--- title: Causal Inference — definition | RGM® Glossary url: https://realgrowthmatters.com/glossary/causal-inference/ updated: 2026-06-10 source_html: https://realgrowthmatters.com/glossary/causal-inference/ --- # Causal Inference caus·al in·fer·encenoun Did the campaign cause the sales, or just attend them — the discipline that separates marketing's effects from its coincidences. Term : Causal Inference Asks : What caused this, not what preceded it Gold standard : Randomized experiments (holdouts) Enemy : Confounders and selection effects ## Forms & parts of speech causal inference · noun Cause-vs-coincidence discipline. "Retargeted users bought 5x more - **causal inference** asked whether retargeting caused that, or just selected buyers." ## Definition in plain terms Causal inference is the discipline of determining what actually caused an outcome — as opposed to what merely correlated with it, preceded it, or was standing nearby when it happened. Marketing analytics is soaked in correlations that flatter: exposed users buy more, email openers retain better, retargeted visitors convert at multiples. Causal inference asks the only question budgets should care about — would the outcome have happened without the marketing — and supplies the methods for answering it honestly. ## The mechanics The central enemy is the comparison that selection built. Retargeted users convert more because retargeting selects people already shopping; email openers retain better because opening email reveals engagement, not because the email caused it (a CONFOUNDING VARIABLE — engagement — drives both sides). The methods ladder by rigor. Randomized experiments sit on top: assign exposure by coin flip and the selection problem dies, which is what AB-TESTS, CONVERSION-LIFT holdouts, and geo-experiments operationalize. Where randomization is impossible, quasi-experimental methods approximate it — difference-in-differences (compare the change in treated markets against the change in untreated ones), regression discontinuity (compare just-above against just-below a threshold), synthetic controls (build a statistical twin of the treated region), and matching methods that pair exposed with comparable unexposed users. Observational causal modeling — the Judea Pearl tradition of causal graphs and adjustment — formalizes which variables must be controlled and which must not (adjusting for the wrong variable creates bias rather than removing it). The marketing applications are everywhere once named: incrementality of channels, true promotion effects against the baseline that would have happened, price elasticity, creative effects. The discipline's posture is humility with teeth — default to 'selection until proven otherwise' for any non-randomized comparison, and let experiments calibrate the observational models that run between them. ## When it matters Causal inference matters wherever spend defends itself with numbers — which is everywhere — because correlational evidence systematically flatters whatever selects engaged audiences. It matters most at scale decisions: doubling a channel that selects buyers wastes the double; killing one that quietly causes them costs the base. The discipline is a hierarchy of evidence — experiments where possible, quasi-methods where not, causal modeling to structure the rest — and one standing question pinned over every dashboard: compared to what would have happened anyway? **Worked example.** A streaming service's analytics show that users who see its win-back ads resubscribe at 5x the rate of those who don't, and the retention team requests triple the budget. The analyst reframes: the ad platform targets users showing return-intent signals, so the comparison bakes in selection - the ads find resubscribers at least as much as they create them. A randomized holdout settles it: 20% of the eligible audience sees no win-back ads for eight weeks. The exposed group resubscribes at 6.1%, the holdout at 4.8% - real lift, but a fraction of the correlational 5x story, and concentrated almost entirely in users lapsed under 60 days. Budget lands at 1.4x rather than 3x, aimed at the recent-lapse window where causation actually lives, and the team adopts the standing rule: any number built on a non-randomized comparison carries a selection warning until an experiment says otherwise. **Failure modes to watch.** Reading exposed-versus-unexposed comparisons as effects when targeting built the difference; adjusting for variables that create bias instead of removing it; quasi-methods applied without checking their assumptions; experiments skipped because the correlational answer was more flattering; and forgetting the only question that matters - compared to what would have happened anyway. ## Synonyms & antonyms ### Synonyms causal inferencecausal analysisincrementality (the question) ### Antonyms correlation analysislast-touch credit ## Origin & history Causal inference matured from statistics and econometrics — randomized experimentation from R.A. Fisher's 1920s designs, the potential-outcomes framework from Neyman and Rubin, and causal graphs from Judea Pearl's work (Causality, 2000) — and marketing adopted the toolkit as incrementality questions outgrew what attribution's correlations could answer. Etymology: [source](https://en.wikipedia.org/wiki/Causal_inference). ## Usage trends Search interest for this term over the last five years: [View interest-over-time on Google Trends →](https://trends.google.com/trends/explore?q=causal%20inference&date=today%205-y) ## Common questions What is causal inference? : The discipline of determining what actually caused an outcome — separating marketing's real effects from correlations created by selection, timing, or confounders. Why do marketing correlations mislead? : Targeting selects engaged audiences, so exposed users were already likelier to convert — retargeted visitors buying more proves the selection worked, not that the ads did. What methods does causal inference use? : Randomized experiments first (holdouts, geo-tests), then quasi-experimental methods — difference-in-differences, regression discontinuity, synthetic controls — structured by causal modeling of what to control. ## Related tools & calculators - tool[A/B test sample size](/tools/a-b-test-sample-size/) ## Resources & people to follow - reference[Wikipedia — Causal inference](https://en.wikipedia.org/wiki/Causal_inference) - referencePearl — causal graphs and the adjustment formalism - referenceRGM analysis — selection until proven otherwise; pin 'compared to what?' over every dashboard Curated, non-competitor resources verified per term. ## Related training - module[Performance marketing](/training/performance-marketing-foundations/) ## Disciplines Areas of marketing where causal inference is a core concern: [Performance marketing](/training/performance-marketing-foundations/)[Growth strategy](/training/growth-marketing-foundations/) ## Read next ## Related terms [Conversion lift](/glossary/conversion-lift/)[Incrementality testing](/glossary/incrementality-testing/)[Confounding variable](/glossary/confounding-variable/)[AB test](/glossary/ab-test/)[Marketing attribution](/glossary/marketing-attribution/) ## Sources 1. trends[Google Trends — "causal inference"](https://trends.google.com/trends/explore?q=causal%20inference&date=today%205-y)