Potential calendar year drift in ALT measures.

[davebridges]

Gap 6: Calendar drift and sex differences in ALT/AST reference intervals not addressed — z-score sensitivity analysis

Labels: reviewer-response analysis priority-low

Reviewer summary

ALT and AST are modeled as raw values. Reference intervals and assay methods for transaminases have drifted over the 2000–2025 cohort window (e.g., ALT ULN was revised from ~40 U/L to 33/25 U/L for men/women per ACG 2017 guidelines). Reviewer argues that without ULN normalization or z-score standardization, the obesity × Cushing's interaction could conflate biology with assay/calendar variation. Proposes either (Option 1) sex- and year-specific robust z-scores, or (Option 2) full LIS metadata harmonization to laboratory-specific ULNs with log(value/ULN) modeling.

My critique of the critique

Legitimate concerns:

• Calendar drift in ALT reference intervals is real. If cases span 2000–2025 and controls are concentrated in 2022–2025, a reference-interval shift propagates into the raw-value comparison. This is mechanistically a variant of the Gap 1 calendar-time concern.
• ALT has well-established sex differences in reference intervals (men's ULN ~30–40% higher than women's). The primary analysis adjusts for gender as a covariate, which handles mean differences, but won't absorb an obesity-sex interaction structure if one exists.

Overstated:

• "ALT/AST assays and reference intervals vary significantly across analyzers" is a multi-institutional concern. This is single-system Michigan Medicine data with centralized clinical pathology. Inter-assay variability within one system over time is much smaller than the reviewer's framing implies. This is the primary defense in the response letter.
• Option 2 (full LIS metadata harmonization via LOINC + analyzer IDs + per-period reference intervals + log(value/ULN) modeling) is a disproportionate engineering effort for marginal payoff. Skip unless a peer reviewer specifically demands it.

Where this overlaps existing work:

• The reviewer's secondary points (hepatic comorbidity ICD adjustment, hepatotoxic medication adjustment, naive-restriction cohort) are already covered by Gap 4. Cross-reference rather than duplicate.

What needs to change

A. Prevalence descriptive — sanity-check the calendar-drift premise

Before any modeling, check how cases are distributed across calendar years and plot raw ALT/AST in the control arm over time. If control ALT is roughly stable across 2022–2025 (which it should be, since all controls are concentrated there), there's little room for calendar drift to have affected the control arm. Cases span 25 years, so a year-of-draw distribution of cases is the key diagnostic.

library(dplyr)
library(ggplot2)
library(lubridate)

alt_year_dist <- alt.data |>
  mutate(lab_year = year(DeID_AdmitDate)) |>
  count(Cushings, lab_year) |>
  ggplot(aes(x = lab_year, y = n, fill = factor(Cushings))) +
  geom_col(position = "dodge") +
  labs(x = "Year of ALT draw", y = "n",
       fill = "Cushing's",
       title = "Distribution of ALT measurements by year") +
  theme_minimal()

# Control-only temporal trend (should be flat if no calendar drift concern)
alt_control_trend <- alt.data |>
  filter(Cushings == 0) |>
  mutate(lab_year = year(DeID_AdmitDate)) |>
  group_by(lab_year) |>
  summarise(median_alt = median(value, na.rm = TRUE),
            n = n())

If controls are all within 2022–2025 and show stable median ALT, and cases show a wide distribution, the asymmetry is the calendar-time concern from Gap 1, not a separate assay-drift issue. Frame accordingly in the response.

B. Sex- and year-specific robust z-scores (Reviewer's Option 1)

Compute z-scores using the control arm as the reference distribution, stratified by sex and — where feasible — calendar year. Given control concentration in 2022–2025, year-stratification of the reference distribution will be limited; sex-stratification is straightforward.

library(dplyr)

# Robust center/scale from controls, by sex
alt_reference <- alt.data |>
  filter(Cushings == 0) |>
  group_by(GenderCode) |>
  summarise(
    alt_median = median(value, na.rm = TRUE),
    alt_mad    = mad(value, na.rm = TRUE),
    .groups    = "drop"
  )

alt_z <- alt.data |>
  left_join(alt_reference, by = "GenderCode") |>
  mutate(alt_z = (value - alt_median) / alt_mad)

# Refit the primary interaction model on z-scored outcome
lm.alt.z <- lm(
  alt_z ~ GenderCode + RaceEthnicity + Cushings * Obesity,
  data = alt_z
)

library(broom)
tidy(lm.alt.z, conf.int = TRUE) |>
  filter(term == "Cushings:ObesityObese")

Add a Table 4 row: "Outcome as sex-stratified robust z-score" for ALT and AST.

Interpretation: if the interaction estimate in z-score units corresponds to a clinically meaningful synergy (e.g., >1 SD above additivity), the raw-units claim is robust. If it collapses, the raw-units result was partly assay/calendar artifact.

C. (Optional) Year-specific reference stratification if case distribution warrants it

If the descriptive from (A) shows cases heavily concentrated in pre-2015 years where ULN was higher, add a second z-score variant using year-specific reference means from controls (if feasible given control year distribution) or from published literature ULN values per era. Defer unless (A) flags this as a concrete problem.

D. Manuscript text

Methods (add one sentence):

"As a sensitivity analysis, ALT and AST were re-expressed as sex-stratified robust z-scores using the matched control distribution as reference, and interaction models were refit on the z-scored outcome (Table 4)."

Results (add one sentence to liver section, conditional on result):

"The obesity × Cushing's interaction for ALT remained significant when the outcome was expressed as sex-stratified robust z-scores (Table 4), supporting robustness against potential assay or reference-interval drift over the 25-year case window."

Limitations (add one sentence):

"Transaminase values were modeled in raw units in the primary analysis; results were similar when the outcome was expressed as sex-stratified robust z-scores. Full ULN harmonization across analyzers and calendar periods was not performed, but within-system standardization at Michigan Medicine limits the scope of inter-assay variability."

Acceptance criteria

• Descriptive: year-of-draw distribution for ALT and AST (by Cushing's status) reported in the supplement.
• Descriptive: control-arm temporal stability of ALT/AST median verified (flat = assay drift is not materially affecting the control comparison).
• Table 4: new row "Outcome as sex-stratified robust z-score" for ALT and AST.
• Methods sentence added.
• Results sentence added (conditional on z-score result).
• Limitations sentence added.
• Skip full LIS metadata harmonization (Reviewer's Option 2). Document in the response letter: "We performed a sex-stratified z-score sensitivity within our single-system data, which is the relevant level of standardization for a within-institution analysis. Full LOINC/analyzer-specific ULN harmonization would be appropriate for a multi-institutional pooled analysis but is not necessary here."
• Cross-reference Gap 4 for hepatic comorbidity and hepatotoxic medication adjustments (the reviewer raised these too, but they're already being addressed there).

Notes / open questions

• Pediatric/adult reference intervals. You've already restricted to ages 18–75 so pediatric ranges aren't relevant.
• Pregnancy. If any control ALT/AST values are from pregnant women, reference ranges differ. Probably a small fraction of the control arm; worth noting but not acting on unless the descriptive flags it.
• Bayesian angle. Given the small n for ALT/AST (84), reporting the Bayesian interaction estimate in z-score units alongside the raw-units estimate (as proposed in Gap 4) gives the cleanest answer to "is this robust to measurement-model choice." Runs on the same brms infrastructure.

References

1. Reviewer's refs on inter-assay ALT variability — these describe multi-institutional or multi-analyzer settings. Cite only if acknowledging the limitation; do not cite as if they directly apply to single-system data.
2. ACG 2017 ALT reference interval update (men 33 U/L, women 25 U/L) — relevant for discussing the calendar-drift concern if needed.

(Locate full cites before response letter.)