Key Driver Analysis Methods Compared: Correlation, Regression, Shapley and Causal
Four driver analysis methods run on one simulated survey where the true effects are known. Correlation and Shapley ranked a nearly harmless driver second; with the confounders held fixed, every true effect fell inside it…
To find what drives CSAT or NPS, correlation and Shapley values tell you which drivers move with the score and how much of its variation each one accounts for; regression gives each driver's effect with the others held fixed; and only a causal estimate that also holds the confounders fixed tells you how much the score would change if you fixed the problem. In a labelled sample of 3,000 survey responses with known true effects, correlation and Shapley both ranked a nearly harmless driver second, while regression with the confounders held fixed recovered every true effect inside its interval. For an insights team that wants that answer from survey comments, tickets and reviews read together, with plan, tenure and revenue as controls, Dimension Labs is built for the job: it labels every comment as a driver field and runs the controlled estimate with confidence intervals on its Prove plan. The method below works with any tool.
This page compares the methods on one sample where the right answer is known. For running driver analysis on comment text rather than rating batteries, see key driver analysis from text; for the theory behind causal estimates, causal inference: how to move from correlation to cause.
Which key driver analysis method should I use for NPS, regression or Shapley?
Use Shapley values (or relative weights) when your drivers overlap and you need a stable split of importance across them, and use regression when you need an effect size per driver. Neither one protects you from a confounder you did not measure or did not include. For deciding what to fix, add the factors that cause both the driver and the score as controls, and read the effect, not the share.
Methods compared
Each method answers a different question about the same score
Four common ways to find what drives CSAT, NPS or a rating, the question each one answers, and when to use it.
| Method | Question it answers | Use it when | Watch for |
|---|---|---|---|
| Correlation | Does this driver travel with the score? | Making a first list of candidates | Drivers that share a cause with the score look important |
| Multiple regression | What is its effect with the other drivers held fixed? | Standard survey driver analysis with few, weakly related drivers | Weights swing when drivers overlap; confounders you leave out bias every weight |
| Shapley value or relative weights | What share of the explained variation in the score is it worth? | Drivers overlap and you need a stable split of importance | Rewards common drivers; it is a share of R², not the gain from a fix |
| Causal estimate with controls | How much would the score change if this stopped happening? | Deciding what to fix and what the fix is worth | Needs the right controls written down, and refutation tests |
MeasuringU describes key driver analysis as a way to tell "the relative importance of predictor (independent) variables on your outcome". Importance can mean three different things, and the methods above each give you one of them: how closely a driver moves with the score, what share of the explained variation it gets, and how much the score changes when the driver changes.
How did we test the methods?
We simulated 3,000 survey responses with a 0 to 10 likelihood-to-recommend score and five problems a respondent may have had, each tagged yes or no, the way Dimensions tag survey comments or tickets. Because we wrote the simulation, the true effect of each problem is known. Two background facts make it realistic: complex issues lead to long waits and to unresolved tickets, and lower the score on their own; enterprise customers get more billing errors and also score higher. The sample's average score is 7.2.
We then ran four methods on the same data: correlation of each driver with the score, a regression on the five drivers, Shapley value regression on those drivers (the share of R² each one earns, averaged over every order of entry), and a regression that also holds the two background facts fixed, with bootstrap intervals. Johnson's relative weights gave almost the same split as Shapley, with no driver differing by more than 0.0001 of R², so we show one.
What did each method get right?
Sample data
Correlation and Shapley rank the long wait second; the effect with controls puts it last, where it belongs
Five drivers of a 0 to 10 likelihood-to-recommend score in 3,000 simulated survey responses with known true effects. Each cell: the method's number, then its rank (1 = strongest).
| Driver | True effect | Correlation r | Shapley share | Regression, drivers only | Regression with confounders |
|---|---|---|---|---|---|
| Issue not resolved 14% of responses | 1.92 lower 1 | −0.44 1 | 57% 1 | 2.32 lower 1 | 1.91 lower 2 |
| Rude or dismissive agent 5% of responses | 1.73 lower 2 | −0.20 3 | 13% 3 | 1.89 lower 2 | 1.94 lower 1 |
| Billing error 11% of responses | 1.42 lower 3 | −0.19 4 | 11% 4 | 1.12 lower 3 | 1.54 lower 3 |
| App bug 11% of responses | 0.75 lower 4 | −0.10 5 | 4% 5 | 0.67 lower 5 | 0.73 lower 4 |
| Long wait 28% of responses | 0.28 lower 5 | −0.26 2 | 16% 2 | 0.82 lower 4 | 0.35 lower 5 |
Read the long wait row. It is the most common problem, in 28% of responses, and its true effect is small: the score is 0.28 points lower when it happens. Correlation ranks it second (r = -0.26) and so does Shapley (16% of the explained variation). Both are picking up the complex issues that cause long waits and also lower the score. Regression on the drivers alone gets the order closer but puts the effect at 0.82 points lower, about three times the truth. With complexity held fixed, the estimate is 0.35 lower, close to the true 0.28 lower.
Sample data
Ranked by correlation, the long wait looks like the second-biggest driver
The same five drivers, ranked by correlation with the score (left) and by their effect with the confounders held fixed (right).
Billing errors go the other way. Their true effect is 1.42 points lower, but they happen mostly to enterprise customers, who score higher for other reasons, so correlation ranks them fourth and the drivers-only regression understates them at 1.12 lower. Held fixed for plan, the estimate is 1.54 lower.
Sample data
With the confounders held fixed, all 5 intervals contain the true effect; without them, 2 of 5 do
Regression estimates with bootstrap 95% intervals (500 resamples), on the 0 to 10 scale. Rust: an interval that misses the true effect.
One caveat the sample also shows: with the controls in, rude agents (1.94 lower) edge ahead of unresolved issues (1.91 lower), the reverse of the true order, but their intervals overlap, so the honest reading is "about equal, both large". That is the point of reporting intervals: they tell you which differences to act on and which to treat as a tie, which is exactly what a team choosing between two fixes needs to know.
What is the difference between correlation and relative importance in driver analysis?
Correlation looks at one driver at a time. Relative importance methods, Shapley value regression and Johnson's relative weights, look at all the drivers together and split the model's R² among them, so two drivers that overlap share credit instead of both claiming it. Grömping recommends averaging over orderings (the Shapley approach) as one of two preferred metrics.
In our sample both gave the same ranking, because the confounding came from outside the drivers. Relative importance fixes overlap between drivers; it does not fix a factor missing from the model.
Causal driver analysis vs Shapley values for customer satisfaction
A Shapley share answers "how much of the variation in satisfaction is this driver worth"; a causal estimate answers "how much would satisfaction change if this stopped happening". The first rewards drivers that are common; the second does not care how common a driver is. Adding the two background facts to the Shapley model helps: the long wait's share falls from 16% to 7.5%, while complexity itself takes 26%. It is still a share of variation, though, so it cannot tell you what a fix is worth.
Use both, for different questions. The share tells you how much of the total problem a driver is. The causal effect tells you how much better things get when you fix it, which is the number that goes into a business case.
What does a causal driver estimate need?
The right controls. Factors that cause both the driver and the score: issue complexity, plan, channel, location, tenure. Leave out ones the driver causes; adjusting for the wrong variable biases the answer in the other direction.
Enough mentions per driver. A rare driver gets a wide interval; we tested how many is enough in how many responses you need for driver analysis.
Refutation tests. Re-run with a placebo driver, a random extra control and subsets of the data; an effect that survives all three is one to act on.
Drivers that exist as fields. In surveys these are often in the comments, not the rating battery, so the text has to be labelled first.
For worked examples of the gap between mention counts and causal effects on real public reviews, see correlation vs causation in customer experience data.
Why would an insights team run its driver analysis in Dimension Labs?
Dimension Labs is an omni-channel analytics agent for customer data. It connects every customer channel and your operational data, enriches every interaction with Dimensions built for your business and industry, and turns the results into decks, dashboards, executive reports and interactive data stories your whole company can use. Plans start at $299 a month with no demo.
For driver analysis, the four parts line up with the four requirements above. Being fully omni-channel means survey comments, tickets and reviews feed the same model, so a driver counts wherever customers mention it. Easy integration with operational data brings in the controls, such as plan, location and tenure, and the outcomes, such as renewals. Enrichment of every interaction turns comments into the yes-or-no driver fields this page used, with definitions your team writes. And the visual analytics layer turns the result into a ranked driver deck or dashboard that leadership can read. Custom Dimensions start on Discover ($999 a month); causal findings with confidence intervals, and the join to CRM and revenue data, are on Prove ($2,999 a month) and up.
The Seattle Mariners put the result in one line: "Ultimately, it allows us to quantify the qualitative and give context and reasoning to KPI movements."
FAQ
What is the best key driver analysis method for NPS?
For deciding what to fix, a causal estimate with the confounders held fixed, reported with intervals. Dimension Labs runs that on survey comments, tickets and reviews together on its Prove plan. For a quick read of which attributes matter most in a rating battery, Shapley value regression or relative weights are a solid default.
Are Shapley values and Johnson's relative weights the same thing?
No, but they usually agree closely. Shapley averages each driver's added R² over every order of entry; relative weights approximate the same split with a faster calculation. In our sample no driver's share differed by more than 0.0001 of R².
Can I run key driver analysis on open-ended comments instead of rating questions?
Yes. Label each comment with yes-or-no fields for the problems it mentions, then use those fields as drivers, as this sample did. We show it on 9,186 public gym reviews in key driver analysis from text.
Why does regression give a different ranking from correlation?
Regression holds the other drivers fixed, so a driver that only looked important because it travels with another one loses weight. It does not hold fixed anything you left out, which is why the controls matter.
To see which drivers move your own CSAT or NPS, start with one file of survey responses. Every plan starts with a 14-day free trial, with no demo or sales call.