Key Driver Analysis

You run a customer survey. Twelve items — support, pricing, speed, reliability, onboarding, and the rest — each rated 1 to 5. The averages come back and they all sit between 3.8 and 4.3. Overall satisfaction is a solid 4.1. Now the executive question lands: if we can only improve one thing next quarter, which one?

The tempting move is to ask people directly — "how important is support to you?" — and rank the answers. But almost everyone rates almost everything as important; stated importance flattens into a wall of 4s and 5s that tells you nothing. The move that works is to stop asking and start measuring: look at how each item's score co-moves with overall satisfaction, and let the ones that genuinely swing the outcome rise to the top.

Why Ranking by Satisfaction — or by Correlation — Misleads

Two shortcuts feel reasonable and both fail. Ranking items by their own satisfaction score tells you where you're weak, not where weakness hurts — a low score on something nobody cares about is not your problem. Ranking by simple correlation with the outcome is closer, but it double-counts: if support quality and response speed are themselves correlated, each looks important partly because of the other, and you can't tell which one to actually move.

Driver analysis fixes this by putting every item into a single multiple regression against the outcome at once. Each item's coefficient is its unique contribution — its effect after holding the other items constant. That's what turns a pile of overlapping ratings into a clean, defensible ranking.

How Driver Analysis Works

The engine is ordinary least-squares regression: overall satisfaction is the dependent variable, and every driver item is a predictor.

Overall satisfaction  =  b0
                      +  b1 x (support quality)
                      +  b2 x (ease of use)
                      +  b3 x (price / value)
                      +  ...  +  error

The raw coefficients (b) come in different units, so you can't compare them directly. The fix is the standardized coefficient (β) — each variable rescaled to standard deviations. A β of 0.34 means: a one-standard-deviation lift in that driver moves overall satisfaction by 0.34 standard deviations, holding everything else fixed. Now every driver is on the same ruler and the ranking is meaningful.

1. 1Pick the outcome (overall satisfaction, likelihood to recommend, renewal intent).
2. 2Regress it on all the driver items simultaneously.
3. 3Standardize the coefficients so drivers are comparable (β).
4. 4Rank by |β| — and convert to a relative-importance % that sums to 100.
5. 5Cross the importance ranking with each driver's own score to decide where to act.

A Worked Example

Say a SaaS product runs the survey and the regression returns these standardized betas. Converting each to a share of the total (positive) importance gives the relative-importance column:

Notice what the ranking exposes. Onboarding scores highest (4.5) but barely moves the outcome — pouring more effort there is wasted. Support quality is both the biggest driver and the weakest score (3.6): that gap is your single highest-return fix. Ranking by satisfaction alone would have sent you to polish onboarding; driver analysis sends you to support.

The Importance-Performance Matrix

The cleanest way to act on driver results is to plot importance (|β|) against performance (the driver's own mean score) and split at the averages. Four quadrants tell you exactly what to do with each driver:

                 high importance
                       |
     PRIORITIZE        |        MAINTAIN
  (important, weak) -- + -- (important, strong)
     SECONDARY         |        MONITOR
  (minor, weak)        |     (minor, strong)
                       |
                 low importance
   low performance --------- high performance

• Prioritize — high importance, low score: your highest-ROI improvements. Act here first.
• Maintain — high importance, high score: strengths that matter. Protect them, don't coast.
• Secondary — low importance, low score: weak spots that don't move the needle. Fix later, if ever.
• Monitor — low importance, high score: fine as-is. You may even be over-investing here.

Common Mistakes

Ignoring multicollinearity

When two drivers are highly correlated (say, "support quality" and "support speed"), regression struggles to assign credit and the coefficients get unstable — sometimes even flipping sign. The diagnostic is the Variance Inflation Factor (VIF). A VIF above 5 is a warning; above 10 is a real problem. The fix is usually to combine the overlapping items into one driver or drop the redundant one.

Trusting stated importance

If you also asked "how important is X," don't be surprised when it disagrees with the derived ranking. People over-claim on socially desirable items (security, reliability) and under-report on emotional ones. Derived importance reflects behavior in the data; stated importance reflects what people think they should say.

Too few responses

With fewer than ~10 responses per driver, the betas are noisy and the ranking can reshuffle on a different sample. If you can't hit the sample size, use fewer, broader drivers rather than many narrow ones.

Confusing a driver with a cause

Driver Analysis in the SKARI Survey Lab

SKARI's Survey Lab runs the whole pipeline from your response data — no formulas to wire up. You map an outcome question and the driver items, and it fits the regression, standardizes the coefficients, and lays out the results the way you'd actually use them.

• Standardized β with 95% confidence intervals — significant drivers (p < .05) highlighted
• Relative-importance % (Pratt's measure) so contributions sum to 100 and are easy to communicate
• The Importance-Performance matrix with automatic Prioritize / Maintain / Secondary / Monitor quadrants
• A VIF flag on every driver so multicollinearity can't quietly distort the ranking
• A gap score — |β| × (avg performance − driver score) — surfacing the highest-ROI fixes directly
• Residual plots and an R² / F-test so you can see whether the model actually fits

It even includes a what-if simulator: nudge a driver's score up and watch the predicted change in the outcome, with a prediction interval — turning the analysis into a planning tool, not just a report.

Frequently Asked Questions

Driver analysis or IPA — what's the difference?

They share the Importance-Performance matrix, but they get importance differently. Classic IPA often uses stated importance (you ask). Key driver analysis uses derived importance (regression). Derived is more robust because it doesn't rely on people accurately judging their own priorities.

What outcome should I use?

Whatever decision you care about: overall satisfaction, likelihood to recommend (NPS), renewal intent, or purchase intent. The drivers are the specific attributes you can act on.

What is relative importance %?

It rescales the driver effects so they sum to 100%, making them easy to present ("support is 34% of what drives satisfaction"). SKARI uses Pratt's measure, which combines each driver's coefficient with its correlation to the outcome.

Can a driver have a negative β?

Yes, and it's worth investigating. It usually signals multicollinearity (check the VIF) or a genuinely inverse relationship — for example, more sales contact lowering satisfaction for self-serve users.

How many drivers is too many?

Keep it to what your sample supports (≈10 responses each) and to distinct concepts. Fifteen overlapping items produce a messier, less stable ranking than six clean ones.

Key Takeaways

When every survey item looks equally important, driver analysis is how you break the tie with evidence. Regress the outcome on all the drivers, rank by standardized effect, cross it with performance, and act on the Prioritize quadrant — while keeping an eye on multicollinearity so the ranking stays honest.