PUBLIC HEALTH · SURVEY DATAAccess to healthcare
Estimate inequities in unmet healthcare need using a complex population survey.
Open walkthrough
Research question
Does the prevalence of unmet need differ by income after accounting for age, gender and region?
Why this method?
The outcome is binary and the survey design includes weights, strata and clusters.
svyset psu [pweight=weight], strata(stratum)
svy: logistic unmet_need i.income age i.gender i.region
margins incomeInterpretation focus
Report adjusted probabilities or contrasts alongside odds ratios, and retain the survey design in variance estimation.
A binary health outcome
Examine whether an exposure is associated with a yes-or-no health outcome.
Open walkthrough
Research question
Is exposure associated with the outcome after adjustment for pre-specified confounders?
Why this method?
Logistic regression respects the binary outcome and estimates adjusted odds.
logistic outcome i.exposure age i.gender comorbidity
margins exposure
marginsplotInterpretation focus
An odds ratio is not a risk ratio. Predicted probabilities often make the magnitude easier to understand.
Counts of adverse experiences
Model the number of adverse experiences reported by each participant.
Open walkthrough
Research question
How does the expected count vary across social or demographic groups?
Why this method?
Poisson regression models non-negative counts; negative binomial regression may be preferable with substantial overdispersion.
summarize adverse_count, detail
poisson adverse_count i.group covariates, irr vce(robust)
nbreg adverse_count i.group covariates, irrInterpretation focus
Translate the incidence-rate ratio into a percentage difference in the expected rate, conditional on the model.
Building a multi-item scale
Evaluate whether several survey items can support a coherent summary measure.
Open walkthrough
Research question
Do the items measure a sufficiently coherent construct, and how should the score be created?
Why this method?
Item distributions, inter-item relationships and reliability should be examined before constructing the scale.
tab1 item1-item6, missing
alpha item1-item6, item std
egen scale_mean = rowmean(item1-item6)Interpretation focus
Cronbach’s alpha is evidence about internal consistency, not proof of unidimensionality or validity.
Differences in an exposure effect
Assess whether an adjusted association differs across population groups.
Open walkthrough
Research question
Does the relationship between exposure and outcome vary by group?
Why this method?
An interaction term represents effect modification on the model’s chosen scale.
logistic outcome i.exposure##i.group covariates
margins group, at(exposure=(0 1))
margins group, dydx(exposure)Interpretation focus
Interpret predicted probabilities and contrasts. The interaction coefficient alone rarely communicates the substantive pattern clearly.
Trade or policy change over time
Use repeated observations to examine within-unit changes in an economic outcome.
Open walkthrough
Research question
How is a policy change associated with an outcome within countries, industries or firms over time?
Why this method?
Fixed effects account for stable unit characteristics while time indicators absorb common period shocks.
xtset unit year
xtreg outcome policy controls i.year, fe vce(cluster unit)
estimates store fixed_effectsInterpretation focus
The coefficient is identified by within-unit change. A causal claim still depends on the policy design and identifying assumptions.