Three downloads

About 204 MB. Every URL, byte size and status is in data/raw/_download_manifest.csv.

The merge, and the trap inside it

read_component() stacks a component across all six cycles and pads any variable absent from a cycle with NA. Then 16 left_joins onto DEMO by SEQN.

N after merge: 59,842. Safe here: SEQN is unique across all six cycles (verified, zero duplicates), and no joined component has more than one row per person.

Step 1 — the four lines

Paper: “The study inclusion criteria were participants aged>18 years with evidence of MASLD.”

⚠ The paper says >, the code says ≥

That is 70 people — and it is not a PDF artefact. The published PDF contains 24 correctly encoded ≥ characters (BMI≥30, WHtR≥0.6, FLI≥60, 1≤PIR≤4). Where the authors meant “at least”, they wrote it. At the age criterion they wrote a plain >.

Counter-evidence: the paper’s Table 1 mean age is 51.0 (16.6) — which is what ≥ 18 gives. Strict > 18 gives 51.4 (16.3). This is a strong lead, not a settled answer.

Step 2 — the Fatty Liver Index

Paper: “A threshold of FLI≥60 was used to determine the presence of hepatic steatosis.”

       exp(L)
FLI = -------- × 100,   L = 0.953·ln(TG) + 0.139·BMI + 0.718·ln(GGT) + 0.053·WC − 15.745
      1 + exp(L)

N: 36,580 → 14,989 (computable) → 6,376 (FLI ≥ 60). Median FLI 85.5 (73.7–94.7).

Reserved codes: there is no global rule

⚠ DESIGN is not missing data

Triglycerides come from the fasting subsample, which carries its own weight WTSAF2YR. The paper never says whether it restricted to that subsample.

• 21,618 of the 36,580 adults fall outside the fasting-weight frame — most were never sampled for it. That is DESIGN.
• 792 are inside the frame and still lack an input. That is MISSINGNESS.

Different problems, different remedies: a weight and a population sentence for the first; a missing-data analysis for the second. Never their sum.

Judgment call 1 — is the glucose cut 100 or 100.9?

The paper sets criterion (ii) in mmol/L. NHANES LBXGLU is in mg/dL. Someone has to convert, and the paper never shows the converted number.

1. Paper: fasting glucose ≥ 5.6 mmol/L
2. Converted: 5.6 × 18.016 = 100.9 mg/dL
3. We coded ≥ 100 — the conventional rounding
4. 212 people have a glucose between 100 and 100.9. Our rule counts them as meeting criterion (ii); a strict reading would not
5. So we re-ran MASLD with ≥ 100.9 and counted who falls out: 0 people

Step 4 — building the alcohol exclusion

Paper: “Individuals with significant alcohol consumption — defined as more than two drinks per day for men and more than one drink per day for women — were excluded.”

The paper says it removed these people, so we built the rule. Getting it right needs two codebook checks that are easy to skip.

1 — “Drinks per day” is frequency × quantity, not ALQ130. ALQ130 is drinks per drinking day. Three drinks every Saturday gives ALQ130 = 3 but an average of 0.4/day. Thresholding ALQ130 on its own removes about 2,000 people, far more than average intake would. So we rebuild it: (ALQ130 × days_per_year) / 365, with ALQ120Q/ALQ120U supplying the frequency (×52 / ×12 / ×1), and ALQ121 replacing that pair in 2017-18.

2 — Strip 777 and 999 before comparing to any number. They mean refused and don’t know. Left in, 9 people here become 777- or 999-drinks-a-day heavy drinkers.

That yields a per-person flag, sig_alcohol. 427 people in the cohort exceed the paper’s threshold. Whether we then drop them is the next slide.

Step 5 — obesity × central adiposity

Paper: “Obesity was defined as BMI≥30kg/m² (≥25 for Asians). High central adiposity was defined as a WHtR≥0.6.”

whtr = BMXWAIST / BMXHT — both in cm, so the units cancel. Five people sit exactly on 0.6 and count as high, matching ≥.

⚠ A cut-point you cannot apply to a third of the cohort

asian comes from RIDRETH3 == 6 — and RIDRETH3 does not exist in 2007-08 or 2009-10. The code recodes the resulting NA to 0, i.e. not Asian, for 2,322 of 6,371 rows (36%).

• Where the variable does exist, the Asian rule moves 121 of 250 Asian participants into the obese arm. This rule has real bite.
• In 2007–2010 the only ethnicity variable is RIDRETH1, whose category 5 is “Other Race — Including Multi-Racial”: it contains Asians but cannot isolate them
• 71 cohort members fall in it; only 17 have BMI in [25, 30) — the window where the rule changes anything

So: at most 17 of 6,371 are misgrouped. The paper states one rule for 2007–2018; the data support it only for 2011–2018.

The paper contradicts itself, and the data settle it

• Methods text: obesity = BMI ≥ 30
• Table 1 and Table 2 column headers: BMI≥25 kg/m² / BMI<25 kg/m²

The headers are wrong, and you can prove it without asking the authors: the paper’s own group III and IV BMI medians are 27.6 and 28.8, sitting in columns labelled “BMI < 25”. Impossible.

Coding to 30 reproduces those medians exactly. Coding to 25 does not.

Finding an internal contradiction and resolving it against the data — rather than picking whichever reading you prefer — is exactly what M2 asks for.

Step 7 — the outcome is not in NHANES

Six fixed-width .dat files from NCHS, joined by SEQN.

Decision: PERMTH_EXM, not PERMTH_INT — the exposure is measured at the MEC exam, so follow-up starts when the exposure was measured. The paper does not say which it used.

⚠ A trap inside our own saved dataset

• README.md says “diseases-of-heart”; reproduction_report.md finding #3 says heart-only
• 04_table2_cox.R:42 quietly defines its own cv_heart and uses that, so the published Table 2 is correct

If you read the analytic file and use cv_death, you will disagree with every table in this repository. Use mortstat == 1 & ucod_leading == 1 until it is fixed.

Step 8 — save, and count what you lost

s0 <- dat                                                    # 59,842
s1 <- filter(s0, age >= 18)                                  # 36,580  ELIGIBILITY
s2 <- filter(s1, !is.na(fli))                                # 14,989  DESIGN + MISSINGNESS
s3 <- filter(s2, steatosis == 1)                             #  6,376  TARGET POPULATION
s4 <- filter(s3, masld == 1)                                 #  6,374  TARGET POPULATION
s5 <- filter(s4, !is.na(group))                              #  6,374  MISSINGNESS
s6 <- filter(s5, eligstat == 1, !is.na(mortstat), !is.na(time_yr))   # 6,371  OUTCOME
s7 <- if (APPLY_ALCOHOL_EXCL) filter(s6, sig_alcohol == 0) else s6   # 6,371  SWITCH

Two files: masld_analytic.rds (6,371 rows) and merged_all.rds (all 59,842, pre-filter, so any step can be re-derived without re-reading the .xpt files).

ELIGIBILITY · DESIGN · MISSINGNESS · TARGET POPULATION · OUTCOME ASCERTAINMENT — if you cannot name a row’s type, you do not yet know what the row means.

Did it reproduce?

Reproduced HRs: unweighted, full 6,371-row file (Model 2 on its complete cases). Published: the paper’s Table 2.

Yes, closely — apart from a residual gap of 71 that is not yet resolved.

Read crudely, Group IV has the poorest survival, and the paper reports that. Under Model 2, Groups III and IV have similar HRs against Group I (ours: 3.00 vs 2.91). The replication succeeds; the disagreement is about which contrast supports the conclusion.

For P1

1. Work the loop, in order: paper sentence → operational concept → NHANES variable → component → that cycle’s codebook → raw .xpt → your code → cohort consequence
2. Write four lines per step: the quoted sentence, the component and variables, the decision and its reason, the N afterwards
3. Type every funnel row. DESIGN, MISSINGNESS and TARGET POPULATION are not interchangeable and do not add up
4. Two decisions here are still open (age, alcohol). Reproducing a number is not the goal — being able to say why yours differs is

One exercise

Open data/derived/merged_all.rds and answer, with code:

1. How many participants are exactly 18? What happens to N, deaths, and the four group sizes if you use age > 18?
2. How many people would the alcohol exclusion remove? Does applying it move you toward or away from the paper’s 6,300?
3. Pick one variable from Step 0’s availability table. Which cycles is it missing from, and what does the NA-padding do to the statistic it feeds?

Write four lines for each: what the paper says · what you did · why · what it cost.