Quick answer
Key driver analysis ranks which factors, such as early behaviors or survey ratings, are most strongly associated with an outcome like retention or satisfaction. It shows association, not cause, so a top driver should be checked for confounders and confirmed with a holdout before you act on it.
What is key driver analysis?
Key driver analysis finds which factors are most strongly associated with an outcome you care about, such as retention, purchase or satisfaction. You list candidate drivers, measure how each one lines up with the outcome and rank them. The result is a ranking of associations, not of causes: it tells you where to look first, and a driver has to be confirmed with a comparison group or a holdout before you spend effort on it.
Two traditions use the name. In market research it usually means a regression on survey ratings, which produces the "derived importance" of each rating for overall satisfaction. In product analytics it means comparing outcomes for people who did and did not do something. This page covers the product version; the logic of ranking and the traps are the same.
Survey and behavioral key driver analysis
| Survey version | Behavioral version | |
|---|---|---|
| Outcome | Overall satisfaction, NPS, purchase intent | Retained at day 30, paid within 30 days |
| Candidate drivers | Ratings of product attributes | Actions taken in the first days |
| Typical method | Regression and relative importance of each rating | Outcome rate with vs without the action, then checks for confounders |
| Output | An importance share per attribute | A rate difference and ratio per action |
| Main trap | Ratings move together, so importance is hard to split | Users who do more of everything make every action look like a driver |
How to run one, step by step
- Define the outcome with a window. "Retained" must be a rule, such as active on at least one day in week 4.
- List candidate behaviors, measured early. Count each action only in a window that ends before the outcome starts, such as the first 7 days. An action taken after the outcome cannot have caused it.
- Compare the outcome rate with and without each action. For every candidate, calculate the retained share among people who did it and among people who did not.
- Rank by the difference, and look at group size. An action that only 10 people took moves its rate by 10 points per person, so a large difference there means little.
- Check confounders. Split the comparison by anything that could drive both the action and the outcome, such as company size or acquisition channel, and see whether the gap survives.
- Validate the top driver. Change the behavior for a random group and keep another group as a holdout. Only that comparison supports a causal claim.
Worked example: ranking four behaviors
Sample data: 1,000 new users, of whom 320 (32%) were still active in week 4. For each of four actions in the first 7 days:
| Action in first 7 days | Did it | Retained, did it | Retained, did not | Difference | Ratio | Correlation (phi) |
|---|---|---|---|---|---|---|
| Invited a teammate | 200 | 60.0% | 25.0% | +35.0 points | 2.40 | 0.30 |
| Connected a data source | 450 | 42.0% | 23.8% | +18.2 points | 1.76 | 0.19 |
| Viewed a report | 700 | 36.0% | 22.7% | +13.3 points | 1.59 | 0.13 |
| Opened the pricing page | 300 | 30.0% | 32.9% | −2.9 points | 0.91 | −0.03 |
Inviting a teammate tops the ranking, and the top three differences are far beyond chance (z between 4.1 and 9.5). The pricing page shows no relationship (p = 0.37). For a yes/no action against a retained yes/no, the correlation is the phi coefficient, which is Pearson's r computed on two 0/1 variables: 0.30 for invites.
Why a ranking is not causation
The top driver deserves a check before anyone builds a campaign around it. Split the same users by company size:
| Company size | Users | Invited | Retained, invited | Retained, not invited |
|---|---|---|---|---|
| Large | 300 | 190 (63.3%) | 118 of 190 (62.1%) | 66 of 110 (60.0%) |
| Small | 700 | 10 (1.4%) | 2 of 10 (20.0%) | 134 of 690 (19.4%) |
Inside each group the invite gap is only +2.1 and +0.6 points. Weighting the two groups by their share of the 1,000 users (30% large, 70% small) gives an adjusted gap of +1.0 point, against the raw +35.0. Company size explains the ranking: large companies both invite more (63.3% vs 1.4%) and stay more (61.3% vs 19.4%). This is a confounding variable, and it is the usual way a top driver turns out to be a bystander. The small-company invite group has only 10 users, so that cell is thin; the point is that the raw ranking would have sent you after the wrong lever. Statisticians know the pattern as Simpson's paradox.
Validating a driver with a holdout
To confirm that invites move retention, split new users at random, nudge one half to invite a teammate and leave the other half alone as the holdout, then compare week-4 retention with a significance test (see p-value). The sample you need is the catch. To detect 25% → 30% retention with 80% power at α = 0.05, you need about 1,248 users per group. To detect 25% → 35%, about 326 per group. With 1,000 new users in total you can only confirm a large lift, so plan for a longer window. See statistical power for how those numbers behave.
Common mistakes
- Counting actions after the outcome. "Opened the billing page" in week 5 is a result of staying, not a reason to stay. Keep the driver window before the outcome window.
- Letting heavy users hide everything. A power user does more of every action, so every action correlates with retention. Compare people with similar activity levels.
- Ranking tiny groups. A behavior only a handful of people did can show any rate you like.
- Testing too many candidates. Check 40 behaviors at α = 0.05 and the chance that at least one looks significant by luck is 1 − 0.95⁴⁰ = 87.1%, if they were independent. Short-list candidates from product knowledge first.
- Treating the ranking as a recipe. Pushing people to do the top action may not change the outcome if both are caused by something else. For the link between correlation and cause, see correlation vs causation.
- Mixing it up with change attribution. Splitting a metric's change across channels or countries answers "where did the change come from", not "which behaviors lead to retention".
Key driver analysis in TapCub
Say you want to know whether inviting a teammate in the first 7 days goes with week-4 retention. TapCub has no automatic key-driver report, so this is a manual job built from segments and retention.
- Build the segments. In All-platform analytics › Users › Segments & tags, create a rule segment for people first seen in a chosen period who did the invite event, and a second for those who did not. Keep the time window on the did condition before the outcome period, and preview the counts before saving. Segments hold only identified people, so anonymous visitors never appear; identify users first.
- Run retention on each. In All-platform analytics › Behavior › Behavior analysis, open the Retention tab and use Analyze on a segment to apply it as a cohort. Retention has no breakdown and no compare period, so run it once per segment and read the rows side by side: each row shows the cohort size and the rate for each period.
- Split by the confounder. Add a company-size condition to each rule and rebuild the four groups from the table above. If the gap shrinks the way it did here, invites were not the lever.
- Look at the whole-metric change separately. Why did it change in All-platform analytics › Operations › AI Assistant splits a metric's change by channel, referrer domain, country, device, entry page or platform. It shows where a change came from, not which behaviors drive retention.
- Keep the holdout on your side. TapCub does not run experiments or hold groups out. Randomize in your own product, then compare the nudged group and the holdout as two segments.
Sources
- Grömping, U. (2006). Relative Importance for Linear Regression in R: The Package relaimpo. Journal of Statistical Software, 17(1).
- Stanford Encyclopedia of Philosophy, Simpson's Paradox.
- NIST/SEMATECH e-Handbook of Statistical Methods, 7.3.3 How can we determine whether two processes produce the same proportion of defectives? (the two-proportion test behind the z values and sample sizes).
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Product Analyst
Focuses on event analysis, conversion funnels, and retention, using clear metric definitions and structured analysis to explain product performance.