Module 4 — Choice and Limited Outcomes

Count Outcomes

Model event rates with exposure, interpret multiplicative effects and diagnose dispersion and excess zeros

Count Outcomes

A count needs a clock and a population

“Twelve asthma visits” is incomplete. Twelve visits among how many residents and over how long?

For municipality ii, let YiY_i be visits and EiE_i be person-years at risk. A log-link model is

E[YiXi,Ei]=Eiexp(Xiβ),E[Y_i\mid X_i,E_i]=E_i\exp(X_i'\beta),

or equivalently logEi\log E_i enters as an offset with coefficient fixed at one.

Clearborough records 240 visits across 20,000 person-years before the clean-air zone and 180 across 21,000 after:

PeriodVisitsPerson-yearsRate per 1,000
before24020,00012.0
after18021,0008.57

The crude incidence-rate ratio is 8.57/12.0=0.7148.57/12.0=0.714, a 28.6% rate reduction. This is descriptive until a counterfactual design handles time trends and composition.

Interpret exponentiated coefficients

In a Poisson model, eβke^{\beta_k} is a multiplicative change in the conditional mean for a one-unit covariate change. If βD=0.25\beta_D=-0.25,

e0.25=0.779,e^{-0.25}=0.779,

so the expected rate is 22.1% lower, conditional on the model. The absolute reduction still depends on the baseline rate.

Poisson can be useful without a Poisson variance

Poisson pseudo-maximum likelihood targets a correctly specified conditional mean and can use robust or design-aligned standard errors when variance differs from the mean. Negative binomial models add a particular overdispersion structure; choosing one is not merely “variance exceeds mean, therefore negative binomial.”

SymptomPossible causeResponse
variance exceeds meanheterogeneity, dependence, wrong meaninspect design; robust/clustered inference; richer mean
many zerosgenuinely no risk, separate participation, rare eventsdistinguish structural process from chance zeros
repeated peoplewithin-person dependencecluster or model dependence at person level
different exposuretime/population at risk variesoffset or explicit rate estimand

Hurdle and zero-inflated models answer different stories

A hurdle model separates whether any event occurs from the positive count. A zero-inflated model posits a latent group that cannot generate events plus a count process that can also produce zeros.

For student support visits:

  • “no need for support” may generate structural zeros;
  • “needed support but did not attend” is a behavioural zero;
  • “one or more visits” has an intensity process.

Use a two-process model only when that story is meaningful and identified by the data. A better likelihood cannot reveal latent zero types without information.

Fixed effects and short panels

With repeated municipality counts, Poisson fixed-effects estimation can target multiplicative within-unit changes under a conditional-mean assumption. Include time effects for common shocks and align clustering with policy assignment. A clean likelihood does not solve staggered-treatment bias; combine the outcome model with an appropriate DID/event-study design.

Quick check

District A has twice as many incidents as B but four times the population at risk. Which district has the higher rate?

Answer
B. A’s rate is half of B’s: (2Y)/(4E)=0.5(Y/E)(2Y)/(4E)=0.5(Y/E). Counts without exposure can reverse the substantive comparison.

Next: Censoring and Selection

Copyright © 2026