Key Driver Analysis Guide — Finding What Moves Satisfaction and NPS

The satisfaction survey is tabulated. Attribute-level scores sit next to overall satisfaction. And the next question is always the same: "We want to raise overall satisfaction. So which items do we lift to make it move?"

This is where most people compute the correlation between each attribute and overall satisfaction and report "the highest correlations are the drivers." That is half a trap. Correlation only shows that two things move together — it does not guarantee that lifting one lifts the other. And when attributes correlate with each other, you end up counting the same importance twice or three times over. Key Driver Analysis (KDA) is the technique for statistically untangling this "apparent importance" and pulling out the factors that truly matter. This guide works through it the way it feels in practice: from the limits of correlation, to multicollinearity in regression, to the Shapley value and relative weights that resolve it, and finally to the most important discipline of all — never confusing correlation with causation.

1. What key driver analysis is — quantifying "what works"

Key driver analysis decomposes how much a single outcome variable (overall satisfaction, NPS, retention intent, and so on) can be explained by multiple driver variables (satisfaction with each attribute) and produces a number for each driver's "pull" — its importance.

For example, in a SaaS satisfaction survey you take "overall satisfaction" as the dependent variable and "support satisfaction," "feature satisfaction," "price satisfaction," and "UI satisfaction" as independent variables, then identify which one moves overall satisfaction most strongly. What comes out is derived importance — importance you did not ask respondents about directly, but inferred statistically from the data.

This derived importance is the very concept that appeared as "how to measure importance" in the previous Importance-Performance Analysis (IPA) guide. KDA computes importance, and IPA plots it onto its four quadrants to set priorities — the two are designed to be used as a pair (the connection is detailed in Section 7).

The foundational tabulation for correlation and regression is covered in the survey aggregation and significance testing guide. This guide builds on that, stepping into the harder problem of separating out "which factor works among many."

2. Why "a list of correlation coefficients" is not enough

The quickest KDA is to compute the Pearson correlation between each driver and overall satisfaction and sort it highest to lowest. You can even do it with Excel's CORREL function. As a first-pass screen for getting your bearings it is useful, but making it your final conclusion gets you wrong for two reasons.

Reason 1: importance gets "double-counted"

Attributes usually correlate with one another. "Speed of support" and "courtesy of support" tend to rise together. Look at the two separately by correlation and both show high correlations. But in reality, a single underlying thing — "the support experience" — may be all that is working. A correlation ranking double-counts the importance of mutually correlated factors, painting a distorted picture in which "support-related items monopolize the top spots."

Reason 2: it ignores every other factor

Correlation looks at two variables only. Even if "price satisfaction correlates highly with overall satisfaction," you cannot tell whether price itself is doing the work or whether some other factor correlated with price (a sense of value for money, met expectations) is. To separate these, you need to consider multiple factors at once and extract each one's "pure" contribution. That is multiple regression, next.

3. Multiple regression and the multicollinearity trap

Multiple regression explains overall satisfaction with all drivers simultaneously and treats each driver's standardized beta coefficient (β) as its importance. It yields a pure contribution one step beyond correlation: "holding the other factors constant, when this factor moves by one standard deviation, overall satisfaction moves by β standard deviations."

But use regression for KDA and you will almost always fall into the multicollinearity trap.

What multicollinearity actually does

When drivers correlate strongly (e.g., speed and courtesy of support correlate at 0.8), the regression cannot decide "whose credit it is," and the coefficients become unstable. Specifically:

• Coefficient signs flip (the coefficient for "speed of support," which ought to be important, turns negative)
• Coefficients swing wildly with small changes to the sample
• Standard errors balloon and the coefficient stops being significant

Report to the executive meeting that "speed of support has a negative impact on overall satisfaction" and no one will believe you. And they are right not to — that negative coefficient is not reality, it is a statistical artifact produced by multicollinearity.

How to detect it

Detect multicollinearity with the VIF (Variance Inflation Factor). As a rule of thumb, a VIF above 5 warrants caution, and above 10 is clearly a problem. Even just looking at the correlation matrix among attributes, any pair correlated at 0.7 or higher is a red flag.

The catch is that customer-satisfaction attributes correlate structurally (satisfied customers rate everything high), so multicollinearity is not something that "occasionally happens" — it "almost always happens." That is why, in KDA, you must never take raw regression coefficients as importance.

4. Solving multicollinearity — Shapley value and relative weights

The methods for producing stable importance while sidestepping multicollinearity are Relative Weights and Shapley value analysis. These are the current practical standard for KDA.

How Shapley value and relative weights think

The two are close in spirit: they produce each driver's fair contribution by averaging "how much explanatory power (R²) increases when a given driver is added" across every possible order in which the variables are entered. The Shapley value comes from game theory — known via Kruskal (1987) and as the LMG method — and Johnson's (2000) relative weights is an approximation that keeps that computational burden in check.

The biggest practical advantage is that importance is always non-negative and sums to the model's R². You get numbers you can interpret intuitively as a share of contribution — "support accounts for 32% of total explanatory power, price for 21%…" — and you are freed from the hell of explaining "negative importance" at the executive meeting.

The computation is hard in standard Excel; do it with R's relaimpo package (Grömping 2006), Python, or a dedicated survey-analysis tool. Tonidandel & LeBreton (2011) lay out, in practitioner terms, how relative importance analysis is a useful supplement to regression — a good place to start.

5. Don't confuse correlation with causation — KDA's biggest misreading

This is the point that is most often, and most fatally, gotten wrong in key driver analysis. What KDA produces is association, not causation.

A result that "support satisfaction strongly explains overall satisfaction" does not guarantee that "improving support will raise overall satisfaction." The following pitfalls always lurk.

Reverse causation (the halo effect)

Customers who are satisfied overall tend to rate individual attributes somewhat high too (the halo effect). That produces a correlation in which "people with high overall satisfaction also rate support highly," but this may not be "support raised the overall" — it may be the reverse relationship, "overall satisfaction lifted the support rating." KDA alone cannot determine which way it runs.

Confounding

It is also possible that both support and overall satisfaction are being lifted simultaneously by an unobserved third factor (e.g., the customer's proficiency, fit).

How to handle it in practice

• Do not rephrase a "driver" as "a lever that will definitely work if improved." Stop at the wording "a factor strongly co-moving with overall satisfaction."
• For the top-importance factors, verify causation with A/B tests or before-and-after comparisons where possible. KDA prioritizes "where to test," it does not prove causation.
• Always add a sentence to the report: "this is correlation-based importance, not a guarantee of causal effect."

Humility protects KDA's credibility. The accurate way to say it is not "we found the factors that work" but "we ranked the hypotheses most likely to be working, strongest evidence first."

6. Choosing the outcome variable and the "asymmetry of satisfaction"

What to put as the dependent variable

In KDA, the conclusion changes with the choice of outcome variable. Overall satisfaction? NPS (recommendation intent)? Retention intent? Repurchase? Each has different drivers.

• Use overall satisfaction as the dependent variable and you get "the factors that make up the current experience."
• Use retention intent / NPS as the dependent variable and you get "the factors that govern future behavior" (see reading NPS and benchmarks).

If you are chasing the "high satisfaction yet they churn" phenomenon, the dependent variable should be retention intent, not satisfaction. When the outcome variable is misaligned with the objective, everything downstream swings and misses.

Don't overlook the asymmetry of satisfaction (Kano)

Ordinary regression assumes a linear relationship in which "as the driver rises, overall rises proportionally," but reality is asymmetric. As noted with Matzler et al. (2004) in the IPA guide:

• Must-be quality: taken for granted. Its absence drops overall sharply, but meeting it does not raise overall.
• Attractive quality: its absence causes no dissatisfaction, but its presence makes overall jump.

To capture this, use penalty-reward contrast analysis, which splits each driver into a "high-rating dummy" and a "low-rating dummy" and regresses on both. It lets you spot factors that are "high in importance, yet actually only work to relieve dissatisfaction (fixing them does not raise satisfaction)."

7. The connection to IPA — plotting KDA's output onto four quadrants

Key driver analysis and Importance-Performance Analysis (IPA) connect as an input-output relationship.

1. Compute derived importance with KDA: produce each driver's contribution share via Shapley value / relative weights (the material for the vertical axis)
2. Tabulate each driver's satisfaction (performance): mean values or Top 2 Box (the material for the horizontal axis)
3. Plot onto the IPA scatter: vertical axis = KDA's derived importance, horizontal axis = satisfaction
4. Read priorities by quadrant: "high importance (KDA) yet low satisfaction" = the priority-improvement zone

Run it this way and KDA's derived importance solves IPA's old weakness — that "measuring importance with a direct question collapses under ceiling effects." KDA builds the vertical axis, and IPA draws the decision map. KDA + IPA = the complete form of improvement prioritization.

For how to measure the satisfaction (performance) side, see the CSAT survey design guide; for the required sample size, see how to determine survey sample size.

8. The editorial view — five things never to do in key driver analysis

From a vantage point that continually tracks industry cases and the voices of practitioners, here are five accidents that recur in KDA.

1. Calling a list of correlation coefficients "driver analysis"

The most common. Reporting nothing more than a correlation ranking as "we did key driver analysis." Because correlation double-counts, clusters of mutually correlated factors (support-related items, etc.) unfairly monopolize the top. First-pass screening with correlation is fine, but always produce the conclusion with relative weights / Shapley.

2. Reporting regression's negative coefficient as-is

Putting the "negative coefficient on a factor that ought to be important" produced by multicollinearity into the report without sanity-checking it. Readers will see through it instantly: "this analysis is broken." Always check the VIF, and switch to relative weights when collinearity is present. Never make raw regression coefficients KDA's final output.

3. Asserting "driver = a lever that works if improved"

KDA is correlation, not causation. An initiative declared on "support is the biggest driver → invest in support and overall satisfaction rises" misses, and "we trusted the analysis" turns into lost trust in the analysis itself. Distinguish "strength of co-movement" from "improvement effect" in your wording, and verify causation for the top factors with A/B tests.

4. Defaulting the outcome variable to "overall satisfaction" out of habit

When the objective is "reduce churn" but you make the dependent variable overall satisfaction, drivers unrelated to churn rise to the top. Choose the outcome variable to match the objective (satisfaction / retention / recommendation / repurchase). Decide this on autopilot and all the refined analysis that follows is wasted.

5. Failing to balance sample size against the number of variables

Run a regression with 30 explanatory variables on N=80 and overfitting makes the importances swing at random. As a rule of thumb you want 10 to 15 samples per explanatory variable. If you have too many variables, compress the dimensions with factor analysis, or group them with domain knowledge before entering them.

9. Running key driver analysis with the survey tool Kicue

KDA splits into "the question design that measures drivers and the outcome" and "the analysis that computes importance via relative weights and the like." Kicue handles the former; the latter pairs with external statistical tools.

• Designing driver and outcome questions: supports a design that measures each attribute's satisfaction (drivers) and overall satisfaction / NPS / retention intent (the outcome) on Likert scales within a single form (question types · Likert scale design guide)
• CSV export with respondent IDs: outputs in a regression-ready structure — one row per response, attribute satisfaction lined up next to overall satisfaction
• GT and cross-tabulation: checking each driver's mean and distribution, and reviewing the raw data before the first-pass correlation screen, are possible on the dashboard

⚠️ What Kicue cannot do

• No function to compute correlation, multiple regression, relative weights, or the Shapley value: run statistical analysis in R (relaimpo, etc.) / Python / SPSS / JASP. Kicue itself does not carry statistical-analysis features.
• No VIF or multicollinearity diagnostics either: run these in your statistical software after export
• No random forest or other machine learning either: run these in Python (scikit-learn, etc.)
• No IPA scatter-plot creation either: build the scatter that places KDA's derived importance on the vertical axis in Excel / R / Python

As related reading, pairing this with the Importance-Performance Analysis (IPA) guide, CSAT survey design guide, reading NPS and benchmarks, survey aggregation and significance testing guide, and VoC program design guide brings the whole analysis pipeline into view — "measure → identify what works (KDA) → prioritize (IPA) → operate."

Summary — six points to make key driver analysis an analysis you can trust

1. Correlation rankings are for first-pass screening only — they double-count, so don't make them the conclusion
2. Suspect multicollinearity in regression — check the VIF and don't take negative coefficients at face value
3. Produce importance with relative weights / Shapley — non-negative, summing to R², interpretable as a contribution share
4. Distinguish correlation from causation — it is "strength of co-movement," not "a guarantee of improvement effect." Verify the top factors with A/B tests
5. Match the outcome variable to the objective — retention intent for churn work, overall satisfaction for experience improvement
6. 10 to 15 samples per explanatory variable — if you have too many variables, compress with factor analysis

In key driver analysis, reliability is decided less by "which statistical method you use" than by not missing the two points of multicollinearity and causation. Hold those two and KDA becomes a powerful weapon for lifting the "where do we start" debate out of gut feel and into a discussion grounded in numbers.

If you want to design a survey that measures drivers and outcomes, why not try the free survey tool Kicue? From Likert question design that measures attribute satisfaction alongside overall satisfaction and NPS on a single form, to a respondent-ID CSV export you can feed straight into regression, you can start the part that builds the input data for key driver analysis with one account (correlation, multiple regression, relative weights, the Shapley value, and VIF diagnostics are run in combination with R / Python / SPSS / JASP).