Potential Fragmentation of Cohort

[davebridges]

Gap 1: Outcome-specific matching + calendar-time misalignment may fragment the cohort and generate spurious cross-outcome interaction patterns

Labels: reviewer-response analysis methods priority-high

Reviewer summary

The paper's headline narrative — additive effects on glucose, synergistic elevation of liver enzymes, and attenuated/protective effects on blood pressure — is a cross-outcome contrast. That contrast is vulnerable because:

1. Cases span 2000–2025; controls are constrained to at least one visit in 2022–2025. This is a calendar-time mismatch; distributions of assays, reference ranges, treatment practices, and patient mix drift over a 25-year window.
2. Matching was re-done separately per outcome. Each outcome therefore uses a different matched control subset, so cross-outcome comparisons are not made within a common cohort.
3. Standard propensity matching treats early and late exposures as exchangeable, which is known to introduce bias when exposure timing varies widely [1], and temporal mismatches between cases and controls can inflate observed associations and create spurious interactions in case-control designs [2].
4. This concern was raised a second time by a separate reviewer specifically for glucose, who additionally framed it as "exposure-trend bias" requiring case-time-control designs. That citation is a misapplication (case-time-control is for within-person time-varying exposures, not cross-sectional propensity-matched studies) — but the underlying calendar-time concern is the same one.

Risk: the "additive / synergistic / attenuated" trichotomy that drives the Discussion could partly be a sampling artifact rather than biology.

Known constraint

Per project notes: exact calendar-year matching is not feasible with the current data structure (recency bias in the control pool). This rules out the reviewer's first suggestion to match on measurement year directly. Workaround: adjust for admit-date year as a covariate and use a common-cohort design.

What needs to change

A. Manuscript text

Three targeted edits:

1. Methods — justify the design. Add near the MatchIt paragraph:

   "Given the limited availability of certain laboratory measures in this rare disease, we matched separately for each outcome to maximize usable sample size. To address potential calendar-time and cohort-fragmentation biases inherent to this design, we conducted time-aligned and common-cohort sensitivity analyses (see below)."

2. Discussion / Limitations — acknowledge the threat. Add:

   "Controls were required to have at least one clinical visit in 2022–2025, whereas cases spanned 2000–2025. Exact calendar-year matching was infeasible due to the recency of the control pool; we instead adjusted for admit-date year as a covariate and repeated the analysis on a common cohort contributing to all outcomes. Because matching was repeated separately for each outcome, cross-outcome contrasts may still be influenced by residual secular changes or cohort fragmentation; we therefore interpret the integrated pattern (Figure 2; Table 3) cautiously unless supported by time-aligned analyses."

3. Abstract — soften the headline claim. Replace:

   "These findings highlight the need to consider obesity status when evaluating the effects of Cushing's disease."

   with:

   "These findings suggest that obesity status is an important contextual factor when evaluating the effects of Cushing's disease. Because matched sets differed by outcome and calendar time was not aligned between cases and controls, cross-outcome contrasts (additive glycemia; synergistic transaminases; attenuated blood pressure) should be interpreted with caution and were corroborated where indicated by time-aligned, common-cohort sensitivity analyses."

B. Analysis — two new sensitivity analyses (add as rows to Table 4)

Admit-year covariate adjustment (replaces reviewer's year-matching suggestion)

Since exact year matching is infeasible, add admit_year as a model covariate and report the interaction estimate as a new Table 4 row.