What is a Dependent Variable?

The dependent variable (DV) is the outcome a study measures: the factor expected to depend on the independent variable's manipulation. In the causal sentence "we vary X to see its effect on Y", the DV is Y: conversion rate in an A/B test, task success and time in a usability comparison, a satisfaction score in a survey experiment. The IV asks the question; the DV is where the answer is read. And because the DV defines what "worked" means, choosing it is not a technical afterthought but the study's value judgment, made operational.

Operationalising the Outcome

Concepts don't fit on dashboards; measures do, and the distance between them is where studies go wrong. "Better onboarding" must become something countable (activation within 7 days, completion of setup, first-week retention), and each candidate captures a different slice of the concept while missing others: the construct validity problem, applied to outcomes. Three properties make a good DV. Sensitivity: it can actually move in the study's timeframe and sample (quarterly churn cannot answer a two-week test). Proximity: it sits close enough to the manipulation for effects to reach it, while still mattering (clicks are proximate and cheap; revenue is distal and slow; the craft is pairing them). Definitional precision: a written rubric for what counts (what is "task success" exactly?), because a DV scored by vibe drifts between raters and studies, the reliability requirement, applied to outcomes.

Primary, Secondary, and Guardrails

Mature experiments name one primary DV before fielding: the outcome the decision rides on, the one the sample was powered for, the one that cannot be swapped when results disappoint. Secondary DVs add texture (segments, mechanisms, related outcomes), explicitly exploratory, their surprises feeding the next study rather than this one's verdict, the multiple-comparisons discipline in outcome form. Guardrail metrics watch for collateral damage: the checkout tweak that lifts conversion (primary) while quietly raising refunds (guardrail). A DV set without guardrails optimises the measured at the expense of the unmeasured, which is how metric wins become product losses.

The Craft Rules

1. Write the DV into the hypothesis, with its rubric.
"Variant B increases setup completion (defined as reaching step 5 within one session)" is testable; "improves onboarding" is a wish.

2. Prefer behaviour, paired with report.
Behavioural DVs (completion, time, errors) resist performance and politeness; attitudinal DVs (ratings, SUS) capture the experience behaviour can't. Mixed instruments collect both in one pass, a task's success metric beside a video "what was difficult?", the design a Ballpark study makes routine, so every movement in the dial ships with its explanation.

3. Measure identically across conditions.
The DV's collection (timing, instrument, scorer) must not vary with the IV, or the measurement becomes a second, hidden manipulation.

4. Report the DV's magnitude with its uncertainty.
Effect size and interval, not just the verdict: the DV's job is to inform a decision, and decisions run on magnitudes.

The Takeaway

The dependent variable is success, operationalised: choose the outcome the decision truly rides on, define it to a rubric, power the study for it, surround it with guardrails, and refuse to renegotiate it after the data arrives. Experiments answer exactly the question their DV asks, which is why writing that question carefully is the first analysis.