Regression Analysis

Diagnostics

• Residual plots — primary diagnostic tool. Plot residuals vs fitted values, vs each predictor, Q-Q plot.
• Influential observations — Cook's distance (> 4/n suggests influence), leverage (hat values), DFBETAS.
• R-squared — proportion of variance explained. Adjusted R-squared penalizes for number of predictors. R-squared alone does not validate the model.
• F-test — overall model significance (at least one predictor is associated with Y).

Logistic Regression

Binary Logistic Regression

Models the probability of a binary outcome:

log(p / (1-p)) = beta_0 + beta_1X1 + ... + beta_kXk

Where p/(1-p) is the odds of the outcome.

Odds Ratio Interpretation

• exp(beta_j) = adjusted odds ratio for predictor X_j
• For a continuous predictor: OR per one-unit increase in X, adjusted for other covariates
• For a categorical predictor: OR comparing category to reference category
• OR = 1: no association; OR > 1: higher odds; OR < 1: lower odds
• Report OR with 95% CI. The CI is more informative than the p-value alone.

OR vs RR

The OR overestimates the RR when the outcome is common (> 10%). For common outcomes, consider:

• Log-binomial regression (directly estimates RR)
• Modified Poisson regression with robust standard errors (Zou, 2004)
• Report both OR and RR when feasible

ROC Curve and Model Discrimination

• ROC curve — plots sensitivity vs (1-specificity) across all possible thresholds
• AUC (C-statistic) — probability that a randomly selected case has a higher predicted probability than a randomly selected non-case
  • AUC = 0.5: no discrimination (coin flip)
  • AUC = 0.7-0.8: acceptable
  • AUC = 0.8-0.9: excellent
  • AUC > 0.9: outstanding
• Calibration — does predicted probability match observed probability? Hosmer-Lemeshow test (limited), calibration plots (preferred), calibration slope and intercept.

Model Fit

• Hosmer-Lemeshow test — tests calibration across deciles of predicted probability. Non-significant = adequate fit. Sensitive to grouping strategy and sample size.
• AIC/BIC — for model comparison (lower is better)
• Pseudo R-squared — Nagelkerke, McFadden, or Cox-Snell. Not directly comparable to linear regression R-squared.

Multinomial Logistic Regression

For nominal outcomes with > 2 categories (e.g., treatment response: complete, partial, none):

• Models log-odds of each category relative to a reference category
• Produces one set of coefficients for each non-reference category
• Assumes independence of irrelevant alternatives (IIA) — adding/removing a category should not change the relative odds of remaining categories. Test with Hausman-McFadden test.

Ordinal Logistic Regression

For ordered categorical outcomes (e.g., disease severity: mild, moderate, severe):

Proportional Odds Model

• Most common ordinal regression
• Estimates a single OR that applies across all cutpoints of the ordinal scale
• Proportional odds assumption — the OR is constant regardless of where the outcome is dichotomized. Test with Brant test or score test.
• If assumption violated: consider partial proportional odds model, multinomial logistic, or treat as continuous

Interpretation

OR represents the odds of being in a higher (worse) category for a one-unit increase in the predictor. A single summary across the entire ordinal scale.

Poisson Regression

For count outcomes (number of events: ED visits, infections, falls):

log(mu) = beta_0 + beta_1X1 + ... + beta_kXk

• exp(beta_j) = incidence rate ratio (IRR) for predictor X_j
• Assumes mean equals variance (equidispersion)
• Can incorporate an offset term (log of person-time) to model rates rather than counts

Overdispersion

When variance exceeds the mean (common in medical data):

• Detection: deviance/df or Pearson chi-square/df >> 1
• Consequences: standard errors too small, p-values too liberal, CIs too narrow
• Solutions: quasi-Poisson (scales SE by dispersion parameter), negative binomial regression, or robust standard errors

Negative Binomial Regression

Extension of Poisson for overdispersed count data:

• Adds a dispersion parameter (alpha or k) to model extra-Poisson variation
• When alpha approaches 0, converges to Poisson
• Preferred over Poisson when overdispersion is present
• Same interpretation: exp(beta) = IRR

Zero-Inflated Models

When excess zeros exist beyond what Poisson/NB predicts (e.g., many patients with zero hospitalizations):

• Zero-inflated Poisson (ZIP) or Zero-inflated Negative Binomial (ZINB)
• Two-part model: (1) logistic model for the probability of being a "structural zero" (never at risk) vs "at risk," (2) count model for those at risk
• Requires theoretical justification for the two processes generating zeros
• Compare with standard Poisson/NB using Vuong test or AIC

Model Building and Selection

Variable Selection Approaches

Subject-matter driven (preferred):

• Select confounders based on clinical knowledge and DAGs
• Include variables known to be associated with both exposure and outcome
• Do NOT use statistical significance of individual predictors to decide inclusion

Data-driven (use cautiously):

• Forward selection — start with empty model, add variables one at a time based on statistical criteria. Tends to include too few variables.
• Backward elimination — start with full model, remove variables. Generally preferred over forward.
• Stepwise — combination of forward and backward. Criticized: inflates significance, biased coefficients, unstable across samples.
• LASSO (L1 regularization) — penalized regression that shrinks coefficients toward zero, performing variable selection. Better than stepwise for prediction models.

Recommendation: Use clinical judgment for explanatory/etiologic models. Use penalized methods (LASSO, elastic net) for prediction models with validation.

AIC and BIC

• AIC (Akaike Information Criterion) — balances model fit and complexity. Lower is better. Penalizes each parameter by 2.
• BIC (Bayesian Information Criterion) — stronger penalty for complexity (log(n) per parameter). Favors simpler models, especially with large n.
• Neither provides an absolute measure of fit — only relative comparison between models on the same data.

Multicollinearity

Definition

High correlation among predictor variables, causing unstable coefficient estimates and inflated standard errors.

Detection

• Variance Inflation Factor (VIF) — VIF > 5 suggests concern; VIF > 10 indicates serious multicollinearity
• Correlation matrix — pairwise correlations > 0.7-0.8 suggest collinearity
• Condition number — > 30 indicates problematic collinearity
• Tolerance — 1/VIF. Values < 0.2 are concerning

Consequences

• Individual coefficients unreliable (high SE, wide CI, instability across samples)
• Overall model prediction may still be adequate
• Cannot interpret individual predictor effects

Solutions

• Remove one of two highly correlated predictors (choose based on clinical relevance)
• Combine correlated predictors into a composite score or index
• Principal component analysis (PCA) to create orthogonal predictors
• Ridge regression (L2 regularization) — shrinks coefficients without eliminating them
• Center continuous variables to reduce structural multicollinearity (with interaction terms)

Interaction and Effect Modification

Interaction

An interaction exists when the effect of one predictor on the outcome depends on the level of another predictor.

Y = beta_0 + beta_1X1 + beta_2X2 + beta_3*(X1*X2) + epsilon

• beta_3 is the interaction coefficient
• The effect of X1 on Y is: beta_1 + beta_3*X2 (depends on X2)
• Test interaction with the interaction term's p-value or likelihood ratio test comparing models with and without interaction

Additive vs Multiplicative Interaction

• Multiplicative interaction — interaction on the scale of the model (log-odds for logistic, log-rate for Poisson). Default when including product terms.
• Additive interaction — interaction on the absolute risk scale. Measured by RERI (Relative Excess Risk due to Interaction), AP (Attributable Proportion), S (Synergy Index). More relevant for public health.
• Presence of multiplicative interaction does not imply additive interaction and vice versa.

Reporting Standards

Following STROBE and TRIPOD guidelines:

1. Report the number of events per variable (EPV). Minimum EPV of 10 is a rough guideline for logistic regression (though recent work suggests 5-10 is context-dependent).
2. Present unadjusted and adjusted estimates.
3. Report model fit statistics.
4. Describe variable selection method.
5. For prediction models: report discrimination and calibration; perform internal validation (bootstrap, cross-validation).
6. Report confidence intervals for all effect estimates.
7. Check and report assumption verification.