Long Response
Use this to pre-register every predictor feature: its source, timing, allowed processing, missing-data handling, and eligibility for inclusion. This reduces analysis flexibility and supports auditability.
| Field | What to write (pre-specify) | Checks / constraints |
|---|---|---|
| Predictor ID | Unique stable identifier (e.g., PRED_001) | Must be immutable across protocol revisions |
| Predictor name (human-readable) | Short descriptive label | Must match lab/CRF naming conventions used in study files |
| Predictor domain | Clinical, lab, imaging-derived, genomics, demographics, prior history, etc. | Allows downstream reporting and governance |
| Source system | CRF, EHR, lab instrument, device, imaging pipeline, sequencer pipeline, manual adjudication, etc. | Must list data lineage reference documents |
| Measurement timepoint | Exact relative time to index event (e.g., baseline, week 12) and clock definition | Must specify “allowed window” for taking the value |
| Population eligibility | Who can contribute this predictor (inclusion/exclusion criteria) | Prevents post hoc conditioning |
| Raw variable definition | Exact raw field(s) used (units, coding, instrument-specific labels) | State units and coding scheme; forbid unit changes without protocol amendment |
| Transformation rules | Allowed transforms (e.g., log, scaling, categorization cutpoints) and parameters | All parameters fixed; no data-dependent tuning |
| Derived features | Formulas for composites/ratios/aggregates (symbol-level) | Specify whether numerator/denominator require nonzero constraints |
| Missingness handling | Pre-specified strategy: indicator, exclusion rule, imputation method family, or “no imputation” | If imputing, specify imputation target, allowed predictors in the imputation model, and random seed policy |
| Outlier rules | Definition of outlier, capping/winsorization rules, or QC rejection rules | Must be based on pre-defined instrument/QC criteria |
| QC checks | Value ranges, consistency checks across visits, duplicate resolution rule | List hard-fail vs soft-flag thresholds |
| Feature inclusion criterion | Rule for whether predictor enters final model candidate set | Prevents selection based on outcome information |
| Admissible data versions | Protocol version & data freeze identifier(s) that define the analysis dataset | Supports reproducibility |
| Primary use (training vs adjudication) | Whether used for model training, calibration, baseline risk score, stratification, or other | Clarify any separation constraints |
| Analysis time window constraints | Prevent leakage: define that predictors must occur strictly before endpoint assessment | Explicit ordering rule |
| Reporting requirement | What summary statistics will be shown for this predictor | Specify metrics and stratification variables |
| Reproducibility note | Dataset construction query/logic statement reference | Must link to immutable code version or spec text |
| Endpoint leakage audit | Rule describing how you will verify no endpoint information is included in predictors | Example checks: time stamps, forbidden fields, join constraints |
Use this to predefine how endpoints are determined from source evidence (events, imaging, lab triggers, clinician notes) and how to resolve disagreement. Endpoint adjudication should be designed to limit observer bias and ensure consistent decision rules, consistent with GCP expectations for reliable data handling .
| Section | Spec fields (what to write) | Decision/quality constraints |
|---|---|---|
| Endpoint identity | Endpoint name, type (binary/time-to-event/continuous), primary vs secondary label | Must match protocol endpoint naming; no renaming without amendment |
| Adjudication triggers | What source signals initiate a review (suspected event codes, imaging flags, lab thresholds, AE reports) | Define trigger list explicitly; avoid free-text triggers |
| Evidence package | Allowed documents for adjudication (imaging reports, raw images, pathology, operative notes, lab results, etc.) | Predefine admissible vs excluded evidence types |
| Inclusion/exclusion window | Time boundaries for accepting evidence relative to index/start date | Specify timezone/clock definition |
| Case classification rules | Criteria mapping evidence to: definite / probable / possible (or analogous categories), with explicit logic | Use deterministic rules; document any clinical judgment allowed |
| Outcome date determination | Rule for endpoint onset date (first day meeting criteria, imaging date, symptom onset date) | Date precedence order must be defined |
| Reviewer blinding | What reviewers can/cannot see (treatment assignment, investigator site, prior outcomes) | Predefine blinding status and exceptions |
| Number of reviewers | Single vs dual review; if dual, specify independent review vs consensus | Specify minimum review quorum |
| Disagreement handling | Conflict resolution algorithm (third adjudicator, majority rule, arbitration panel) | Specify how ties are broken |
| Adjudication adjudicator training | Calibration process and competency checks (without using outcome labels from participants outside allowed training scope) | Predefine retraining triggers |
| Quality metrics | Planned reliability reporting (e.g., agreement statistics) and how they will be computed | Specify metrics and denominators; avoid outcome-informed tuning |
| Resolution documentation | Required documentation fields: evidence summary, rule(s) used, category selection, reviewer rationale | Rationale completeness criteria |
| Handling missing evidence | Rule for when evidence is incomplete: treat as non-case, probable, or request additional info | Define max time to obtain missing evidence |
| Adjudication data outputs | Final adjudicated endpoint label, plus metadata (category, date, evidence completeness flag) | Specify variable schema for analysis datasets |
| Estimand alignment | How endpoint adjudication decisions correspond to the pre-specified estimand (e.g., event under ideal/real-world definition) | State any sensitivity plans if adjudication uncertainty exists |
| Sensitivity analyses | Planned alternatives: include “probable” vs only “definite”; censoring or reclassification rules | Must be prespecified to avoid post hoc endpoint redefinition |
Keep Exploring
DiscussionKnow what changed, what holds up, and what remains uncertain. Every Friday. No ads.
Explore by Goal
Ask a Follow-Up
Science Movie