AI Causal Inference Analyst Interview Questions

A strong answer references the counterfactual framework - we never observe both potential outcomes for the same unit - and explains how selection bias confounds naive comparisons.

The answer should define each node type precisely and describe the correct adjustment strategy for each - condition on confounders, never on colliders, and clarify whether mediation analysis is the goal.

A great answer explains how aggregated data can reverse the direction of an association found in disaggregated data, and why DAGs help determine the correct conditioning strategy.

The answer should reference how randomization breaks confounding by making treatment assignment independent of potential outcomes, eliminating systematic bias.

A good answer defines propensity score as the probability of treatment given covariates and explains its use in matching, weighting (IPTW), or stratification to achieve balance.

A strong answer covers parallel pre-trends, no anticipation, SUTVA, and shows event-study plots to validate pre-treatment trend equivalence.

The answer must cover relevance (instrument correlates with treatment), exclusion restriction (instrument affects outcome only through treatment), and independence, plus practical justifications.

A strong answer defines the running variable, cutoff, and local randomization argument, then explains sharp RDD assumes perfect compliance while fuzzy RDD uses IV-style estimation for imperfect compliance.

The answer should explain that doubly robust estimators are consistent if either the propensity score model or outcome model is correctly specified (but not necessarily both), providing a safety net.

A good answer discusses issues like poor overlap, sensitivity to matching algorithm, loss of sample, and suggests alternatives like IPW, doubly robust methods, or redefining the estimand.

The answer should define Stable Unit Treatment Value Assignment and cite examples like network spillovers, market-level interventions, or congestion effects where interference breaks SUTVA.

A strong answer discusses observational methods like DiD if there is a clear pre/post period, synthetic controls if there are comparable units, or RDD if there is a rollout threshold.

The answer should define each estimand clearly and discuss when policy relevance or generalizability drives the choice - e.g., ATT for evaluating a program that already enrolled participants.

A great answer explains donor pool construction, the matching on pre-treatment outcomes, placebo tests, and the requirement that no comparable unit is simultaneously treated.

The answer should reference false discovery rate control (Benjamini-Hochberg), Bonferroni correction, pre-registration of primary outcomes, and the distinction between exploratory and confirmatory analysis.

A strong answer covers honest estimation, splitting on treatment effect heterogeneity rather than outcome prediction, and the advantage of data-driven subgroup discovery with valid confidence intervals.

The answer should explain the two-step targeting procedure, the substitution principle, how it uses the efficient influence curve, and why it is doubly robust and semi-parametrically efficient.

A great answer covers Rosenbaum bounds, E-values (VanderWeele and Ding), and stress-testing via the Cinelli and Hazlett framework with partial R-squared measures of omitted variable bias.

The answer must discuss time-varying confounders affected by prior treatment (g-estimation, marginal structural models, IPTW for longitudinal data) and why standard regression induces collider or overadjustment bias.

A strong answer explains that the front-door criterion identifies causal effects through a mediator that fully transmits the treatment's effect and is not directly confounded, using Pearl's graphical criteria.

The answer should discuss differences in effect modification across populations, covariate shift, and methods like reweighting, meta-analysis, and covariate-adjusted transportability frameworks.

A great answer references path-specific effects, direct vs. indirect discrimination through protected attributes, counterfactual fairness (Kusner et al.), and mediation analysis for discrimination auditing.

The answer should address SUTVA violations, interference via network structure, cluster randomization, partial identification under interference, and the challenges of learning causal effects in a feedback-rich environment.

A strong answer covers constraint-based (PC algorithm), score-based (GES), and hybrid methods, the faithfulness assumption, identifiability limitations, and why domain knowledge remains essential.

The answer should discuss preventing p-hacking and HARKing, internal pre-registration templates, analysis plans locked before data inspection, and the cultural shift needed in industry analytics teams.

A strong answer would evaluate DiD if there are pre/post periods and a clear treatment group, synthetic control for market-level comparison, check for spillovers between treated and control markets, and present robustness checks.

The answer should discuss propensity score matching or IPW to address selection on observables, sensitivity analysis for unmeasured confounders (e.g., motivation), and possibly bounds or IV approaches.

A great answer addresses confounding by indication, discusses using physician prescribing tendencies as an instrument (provider preference IV), or propensity score methods with rich EHR covariates, plus sensitivity analysis.

The answer should discuss noncompliance in experiments, intention-to-treat vs. complier average causal effect (CACE/LATE), and how to model spillover effects, potentially using network-aware randomization.

A strong answer discusses seasonality controls, DiD with appropriate comparison groups, synthetic control using markets without the campaign, and the importance of understanding the timing confound explicitly.

The answer should identify threats like regression to the mean, seasonality, concurrent changes, and suggest a synthetic control or DiD approach with proper counterfactuals and robustness checks.

A great answer covers ability bias and omitted variable bias, discusses compulsory schooling laws as instruments (Angrist-Krueger style), returns to education literature, and the limitations of IV extrapolation.

The answer should discuss the absence of a control group, potential use of staggered rollout for DiD, synthetic controls using pre-rollout trends, mediation analysis, and the political sensitivity of the findings.

A strong answer discusses mediator-induced confounding, whether the price change is part of the treatment mechanism or a post-treatment confounder, principal stratum effects, and how to decompose direct and indirect effects.

The answer should cover two-way fixed effects DiD, Callaway and Sant'Anna or Sun and Abraham methods for staggered adoption bias, event study visualization, and sensitivity to parallel trends violations.

A strong answer covers model (create causal graph), identify (compute estimand using do-calculus), estimate (apply causal estimator), and refute (run robustness refutation tests).

The answer should cover the double machine learning approach, feature importance for treatment effect heterogeneity, SHAP values on CATE, and how to segment populations by predicted treatment effect.

A great answer discusses LLM use for literature review, generating candidate DAGs from domain knowledge interviews, code generation for analysis pipelines, and the critical need to validate LLM suggestions against causal theory.

The answer should cover dbt models for covariate tables and treatment assignment, Python scripts for DiD or matching estimation, parameterized notebooks, and dashboard templates for treatment effect reporting.

A strong answer covers iterative DAG construction, implied conditional independence tests, identification of valid adjustment sets, and feedback loops with domain experts to refine the structure.

The answer should discuss online propensity estimation, doubly robust estimation with streaming data, monitoring for covariate shift, and the difference between off-policy evaluation and causal effect estimation.

A great answer discusses prior specification for treatment effects, posterior distributions over causal quantities, handling of uncertainty propagation, and the ability to incorporate domain knowledge directly.

The answer should cover versioned analysis scripts, data versioning (DVC), automated testing of causal assumptions (balance checks, parallel trends), and documentation of identification strategy in the repository.

A strong answer discusses NLP extraction of mediator or confounder variables from clinical notes, product reviews, or support tickets, embedding-based propensity scores, and the risk of introducing bias through model misspecification.

The answer should cover experiment randomization infrastructure, sample size computation, frequentist hypothesis testing alongside Bayesian posterior estimation, and dashboards for sequential monitoring with proper alpha-spending.

A strong answer demonstrates intellectual courage, clear communication of assumptions and limitations, willingness to revisit analysis for robustness, and the ability to present uncomfortable truths diplomatically.

The answer should discuss a framework of assumption plausibility, sensitivity analysis results, cost of being wrong, and the decision-theoretic perspective on acting under uncertainty.

A great answer shows structured learning (papers, textbooks, implementations), seeking expert guidance, building small test cases, and validating understanding through simulation before applying to real data.

The answer should discuss translating statistical uncertainty into business terms, using scenarios and ranges, visual confidence intervals, and framing the decision as risk management rather than binary truth.

A strong answer covers following key researchers (Athey, Imbens, Pearl, VanderWeele), reading top journals, testing methods on simulated data before production use, and evaluating practical assumptions alongside statistical properties.