What this tool does
Key Driver Analysis is Module 9 of Analysis Studio. It ranks the predictors of a continuous outcome by importance, where importance has a precise meaning: each predictor share of the explained variance, computed so the shares add to the model R squared exactly.
- For fifteen or fewer predictors, importance is the LMG decomposition, the Shapley value of R squared: each predictor is credited with its average marginal contribution across every possible order of entry into the model.
- For larger sets, the module uses Johnson relative weights, which approximate the same decomposition at a fraction of the cost.
- Intervals are bootstrap percentile intervals from one thousand seeded replications.
- Survey designs take point estimates from the weighted correlation matrix, with intervals from replicate weights or a rescaling bootstrap.
Outputs are an importance table with shares, ranks, and intervals; a bar chart with the intervals drawn; and the importance-by-performance quadrant chart, with quadrants labeled Priorities to improve, Strengths to maintain, Lower priority, and Possible over-investment, for the satisfaction and loyalty studies where this analysis usually lives. Every narrative line in the output states that importance reflects association, not cause and effect, and the page repeats that here because the sentence matters.
Promotion, stated with its date: through March 31, 2027, this module runs free on screen, like every module. Downloads and Copy Results stay with the subscription throughout. After the promotion, the module returns to the subscription set at $299 per year.
The method, in plain terms
Correlated predictors break naive importance measures. A standardized coefficient rewards a predictor for variance it shares with others only when it wins the regression arithmetic, and a simple correlation credits shared variance to every predictor at once. The Shapley approach settles the dispute fairly: a predictor contribution is averaged over every ordering, so shared variance is divided among the predictors that share it, and the shares sum to the total explained variance with nothing left over and nothing double-counted.
Worked example
Twelve cases, three drivers d1, d2, and d3, and an outcome built mostly from the first two. The outcome values, in case order: 10.3, 14.1, 13.4, 17.4, 14.3, 20.9, 13.9, 18.1, 15.9, 23.0, 17.4, 21.7.
- Model R squared: 0.9906
- Importance share of d1: 58.1 percent of the explained variance
- Importance share of d2: 36.2 percent
- Importance share of d3: 5.7 percent
The read for a decision meeting: effort spent improving d1 buys roughly ten times the outcome movement of effort spent on d3, per unit of improvement, within the ranges the data cover. The bootstrap intervals decide how much confidence that ranking has earned, and the quadrant chart then asks the second question: where does performance on the top driver currently stand?
When the tool fits, and when it does not
It fits when:
- The question is prioritization: a fixed improvement budget, several candidate drivers, and one outcome that matters.
- Predictors are correlated, as satisfaction attributes always are, and the ranking has to survive that correlation honestly.
- The audience needs shares that add up, in a chart a meeting can read in one minute.
It does not fit when:
- A causal claim is on the table. Importance is explained variance, not leverage proven by experiment. The module says so in its own output, and reports should keep the sentence.
- The outcome is binary or a count. Importance decomposition here is built for continuous outcomes; the regression modules carry the other outcome types.
- Predictors number in the dozens with a modest sample. Intervals widen to the width of the ranking itself, and the honest output is that the data cannot rank them.
Run it
Module 9: Key Driver AnalysisOpen in Analysis Studio →Load your data, choose the outcome and two to thirty predictors, and read the intervals before the ranks.
Questions
Why not just rank by correlation with the outcome? Because correlated drivers each claim the shared variance in full, and the claims add to more than the model explains. The Shapley decomposition divides shared variance among the drivers that share it, so the shares sum to the R squared and the ranking is a fair division rather than a pile of overlapping claims.
What do the bootstrap intervals change? They decide whether a ranking is a finding or a coin flip. Adjacent drivers with overlapping intervals are tied for practical purposes, whatever the point estimates say, and the table prints the intervals beside the shares so that reading is unavoidable.
How does the quadrant chart work? Importance on one axis, current performance on the other. High importance with low performance is the improvement priority. High importance with high performance is a strength to protect. Low importance quadrants argue against spending, whatever their performance, because moving them barely moves the outcome.
Can it run on survey data? Yes. Point estimates come from the weighted correlation matrix, and intervals come from the replicate weights where the design supplies them, or a rescaling bootstrap where it does not.
Is a top driver a cause? No. A driver can top the ranking because it proxies for an unmeasured cause, or because the outcome colors how respondents rate it. The analysis ranks association with discipline; experiments and design decide causes.