Predictive Etiologic Research vs Clinical Prediction Models: Why TRIPOD Re-labels Your “Etiology Model” as a Prognosis CPM
- Mayta
- 3 days ago
- 4 min read
Abstract
In CECS language, etiologic research starts with people who have not yet experienced the target event, while prognostic research starts with people who already have a defining event (e.g., diagnosis) and asks what happens next. However, TRIPOD classifies multivariable prediction models differently: the key split is whether the model predicts a condition now (diagnosis) or predicts an outcome later (prognosis). This creates an apparent “label switch”: a non-causal etiologic risk score (predicting incident disease in currently disease-free people) becomes, under TRIPOD, a prognosis-type CPM because it predicts a future outcome from predictors available at the point of use. This article clarifies that this is not a contradiction—it is two taxonomies using different axes.
1. Terms: Three Different Things People Mix Up
Before the TRIPOD issue, separate these three “non-causal” activities:
Exploratory etiologic research (non-causal association finding)
Goal: identify candidate factors associated with an event (often many X’s, reported as OR/RR/HR).
Output: a list/table of associations; not necessarily a deployable tool.
Predictive etiologic research (non-causal prediction in an etiologic domain)
Goal: predict incident outcome among people who do not yet have the event at baseline.
Output: may still be “just predictors,” but often moves toward a multivariable equation/score because prediction needs combined information.
Clinical Prediction Model (CPM)
Goal: provide an individualized absolute risk/probability (or classification) for a single person at a defined “point of prediction.”
Output: final equation, score, nomogram, calculator, or algorithm—something usable at the bedside or system level.
Exploratory etiologic (non-causal) is not the same as predictive etiologic (non-causal). They share “non-causal intent,” but differ in objective and deliverables.
2. CECS DEPTh Taxonomy Uses “Domain at Time Zero.”
CECS distinguishes etiologic vs prognostic primarily by domain:
Etiologic domain (etiognostic): at time zero, the person has not experienced the target event yet.
Prognostic domain: at time zero, the person already has experienced the defining event (diagnosis/index event), and we predict what happens next.
3. TRIPOD Uses a Different Axis: “Now vs Later.”
TRIPOD’s practical split is about when Y exists, relative to when predictors X are available:
Diagnostic CPM: predicts the current presence/absence of a condition (X and Y are essentially at the same time window).
Prognostic CPM: predicts a future outcome after follow-up (X at prediction time; Y occurs later).
This is why I say your etiologic prediction model “turns into” a prognosis CPM under TRIPOD: TRIPOD doesn’t care whether baseline patients are “diseased” vs “not diseased” in the CECS sense; it cares whether the outcome is future vs current.
4. The Key Reconciliation: Two Taxonomies, Two Labels, Same Study
Consider a model that predicts incident diabetes in currently non-diabetic adults using baseline labs and BMI.
CECS domain label: etiologic domain (people do not have the event at time zero; you are studying occurrence).
TRIPOD model type label: prognostic CPM (you are predicting a future outcome using predictors available now).
These labels can coexist because they answer different classification questions.
A useful one-line reconciliation:
DEPTh calls it etiologic because of who you start with.
TRIPOD calls it prognosis because of when the outcome happens.
5. Where PROGRESS III Fits (and Where It Doesn’t)
PROGRESS III (“prognostic model research”) is anchored in prognosis research—classically, people with a condition followed forward. [5]But methodologically, the modeling workflow (multivariable prediction, validation, calibration/discrimination) is shared across “future-risk” CPMs whether the baseline cohort is diseased or disease-free. That is why prediction-model guidance in practice often treats “risk prediction” and “prognostic prediction” similarly once you commit to building a tool.
So, my operational point is pragmatic:
The moment you produce a model/score intended for individual future risk, you should behave like CPM research (performance, validation, reporting), even if your cohort started in an etiologic domain.
6. The Moment a Non-Causal Etiologic Study Becomes a CPM Study
A non-causal etiologic study becomes CPM research when the deliverable shifts from “associations” to “individual prediction tool.”
Not yet CPM (typical exploratory etiologic output):
“X1, X2, X3 are associated with Y” with effect estimates.
Now CPM (TRIPOD-relevant output):
A final multivariable equation or a points-based score that gives a person’s predicted probability of Y over a time horizon (e.g., 2-year risk).
Once you cross that line, you must focus on CPM requirements: clearly defined prediction time, predictor availability at that time, validation, and performance assessment (not only regression coefficients).
7. What Changes in Analysis When You Move to CPM Thinking
If your goal is prediction (CPM), the analysis is no longer “Which X is significant?” but “Does the model predict well and will it transport?”
Typical CPM analysis focus:
Discrimination (can the model separate higher vs lower risk?)
Calibration (are predicted risks numerically correct?)
Internal validation (optimism control via resampling, cross-validation, bootstrapping)
External validation (temporal/geographic/domain), and recalibration if needed.
In contrast, etiologic association work emphasizes estimating relationships and dealing with bias/confounding depending on whether causal intent exists.
8. Practical Decision Rules You Can Use in Meetings
Use these three questions to classify any project:
Is Y “now” or “later”?
Now → Diagnostic CPM logic
Later → Prognostic CPM logic (TRIPOD’s bucket), regardless of DEPTh domain
Is the output a tool that returns an individual risk/probability?
Yes → CPM research (TRIPOD-style reporting/performance/validation)
No → likely factor/association research
At time zero, does the cohort already have the defining event?
No → etiologic domain (DEPTh)
Yes → prognostic domain (DEPTh)
This triad avoids the “label fight” by letting both taxonomies be true simultaneously.
Key takeaways
Exploratory etiologic (non-causal) and predictive etiologic (non-causal) are different: one hunts associations; the other aims to predict.
TRIPOD’s main split is “predict now” (diagnosis) vs “predict later” (prognosis), not “etiology vs prognosis domain.”
An etiologic-domain risk score for incident disease is often a “prognosis CPM” under TRIPOD because the outcome is future.
Once you output a score/model for individual future risk, you should adopt CPM evaluation: discrimination, calibration, and validation.
Comments