Logistic regression for dose finding

Picking a method, and picking the right diagnostics

Index of the documents in the logistic-regression project: comparing glm against Bayesian fitting methods, and a minimal diagnostic checklist for a fitted model, both working documents.

A logistic regression of dose against adverse event rate or overall response rate is a common early analysis in dose finding. Two questions come up every time it is run: which function to fit it with, and whether the fit is any good. Neither is fully worked out; both are drafts, to be extended as time allows.

Documents

• Picking a Method. Comparing stats::glm against two Bayesian alternatives, rstanarm and brms. The glm fit runs live on every build; the two Bayesian fits are kept as code and as Andy’s earlier conclusions, not re-verified, because neither package is installed in the environment that builds this site.
• Diagnostics. Calibration, residuals and a linearity check for a fitted model, all computed live, plus the checklist of what is still missing: a Hosmer-Lemeshow test, an \(R^2\) measure, a Brier score.

Data

• dose_tumor_resp.rds, in data/synthetic/dose-response. 96 patients: assigned dose, baseline covariates (ECOG performance status, baseline tumor burden), and RECIST-derived best overall response. Used in both documents’ worked examples.