Key driver analysis: what actually moves your satisfaction score
Ask customers what matters and they say everything. Key driver analysis looks at what moves the score instead. Here is how it works, where correlation and regression mislead you, and how to read a driver chart without fooling yourself.
By SurveyLane · The team building SurveyLane
Your satisfaction survey has an overall score and eight attribute ratings: price, speed, support, onboarding, reliability and a few more. Then someone asks which of them to fix first. Key driver analysis is the standard answer, and it is easy to get confidently wrong. The chart looks precise. Whether it means anything depends on choices most people never notice they made.
Why you cannot just ask what matters
The obvious move is to ask. "How important is price to you?" "How important is support?" This is stated importance, and almost everything comes back as important.
Nobody tells you reliability doesn't matter. On a five-point importance scale most attributes land between 4 and 5, and the ranking inside that band is mostly noise. You've learned that customers would like everything to be good.
There's a second problem. People report what they think should matter. Ask how important your invoice layout is and you get a shrug. Then look at satisfaction in the accounts that got a confusing invoice. Respondents aren't lying. They just can't see what drives their own judgement, and the importance question asks them to look.
Derived importance: let the data show what moves the score
Derived importance turns the question around. You look at which attribute ratings rise and fall with the overall score across respondents. If people who rate support highly also rate the whole experience highly, and people who rate support poorly don't, support is a candidate driver.
The price is that you're now doing statistics on observational data. The simplest version is one correlation per attribute. The next step is a regression of the overall score on all attributes together. Both are fast and common, and both mislead in ways you need to know before the result reaches a product team.
Correlation inflates every driver because of halo
Correlate each attribute with overall satisfaction and you'll usually see all of them come out positive, often between 0.5 and 0.7. That looks like eight strong drivers. It's mostly one effect, counted eight times.
Edward Thorndike described it in 1920 in the Journal of Applied Psychology, in "A constant error in psychological ratings". Officers rated soldiers on separate qualities such as intelligence, physique and leadership, and the ratings correlated far more than the qualities plausibly could. A general impression leaked into every specific rating. He called it the halo effect.
Your respondents do it too. A happy customer rates your support, onboarding and pricing page more kindly, partly because they're happy. The ratings share a large common component, and that component is the overall judgement you were trying to explain. A correlation table tells you satisfied customers rate everything higher. It can't tell you what to change.
Regression breaks the other way
The textbook fix is multiple regression. Each coefficient shows one attribute's effect with the others held constant, which in theory strips out the shared halo.
In practice your attributes correlate so strongly with each other that the model can't separate them. That's multicollinearity. Add 50 responses or drop one attribute and the ranking reshuffles. One attribute grabs the shared variance and looks dominant, while its near-twin gets a coefficient around zero, sometimes negative.
That negative sign is the tell. "Better onboarding ratings lower overall satisfaction" is no finding. It means onboarding and another attribute measure nearly the same thing and the model split the credit at random. If support was your top driver in Q2 and barely showed up in Q3 while nothing changed in the team, this is the likely reason.
Shapley values and dominance analysis share the credit
The methods built for this problem ask a different question. Average over every order in which you could add attributes to the model: how much does each one add to the explained variance?
David Budescu set out dominance analysis in Psychological Bulletin in 1993. Stan Lipovetsky and Michael Conklin came at it from cooperative game theory in 2001, in Applied Stochastic Models in Business and Industry. They treat each attribute as a player and the model's R² as the payoff to be shared. That's Shapley value regression. For a linear model's R², Budescu's general dominance weights and the Shapley values are the same numbers.
You get importance shares that add up to the total explained variance, are never negative and stay much steadier when attributes overlap.
Two limits. The halo is still in every rating. And the result is still a description of your data, with no causal claim attached.
Computation sets a practical cap. Shapley values need a model for every subset of attributes: 2 to the power of the attribute count. Eight attributes means 256 models, which is nothing. Thirty means over a billion, so you need an approximation or a shorter list.
Good driver data starts in the questionnaire
Most of a driver analysis is decided when you write the survey.
Keep the attribute list short and distinct. "Ease of use" and "intuitiveness" are one attribute asked twice. So are "speed of support" and "response time". If you can't say what a customer would experience differently between two items, merge them.
Ask the overall question first. When attribute ratings come before it, the overall score turns into a summary of whatever the respondent just thought about. The post on question order effects explains the mechanism.
Use one scale, in one direction, for the attributes and the outcome. The post on designing rating scales covers which choices matter.
And cover what customers actually run into. The analysis ranks only what you asked about. If the real problem isn't on the list, your top driver will simply be whichever attribute correlates best with the missing one.
Grid questions damage driver data
Attribute ratings nearly always end up in a matrix: eight rows, one scale, one screen.
In a grid people straightline. They give the same answer down the column because it's faster than reading each row. Every straightlined response is pure halo, identical ratings that track the overall score and say nothing about which attribute matters. Get enough of them and you're analysing respondent fatigue.
Count how many respondents gave every attribute the same rating, and run the model with and without them. The post on spotting bad survey responses has more checks. On a phone, grids get worse, as the post on designing surveys for phones shows.
Some attributes only matter when they fail
A linear model assumes a point on reliability is worth the same from 2 to 3 as from 4 to 5. Customers don't work like that.
Noriaki Kano and colleagues described this in 1984 in the journal of the Japanese Society for Quality Control, in "Attractive quality and must-be quality". Must-be attributes earn no praise when they work and cause anger when they fail: your data saving correctly, say. Attractive attributes go unmissed when absent and delight when present. One-dimensional ones are simply better when better.
A must-be attribute fools a driver analysis. When almost everyone is content with it, its ratings barely vary, and a variable that barely varies can't explain much. So it ranks low. Then it breaks and nobody talks about anything else.
Plot overall satisfaction against low and high ratings for each attribute. If a low rating drags the score down hard and a high one adds little, you've found a must-be. Fix its failures and don't expect polish above the threshold to pay back.
Plot importance against performance
A ranked list is half the answer. Your top driver might already score 4.7 out of 5, where every further gain is expensive. A mid-ranked driver at 2.9 might be cheap to lift.
John Martilla and John James published importance-performance analysis in the Journal of Marketing in 1977, three pages built on car dealer service data. Importance on one axis, performance on the other, split at the means. Four quadrants: concentrate here, keep up the good work, low priority, possible overkill.
Use derived importance for the importance axis. With stated importance everything sits in the top band and the grid collapses. And remember the split lines are arbitrary. Move them from mean to median and an attribute can change quadrant. Use the chart to start a discussion and leave the budget decisions to the evidence.
What a driver chart cannot tell you
Survey data is a snapshot. A driver analysis tells you which ratings go with a higher overall score in this sample, right now. Whether improving an attribute will raise satisfaction is a separate question it can't answer.
Heavy users may rate both support and overall satisfaction highly because heavy use builds familiarity and goodwill. Support could then be a symptom of engagement. The model can't tell.
So treat drivers as hypotheses. Change something and watch the score: improve onboarding for one cohort and compare it with another, or follow the attribute and the outcome over several waves. The post on repeated surveys and trends covers keeping those waves comparable.
Report the uncertainty as well. With 150 responses the gap between driver three and driver five is usually noise. Bootstrap it: resample your respondents a few hundred times, rerun the analysis, and report how often each attribute lands in the top three. "Support makes the top three 95% of the time, price 40%" tells a product team far more than a bar chart.
When you ask an AI to find your drivers
An AI assistant connected to your survey data, for instance through the SurveyLane MCP server described in analysing surveys with AI, runs a driver analysis in seconds. It will also skip every judgement call above unless you ask.
Ask which method it used. "Key drivers" in a casual prompt often becomes a correlation table, the method with the halo problem. Ask for Shapley values or dominance analysis, with the share of explained variance per attribute.
Ask for the correlations between attributes before the ranking. Two attributes above 0.8 shouldn't be ranked against each other.
Ask it to flag straightliners and rerun without them, and to report the model's total R². At 0.25, three quarters of satisfaction comes from things you never asked about.
Then ask what the results don't show. If the summary says "price drives satisfaction" without a word on halo, overlap or causality, it's claiming more than the method can carry. It's the same failure as an AI inventing themes in open text, covered in the post on analysing open-ended responses.
Frequently asked questions
How many responses do I need for a key driver analysis?
There's no fixed number. A common rule of thumb is ten to twenty complete responses per attribute in the model, more when attributes overlap heavily, so 80 to 160 as a floor for eight attributes. Below that, a bootstrap will show you that most of the ranking isn't stable. Comparing drivers between segments needs that number in each segment.
Should I still ask stated importance questions?
Sometimes, next to derived importance. They help you spot must-be attributes that barely vary in your data and so rank low on derived importance. When customers call an attribute very important and the analysis puts it near the bottom, investigate the gap instead of picking a winner. Don't use stated importance as your main ranking, because nearly everything comes back as important.
Can I use NPS as the outcome variable?
Use the raw 0 to 10 likelihood-to-recommend score as the outcome. Collapsing eleven points into promoters, passives and detractors throws away variation the model needs.
Is Shapley value regression the same as SHAP in machine learning?
They share the game theory idea of dividing a payoff fairly among players, but they answer different questions. Shapley value regression divides a model's explained variance across attributes and gives one figure per attribute for the whole dataset. SHAP explains individual predictions and gives a contribution per attribute per respondent. For a survey driver chart you want Shapley value regression or dominance analysis.
