Key driver analysis (KDA)

Key driver analysis (KDA) finds which attributes of a product, service or workplace are most strongly linked to an overall outcome such as satisfaction, likelihood to recommend or employee engagement. It tells you where an improvement is most likely to move the main score, so you can rank what to fix first.

The inputs come from one survey: an outcome question (for example the 0–10 NPS question) and ratings of several attributes from the same respondents, such as support, price, ease of use or delivery speed. In HR the same method is used to find engagement drivers.

Stated and derived importance

You can ask people directly "How important is price to you?". That is stated importance, and it tends to rate almost everything as important. Key driver analysis uses derived importance instead: it measures how closely each attribute moves with the outcome across respondents. The two often disagree. Derived importance is based on how people actually rate you rather than on what they say matters, which is why KDA uses it to set priorities.

Methods, from simple to advanced

Correlation

Compute the correlation between each attribute and the outcome. Use Pearson's r for scores you treat as interval and Spearman's rho for ranks or short ordinal scales. It is quick and easy to explain, but when two attributes overlap (support speed and support quality), both get full credit. See correlation analysis.

Multiple regression

Regress the outcome on all attributes at once and compare the standardized coefficients (betas). Regression accounts for overlap, but when attributes are strongly correlated with each other (multicollinearity), the betas become unstable and can even turn negative for an attribute that clearly helps. See regression analysis.

Relative importance methods

Johnson's relative weights (Johnson, 2000) and Shapley value regression (Lipovetsky and Conklin, 2001) split the model's explained variance (R²) among the attributes. Each weight is positive and the weights add up to R², which makes them easier to read and more stable with correlated drivers. You need a statistics package or a short script to compute them.

Worked example

A software company surveys 400 customers (illustrative numbers, not real data). Each person answers the 0–10 recommendation question and rates five attributes from 1 to 5.

Plot importance (r) against performance (average rating) and split the chart at the averages of both, here r ≈ 0.39 and rating ≈ 3.5. The four quadrants read like this:

• High importance, low performance: fix first. Here, speed and reliability.
• High importance, high performance: strengths to protect. Support and ease of use.
• Low importance, low performance: secondary issues. Value for money and integrations.
• Low importance, high performance: possible over-investment.

This chart is also called an importance–performance matrix, and it pairs well with gap analysis.

How many responses you need

Green (1991) proposed a widely cited rule of thumb for regression: at least 50 + 8k respondents to test the model as a whole and 104 + k to test individual predictors, where k is the number of attributes. With 5 attributes that is 90 and 109. If you want drivers per segment, each segment needs that many on its own.

Common mistakes

• Reading drivers as causes. KDA shows association. Customers who like your support may simply like everything about you (the halo effect). Confirm big decisions with an experiment or follow-up research.
• Overlapping attributes. "Fast support" and "helpful support" measure almost the same thing and split the credit. Check overlap with factor analysis or group similar items.
• Ignoring performance. An attribute with high importance and a perfect score needs no action. Always read importance together with the average rating.
• Too few responses or too many attributes for the sample, which makes the ranking change from wave to wave.
• Treating r of 0.47 and 0.49 as different. Small gaps between drivers are often within noise. Group drivers into tiers rather than an exact order.

Running key driver analysis on survey data

1. Ask the outcome question and the attribute ratings in the same survey, using the same scale for all attributes. A matrix question keeps them compact; see matrix question.
2. Make the attribute list cover the whole experience and keep each item about one thing.
3. Export the responses. SurveyNinja exports to CSV, XLSX and SPSS on every plan, including Free.
4. For a quick read, paste the outcome and one attribute into the Pearson correlation calculator to get r, r² and the p-value, or use the Spearman correlation calculator for ranks. For regression or relative weights, use a statistics package.
5. Build the importance–performance chart and agree on owners for the "fix first" quadrant.