Your satisfaction survey has 15 attribute ratings and one overall score, and the team wants to know which attribute to work on first. Ranking the attributes by their average score tells you where you are weak, not where weakness hurts. Key driver analysis answers the second question. This guide covers stated vs derived importance, the main driver analysis methods, a worked example and how to read the results by segment. It sits within our wider guide to customer segmentation analysis, because drivers often differ from one group of customers to the next.
Key takeaways
- Key driver analysis estimates how much each attribute influences an outcome such as overall satisfaction, NPS or purchase intent.
- Stated importance asks people what matters; derived importance infers it from how ratings move with the outcome. They often disagree, for good reasons.
- Correlation is easy but double-counts overlapping attributes; Shapley value regression and relative weights split shared influence fairly.
- Quadrant analysis plots importance against performance to show what to fix, keep, deprioritize or stop overinvesting in.
- Drivers describe association, not cause. Treat them as a ranked list of hypotheses to act on and test.
What is key driver analysis and why does it matter?
Key driver analysis links a set of attribute ratings (speed, price, support, ease of use) to one outcome measure, usually overall satisfaction, likelihood to recommend, purchase intent or brand preference. Each attribute gets an importance score, which shows how strongly it is associated with the outcome.
It matters because budgets are finite. Teams rarely ask “where do we score lowest?” for long; they ask “what should we fix first to raise satisfaction?” Combined with performance scores, a driver analysis turns a long list of ratings into a short, ordered list of priorities.
What is the difference between stated and derived importance?
There are two ways to measure importance. You can ask customers directly, or you can infer it from their answers.
| Stated importance | Derived importance | |
|---|---|---|
| How it is measured | Ask “How important is X to you?” or have people rank or trade off attributes | Relate each attribute rating to the outcome statistically |
| What it tells you | What customers believe matters | What actually moves the outcome in your data |
| Strengths | Simple to explain; captures must-have basics | No extra questions; reflects how customers actually rate you |
| Weaknesses | Everything gets rated important; social desirability | Sensitive to correlated attributes and sample size |
| Survey length | Adds a question per attribute | None beyond the ratings |
Why do stated and derived importance disagree?
Disagreement is normal. Basics such as safety or billing accuracy are often stated as very important but show low derived importance, because almost everyone rates you acceptably on them. With little variation, they cannot explain differences in satisfaction. The reverse also happens: people rarely say “friendly staff” matters most, yet it can track overall satisfaction closely. In our experience, stated vs derived importance is most useful read together. A basic with high stated and low derived importance is a hygiene factor: keep it at standard, but do not expect gains from pushing it higher.
Which driver analysis methods should you use?
All derived methods start from the same data, attribute ratings and an outcome for each respondent. They differ in how they treat attributes that overlap.
Correlation. The simplest derived importance is the correlation between each attribute and the outcome, explained in our guide to correlation analysis. It is easy to compute and explain. Its weakness is that it looks at each attribute alone, so two attributes that measure similar things both get full credit.
Multiple regression. Regression estimates every attribute’s effect while holding the others constant. The problem is multicollinearity: survey attributes are usually correlated with each other, partly because satisfied customers rate everything a little higher. When they are, coefficients become unstable, can flip sign, and can change sharply with a small change in sample.
Relative importance methods. Shapley value regression asks how much the model’s explained variance (R²) increases when an attribute is added, averaged over every possible order in which attributes could enter. Relative weights reach a similar answer faster by transforming the attributes into uncorrelated versions, then mapping the explained variance back to the originals. Both split shared influence fairly, never go negative, and add up to the model’s total R².
| Method | Pros | Cons |
|---|---|---|
| Correlation | Simple, transparent, stable | Double-counts overlapping attributes; one attribute at a time |
| Multiple regression | Controls for other attributes | Unstable with correlated attributes; negative or tiny weights that make no sense |
| Shapley value regression | Fair split of shared variance; sums to R² | Computation grows fast with many attributes |
| Relative weights | Similar results to Shapley; quick with many attributes | Harder to explain to stakeholders |
What does a worked key driver analysis look like?
The numbers below are illustrative, not from a real study. Suppose 400 customers of a home internet service rated five attributes and overall satisfaction on a 0 to 10 scale. For each attribute, you have the mean performance score and its correlation with overall satisfaction.
To turn correlations into percent importance, divide each one by the sum of all five. The sum is 0.62 + 0.50 + 0.36 + 0.32 + 0.20 = 2.00.
| Attribute | Mean score (0–10) | Correlation with overall satisfaction | Normalized importance |
|---|---|---|---|
| Connection reliability | 6.2 | 0.62 | 0.62 ÷ 2.00 = 31% |
| Customer support | 7.9 | 0.50 | 0.50 ÷ 2.00 = 25% |
| Value for money | 6.0 | 0.36 | 0.36 ÷ 2.00 = 18% |
| Speed | 8.3 | 0.32 | 0.32 ÷ 2.00 = 16% |
| Billing clarity | 6.8 | 0.20 | 0.20 ÷ 2.00 = 10% |
| Total | 2.00 | 100% |
How do you read the quadrant analysis?
Quadrant analysis splits each axis at its average. Average importance is 100% ÷ 5 = 20%. Average performance is (6.2 + 7.9 + 6.0 + 8.3 + 6.8) ÷ 5 = 35.2 ÷ 5 = 7.04.
| Quadrant | Rule | Attributes | What to do |
|---|---|---|---|
| Priorities to fix | Importance above 20%, score below 7.04 | Connection reliability (31%, 6.2) | Invest here first |
| Strengths to keep | Importance above 20%, score above 7.04 | Customer support (25%, 7.9) | Protect and communicate |
| Low priorities | Importance below 20%, score below 7.04 | Value for money (18%, 6.0); billing clarity (10%, 6.8) | Monitor; fix cheaply if possible |
| Possible overkill | Importance below 20%, score above 7.04 | Speed (16%, 8.3) | Check whether spend can move elsewhere |
Two cautions. Reliability and support are probably correlated, since outages drive support calls, so simple correlation may overstate both; a relative importance method would split their shared influence. And value for money sits close to the line, so treat it as borderline rather than safe to ignore.
How do you read drivers by segment and across waves?
A total-sample driver analysis averages over customers who may want different things. New customers might be driven by setup and onboarding, long-tenure customers by reliability and price. Run the analysis within segments you already use, such as tenure or usage groups from behavioral segmentation, or attitudinal groups from psychographic segmentation. If you have built segments with cluster analysis, drivers are a good way to describe what each cluster needs.
Before comparing segments, check that performance gaps are real. Crosstabs and banner tables with weighting and significance testing show whether one segment rates reliability lower than another, and whether a score moved between waves. In a tracker, rerun the drivers each wave or on pooled waves, and react to a driver that rises steadily over several waves rather than to one wave’s jump.
What are the common mistakes in key driver analysis?
- Using causal language. Derived importance shows association. “Reliability drives satisfaction” is shorthand; test the fix before you promise a satisfaction gain.
- Including too many correlated attributes. Five near-duplicate support items will split or double-count support’s influence. Group or trim attributes that measure the same thing.
- Reading small bases. The standard error of a correlation is roughly (1 − r²) ÷ √n. At r = 0.35 and 100 respondents, that is 0.88 ÷ 10 = 0.088, so the 95% range is about ±0.17. Drivers at 0.30 and 0.40 cannot be told apart.
- Ignoring the outcome’s reliability. A single overall rating carries measurement noise, which weakens every correlation. Where possible, use an index of overall satisfaction, recommendation and repurchase intent.
What should you do next?
Start with the outcome you care about and the attribute list you already field. Compute simple correlations first, then a relative importance method if attributes overlap, and plot the quadrant chart. If your goal is to see what drives satisfaction and loyalty by segment and journey, repeat the key driver analysis for each segment and wave, and take the top fix-first item to the team that owns it.