How I choose methods

Which tool for which data

Interviewers ask "why this method?" more often than "how does it work?" Here is my actual decision map: seven data situations, the tool each one calls for, and a worked example from this portfolio for every route. The rule behind all of them: the method is chosen after meeting the data.

7 routes, each with a worked example ~3 minute read
Individuals answering a one-time survey, with design weights attached Weights exist because some households are harder to reach; ignoring them biases everything. The survey's own replicate weights drive the uncertainty. Worked example: rent & mental health.
survey-weighted regression
A monthly series around a policy event, with a group the event missed The untouched group absorbs seasonality and citywide shocks; a placebo test proves the machinery is honest. Worked example: the 2023 shelter crisis.
controlled interrupted time series
A series around a nationwide shock, with no one left untouched No comparison group exists, so the design is weaker and the write-up says so instead of dressing it up. Worked example: opioid deaths at the pandemic.
single-series before-after
Prices of things that differ in many ways at once Compare like with like: fixed effects hold the location constant, controls hold size and quality constant, and the remaining gap is the thing you care about. Worked example: the 28-day loophole.
hedonic regression, fixed effects
A "how long until it happens" question, where only "still open today" is recorded Time-to-event methods need event dates. When the data has none, each cohort's share still open traces the curve instead, and the limit is stated. Worked example: the planning queue.
current-status survival design
Two-way causality suspected between the exposure and the outcome When X plausibly moves Y and Y plausibly moves X, ordinary regression cannot untangle them; an instrument that shifts only X can. Worked example: my MA thesis on Airbnb and rents.
instrumental variables (2SLS)
A question the data can answer by counting When the finding is a share, a median, or a flat line, a model adds nothing except false precision. Full-scale axes keep small differences looking small. Worked examples: heat risk and the flat income gradient.
careful descriptives

One model in different costumes

Underneath the map sits the insight that organizes it: most of these routes are the same model. Ordinary regression, the logistic models, the hedonic model, and both interrupted time series are all the general linear model wearing different costumes. Change the outcome's distribution and link and you have the logistic version; build the predictors out of time, a break, and seasonal terms and you have an interrupted time series; add fixed effects and you have the hedonic version. Seeing one family instead of a menu of rituals is why moving between them is cheap, and why the real work in every study above was the design and the data rather than the fitting.

Two habits sit on top of the map. First, the plan bends to the data: the planning-queue study was designed as a time-to-decision analysis until first contact showed the decision dates do not exist, and the design changed rather than the data being forced. Second, functional form gets checked rather than assumed: a straight line was rejected for visible underfit in the shelter study, and the same check ran as a published sensitivity analysis in the opioid study.

What is missing from this map is as deliberate as what is on it: methods I have not used in real work do not appear. The claims-and-evidence page holds that line for the whole portfolio.

Browse the six studies The pre-flight checklist →