Confounding Variable
A confounding variable is a pre-treatment factor associated with both an intervention or exposure and its outcome, creating a misleading association when it is not addressed.
Download Markdown (.md)Soroush’s Product Wiki · https://wiki.soroush.uk/topic/confounding-variable/
01What it means
A confounder is a common cause, or proxy for one, that influences both the exposure being compared and the outcome. For example, account maturity may affect both feature adoption and renewal, making adopters look more successful even if adoption itself has no causal effect.[1][2]
Association of interest: adoption → renewal
An observed association between exposure and outcome can mix the effect of interest with the common cause.
02Why it matters
Confounding is a central reason observational product analytics cannot be read as impact. Naming likely confounders makes causal assumptions inspectable and prevents teams from targeting an association that will not change outcomes.
03When to use it
Assess confounding before interpreting nonrandomized comparisons, including cohorts, feature-adopter analyses, and campaigns. Use randomization where feasible; for observational work, involve appropriate expertise and define variables from a causal model rather than adjusting every available field.
04How it works
- Specify the intervention, outcome, comparison, and time ordering.
- Use domain knowledge to identify plausible common causes before treatment.
- Draw or state the assumed causal relationships.
- Choose a design or adjustment strategy justified by those assumptions.
- Check overlap, measurement quality, sensitivity, and remaining uncertainty.
05A practical example
A team observes that users who attend onboarding webinars convert more. Because motivated customers are more likely to attend and buy, it randomizes invitations to estimate the effect of encouragement instead of attributing the raw difference to the webinar.
06Where it breaks
Unmeasured confounding can remain after adjustment, and controlling for variables caused by treatment can introduce bias. A statistical model cannot determine causal roles without substantive knowledge and a defensible design.
Common mistakes
- Adjusting for every available variable without a causal rationale.
- Controlling for a mediator or post-treatment outcome.
- Assuming a correlated demographic field fully removes an underlying confounder.