Exercise 4: Analytical applications

Work in progress. These tutorial materials are still under active development and will continue to change until the tutorial takes place on October 20, 2026. Content, links, and exercises may be incomplete or shift without notice.

Goal

Design an analysis-ready exposure feature and identify threats to validity and federated execution requirements. Exposure records are longitudinal facts; models require index-aligned covariates, and turning one into the other requires deliberate choices that can be gotten wrong in specific, checkable ways.

Steps

Choose a research question that uses the exposure metric you calculated in Exercise 3 (or a similar one).
Specify the feature: lookback window, lag, aggregation method, missingness handling, unit harmonization, geography, and update frequency.
Decide whether this feature belongs in cohort eligibility or as a post-index predictor — these must stay separate. If your feature involves a spatially informed cohort, check it against appropriate vs. inappropriate uses.
Identify three specific threats to validity for your feature, drawn from: ecological fallacy, exposure misclassification, residential mobility, MAUP (modifiable areal unit problem), temporal mismatch, collider/selection bias, spatial autocorrelation/clustering, site effects, or region-level confounding. For each, state concretely how it could bias your result, not just name it.
Note what would need to stay local vs. what could travel in a federated setting: standardized concepts and code travel; precise locations and patient-level geographies remain local; only diagnostics and aggregate estimates are shared.
Write the feature definition as reproducible SQL (or pseudocode if SQL access isn’t available in your environment).

Deliverable

One feature specification, plus three anticipated biases (with the concrete mechanism for each, not just the name).

Check Your Work

SQL (or equivalent logic) is reproducible from the specification alone
Vocabulary release used is fixed/pinned, not “whatever is currently installed”
Temporal alignment to the index date is explicit and correct
Missingness policy is stated
Output is privacy-safe — no exact coordinates or other re-identifying detail in participant-facing output

Closing synthesis

Dataset → metadata → source layer → location interval → exposure fact → covariate → evidence. Each session in this tutorial produced one link in that chain; the chain only holds if every link stays traceable back to its source.

Previous: Exercise 3