Module 7: Choosing Between Logistic, Log-Binomial, Poisson, and Probit Models

Module 7: Choosing Between Logistic, Log-Binomial, Poisson, and Probit Models

STATA playlist

Open the full YouTube playlist

This module teaches model choice as a practical decision rather than a theoretical contest. Your uploaded comparative notes distinguish logistic regression for odds ratios, log-binomial regression for direct risk ratios, Poisson regression with robust standard errors as a practical workaround, and probit regression as a latent-variable probability model.

A particularly valuable part of your teaching notes is the real mistreatment example, where a log-binomial model was attempted but failed to converge. That example makes the teaching point clear: students must learn to balance interpretability, outcome prevalence, and whether the model actually behaves well in the data.

Decision framework

  • Logistic: stable and widely used when odds ratios are acceptable.
  • Log-binomial: attractive for direct risk ratios, but convergence can be fragile.
  • Poisson with robust standard errors: often a practical alternative when log-binomial fails.
  • Probit: useful when a latent normal framework is preferred.

STATA examples

logit outcome x1 x2 x3
glm outcome x1 x2 x3, family(binomial) link(log)
poisson outcome x1 x2 x3, vce(robust)
probit outcome x1 x2 x3

Comments

Post a Comment

Popular posts from this blog

Module 5: Logistic Regression (Core Module)

Image
Module 5: Logistic Regression (Core Module) STATA playlist Open the full YouTube playlist Logistic regression is the central model in this course because it is the standard approach for binary outcomes. It estimates the probability that an event occurs while keeping predicted values within the 0 to 1 range, and it is commonly interpreted using odds ratios. This module explains the model in applied terms: what the coefficients mean, why exponentiating them gives odds ratios, and how to distinguish a change in odds from a change in probability. What students should understand Use logistic regression when the dependent variable is coded 0/1. logit reports coefficients in log-odds; logistic reports odds ratios directly. Odds ratios are not the same as risk ratios. STATA commands logit outcome x1 x2 x3 logistic outcome x1 x2 x3 predict phat margins
Read more

Module 4: Regression with Dummy Variables

Image
Module 4: Regression with Dummy Variables STATA playlist Open the full YouTube playlist Dummy variables are essential whenever categorical predictors appear in a model, including sex or gender, hospital type, race or ethnicity, treatment groups, or place of residence. This module explains how to include those predictors correctly and why one category must be omitted and treated as the reference group. Students learn the dummy variable trap, the logic of the baseline category, and how to interpret coefficients as differences relative to that reference group. Key points Use one fewer dummy than the total number of categories. The omitted category becomes the reference group. Interpretation is always relative to that baseline. STATA examples tab group, gen(d_) regress y x1 d_2 d_3 regress y i.group x1
Read more

Module 8: Interpretation for Policy and Health Research

Image
Module 8: Interpretation for Policy and Health Research STATA playlist Open the full YouTube playlist This final module focuses on translation rather than estimation. Students learn how to move from model output to clear, defensible statements for clinicians, public health practitioners, policymakers, and other non-technical audiences. The goal is to teach careful language: distinguish odds from risks, statistical significance from substantive importance, and model stability from policy relevance. Good interpretation also means acknowledging limitations such as convergence problems, uncertain fit, or the possibility that a statistically significant effect is still modest in practical terms. Interpretation checklist Name the outcome clearly. State whether the estimate is an odds ratio, risk ratio, probability, or marginal effect. Translate the estimate into plain language. Note important caveats about prevalence, model assumptions, and generalizability. Example phras...
Read more