How to Choose the Right Causal Inference Method
July 2026 | Applied Economics | Causal Inference | Industrial Organization | Marketplace Economics | ← Back to Blog
The hardest part of causal inference is not choosing a method. It is constructing a credible counterfactual. This post walks through a practical decision framework for selecting the right empirical strategy when randomization is not possible.
The Core Problem
The hardest part of causal inference is not choosing a method. It is constructing a credible counterfactual.
In most real-world business settings, we cannot run every experiment we want. The question is what we can credibly estimate with the data and setting we have.
Without a credible counterfactual, we cannot separate correlation from causation.
Randomization is not always possible because of budget, timing, ethics, or simply because the intervention has already happened. A merger has closed. A pricing initiative has already launched. A regulatory change has already taken effect.
So the real question becomes: which method can I credibly use here?
"There is no universally best method. Every method trades stronger assumptions for broader applicability. The best method is the one whose assumptions are most credible for your setting."
The Decision Framework
The answer depends on which assumptions are credible, what data are available, and what is feasible in practice. Here is how I think about it.
If you can randomize, an A/B test is usually the cleanest answer. Random assignment breaks the link between treatment and confounders. The identification assumption is satisfied by design. Standard inference applies.
If network effects prevent user-level randomization, a switchback experiment randomizes over time periods instead. When a price change in one market affects supply or demand in adjacent markets, user-level A/B testing introduces interference. Switchbacks handle this by alternating treatment and control periods rather than treatment and control units.
If a comparable control group exists and parallel trends is credible, Difference-in-Differences is often the right tool. DiD compares the change in outcomes over time between a treated group and a control group. The key assumption is that absent the intervention, the two groups would have followed the same trend. Common macroeconomic shocks are differenced out. What matters is whether differential trends between the groups are plausible.
If no single control unit is defensible, Synthetic Controls constructs a weighted combination of donor units that tracks the treated unit before the intervention. Rather than selecting one comparison unit and arguing it is similar, synthetic controls builds a data-driven counterfactual whose pre-treatment fit is directly visible and verifiable. This is particularly useful in antitrust cases where the yardstick market is contested.
If there are many observed confounders and the relationships are complex, Double Machine Learning helps isolate the treatment effect after accounting for those confounders. DML partials out the confounders from both the treatment and outcome using flexible machine learning models, then estimates the causal effect on the residuals. It is particularly useful when there are many potential confounders and the relationship between those confounders, treatment, and outcome is complex or unknown.
The Principle Behind the Framework
Good causal inference is less about knowing every method and more about understanding the assumptions each method requires.
Every causal inference method relies on different identifying assumptions and makes different trade-offs between credibility and applicability. The best method is the one whose assumptions are most credible for your setting.
This applies whether you are evaluating a pricing change at a marketplace, simulating the competitive effects of a merger, measuring overcharge damages in an antitrust case, or assessing the impact of a platform policy on buyer behavior.
The challenge is always the same: constructing a credible counterfactual. The tool matters less than the identification strategy behind it.
Further Reading
Angrist, J. and Pischke, J.S. (2009). Mostly Harmless Econometrics: An Empiricist's Companion. Princeton University Press. The standard reference for applied causal inference.
Card, D. and Krueger, A. (1994). Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania. American Economic Review, 84(4), 772–793. The landmark DiD application.
Abadie, A., Diamond, A., and Hainmueller, J. (2010). Synthetic Control Methods for Comparative Case Studies. Journal of the American Statistical Association, 105(490), 493–505.
Abadie, A. (2021). Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects. Journal of Economic Literature, 59(2), 391–425. The practitioner guide to synthetic controls.
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., and Robins, J. (2018). Double/Debiased Machine Learning for Treatment and Structural Parameters. The Econometrics Journal, 21(1), C1–C68. The foundational Double ML paper.
Kohavi, R., Tang, D., and Xu, Y. (2020). Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing. Cambridge University Press. The practitioner guide to A/B testing at scale.
