# Key Driver Analysis: How to Find What Really Drives Satisfaction in Survey Data

How key driver analysis finds what drives overall satisfaction: derived vs stated importance, correlation vs regression, a worked example and its limits.

Published 2026-10-30 · Updated 2026-10-30
Canonical: https://hercules.works/blogs/key-driver-analysis-survey

Key driver analysis estimates how much each rated attribute (delivery, price, support) moves an overall rating such as satisfaction. You collect one overall question plus attribute ratings, then relate them statistically. This is derived importance, unlike stated importance, where people say what matters. Correlated attributes mislead plain correlation and regression, so use methods that share the overlap fairly.

## Which Attributes Actually Move the Overall Score?

A satisfaction survey usually ends with a table of attribute scores and a single overall score. The table tells you what customers rate low. It does not tell you which of those low scores is worth fixing. Key driver analysis closes that gap.

This guide explains the method, the difference between derived and stated importance, why correlated attributes break the simple versions, and a worked example you can check by hand. The example numbers are illustrative, not survey results.

## What Is Key Driver Analysis?

Sawtooth Software describes "driver analysis" as also known as derived importance analysis, a staple deliverable in satisfaction and loyalty research ([Sawtooth Software](https://sawtoothsoftware.com/resources/blog/posts/driver-analysis-how-to-do-it-badly-and-well)). It estimates how much each attribute rating contributes to an overall outcome rating.

The data you need is simple:

1. One overall outcome question, for example overall satisfaction or likelihood to recommend, asked on the same scale for everyone.
2. Ratings of the same respondents on a list of attributes, such as delivery speed, price and support helpfulness.
3. Enough completed answers that each attribute has real variation. Attributes where nearly everyone gives the same score cannot show an effect.

The output is a ranked list that says which attributes account for the most variation in the overall rating. Pair it with each attribute's average score and you have the inputs for an [importance-performance grid](/blogs/importance-performance-gap-analysis/).

## Derived or Stated Importance: What Is the Difference?

Grigoroudis and Spyridaki define stated importance as the straightforward customer preferences for the weight of a satisfaction criterion, and derived importance as estimated statistically from customers' judgments of performance ([Grigoroudis and Spyridaki, *Operational Research*, 2003](https://iaorifors.com/paper/51642)). Their paper plots both on one "dual importance diagram" rather than choosing between them.

| | Stated importance | Derived importance |
|---|---|---|
| How collected | Ask: "How important is X?" | Compute from attribute ratings and the overall rating |
| Extra questions needed | Yes, one per attribute | No |
| Typical weakness | Respondents tend to call nearly everything important ([Koji](https://www.koji.so/docs/importance-performance-analysis-guide)) | Needs enough variation and care with overlapping attributes |
| Use it for | What people say they value | What actually moves the overall score |

A practical rule is to collect both and look at the gaps. An attribute that is rated important but has little derived effect is often a basic expectation that nobody notices until it fails. The [Kano model](/blogs/kano-model-survey-tool/) is built to separate those cases directly.

## Why Do Correlation and Regression Mislead?

Attributes in a survey overlap. People who rate delivery well also tend to rate packaging well. Sawtooth explains the consequence: correlation "over-counts the variation that the predictors share with the dependent variable", so squared correlations can add up to more than 100% of the outcome's variance (its example reaches 188%). Regression has the opposite problem. It "ignores all the variance that multiple predictors share with the dependent variable", and its coefficients can become unstable or even negative.

Sawtooth names two fixes that share the overlap fairly. The first averages each attribute's incremental R-squared over every possible order of entering the model. That average is equivalent to the Shapley value, a point it credits to Cox (1985). The second is Johnson's relative weights, which it says give very similar results, run quickly with many attributes and add up to 100%.

Treat any single-number ranking with caution. A rank-order of drivers is usually safer to report than a precise percentage.

## Worked Example: Three Attributes, One Overall Score

These correlations are illustrative. Suppose three attributes are measured with an overall satisfaction score.

| Pair | Correlation |
|---|---|
| Delivery with overall | 0.60 |
| Support with overall | 0.55 |
| Packaging with overall | 0.30 |
| Delivery with support | 0.70 |
| Delivery with packaging, support with packaging | 0.20 each |

Squared correlations are 0.36, 0.30 and 0.09, a total of 0.753. Yet a regression on all three explains only R-squared = 0.422. Correlation alone credits the same variation more than once.

| Attribute | Standardised regression coefficient | Share of R-squared (average over orderings) |
|---|---|---|
| Delivery | 0.401 | 0.213 (50.5%) |
| Support | 0.235 | 0.157 (37.3%) |
| Packaging | 0.173 | 0.051 (12.2%) |

The right-hand column adds up to 0.422 by construction. Delivery alone explains 0.360 and support alone 0.303, but together they explain only 0.393, because 0.270 of the variation is shared between them. The averaging method splits that shared part between the two instead of giving it all to one.

Read the result as a ranking: delivery first, support second, packaging a distant third. If packaging has the best average score, it is a candidate for the "possible overkill" box on a grid.

## How Do You Run a Driver Analysis Step by Step?

1. **List the attributes first.** Use earlier interviews or open-ended answers, and keep each attribute to one idea. A double-barrelled "price and value" item cannot be interpreted.
2. **Use the same rating scale for every attribute and the outcome.** Mixed scales add noise. See our [Likert scale guide](/blogs/likert-scale-survey-creator/).
3. **Check how attributes correlate before modelling.** Very high correlations mean the ranking will be unstable.
4. **Choose a shared-variance method** such as averaging over orderings or relative weights, not raw coefficients.
5. **Report the base and the interval.** Resample the data many times (bootstrapping) to see how often the ranking changes. If rank 2 and rank 3 swap often, say they are tied.
6. **Put derived and stated importance side by side**, then add average performance.

## What Key Driver Analysis Cannot Tell You

It shows association, not cause. An attribute may rank high because it is the easiest thing for customers to judge, or because it stands in for a feeling about the whole experience.

It only ranks the attributes you measured. A driver you never asked about cannot appear. Use open-ended answers to check for missing themes.

It describes the sample you surveyed. Weak sampling or heavy weighting changes who is being described, so read our [survey weighting guide](/blogs/survey-weighting-effective-sample-size/) before trusting subgroup rankings.

Last, the method says what moves the score now, not what a new feature would do. For choosing between options, a trade-off method such as [MaxDiff](/blogs/maxdiff-survey-tool/) is a better fit.

## Running This Study on Hercules

Hercules Works supports rating scales and grids in a survey, and Poseidon, our analytics engine, selects formulas by question type; its list includes driver analysis. Ask any provider which method produces the driver ranking, whether it reports the overlap problem, and whether it shows intervals.

The panel is app-based, so it reaches smartphone users only. You can test the questionnaire on the Free plan with 30 responses before you commit to a larger sample.

## Frequently asked questions

### What is key driver analysis in a survey?

It is a statistical method that estimates how much each attribute rating contributes to an overall rating such as satisfaction. It is also called derived importance analysis.

### What is the difference between derived and stated importance?

Stated importance comes from asking people how important each attribute is. Derived importance is estimated statistically from their attribute ratings and an overall rating. Collecting both lets you spot attributes that people say matter but that do not move the overall score.

### Why not just use correlation with the overall score?

Attributes overlap, and correlation counts the shared variation for each attribute separately. In the example above the squared correlations total 0.753 while the full model explains 0.422.

### Which method handles overlapping attributes?

Shapley-style averaging of incremental R-squared across all orderings, and Johnson's relative weights, both described by Sawtooth Software. They tend to give similar rankings.

### How many attributes can I include?

We have not found a standard published limit. The practical limit is that each attribute needs real variation and a clear meaning, and highly correlated attributes make rankings unstable. Fewer, well-defined attributes are easier to interpret.

### Does a high derived importance prove the attribute causes satisfaction?

No. It shows association in the data. Treat it as a guide to where to look and test changes before assuming impact.