Driver Analysis

What is key driver analysis?

A family of methods that relates a set of measured attributes to an outcome such as satisfaction, loyalty or intention, and ranks the attributes by how strongly they are associated with it. The important qualification is in the name: it identifies statistical associations, not causes. Attributes and outcome are usually measured at the same moment from the same respondent, so temporal order is unknown, reverse causation is unexcluded and confounding is untreated.

Why do stated importance and derived importance disagree?

Because they measure different things, and the disagreement is informative rather than a problem to fix. Stated importance reflects what people believe matters and what is acceptable to say matters; it over-reports price, reliability and safety. Derived importance reflects what covaries with the outcome in this sample, and it favours whatever is caught up in the general favourability running through all the ratings. The most useful cell is low stated and high derived, where something matters that people do not articulate. The most dangerous is high stated and low derived, which usually means the attribute is delivered uniformly well and therefore cannot discriminate.

Why is my best-performing attribute ranked least important?

Because a model detects an attribute through its variation, and an attribute everyone rates highly has almost none. Security, safety and basic reliability routinely rank near the bottom of derived importance for exactly this reason. They are hygiene factors: they do not distinguish satisfied customers from dissatisfied ones while they are working, and they would dominate the model if they failed. Annotate them explicitly, because an instruction to divest from them can be very expensive.

What does multicollinearity do to a driver analysis?

Survey attribute batteries typically correlate at 0.4 to 0.7 with each other. When predictors share that much variance, the regression cannot determine how to split the shared association between them, so individual coefficients become unstable: standard errors inflate, signs can flip, and dropping one attribute reorders the rest. The ranking then depends on which attributes happened to be in the battery. Report a collinearity measure per attribute, and test the ranking by resampling rather than trusting the point estimates.

How do I know whether my driver ranking is stable enough to report?

Perturb it. Re-estimate on bootstrap resamples and record how often each attribute lands in the top three; split the sample in half and compare; drop an attribute and see whether the rest reorder. If the leading attributes hold their positions in the large majority of resamples, the ranking is usable. If an attribute moves between second and ninth depending on the resample, there is no ranking to report, however precise the point estimates look, and the honest output is a grouped set rather than an ordered list.

Can driver analysis tell me what to fix?

It can tell you which of the things you measured are most strongly associated with the outcome, which is a real and useful input. It cannot tell you that fixing one of them will move the outcome, because the design does not establish causation, and it cannot see anything that was not in the questionnaire. It also says nothing about cost, feasibility or how much room for improvement exists. Prioritisation needs all of that, and it needs a person.

What if an important driver was not asked about?

It cannot appear in the model, and its association is absorbed by whichever measured attributes correlate with it. If price was not asked and price is what matters, the model will credit value perception or fairness instead, and the output will look exactly the same as a correct one. This is the largest and least visible risk in the method. List what was not measured alongside the ranking every time, and where a known material attribute is missing, say the analysis cannot answer the question.

How much of the outcome should a good driver model explain?

There is no target, and chasing a high figure usually means adding items that duplicate the outcome. A substantial part of any figure from a single-questionnaire model is the general favourability that colours all of a respondent's ratings, so a high value is not evidence that the business understands its customers. A low value often means the outcome depends on things the study did not measure, which is a genuine finding rather than a defect.

What is the importance and performance matrix used for, and what goes wrong with it?

It plots derived importance against current performance to identify where improvement effort is best placed. Three things go wrong. It invites the reading that raising an attribute's performance will move the outcome in proportion to its importance score, which the design does not support. Attributes near the axis crossing are not distinguishable from each other, so the quadrant boundary gets treated as a decision line when it is not. And well-delivered hygiene factors sit in the low-importance region precisely because they are working, so the matrix can appear to recommend abandoning them.

Can AI run driver analysis reliably?

It applies the diagnostics consistently and does not get bored of running the stability checks, which is a genuine advantage. The risks are specific: producing a ranked list with no model behind it, presenting ranks that the data will not support, letting causal language reappear in a heading or an executive summary after being removed from the analysis, and converting importance shares into projected gains as though they were rates of return. Used properly, AI should be required to report model fit, base and attribute set with every ranking, attach a stability result, state what was not measured, and place the association statement in the same document as the numbers.