Group: Demonstration (instructor’s worked example) Members: AB, CD, EF Date: 22 September 2026
Kueh MTW, Chew NWS, et al. Body weight categories and fat distribution in relation to all-cause mortality among adults with metabolic dysfunction-associated steatotic liver disease. BMJ Open 2026;16:e113719. doi:10.1136/bmjopen-2025-113719
Open access (CC BY-NC). Built entirely on public NHANES data with NCHS linked mortality.
Our estimands, fixed for the term:
Scientific (locked) estimand — IV vs III. Among adults meeting FLI-based MASLD criteria, do non-obese adults with high central adiposity have higher all-cause mortality than non-obese adults with low central adiposity — and how robust is that to cohort construction, confounding, effect heterogeneity, survey design, and incomplete observation?
Replication estimand — IV vs I. The paper references its contrasts to Group I (obese, low central adiposity). We reproduce that, because reproducing it is what Replicate means.
Both are contrasts within one fitted model; re-levelling is reparameterization, not reanalysis. Every estimate we report says which estimand it belongs to. The locked contrast is the one the term interrogates, because it holds body-mass category fixed and so isolates central adiposity.
Each claim below is confirmed with a specific pointer: a file name, a variable name, or a number we counted ourselves.
Six NHANES cycles, all public, no data-use agreement required:
| Cycle | Suffix | Cycle | Suffix |
|---|---|---|---|
| 2007–2008 | _E |
2013–2014 | _H |
| 2009–2010 | _F |
2015–2016 | _I |
| 2011–2012 | _G |
2017–2018 | _J |
Download pattern:
https://wwwn.cdc.gov/Nchs/Data/Nhanes/Public/<year1>/DataFiles/<COMPONENT>_<suffix>.XPT
Components we need: DEMO, BMX, BPX, TRIGLY, BIOPRO, GHB, GLU, HDL,
TCHOL, ALQ, SMQ, BPQ, DIQ, MCQ, KIQ_U, PAQ, ALB_CR, RXQ_RX.
Mortality linkage: NHANES_<cycle>_MORT_2019_PUBLIC.dat from the NCHS linked mortality
files (fixed-width, public-use).
Merge key: SEQN, present in every file including the mortality .dat.
We downloaded all six cycles and merged them before writing this memo — the funnel in criterion 4 is counted from our own merged file, not read off the paper.
Exposure — a 2×2 phenotype, not a continuous variable:
BMXBMI ≥ 30 (≥ 25 if RIDRETH3 = Asian)BMXWAIST / BMXHT ≥ 0.6giving Groups I (obese, low central), II (obese, high central), III (non-obese, low), IV (non-obese, high). Group IV is the paper’s headline.
Outcome — all-cause mortality: MORTSTAT = 1, follow-up time PERMTH_EXM, censored
31 December 2019.
Eligibility — MASLD by Fatty Liver Index ≥ 60 plus ≥ 1 cardiometabolic criterion.
FLI is computed from LBXTR (triglycerides), LBXSGTSI (GGT), BMXBMI and BMXWAIST.
Measured and available on every cycle:
| Confounder | Variable(s) |
|---|---|
| Age | RIDAGEYR |
| Sex | RIAGENDR |
| Race / ethnicity | RIDRETH3 |
| Income-to-poverty ratio | INDFMPIR |
| Smoking | SMQ020, SMQ040 |
| Type 2 diabetes | DIQ010 + LBXGH (HbA1c) |
| Hypertension | BPQ020 + measured BPXSY1–BPXSY4 |
| Kidney function | LBXSCR (CKD-EPI 2021), ALB_CR (ACR) |
| Medication classes | RXQ_RX |
A judgment we are flagging now, not at M1: several of these — T2DM, hypertension, CKD, prior MI — could be descendants or intermediates under plausible causal structures, but their causal roles and temporal ordering are not established here. The paper adjusts for all of them. We will reproduce that adjustment set because replication requires it, and we expect to argue at M1 that the crude-to-adjusted attenuation cannot be attributed to confounding control without a prespecified causal model supporting the adjustment set. Temporal ordering supports adjusting for age, sex, race, income and smoking without reservation.
Counted from our merged file. This is the chain to the paper’s own analytic file (the full analytic file), which requires a computable FLI but not a positive fasting weight, so the fasting-subsample step (DESIGN) and missing FLI inputs (MISSINGNESS) are fused into one row here:
| Step | N |
|---|---|
| All NHANES 2007–2018 records | 59,842 |
| Age ≥ 18 | 36,580 |
| FLI computable (TG, GGT, BMI, waist present) | 14,989 |
| Steatosis, FLI ≥ 60 | 6,376 |
| MASLD (≥ 1 cardiometabolic criterion) | 6,374 |
| Mortality-eligible and linked | 6,371 |
Criterion 6 keeps the two steps apart by counting inside the fasting subsample. Counted that way, the chain runs 14,962 adults sampled, 14,170 with a computable FLI, and ends at 6,048 mortality-eligible and linked records with a valid fasting weight. The two chains are different row sets, not a discrepancy.
On the 6,371 rows, group sizes are 391 / 4,585 / 787 / 608; 586 deaths.
That is ample for the main analysis. The caveat is Group I, the reference group, at n = 391. A sex-stratified interaction analysis (P3) splits it further, so stratum-specific estimates in Group I will be imprecise and we should expect wide confidence intervals rather than treat them as null findings. We are proceeding, with that stated in advance.
All present in DEMO and TRIGLY:
WTMEC2YR — MEC examination weightWTSAF2YR — fasting subsample weight, which is the correct one here, because FLI
requires fasting triglyceridesSDMVSTRA — stratumSDMVPSU — primary sampling unitSDDSRVYR — cycle indicatorBecause we pool six two-year cycles, the weight must be divided by 6. The published analysis appears to be unweighted, which is one of the improvements we expect to propose.
We claim only that the paper can support an incomplete-observation analysis. We do not claim the problem is large, because we have not yet measured whether it matters.
First, what does not count. Four kinds of loss shrink this cohort, and none of them satisfies criterion 6 on its own, because P5 does not address them with the methods this course teaches. We must be able to tell each apart from missingness, and at P5 we report it and scope the conclusion around it.
WTSAF2YR exists precisely to account for that selection stage and its nonresponse.
Their absence outside the subsample is a design feature, not by itself a missing-data
problem or evidence of MNAR — and the syllabus is explicit that designed subsampling
alone does not satisfy this criterion. Inference does remain conditional on the
weighting assumptions.What does count, measured inside the fasting frame:
WTSAF2YR and by a sentence naming who our
estimate is about, not by anything we would impute.We are not arguing from magnitude. 5.3% is not self-evidently serious, and it may turn out to change nothing. The criterion asks whether there is a real incomplete- observation issue that requires explicit assumptions and can support a substantive sensitivity analysis. There is, and whether it matters is P5’s job to measure rather than ours to presume.
All six criteria are met. We are locking this paper.
The thing we already expect to be interesting is that the FLI eligibility index shares its inputs (BMI and waist) with the exposure definition — cohort selection and exposure definition are built from the same measurements. That is a candidate threat we will develop across P1–P5. Which threat actually dominates is a judgment for M1, once all five increments have been run; naming it now would be guessing.
Contribution. AB: data download, merge, funnel counts. CD: variable mapping for criteria 2–3, confounder table, temporal-ordering argument. EF: design variables, missingness quantification, memo assembly.
External feedback. None received.
AI Use Statement. Tool and version: Claude (Opus 5), used in RStudio and in a browser. Used for: drafting the download loop across the six cycle suffixes, and for checking the NHANES variable names for GGT and creatinine against the codebooks. Not used for: the six eligibility judgments. Each was verified by us against the actual codebooks and the actual downloaded files, and every N in criterion 4 was counted from our own merged data. An AI’s guess about what NHANES contains is not a feasibility check.