A claims database records diagnoses, procedures, dispensings and demographics; it does not record lab values, vitals,
imaging measurements, severity scores or lifestyle. A propensity score fitted there is therefore error-prone .
If you also have a smaller dataset holding both the claims variables and the missing clinical ones, you can
fit a gold-standard propensity score in it, learn how the two relate, and correct the large study's effect estimate.
That is propensity-score calibration — regression calibration with the propensity score as the mismeasured covariate
(Stürmer T, Schneeweiss S, Avorn J, Glynn RJ. Am J Epidemiol 2005;162(3):279-89).
The fitted equation is the deliverable: a later claims-only study can cite it and apply it without repeating the validation work.
Where the numbers come from. This is a protocol builder — it holds no data and imports none. Every coefficient below
is the output of fitting two propensity models in your validation dataset: an error-prone score from the claims variables alone,
and a gold-standard score from those plus the clinical ones. Fit them wherever you like, or use
RWE Studio , which runs real R in your browser on
a dataset that never leaves it and can emit a calibration handoff file you import here instead of retyping a dozen numbers.
The treatment coefficients from the claims-only study are the one thing no validation dataset can supply — those stay hand-entered.
Include a propensity-score-calibration validation sub-study in this protocol. Adds Section 7b (variable split, calibration equation, assumptions, R code) to the Word and Markdown exports. Leave unticked and nothing below appears in the protocol.
A small metadata-and-model-output file from RWE Studio (kind: "rwe-studio/psc-handoff"). It carries the two variable lists, the PS scale,
α₀/α₁/α₂ with standard errors, R², and the surrogacy result — never rows or cell values . A payload containing anything row-shaped is refused.
Import
Main (error-prone) data source The large claims database the main analysis runs in. Validation dataset Holds the claims variables and the clinical ones. Validation sample size Validation sample type Internal (a subset of the main study cohort) External (a separate but comparable population) PS scale (must match the outcome model) Probability scale, 0-1 — as in the original method logit(PS), linear predictor — a deviation; justify it Stürmer's method uses the propensity score as a probability and a linear calibration model; keep the default unless you have a reason. Either way the correction is scale-dependent, so both models must use the same scale.
Which covariates does the claims database actually have?
🤖 Let AI classify Reset to rules All claims All clinical
Every covariate ticked in Section 8, plus anything you add below. Click a chip to move a variable between the two propensity scores.
Claims variables go into both scores; clinical-only variables go into the gold-standard score alone; exclude drops
identifiers, outcomes and post-baseline measurements from both.
The keyword rule is only a starting position, and it is wrong often enough to matter. It reads variable names, so it cannot tell a
measurement from a billing code, and it misses instruments it has never seen. Run Let AI classify and check every row it marks AI changed —
a clinical variable left in the claims list quietly collapses the correction to zero, which looks exactly like "no unmeasured confounding".
What the AI button sends. Variable names from Section 8 and from the two boxes below, your codebook notes, the name of the claims
database, and the four research-question fields (population, exposure, comparator, outcome) — all of it verbatim. Nothing else leaves the browser:
no patient data, no rows, no values. Keep patient-identifiable text out of every one of those fields.
Variables only the validation dataset has (one per line) Codebook notes for the AI (Variable | definition) Definitions only — see the note above on what is sent.