Driver analysis is a set of statistical techniques that shows which factors have the biggest influence on an outcome you care about, such as customer satisfaction, Net Promoter Score (NPS), or retention. It turns a long list of survey scores into a short, ranked list of priorities.
Most teams already collect plenty of feedback. The hard part is deciding what to fix first. The lowest score is rarely the right answer, because a weak area that customers barely care about won’t move the overall number.
In this blog, we’ll break down how driver analysis works, compare the main methods, walk through a worked example, and cover the data checks that keep your results trustworthy.
What is driver analysis?
Driver analysis measures how strongly each part of an experience relates to one overall outcome. It answers “what matters most?” with evidence from your data instead of opinion.
Every driver analysis has two building blocks:
- Outcome variable: The single result you want to improve, such as overall satisfaction, likelihood to recommend, renewal intent, or spend.
- Drivers: The individual attributes that might explain the outcome, such as price, product quality, wait time, or ease of use. Statisticians call these independent variables.
The analysis then answers three practical questions. What do customers care about most? Which factors actually make a difference to their rating? And what would likely happen to the outcome if you improved one of them?
The technique works anywhere you have an outcome and its possible causes in the same dataset. CX teams use it to explain NPS, CSAT (customer satisfaction score), and CES (customer effort score). HR teams use it on engagement surveys. Product and brand teams use it to understand purchase intent or brand preference.
Driver analysis is the umbrella term, while several related techniques get confused with it. The table below shows where each one fits.
| Method | What it answers | How it differs from driver analysis |
|---|---|---|
| Driver analysis | Which factors relate most strongly to an outcome? | The umbrella term covering correlation, regression, relative weights, and model-based methods |
| Key driver analysis (KDA) | Which survey questions best predict one outcome? | A specific, usually regression-based form of driver analysis; also called relative importance analysis |
| Correlation analysis | How closely do two variables move together? | Looks at one driver at a time and ignores overlap between drivers |
| Root cause analysis | Why did a specific problem happen? | Traces one failure back to its source, often with qualitative tools such as the 5 Whys |
| Conjoint analysis | Which feature trade-offs do buyers prefer? | Uses designed choice tasks to measure value, not ratings of an existing experience |
If you want a step-by-step regression walkthrough, see our guide on how key driver analysis works. For pricing and feature trade-offs, conjoint analysis is the better tool.
Why driver analysis matters for CX and research teams
Driver analysis matters because budgets are finite and every survey produces more problems than a team can fix. It tells you where an extra dollar or hour will move the outcome most.
Here is what it changes in practice:
- Better decisions: You rank improvements by their likely effect on the outcome, not by gut feel or the loudest complaint.
- Efficient resource use: You stop funding fixes for low-impact areas, even when they score badly.
- Stronger customer experience: You focus on the moments that shape loyalty and lower the risk of churn.
- Proactive problem solving: A driver that slips quarter over quarter warns you before the overall score drops.
- Sharper messaging: Marketing can lead with the attributes customers weigh most when they choose.
- Defensible planning: Leaders get a data-backed reason for each priority, which makes budget conversations shorter.
What are the main driver analysis methods?
The main driver analysis methods are correlation, multiple regression, relative importance analysis, path modeling, and machine learning. They differ in how they handle drivers that overlap with each other, which is the norm in survey data.
Correlation analysis
Correlation analysis measures how closely each driver moves with the outcome, one pair at a time. Two versions are common:
- Pearson correlation: Suits rating scales treated as numbers with a roughly even spread.
- Spearman rank-order correlation: Works on ranks, so it handles skewed or ordinal data better.
Correlation is easy to explain and tolerates missing data well. Its weakness is overlap. If “staff friendliness” and “staff knowledge” both correlate with satisfaction, correlation can’t tell you which one is doing the work.
Multiple regression
Regression analysis models the outcome with all drivers at once. Each coefficient shows a driver’s unique effect while the model holds the others constant.
- Strength: It isolates unique effects and reports how much of the outcome the drivers explain together (R-squared).
- Strength: Most statistics tools and spreadsheets can run it.
- Limit: Correlated drivers make coefficients unstable, so rankings can flip between survey waves.
- Limit: A clearly important driver can even show a negative coefficient when it overlaps heavily with another one.
Shapley value regression and relative weights
Relative importance analysis splits the model’s R-squared fairly across drivers, even when they overlap. The result is a share of explained variance for each driver that adds up to 100%, which makes it easy to present.
Shapley value regression gets there by averaging each driver’s contribution across every possible combination of drivers. That is thorough but slow, since 20 drivers create over one million combinations. Johnson’s relative weights reach a very similar answer in a fraction of the time, so many analysts use it for large driver lists.
Partial least squares path modeling
Partial least squares (PLS) path modeling groups related questions into themes and estimates how those themes connect to the outcome. It suits long questionnaires with natural question bundles.
- Best fit: Surveys where questions roll up into areas like “support,” “product,” and “billing.”
- What you get: Theme-level importance plus the strength of each question within its theme.
- Trade-off: It needs more statistical skill to set up and explain to stakeholders.
Machine learning models
Decision trees and random forests find drivers by testing which variables best split respondents into high and low outcome groups. They catch nonlinear patterns, such as a driver that only matters once it falls below a threshold. The output is harder to explain, and these models usually need larger samples than regression.
How to choose the right driver analysis method
Choose the method that matches your sample size, the number of drivers, and how much your drivers overlap. When in doubt, start simple and confirm with a relative importance method.
| Your situation | Recommended method | Why |
|---|---|---|
| First analysis, small sample, quick read needed | Correlation | Easy to run and explain; a solid directional check |
| 5 to 15 fairly distinct drivers | Multiple regression | Shows unique effects with familiar output |
| Drivers that clearly overlap | Johnson’s relative weights or Shapley | Stable rankings that sum to 100% |
| 30+ questions in natural themes | PLS path modeling | Rolls questions into themes before ranking |
| Large dataset, suspected thresholds | Random forest or decision tree | Finds nonlinear effects regression misses |
| Choosing features or price for a new offer | Conjoint analysis | Measures trade-offs before the experience exists |
How to design a driver analysis survey
A driver analysis survey needs one clear outcome question, a focused battery of driver questions, and enough responses to support the method. Design decisions made here limit what the analysis can tell you later.
- Write one outcome question.
Use a single overall measure, such as “Overall, how satisfied are you with us?” or the 0 to 10 likelihood-to-recommend question. - List candidate drivers from real touchpoints.
Pull them from journey maps, support tickets, and past open-ended comments. Aim for roughly 8 to 20 drivers. - Make each driver specific and distinct.
“Speed of checkout” is usable. “Overall website experience” overlaps with the outcome and muddies the result. - Use the same rating scale throughout.
Rate all drivers on one scale, such as 1 to 5, so scores stay comparable. - Add segment questions.
Include customer type, region, tenure, or plan tier so you can compare drivers across groups. - Include one open-ended “why” question.
It captures drivers you didn’t think to list. - Plan the sample size.
A working rule many analysts use is at least 10 complete responses per driver. Plan more if you will split results by segment.
Driver analysis example: A retail bank’s satisfaction survey
Here is a worked example with illustrative numbers. A regional bank surveys 1,200 customers about six drivers and one outcome, overall satisfaction. Together, the six drivers explain 58% of the variation in satisfaction (R-squared of 0.58), and relative weights split that total across them.
| Driver | Share of importance | Average score (1 to 5) | Priority |
|---|---|---|---|
| Issue resolved on first contact | 31% | 3.2 | Fix first |
| Wait time to reach support | 22% | 3.0 | Fix first |
| Mobile app ease of use | 18% | 4.3 | Protect |
| Clarity of fees | 14% | 2.6 | Monitor |
| Branch staff friendliness | 10% | 4.5 | Maintain efficiently |
| Statement design | 5% | 2.9 | Low priority |
Fees score lowest, so a score-only review would put them first. The driver analysis says otherwise. First-contact resolution and wait time carry more than half of the explained importance and both score below average. That is where the bank should invest first.
The app is a strength that customers weigh heavily, so protecting it matters. Branch friendliness scores well but carries little weight. The bank can hold that level without adding more spend.
How to read driver analysis results
Read results by crossing importance with current performance. The most useful output is a priority matrix, often called an importance-performance matrix, with four quadrants.
- Fix first: High importance, low performance. These are your biggest opportunities.
- Protect: High importance, high performance. Keep these strong, because any slip will hurt the outcome.
- Monitor: Low importance, low performance. Watch them, but don’t let low scores pull budget away from the first quadrant.
- Maintain efficiently: Low importance, high performance. You may be overinvesting here.
Also check the direction of each relationship. A positive link means higher driver scores go with a higher outcome, so improving that area should help. A negative link is rarer and often signals overlap between drivers or a reverse-worded question, so check the data before acting on it.
Finally, look at R-squared. A low value, say under 0.30, means your drivers explain little of the outcome. The survey has probably missed something important, and the open-ended answers are the place to look.
Data checks before you trust the results
Driver analysis is only as reliable as the data behind it. Run these checks before you share rankings with leadership.
- Test for multicollinearity.
Multicollinearity means two or more drivers overlap so much that the model can’t separate them. A common rule of thumb says variance inflation factors above 4 warrant investigation and above 10 signal serious multicollinearity. Merge near-duplicate questions or switch to relative weights. - Check response variance.
A driver where nearly everyone answers 5 can’t explain differences in the outcome. Flag items with very little spread. - Handle missing and “not applicable” answers.
Decide whether to drop, impute, or analyze subgroups separately. Silent deletion can shrink your sample fast. - Screen out low-quality responses.
Remove straight-liners and speeders before you model. They add noise that weakens every driver. - Confirm the sample supports splits.
If a segment has too few responses for your driver count, report it as directional only. - Remember that correlation is not causation.
Driver analysis shows association. Test major investments with a pilot or an A/B test before a full rollout.
Driver analysis best practices
The strongest driver analysis programs combine methods and repeat the work over time. A single run is a snapshot, while a tracked program shows how priorities shift.
- Start with correlation: It gives a quick, assumption-light view. Confirm the ranking with relative weights before you act.
- Pair derived and stated importance: Derived importance comes from the data, while stated importance asks customers how much each factor matters. The gap between the two often reveals hidden drivers that customers don’t name but clearly respond to.
- Reduce long questionnaires first: Factor analysis, a dimension reduction technique, groups related questions into themes and cuts overlap before you rank anything.
- Move to path modeling for themes: Use PLS path modeling when you want to see how themes feed into each other and into the outcome.
- Segment the results: Your highest-value customers may care about different things than the average respondent.
- Rerun on a schedule: Repeat the analysis each quarter or after major changes, such as a pricing update or a new app release.
How to add open-ended feedback to driver analysis
Open-ended comments explain the “why” behind a driver and surface drivers your rating questions missed. Combining text with ratings gives you both the ranking and the story behind it.
The process is simple. Code comments into themes, such as “billing errors” or “long hold times.” Then treat each theme as a yes or no variable and include it in the analysis alongside rated drivers. A theme that appears often among detractors and rarely among promoters is a strong candidate for a missing driver.
Manual coding is slow at scale, so AI tools now do most of the work. Text analytics in QuestionPro BI detects themes and sentiment in free-text responses and connects them to KPIs such as satisfaction and loyalty.
How QuestionPro supports driver analysis
QuestionPro builds driver analysis into its dashboards, so teams can rank drivers without exporting data to a statistics package. The setup differs slightly between customer and employee programs.
In QuestionPro Customer Experience, the NPS Key Driver Analysis widget shows which attitudes and interactions have the strongest effect on your overall NPS. It uses Spearman rank-order correlation between each driver question and the NPS question. A coefficient around 0.5 typically signals a strong driver, while most drivers land between 0.3 and 0.4.
In QuestionPro Employee Experience, a Driver Analysis widget does the same job for engagement surveys:
- Choose the outcome: Select the outcome question, such as eNPS, the employee version of NPS.
- Pick the drivers: Use the matrix questions in the same survey as candidate drivers.
- Read the dashboard: Ranked drivers update as new responses arrive.
For an HR walkthrough, see driver analysis in employee experience.
Driver analysis turns scores into priorities
A survey report tells you how you’re doing. Driver analysis tells you what to do about it. The difference matters most when the lowest score and the biggest opportunity point in different directions.
The teams that get the most from it keep three habits:
- One clear outcome: They measure a single overall result against a focused list of distinct drivers.
- A fitting method: They match the method to the sample size and the overlap between drivers, then check the data before trusting the ranking.
- Action on the right quadrant: They invest first in drivers that matter most and score below par, and they rerun the analysis to see if it worked.
Frequently Asked Questions (FAQs)
There is no single cutoff, but many analysts plan for at least 10 complete responses per driver tested. A survey with 15 drivers would need about 150 responses. Plan for more if you want to compare drivers across customer segments.
Yes. Excel’s Analysis ToolPak runs correlation and multiple regression, which covers a basic driver analysis. Shapley values and relative weights need a statistics package, a script in R or Python, or a survey platform that calculates driver importance automatically.
Derived importance is the weight of a driver calculated from response patterns rather than asked directly. If customers who rate support highly also give high overall ratings, support has high derived importance, even when customers never call it important.
Yes. Use the raw 0 to 10 likelihood-to-recommend score as the outcome rather than the promoter, passive, and detractor groups. The raw scale keeps more detail, which gives correlation and regression models more information to work with.
Most CX teams rerun it quarterly or after a major change, such as a product launch, pricing update, or new support process. Drivers shift as customer expectations change, so a ranking from last year may point teams at the wrong priorities.



