Prof. Dr. Larry AdamsAcademic, Author & Researcher
How to Conduct Quantitative Research · Chapter 19 of 33

Part III — DATA ANALYSIS AND INTERPRETATION

Chapter 15. Regression, Mediation, Moderation and Advanced Models

15.1 Linear regression

Linear regression estimates the conditional mean of a continuous outcome as a function of predictors. A coefficient represents the expected change in the outcome for a one-unit increase in a predictor, holding included variables constant, under the specified model. For categorical predictors, interpretation depends on the reference category and coding. Report unstandardised coefficients with confidence intervals; standardised coefficients can supplement but should not replace meaningful units.

Assess linearity, residual patterns, heteroscedasticity, influential observations, multicollinearity, independence and model specification. VIF is a diagnostic, not a mechanical universal cutoff. If heteroscedasticity is present, robust standard errors may help with inference but do not repair all forms of misspecification. Consider transformations or alternative models only with theoretical and diagnostic justification.

15.2 Logistic regression

Logistic regression models the log odds of a binary outcome. Exponentiated coefficients are odds ratios, not risk ratios. An odds ratio of 1.50 means 50% higher odds, not necessarily 50% higher probability. Report confidence intervals and, where possible, predicted probabilities or marginal effects for practical interpretation. Check coding, sparse cells, separation, influential observations, nonlinearity of continuous predictors on the logit scale, calibration and discrimination.

15.3 Mediation

Mediation proposes that an exposure or predictor relates to an outcome through an intermediate variable. The indirect effect is commonly estimated as a product of paths and assessed with a bootstrap confidence interval or suitable model-based interval. Statistical mediation does not by itself prove a causal mechanism. Cross-sectional data cannot establish temporal order, and mediator–outcome confounding may remain. Strong causal mediation claims require careful design and assumptions.

15.4 Moderation

Moderation asks whether a relationship differs across levels of a moderator. In regression, it is commonly represented by an interaction term, such as X × W. Interpret the interaction with predicted values, simple slopes or marginal effects at meaningful moderator values. Avoid concluding that an effect differs merely because it is significant in one subgroup and not significant in another; test the interaction directly.

15.5 Multilevel and clustered data

Employees nested within teams, branches or organisations are not fully independent. Multilevel models can represent within- and between-group variation; generalised estimating equations can estimate population-average relationships with robust inference under suitable conditions. Report the number of clusters as well as individuals. A large number of individuals within only a few clusters does not necessarily provide strong information about cluster-level effects.

15.6 Structural equation modelling (SEM)

SEM can estimate measurement and structural relationships together, but it demands theoretical clarity, adequate data, model identification, appropriate estimators and transparent fit reporting. Fit indices such as CFI, TLI, RMSEA and SRMR should be considered collectively and in context, not as universal pass/fail thresholds. Avoid repeatedly modifying a model based only on modification indices without theoretical justification and independent validation.

15.7 Predictive modelling

Prediction focuses on performance for new cases, not necessarily causal explanation. Separate training, tuning and test data appropriately; use cross-validation where suitable; prevent leakage; and compare against meaningful baselines. Report metrics aligned to the problem, such as RMSE/MAE for continuous outcomes, calibration and discrimination for classification, and precision/recall when class imbalance matters. A high in-sample R² is not evidence of out-of-sample usefulness.