Outcome Type Lock

Applying ordinary least squares regression to a proportion bounded between 0 and 1. Using a t-test on count data with many zeros. Fitting a linear model to survival times. Each of these is a model misspecification —and each produces results that look valid but are not. Outcome Type Lock blocks these errors at the source by requiring explicit declaration of what kind of outcome you are analyzing.

STEP 1 —The Pitfall: Model Misspecification by Default

Model misspecification —using a statistical model whose assumptions do not match the data generating process —is systematically underreported in science. The reason is not that researchers are unaware that different outcome types require different models. The reason is that modern analysis software applies the default model (typically OLS or a t-test) unless the researcher actively overrides it. Most don't.

Common misspecification patterns by outcome type:

• Proportion data (0 to 1 or 0 to 100%). OLS regression applied to proportions can predict values below 0 or above 1, produces heteroscedastic residuals, and is inappropriate when values cluster near the boundaries. Logistic or Beta regression is correct.
• Count data. Count data with many zeros or with a Poisson structure is routinely analyzed with t-tests, which assume a continuous normal distribution. Poisson or negative binomial regression correctly handles the integer, non-negative structure.
• Binary outcomes (yes/no, success/failure). Applying a t-test to a binary outcome (coded 0/1) produces a linear probability model with unconstrained predictions. Logistic regression is the appropriate model.
• Compositional data (cell proportions summing to 100%). When components are analyzed independently, the compositional constraint (they sum to 1) creates spurious negative correlations. Dirichlet regression or log-ratio approaches are required.
• Survival/time-to-event data. Applying ANOVA or t-tests to survival times ignores censoring and produces biased estimates. Kaplan-Meier and Cox regression are required.

STEP 2 —Journal Requirement

JAMA's statistical reporting guidelines require that "the statistical methods used, including the software package, version, and any special features or parameters" be described, with an explicit statement of "the outcome data type and the rationale for the chosen model." The Lancet's statistical guidelines state that "the statistical methods should be appropriate for the type of outcome data."

For clinical trials, ICH E9 requires that the Statistical Analysis Plan specify the outcome type and the primary statistical model, along with the rationale. The FDA's statistical guidance for binary endpoints requires logistic regression or exact methods; applying a t-test to a binary endpoint would be cited as a statistical concern in an FDA review.

In practice, a reviewer who sees a t-test applied to data labeled as proportions is likely to request either a correction or a justification. Licklider's Outcome Type Lock prevents this scenario by blocking inappropriate model applications before they happen.

STEP 3 —Licklider's Solution

Input

• An interactive declaration selecting the outcome type from: Continuous (unbounded) / Count (non-negative integer) / Binary (0/1) / Proportion (0 to 1 or 0 to 100%) / Compositional (components summing to 100%) / Ordinal / Survival (time-to-event).
• A second declaration of analysis intent: Descriptive / Inferential / Exploratory / Confirmatory / Publication-ready. Analysis intent governs which guard levels apply and what disclosure is required.
• Both declarations are made before any test or figure is generated. They cannot be changed mid-session without starting a new session.

Output

• Filtered test menu. The available statistical tests and figure types shown in the Licklider interface are filtered based on the declared outcome type. Inappropriate options are hidden, not just warned about.
• Recommended model hierarchy. For each outcome type, Licklider presents a ranked list of appropriate models with brief explanations. The default selection is the most broadly applicable model for the declared type.
• Analysis intent badge. The declared intent (e.g., "Confirmatory") is displayed on every figure panel and export preview, providing a persistent reminder of the analytical context.
• Methods text contribution. The declared outcome type and analysis intent flow directly into the auto-generated Methods paragraph (see STEP 4).

Guard

STEP 4 —Draft Output (Draft / Needs review)

Example: Proportion outcome, confirmatory analysis

The primary outcome was the proportion of CD8+ T cells as a fraction
of total live lymphocytes (continuous proportion, range 0—). The analysis
intent was confirmatory, as pre-specified in the study plan. Beta regression
was used as the primary model, appropriate for proportion outcomes bounded
between 0 and 1 (Ferrari & Cribari-Neto, 2004). Ordinary least squares
regression was not used due to the bounded nature of the outcome
(Licklider Data Contract: outcome type lock ID: lock-3f1a, intent: confirmatory).

Example: Survival outcome, exploratory analysis

The primary outcome was time-to-tumor-reaching-1000-mm^3 (survival outcome,
right-censored). The analysis intent was exploratory. Kaplan-Meier survival
curves were generated and compared using the log-rank test. Cox proportional
hazards regression was used to estimate hazard ratios adjusting for body
weight at baseline. No parametric survival models or t-tests were applied
to survival times (Licklider Data Contract: outcome type lock ID: lock-9b7c,
intent: exploratory).

Related Features