Introduction: Understanding Causal Inference
Causal inference is the process of determining the actual effect of a particular phenomenon (the cause) on another phenomenon (the effect). Unlike traditional statistical methods that focus on correlation and association, causal inference seeks to establish whether and how a change in one variable directly influences another. This is essential for answering “what if” questions, making effective interventions, and developing policies based on true cause-effect relationships rather than mere associations. This cheatsheet provides a structured overview of key causal inference frameworks, methodologies, assumptions, and applications for researchers, data scientists, and analysts working with observational and experimental data.
Core Causal Frameworks and Concepts
Fundamental Causal Models
| Framework | Key Concepts | Primary Contributors | Best For |
|---|---|---|---|
| Potential Outcomes (Rubin Causal Model) | Counterfactuals, treatment effects, ignorability | Donald Rubin, Jerzy Neyman, Paul Holland | Experimental design, treatment effect estimation |
| Structural Causal Models (SCM) | Directed graphs, structural equations, do-calculus | Judea Pearl, Peter Spirtes | Complex causal relationships, identification |
| Graphical Models (DAGs) | Directed acyclic graphs, d-separation, Markov condition | Judea Pearl, James Robins | Visualization, bias identification, confounder selection |
| Granger Causality | Time series precedence, predictive causality | Clive Granger | Time series data, economic forecasting |
| Instrumental Variables | Exogenous variation, exclusion restriction | Philip Wright, Joshua Angrist | Natural experiments, econometrics |
Key Terminology in Causal Inference
| Term | Definition | Relevance |
|---|---|---|
| Counterfactual | What would have happened to a unit under alternative treatment condition | Foundation of potential outcomes framework |
| Causal Effect | Difference between potential outcomes under different treatments | Primary quantity of interest in causal analysis |
| Confounding | Bias due to common causes of treatment and outcome | Major threat to causal inference from observational data |
| Selection Bias | Systematic differences between compared groups | Compromises validity of causal conclusions |
| Mediator | Variable in causal path between cause and effect | Important for understanding mechanisms |
| Collider | Common effect of two variables | Can create spurious associations when conditioned upon |
| Instrumental Variable | Affects outcome only through treatment | Enables causal inference despite unmeasured confounding |
| do-operator | Mathematical notation for intervention | Essential for calculating intervention effects |
| Backdoor Path | Non-causal path creating association | Must be blocked to estimate causal effects |
| Identifiability | Whether causal effect can be estimated from observed data | Determines feasibility of causal inference |
Treatment Effect Types
| Effect Type | Definition | Formula (Potential Outcomes) | When to Use |
|---|---|---|---|
| Average Treatment Effect (ATE) | Average effect across entire population | E[Y(1) – Y(0)] | General policy evaluation |
| Average Treatment Effect on Treated (ATT) | Average effect for those who received treatment | E[Y(1) – Y(0) | T=1] | Program evaluation, selective treatments |
| Average Treatment Effect on Untreated (ATU) | Average effect for those who didn’t receive treatment | E[Y(1) – Y(0) | T=0] | Policy extension considerations |
| Conditional Average Treatment Effect (CATE) | Average effect for subgroup defined by covariates | E[Y(1) – Y(0) | X=x] | Heterogeneous treatment effects, personalization |
| Local Average Treatment Effect (LATE) | Average effect for compliers (IV context) | E[Y(1) – Y(0) | compliers] | Instrumental variable analysis |
| Intention-to-Treat (ITT) | Effect of being assigned to treatment | E[Y | assigned to T=1] – E[Y | assigned to T=0] | Randomized trials with non-compliance |
Causal Inference Methods and Approaches
Experimental Methods
| Method | Key Features | Strengths | Limitations | Example Application |
|---|---|---|---|---|
| Randomized Controlled Trials (RCT) | Random assignment to treatment/control | Gold standard for causal inference | Costly, ethical limitations, external validity concerns | Clinical drug trials, A/B testing |
| Factorial Designs | Multiple treatments tested simultaneously | Efficiency, interaction effects | Complex analysis, larger sample requirements | Agricultural experiments, multi-factor interventions |
| Crossover Designs | Subjects receive multiple treatments in sequence | Within-subject comparison, higher power | Carryover effects, limited to reversible treatments | Bioequivalence studies, short-term interventions |
| Cluster Randomization | Groups rather than individuals randomized | Accounts for intra-group dependencies | Reduced statistical power, implementation challenges | Community interventions, classroom-based studies |
| Stepped Wedge Design | Sequential rollout of intervention | Practical for universal implementation, temporal controls | Complex analysis, temporal confounding | Health system changes, policy implementations |
Quasi-Experimental Methods
| Method | Key Features | Assumptions | Limitations | Example Application |
|---|---|---|---|---|
| Difference-in-Differences (DiD) | Before-after comparison between treated and control groups | Parallel trends, no anticipation | Sensitive to time-varying confounders | Policy evaluation, natural experiments |
| Regression Discontinuity (RD) | Treatment assigned based on threshold of running variable | Continuity around threshold | Limited to threshold vicinity, requires threshold-based assignment | Scholarship effects, age-based eligibility |
| Instrumental Variables (IV) | External variable influences treatment but not outcome directly | Relevance, exclusion restriction, monotonicity | Finding valid instruments, weak instrument bias | Mendelian randomization, lottery-based assignments |
| Synthetic Control | Weighted combination of control units to match treated unit | Convex hull, no anticipation | Few treated units, requires long pre-treatment period | State/country policy changes, single-unit interventions |
| Interrupted Time Series | Before-after comparison with multiple time points | No contemporaneous events, stability | Confounding by time-varying factors | Policy changes, interventions with defined start times |
| Propensity Score Methods | Balancing treatment groups based on probability of treatment | No unmeasured confounders, positivity | Sensitivity to specification, cannot adjust for unobserved confounders | Observational healthcare data, program evaluations |
Observational Methods
| Method | Key Approach | Critical Assumptions | Advantages | Limitations |
|---|---|---|---|---|
| Matching | Pair treated units with similar controls | Conditional exchangeability, common support | Intuitive, transparent | Dimensionality issues, exact matches often impossible |
| Covariate Adjustment | Regression controlling for confounders | No unmeasured confounding, correct model specification | Familiar, adjusts for multiple variables | Model dependence, extrapolation dangers |
| Inverse Probability Weighting (IPW) | Reweighting to create balanced pseudo-population | Positivity, correct propensity score model | Can estimate various treatment effects | Extreme weights, model sensitivity |
| Doubly Robust Methods | Combine outcome modeling and weighting | Either propensity or outcome model correct | Protection against model misspecification | Computational complexity, implementation challenges |
| G-Methods (G-Computation) | Standardization using predicted counterfactuals | Correct specification of outcome model | Handles time-varying confounding, treatments | Computational intensity, “g-null paradox” |
| Targeted Maximum Likelihood (TMLE) | Efficient estimation with targeted bias correction | Initial estimators correctly specified | Efficiency, double robustness | Implementation complexity, computational demands |
Causal Discovery Methods
| Method | Approach | Assumptions | Strengths | Limitations |
|---|---|---|---|---|
| PC Algorithm | Constraint-based learning of graph skeleton | Causal sufficiency, faithfulness, no latent confounders | Computationally efficient, handles high dimensions | Sensitive to statistical decisions, assumes no latent variables |
| FCI Algorithm | Extension of PC allowing latent confounders | Acyclicity, faithfulness | Can work with latent confounders | Computational complexity, often returns equivalence classes |
| GES (Greedy Equivalence Search) | Score-based search over graph space | Acyclicity, correct scoring function | Often more accurate than constraint-based | Computationally intensive for many variables |
| LiNGAM | Exploits non-Gaussian distributions | Linear relationships, non-Gaussian errors | Can identify unique DAG, not just equivalence class | Requires non-Gaussian distributions, assumes linearity |
| Additive Noise Models | Identifies direction via independence of residuals | Additive noise, no latent confounders | Can identify direction between two variables | Strong functional form assumptions |
| Invariant Causal Prediction | Uses stability across environments | Causal sufficiency, invariance principle | Robust to certain forms of model misspecification | Requires multiple environments/datasets |
Step-by-Step Causal Inference Process
1. Define Causal Question
- Clearly specify treatment/exposure and outcome
- Define target population
- Formulate precise causal estimand (ATE, ATT, etc.)
- Determine whether average effects or heterogeneous effects are of interest
2. Represent Causal Knowledge
- Develop causal diagram (DAG) based on domain knowledge
- Identify potential confounders, mediators, colliders
- Determine minimum adjustment set(s) for identification
- Check for possible violations of assumptions
3. Design Data Collection Strategy
- Choose between experimental, quasi-experimental, or observational approaches
- Determine sample size requirements for adequate power
- Plan for collecting data on all necessary confounders
- Consider instrumental variables or other identification strategies
4. Select Appropriate Method
- Match method to data structure and causal question
- Consider robustness vs. efficiency tradeoffs
- Plan for sensitivity analyses to assess assumption violations
- Determine whether bounds rather than point estimates are needed
5. Implement Analysis
- Check balance between treatment groups
- Apply selected causal inference method(s)
- Estimate treatment effects and construct confidence intervals
- Consider multiple methods for robustness
6. Conduct Sensitivity Analysis
- Assess robustness to unmeasured confounding
- Vary model specifications and assumptions
- Apply bounds or partial identification approaches if needed
- Determine thresholds at which conclusions would change
7. Interpret and Communicate Results
- Distinguish between statistical and causal inference
- Clarify assumptions underlying causal claims
- Discuss limitations and generalizability
- Translate findings into practical implications
Key Assumptions and Their Assessment
Fundamental Causal Assumptions
| Assumption | Definition | Assessment Methods | Implications if Violated |
|---|---|---|---|
| Consistency | Well-defined treatments lead to same outcome regardless of assignment mechanism | Clarify treatment definition, assess variation in implementation | Undefined treatment effects, misleading estimates |
| Positivity (Overlap) | Positive probability of each treatment level for all covariate values | Check propensity score distributions, identify sparse regions | Extreme weights, unreliable estimates, extrapolation |
| Exchangeability (No Unmeasured Confounding) | Treatment assignment independent of potential outcomes given observed covariates | Sensitivity analysis, negative controls, bias formulas | Biased treatment effect estimates |
| Stable Unit Treatment Value Assumption (SUTVA) | No interference between units, no hidden treatment variations | Assess spillover potential, treatment implementation consistency | Biased effect estimates, unclear interpretation |
| Instrument Relevance | Instrumental variable meaningfully affects treatment assignment | F-statistic in first stage, partial R² | Weak instrument bias, large standard errors |
| Exclusion Restriction | Instrument affects outcome only through treatment | Domain knowledge, falsification tests | Biased IV estimates |
| Monotonicity | Instrument affects treatment in same direction for all units | Theoretical justification, empirical checks where possible | Biased LATE estimates |
| Parallel Trends (DiD) | Treatment and control would follow same trends without intervention | Visual inspection of pre-treatment trends, placebo tests | Biased DiD estimates |
Testing Assumptions with Data
- Balance checks: Compare covariate distributions between treatment groups
- Placebo tests: Analyze outcomes known to be unaffected by treatment
- Sensitivity analysis: Quantify how strong unmeasured confounding would need to be to invalidate results
- Negative controls: Test effects on outcomes that shouldn’t be affected or from exposures that shouldn’t have an effect
- Over-identification tests: When multiple instruments are available, test consistency of estimates
- Cross-validation: Compare estimates from different methods or data subsets
- E-values: Minimal strength of unmeasured confounder needed to explain away an effect
Common Challenges and Solutions in Causal Inference
Challenge: Unmeasured Confounding
- Potential Causes:
- Unavailable data on important confounders
- Unknown confounding mechanisms
- Limitations in measurement
- Solutions:
- Instrumental variable approaches
- Difference-in-differences if confounders are stable over time
- Sensitivity analysis to quantify potential bias
- Bounded estimates incorporating uncertainty
- Proxies for unmeasured confounders
- Negative control outcomes for bias detection
Challenge: Selection Bias and Missing Data
- Potential Causes:
- Non-random sampling
- Differential attrition/dropout
- Self-selection into treatment
- Incomplete observation of outcomes
- Solutions:
- Inverse probability of selection weighting
- Multiple imputation for missing data
- Heckman selection models
- Bounds analysis incorporating range of possible values
- Pattern-mixture models for sensitivity analysis
- Addressing selection as part of causal model
Challenge: Treatment Heterogeneity
- Potential Causes:
- Individual variation in treatment response
- Differences in treatment implementation
- Interaction with observed or unobserved characteristics
- Solutions:
- Subgroup analysis with pre-specified groups
- Causal forests or Bayesian additive regression trees
- Varying coefficient models
- Meta-regression across studies or contexts
- Hierarchical models incorporating group structure
- Testing for qualitative interactions
Challenge: Interference Between Units
- Potential Causes:
- Network effects
- Spillovers between treated and control units
- General equilibrium effects
- Solutions:
- Cluster randomization
- Network exposure models
- Partial identification approaches
- Spatial models incorporating distance
- Two-stage randomization designs
- Explicit modeling of interference patterns
Applied Causal Inference by Field
Economics and Policy Evaluation
| Common Methods | Key Considerations | Example Applications |
|---|---|---|
| Instrumental Variables | Finding valid instruments, relevance/exclusion | Estimating returns to education, policy impacts |
| Difference-in-Differences | Unit and time fixed effects, staggered adoption | Minimum wage effects, healthcare policy changes |
| Regression Discontinuity | Bandwidth selection, manipulation tests | Program eligibility effects, election outcomes |
| Synthetic Control | Pre-treatment fit, placebo tests | State-level policy evaluation, country-level interventions |
| Structural Models | Behavioral assumptions, parameter stability | Labor market interventions, tax policy changes |
Healthcare and Epidemiology
| Common Methods | Key Considerations | Example Applications |
|---|---|---|
| Propensity Score Methods | Confounder selection, balance assessment | Treatment effectiveness, medication side effects |
| Marginal Structural Models | Time-varying confounding, treatment adherence | Long-term treatment effects, dynamic treatment regimes |
| Mendelian Randomization | Genetic confounding, pleiotropy | Causal effects of biomarkers, exposures on disease |
| Target Trial Emulation | Protocol specification, eligibility criteria | Comparing treatment strategies, preventive interventions |
| Risk-Score Matching | Prognostic score development, overlap | Comparative effectiveness, rare disease studies |
Computer Science and Machine Learning
| Common Methods | Key Considerations | Example Applications |
|---|---|---|
| Causal Forests | Feature selection, honest estimation | Heterogeneous treatment effects, personalization |
| Double/Debiased Machine Learning | Sample splitting, regularization | High-dimensional confounder adjustment |
| Invariant Prediction | Environment definition, stability testing | Domain adaptation, transferable predictions |
| Neural Networks for Causal Effects | Representation learning, balance metrics | Individual treatment effect estimation, image data |
| Causal Reinforcement Learning | Exploration strategy, off-policy evaluation | Policy learning, sequential decision making |
Social Sciences
| Common Methods | Key Considerations | Example Applications |
|---|---|---|
| Mediation Analysis | Sequential ignorability, direct/indirect effects | Understanding mechanisms, intervention components |
| Fixed Effects Models | Time-invariant confounding, clustering | Panel data studies, educational interventions |
| Matching Methods | Distance metrics, common support | Program evaluation, observational studies |
| Interrupted Time Series | Seasonality, autocorrelation | Policy implementation, social media interventions |
| Instrumental Variables via Natural Experiments | Exogeneity, local effects | Voting behavior, peer effects |
Best Practices for Credible Causal Inference
Research Design Principles
- Pre-specify analysis plan: Document hypotheses, methods, variables before analysis
- Power analysis: Ensure adequate sample size for meaningful causal estimates
- Diversify methods: Apply multiple approaches with different assumptions
- Triangulate evidence: Combine results across methods, datasets, contexts
- Transparency: Share data, code, and detailed methodological decisions
- Consider mechanisms: Test mediating pathways, not just overall effects
- Plan for heterogeneity: Design to detect variation in treatment effects
- Specify counterfactuals: Clearly define the alternative treatment scenario
Analytical Best Practices
- Balance checks: Verify similarity between treatment and comparison groups
- Covariate selection: Use domain knowledge and causal graphs to select confounders
- Robustness checks: Vary specifications to assess stability of findings
- Sensitivity analysis: Quantify vulnerability to assumption violations
- Avoid overfitting: Use cross-validation or sample splitting
- Report uncertainty: Provide confidence intervals, not just point estimates
- Check functional form: Test for nonlinearities in confounder relationships
- Beware of colliders: Avoid conditioning on variables affected by treatment/outcome
Reporting and Interpretation Guidelines
- Distinguish descriptive and causal claims: Clearly separate association from causation
- State assumptions explicitly: Acknowledge required assumptions for causal interpretation
- Contextualize effect sizes: Relate findings to meaningful benchmarks
- Address generalizability: Discuss external validity and transportability
- Present visual evidence: Use graphs to support causal reasoning
- Acknowledge limitations: Discuss threats to validity transparently
- Consider alternative explanations: Address competing interpretations
- Connect to theory: Link empirical findings to theoretical mechanisms
Resources for Further Learning
Foundational Books
- “Causal Inference in Statistics: A Primer” by Judea Pearl, Madelyn Glymour, and Nicholas Jewell
- “Causal Inference for Statistics, Social, and Biomedical Sciences” by Guido Imbens and Donald Rubin
- “Causality: Models, Reasoning, and Inference” by Judea Pearl
- “Counterfactuals and Causal Inference” by Stephen Morgan and Christopher Winship
- “Causal Inference: What If” by Miguel Hernán and James Robins (freely available online)
Specialized Resources
- “Mostly Harmless Econometrics” by Joshua Angrist and Jörn-Steffen Pischke
- “Elements of Causal Inference” by Jonas Peters, Dominik Janzing, and Bernhard Schölkopf
- “Statistical Methods in Epidemiology” by James Robins and Miguel Hernán
- “Targeted Learning” by Mark van der Laan and Sherri Rose
- “Explanation in Causal Inference” by Tyler VanderWeele
Software Packages
| Language | Package | Specialization | Features |
|---|---|---|---|
| R | MatchIt | Matching methods | Various matching algorithms, balance diagnostics |
| R | grf | Causal forests | Heterogeneous treatment effects, honest estimation |
| R | CausalImpact | Time series | Bayesian structural time series for intervention analysis |
| Python | DoWhy | Causal graph-based inference | Unified API for multiple causal methods, graphical models |
| Python | EconML | Machine learning approaches | Double ML, metalearners, heterogeneous effects |
| Python | CausalML | Uplift modeling | Tree-based and neural network methods |
| R | mediation | Mediation analysis | Causal mediation, sensitivity analysis |
| R | dagitty | Graphical models | DAG creation, testable implications |
| R | twang | Propensity methods | GBM for propensity score estimation |
| Stata | teffects | Treatment effects suite | Comprehensive treatment effect models |
Online Courses and Tutorials
- “A Crash Course in Causality” on Coursera by Jason Roy
- “Causal Diagrams” on edX by Miguel Hernán
- “Causal Inference” at Stanford Online by Jas Sekhon
- “Introduction to Causal Inference” from Brady Neal (freely available)
- “Causal Inference with Python” tutorials from Microsoft Research
This cheatsheet provides a structured overview of causal inference concepts and methods, but the field continues to evolve rapidly. Always consult the latest research and domain-specific literature when applying these techniques. Causal claims require careful consideration of underlying assumptions and limitations, alongside substantive knowledge of the study context.