Handbook of Regression Modeling in People Analytics
One book, placed in its field — what Keith McNulty's book gets right, where it goes further than the rest, and what the rest of the field adds.
Mike West, Editor · Leading With Data · peopleanalyst.com/guides/profile-regression_modeling_in_people_analytics
Arrow keys or swipe to move · N for notes · F for full screen
Handbook of Regression Modeling in People Analytics
One book, placed in its field — what Keith McNulty's book gets right, where it goes further than the rest, and what the rest of the field adds.
Mike West, Editor · Leading With Data · peopleanalyst.com/guides/profile-regression_modeling_in_people_analytics
The problem
Valid Statistical Inference erodes when it is left to instinct instead of method.
You were taught the moves piecemeal, never the whole model.
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data
The model
- 01Outcome Variable Type
- 02Data Structure and Hierarchy
- 03Regression Method Selection
- 04Coefficient Interpretation
- 05Assumption Validation
- 06Model Fit and Parsimony
- 07Statistical Power and Sample Adequacy
- 08Valid Statistical Inference
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data
Step 1 of 5 · Outcome Variable Type
Classify the outcome's measurement nature before opening any modeling software — it is the single decision that constrains all others.
Misconception
Practitioners often treat any numeric-coded outcome as continuous, running ordinary linear regression on a 1–5 engagement rating or a 0/1 termination flag because the column contains numbers.
Reality
The numbers stored in a column are labels, not measurements; a five-point Likert scale is ordinal because you cannot assume 'agree' is exactly one unit more than 'neutral', and a binary flag violates linear regression's constant-variance and unbounded-range assumptions.
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data
Step 2 of 5 · Data Structure and Hierarchy
Grouping hierarchy determines whether you need a mixed model; check the intraclass correlation before assuming independence.
Misconception
Analysts assume that as long as they have thousands of employee rows, more data guarantees reliable inference regardless of how those rows relate to one another.
Reality
Employees within the same manager or office are correlated, so 5,000 employees in 50 teams carry far less independent information than 5,000 unrelated observations; the effective sample size — and thus your confidence — is governed by structure, not row count.
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data
Step 3 of 5 · Regression Method Selection
Let outcome type and data structure narrow the method before personal preference enters the decision.
Misconception
Many practitioners believe method selection is a matter of preference or familiarity — reaching for logistic regression or linear regression because that is what they know best.
Reality
Method selection is largely determined for you once you have honestly diagnosed the outcome type and data structure; the analyst's judgment operates at the margins, such as choosing between proportional-odds and multinomial when the proportional-odds assumption is borderline.
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data
Step 4 of 5 · Coefficient Interpretation
Always attach the coefficient's scale (log-odds, odds ratio, hazard ratio, raw units) to any spoken or written claim.
Misconception
Practitioners treat a logistic regression coefficient (or its odds ratio) as if it were a change in probability or percentage points of the outcome.
Reality
Logistic coefficients live on the log-odds scale and odds ratios are multiplicative on odds, not probability; the same odds ratio implies a large probability change near a 50% base rate and a tiny one near 5%, so the marginal effect depends on where each employee sits.
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data
Step 5 of 5 · Assumption Validation
Run the assumption checks specific to your model family, not a generic OLS residual scan applied to everything.
Misconception
Analysts believe a high R-squared or a good fit statistic means the model's assumptions are satisfied and the inferences are safe.
Reality
Fit quality and assumption validity are orthogonal — a model can predict well overall while violating homoscedasticity or proportional hazards, which corrupts the standard errors and confidence intervals even when the point predictions look fine.
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data
The cheat sheet
- Outcome Variable TypeClassify the outcome's measurement nature before opening any modeling software — it is the single decision that constrains all others.
- Data Structure and HierarchyGrouping hierarchy determines whether you need a mixed model; check the intraclass correlation before assuming independence.
- Regression Method SelectionLet outcome type and data structure narrow the method before personal preference enters the decision.
- Coefficient InterpretationAlways attach the coefficient's scale (log-odds, odds ratio, hazard ratio, raw units) to any spoken or written claim.
- Assumption ValidationRun the assumption checks specific to your model family, not a generic OLS residual scan applied to everything.
- Model Fit and ParsimonyAdjusted R-squared and AIC, not raw R-squared, are the metrics that reward economy alongside explanation.
- Statistical Power and Sample AdequacyPower is driven by events per predictor, so rare HR outcomes demand disproportionately large samples.
- Valid Statistical InferenceValid inference is a conjunction of conditions, not a single significant coefficient.
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data
Sources
- Handbook of Regression Modeling in People Analytics — Keith McNulty
Handbook of Regression Modeling in People Analytics · Mike West, Editor · Leading With Data