Key-driver analysis: choose the improvement behind the score

A customer-experience dashboard says communication has the strongest relationship with satisfaction. Your team can afford one improvement this quarter. Should it redesign notifications, make arrival times clearer or improve the chance that the first visit resolves the problem?

Key-driver analysis helps investigate that choice by examining relationships between a defined outcome and potential explanatory attributes. Its value is identifying a more useful question about the experience. A driver ranking alone does not tell you which intervention will work or where the next dollar should go.

Two highly ranked attributes may describe almost the same experience. A lower-ranked attribute may add information the others miss. The next decision depends on those relationships, current performance and what the business can change.

Give the outcome a specific experience

Consider Alder, a fictional appliance-repair company deciding between better appointment communication and better preparation for first visits. All numbers in this example are constructed teaching assumptions, rather than customer findings or results from a Sapien study.

Alder wants to ask customers seven days after their first technician visit, including requests that remain unresolved. Its outcome question is: “Thinking about this repair request so far, how satisfied are you with the service?” The scale runs from 0, completely dissatisfied, to 10, completely satisfied.

The candidate attributes are separate statements rated from 1, strongly disagree, to 5, strongly agree:

• “I was kept informed about the appointment.”
• “It was clear when the technician would arrive.”
• “I feel confident that the fault has been resolved.”

These are perceptions of a particular request. Confidence that a repair worked is not the same as a confirmed first-time fix. Alder also needs operational records, such as repeat visits, unresolved faults and parts availability, to investigate what happened.

Keep the outcome separate from the attributes. Putting “overall service quality” among the predictors of overall satisfaction risks rediscovering the same broad judgment under another name. If the team needs help selecting the outcome, our NPS, CSAT and CES guide explains the different questions those measures answer.

Read the first ranking, then inspect the overlap

Suppose the constructed example has the following summary:

The communication items have the strongest individual associations with satisfaction. The lower repair-confidence average suggests a different experience to investigate. Neither observation, by itself, establishes the cause of dissatisfaction.

Pearson's correlation describes a linear association and runs from −1 to +1. A positive value means higher ratings tend to occur together. A correlation of +0.80 does not mean that 80% of customers are satisfied, or that communication explains 80% of satisfaction. Squaring it gives 0.64, the R² of the corresponding one-predictor linear model with an intercept.

Now add the relationship between the attributes. In Alder's example, the two communication items correlate at +0.95. They share much of the same signal. “Kept informed” and repair confidence correlate at +0.40, leaving more room for distinct information.

That changes the model comparison. Using the same valid records and ordinary linear regression with an intercept:

• “Kept informed” alone gives R² = 0.640.
• Adding arrival clarity raises it to approximately 0.644.
• Adding repair confidence instead raises it to approximately 0.770.

Arrival clarity adds little fit after “kept informed” is included, even though it correlates with satisfaction more strongly than repair confidence does. Repair confidence adds more under this particular comparison. These R² values describe variation accounted for in the constructed example; they are not percentages of people predicted correctly or estimates of how much an intervention will improve satisfaction.

Azen and Budescu's methods paper explains why predictor importance becomes more complicated when attributes are correlated. It examines contributions across different combinations of predictors, rather than treating each isolated association as an independent contribution.

Ask what the importance measure actually means

“Key-driver analysis” describes a research job, rather than one universal calculation. A report might show individual correlations, regression coefficients or a relative-importance method that allocates a model's explained variation. Those outputs require different interpretations.

Ask which method produced the ranking, which attributes were included and how shared information was handled. Simply squaring each correlation and converting the results into percentages can produce a tidy ranking, but it does not separate overlapping contributions. A percentage label does not make it a budget allocation.

Specify a model appropriate to the outcome and response scales. Treating five-point items as numeric in Alder's linear illustration is a modeling choice, not a rule for every questionnaire. A model of renewal or cancellation needs its own outcome definition and interpretation.

Multiple regression examines relationships conditional on the other included predictors. Strong overlap can make individual coefficients sensitive to the model specification and data. NIST's variance-inflation guidance describes that instability and explains why pairwise correlations alone can miss more complex dependencies.

For Alder, the useful question is whether the two communication items capture distinct operational problems or mostly repeat one another. The team should investigate their meaning before combining them or removing either. Deleting an item solely to make a diagnostic look better can discard a real distinction.

Check whether the relationships can support the decision

Keep each response attached to the same customer request and survey timing. A recent first visit, an unresolved repeat repair and a routine maintenance appointment are different experiences. If their mix changes, the relationships can change too.

Preserve genuine missingness. A customer who cannot judge the repair should not automatically receive a neutral confidence rating. Report the eligible and valid bases, missing-answer rules and whether the model comparisons use the same records. Otherwise, a change in fit could partly reflect a change in who was analyzed.

Inspect distributions and unusual cases before trusting the ranking. An attribute almost everyone rates highly has limited variation; a small correlation does not establish that the capability is dispensable. Look for curved relationships, outliers and problems in the residuals, the differences between observed and fitted values. NIST's regression-diagnostics documentation describes checks that can expose weaknesses hidden by a summary fit measure.

The analysis should also report uncertainty and stability appropriate to its data and method. Does the apparent order persist across reasonable model choices and relevant customer groups? Does a model predict adequately on data not used to fit it? A close ranking should remain a close ranking, rather than becoming a confident first, second and third priority.

Turn the result into a testable improvement

For Alder, the worked comparison gives repair confidence a stronger reason to remain in the investigation. It adds information beyond communication and has a lower example rating. That supports examining unresolved faults and repeat-visit accounts before spending the entire quarter on another notification feature.

The next step is operational. Are customers uncertain because the technician could not complete the repair, because a symptom returned, or because nobody explained what to expect afterward? Those explanations lead to different changes: parts preparation, repair quality or a clearer explanation.

Choose a feasible candidate only after identifying that mechanism. Define the customers and requests it applies to, the change they will receive, the comparison and the outcome window. Record satisfaction alongside actual resolution, repeat visits and delivery cost. A suitable controlled intervention can provide evidence about the change's effect; the original association cannot supply it. NIST's guidance on association makes that causal boundary explicit.

Sapien can help explore proposed experience changes with a data-grounded simulated audience and compare reactions across relevant groups. Those results can develop hypotheses about expectations, communication or offer design. Relationships generated within a simulated setup should remain identified as modeled relationships, rather than become claims about what drove actual customers' satisfaction. Agree the analysis method, evidence source and execution responsibilities for any particular driver project.

A useful readout leaves Alder with a specific next investigation and a candidate improvement to test. It preserves the overlap between the communication measures, the additional information in repair confidence and the uncertainty about the underlying cause. That is a firmer basis for investment than funding whichever row happens to rank first.