Before you fit a GLM, you look. A few well-chosen plots tell you whether a covariate is skewed, whether a target is over-dispersed, and whether a categorical level is too sparse to trust. MAS-I rewards candidates who can read these plots fluently.
Why EDA before modeling. A GLM assumes a link, a variance function, and (often) a particular distributional family. EDA checks those assumptions cheaply. A histogram that looks log-normal makes a gamma GLM with log link plausible. A box plot showing fat upper tails warns that a Poisson model will be over-dispersed and a negative binomial may fit better.
KEY: EDA is descriptive, not inferential. You are not testing hypotheses yet. You are deciding which model family to try and which covariates need transforming.
Histograms for continuous data. A histogram bins observations and draws a bar whose height equals frequency (or relative frequency, or density). You read three things off a histogram:
- Shape. Symmetric, right-skewed, left-skewed, or multi-modal. Insurance losses are almost always right-skewed.
- Center and spread. Roughly where the mass sits and how wide it is.
Common mistakes
- Reading skewness from the box only and ignoring the whiskers. A symmetric box with a long upper whisker is still right-skewed. Always check whisker asymmetry.
- Using histograms for categorical data. Histograms require an ordered numeric axis; categorical data uses a bar plot. Conflating the two leads to meaningless bin widths.
- Trusting one bin width. A histogram with the default 30 bins can hide a second mode that appears clearly at 15 bins. Always view at least two bin widths before reporting "unimodal."
Bottom line
- Histogram reveals shape, modality, skew, and gaps for numeric data, and bin width controls the story (view at least two widths before declaring unimodal).
- Box plot flags center (median), spread (IQR ), and outliers beyond or ; it is the fastest skewness screen.
- Right-skewed loss data (median below mean, long upper tail) usually wants a log transform before modeling.
- Bar plot or frequency table is the univariate display for categorical variables; merge or pool levels with under about 30 observations.
Exam shortcut
If the question describes a tall left side of the box, a short whisker below, and a long whisker above, answer "right-skewed" and move on; do not search for additional statistics. When a box plot or histogram shows variance that grows with the mean, the exam usually wants a log link or a gamma family next, not deletion of points.
The full lesson (about 2,551 words, 17 min read) adds 2 worked examples, all 6 common mistakes, a self-check, free in the app.
Learning objectives
- C9
Browse all free MAS-I lessons or jump into free MAS-I practice questions.