating that comparable groups would have similar outcome distributions under the same treatment. These assumptions are central to many methods that estimate causal effects from observational data.
5.4 Positivity and overlap
Positivity requires that every relevant subgroup has a nonzero chance of receiving each exposure level. Overlap refers to the practical extent to which treated and untreated units resemble one another in covariate space. Without sufficient positivity, the effect for some groups cannot be reliably estimated.
5.5 Consistency and SUTVA
Consistency means that the observed outcome for a unit equals the potential outcome under the treatment actually received. The Stable Unit Treatment Value Assumption, or SUTVA, adds that one unit’s outcome is unaffected by another unit’s treatment and that treatment versions are well defined. These assumptions are foundational but often only approximately true in practice.
6 Estimation methods
Once identification is established, estimation methods are used to compute the causal quantity from data. Different estimators emphasize different tradeoffs among bias, variance, model dependence, and interpretability. The best choice depends on the study design and the assumptions that can reasonably be defended.
6.1 Regression-based methods
Regression-based methods estimate causal effects by modeling the outcome, the exposure, or both as functions of covariates. They are widely used because they are familiar and computationally efficient. Their validity depends on correct model specification and appropriate adjustment for confounding.
6.2 Propensity score methods
Propensity score methods summarize the probability of receiving treatment given observed covariates. These scores can be used for matching, stratification, or weighting to improve balance between groups. They are especially useful when many covariates must be managed in a structured way.
6.3 G-computation
G-computation estimates what the average outcome would be under each treatment scenario by modeling the outcome as a function of covariates and exposure. The method then averages predicted outcomes across the observed covariate distribution. It is a direct way to estimate population-level causal effects.
6.4 Inverse probability weighting
Inverse probability weighting creates a weighted sample in which treatment assignment is more balanced with respect to covariates. Observations that received an unlikely exposure are given greater weight. This approach can recover causal effects when the treatment model is correctly specified and overlap is adequate.
6.5 Doubly robust estimation
Doubly robust estimators combine an outcome model with a treatment or propensity model. They remain consistent if at least one of the two models is correctly specified, which provides a practical safeguard against misspecification. This flexibility has made them popular in applied causal research.
6.6 Machine learning approaches
Machine learning approaches can improve estimation by capturing nonlinearities, interactions, and high-dimensional patterns. In causal analysis, these tools are typically used within a framework that preserves identification assumptions and avoids overfitting. Their strength lies in prediction, but careful integration is needed to ensure valid causal interpretation.
7 Mediation and path analysis
Mediation and path analysis examine how causal effects are transmitted through intermediate variables. These methods aim to separate a total effect into components that act through different channels. They are useful for studying mechanisms, policy pathways, and system dynamics.
7.1 Direct and indirect effects
A direct effect is the part of the causal impact not transmitted through a specified mediator, while an indirect effect operates through that mediator. Together, these components describe how an intervention influences the outcome. Defining them precisely requires a clear causal model and careful handling of intermediate variables.
7.2 Causal pathways
Causal pathways describe sequences of variables through which one factor influences another. Mapping these pathways can clarify mechanisms and reveal where interventions might be most effective. However, pathway analysis depends heavily on assumptions about ordering and the absence of hidden feedback.
7.3 Decomposition of effects
Effect decomposition breaks a total causal effect into meaningful parts, such as pathway-specific contributions or sequential stages. This can be useful for identifying how much of an intervention’s impact is attributable to each mechanism. The validity of decomposition depends on strong structural assumptions and precise variable definitions.
7.4 Mediation assumptions
Mediation analysis typically assumes correct temporal ordering, no unmeasured confounding of exposure-mediator, mediator-outcome, and exposure-outcome relations, and well-defined interventions. These assumptions are often difficult to verify fully. As a result, mediation findings are usually interpreted with greater caution than simple total effect estimates.
8 Sensitivity and robustness
Sensitivity and robustness analyses assess how much a causal conclusion depends on assumptions or analytic choices. They do not prove causality, but they help gauge the stability of the result. Such checks are an important part of responsible inference.
8.1 Sensitivity analysis for unmeasured confounding
Sensitivity analysis evaluates how large an unmeasured confounder would need to be to alter the conclusion. This provides a way to judge whether the estimated effect is fragile or fairly robust. The approach helps translate abstract concern about hidden bias into a more concrete assessment.
8.2 Placebo and falsification tests
Placebo and falsification tests examine whether an apparent effect also appears where no effect should exist. For example, one may test outcomes or time periods that should not be affected by the exposure. Failure in these tests can indicate bias, poor design, or model problems.
8.3 Robustness checks
Robustness checks repeat the analysis under alternative specifications, samples, or adjustment sets. If the causal estimate remains similar, confidence in the result increases. Large swings across reasonable alternatives suggest that the inference is unstable.
8.4 Model misspecification
Model misspecification occurs when the chosen statistical form does not match the underlying data-generating process. In causal work, this can distort estimates even if the identification assumptions are otherwise sound. Diagnostics and flexible modeling can reduce, though not eliminate, this risk.
9 Applications
Causal inference is widely used in disciplines that require evidence about interventions, policy effects, or mechanisms. Its methods help researchers answer practical questions in settings where controlled experimentation is limited or impossible. The field continues to expand because many domains demand conclusions about what causes what.
9.1 Medicine and public health
In medicine and public health, causal inference supports evaluation of treatments, preventive measures, and risk factors. It helps estimate the effects of drugs, screening programs, and lifestyle interventions. The approach is especially important when randomized trials are infeasible, too costly, or ethically constrained.
9.2 Economics and policy evaluation
Economists use causal methods to assess the effects of taxes, subsidies, labor rules, and other policies. These analyses help estimate how interventions influence employment, income, consumption, and market behavior. Natural experiments and quasi-experimental designs are especially common in this area.
9.3 Education research
Education researchers apply causal tools to study instructional programs, class size, school choice, and student support services. The aim is often to determine which interventions improve learning outcomes and for whom. Because assignment is frequently nonrandom, careful design and adjustment are essential.
9.4 Social science
In social science, causal inference is used to investigate family structure, migration, neighborhood context, inequality, and many other topics. The field helps distinguish systematic effects from background correlations in complex social environments. Since many variables are interdependent, transparent assumptions are particularly important.
9.5 Computer science and machine learning
Computer science and machine learning use causal inference to improve decision systems, interpret models, and reason about interventions. This includes fairness analysis, recommendation systems, and policy optimization. Causal thinking is increasingly valuable where predictive accuracy alone is not enough.
10 Limitations and challenges
Causal inference is powerful, but it is rarely definitive. Many real-world questions involve incomplete data, hidden factors, or settings that do not fit clean experimental logic. As a result, causal conclusions often remain provisional and context dependent.
10.1 Unobserved confounding
Unobserved confounding remains one of the most serious threats to validity. When an important common cause is missing from the analysis, estimated effects may be biased in unpredictable ways. Analysts often address this concern through design, sensitivity analysis, or instrumental-variable strategies.
10.2 Temporal ambiguity
Temporal ambiguity arises when the order of cause and effect is unclear. If exposure and outcome are measured at similar times, it may be difficult to determine which came first. Establishing a credible timeline is therefore essential for causal interpretation.
10.3 Interference and spillover effects
Interference occurs when one unit’s exposure influences another unit’s outcome, violating the assumption that units are independent. Spillover effects are common in networks, households, schools, and communities. These phenomena complicate both design and estimation because treatment effects may extend beyond the directly exposed individual.
10.4 Generalizability and transportability
Generalizability concerns whether findings from one sample apply to a broader population, while transportability refers to moving results across settings with different characteristics. A causal effect estimated in one context may not hold elsewhere if populations, environments, or implementation conditions differ. External validity is therefore a major concern.
10.5 Ethical constraints
Ethical constraints limit the kinds of experiments that can be performed. Researchers cannot always assign exposures randomly, especially when harms are possible. Ethical limits make observational and quasi-experimental methods indispensable, but they also require careful judgment about acceptable evidence.