MAST90139 · University of Melbourne · 2023 Semester 1

GLM Playground

Six case studies from my Statistical Modelling for Data Science assignments. Logistic, log-linear, multinomial, proportional-odds and GEE models are refitted live in your browser, and checked against the R code I submitted in 2023.

Start with dose–response 49/49 parity checks pass

Take the guided tour · three short recorded walkthroughs

g\big(\mathbb{E}[y_i]\big) = \mathbf{x}_i^{\top}\boldsymbol\beta, \qquad \operatorname{Var}(y_i) = \phi\, V(\mu_i)

Fig. 1 — Proportion of beetles killed at eight doses (dots) with the fitted logistic curve and 95% band. Computed by the TypeScript IRLS port; matches R's glm() to 10⁻⁹.

The coursework

What the assignments asked

MAST90139 takes regression beyond the normal linear model: binary, count and ordered responses, contingency tables and correlated data. The three individual assignments (written in R Markdown) asked me to fit, test and interpret these models on small published data sets.

Assignment 1

Choosing a logistic model

Select a parsimonious logistic regression for a binary survey outcome with forward and backward steps, analysis of deviance and AIC, then interpret its odds ratios — including an interaction.

Assignment 2

Binomial and Poisson GLMs

A dose–response model with intervals, LD50 and goodness-of-fit tests; additive versus interaction models for survey attitudes; and log-linear tests of independence in a three-way table, with their logistic equivalents.

Assignment 3

Multicategory and correlated responses

Write down and interpret nominal (baseline-category) and ordinal (proportional-odds) models for disease severity, and a GEE model with exchangeable correlation for repeated binary measurements.

Interactive

Six case studies

Each page states the model, lets you change its inputs, and recomputes estimates, intervals and tests on the spot.

Assignment 2 · Q1
Beetle mortality and insecticide dose
Binomial GLM, logit link
Drag the dose, watch the fitted kill probability and its interval, find the LD50 and test a quadratic term.
LD50 1.771
95% CI (1.764, 1.779); quadratic term p = 0.0035
Assignment 2 · Q2
Education, gender and attitudes to women's role
Binomial GLM with factors and interactions
Six candidate models side by side, from saturated to the sex-specific education slope I chose.
× 0.708
odds of agreeing per extra year of school (women; men × 0.768)
Assignment 2 · Q3
Race and the death penalty: a three-way table
Poisson log-linear models
Test four independence hypotheses, see their logistic twins, and meet Simpson's paradox.
p = 0.3903
only [VD][VP] fits: the victim's race, not the defendant's, predicts the sentence
Assignment 3 · Q1–Q2
Coal-face exposure and lung disease severity
Multinomial vs proportional-odds logit
Slide the years of exposure and compare a nominal and an ordinal model category by category.
AIC 422.9
proportional odds beats the nominal model (425.4) with one fewer parameter
Assignment 3 · Q3
Children's wheeze, age and maternal smoking
GEE, exchangeable working correlation
Repeated binary outcomes: robust sandwich errors versus a naive GLM, and an odds-ratio calculator.
α = 0.354
within-child correlation; robust SE for smoking 0.178 vs naive 0.123
Assignment 1
Stepwise logistic model selection
Logistic regression, analysis of deviance
The selection path and final odds ratios, published as summaries only because the survey is sensitive.
1.70×
odds for smokers in the final 18-parameter model (summaries only)

Added in 2026

Checked, documented and explained

The coursework fitted and interpreted the models. The revival adds what I would expect of the same analysis at work: checks of each model, written reasons for each choice, and an optional AI reading aid that is labelled, reviewed and logged.

Model checking
Simulated residual envelopes, influence, over-dispersion, profile-likelihood intervals, a proportional-odds test and GEE sensitivity, every simulation seeded.
Methods and decision records
Data provenance, evaluation design, assumptions and limitations, three decision records and a model card, with the weak numbers stated.
Optional AI, your own key
Explanations grounded only on each page's numeric summary, labelled as AI-generated, checked for untraceable numbers and recorded in an exportable audit log.

Faithful ports

Same algorithms, same answers

The R functions I used — glm, nnet::multinom, MASS::polr and geepack::geeglm — are reimplemented in around 1,300 lines of dependency-free TypeScript: IRLS with Householder QR, Newton–Raphson for the multinomial and cumulative logit likelihoods, and GEE with a sandwich estimator.

A script re-runs the original model calls and records R's output; the test suite holds the ports to it, typically within 10⁻⁹. Where the 2023 write-up contained a slip, the page says so and shows the recomputed value.

49/49

parity checks on /verification

10⁻⁹

typical agreement with glm()

published data sets, no raw survey rows

0.028 → 0.136

smoking p-value, naive GLM → GEE

See the parity table

About this project

From R Markdown to an interactive site

Subject: MAST90139 Statistical Modelling for Data Science
University: University of Melbourne
Year / semester: 2023 Semester 1
Format: Three individual assignments (no team)
Author: Sunchuangyu (Rin) Huang
Original stack: R 4.2.2, R Markdown, glm, nnet, MASS, geepack, dplyr, ggplot2
Revived stack: Next.js 16, React 19, TypeScript ports with vitest parity tests, Tailwind CSS, KaTeX
Hosting: Static pages on Vercel. Fits are computed at build time and refitted in your browser in the interactive panels
Repository: rNLKJA/Unimelb-Master-2023-MAST90139 on GitHub, kept private because coursework/ holds University of Melbourne material and the Assignment 1 survey.

Academic integrity. The original submissions are preserved unchanged in the (private) repository under coursework/ for reference. Assignment specifications, model answers and lecture material belong to the University of Melbourne and are not reproduced on this site; the tasks above are paraphrased.

Data. The beetle, attitudes, death-penalty, pneumoconiosis and Ohio wheeze data are classic published teaching data sets, cited on each page. The Assignment 1 survey is sensitive and appears only as model summaries.