A five-minute NHANES mental model

Cycle
One two-year survey wave, such as 2011–2012

Component
A topic-specific file, such as demographics, body measurements or laboratory tests

Variable
A column inside a component, such as age or waist circumference

SEQN
The participant identifier used to join files

This reproduction combines six cycles from 2007–2008 through 2017–2018.

Stack cycles first, then join components

For each component:

1. read the same variables from all six cycles
2. stack the cycles into one long file
3. join components by SEQN

demo_all <- bind_rows(demo_E, demo_F, demo_G, demo_H, demo_I, demo_J)
dat <- left_join(demo_all, bmx_all, by = "SEQN")

The join choice defines which people can remain in the analysis.

An absent variable can look like nonresponse

Padding an unavailable variable with NA allows the code to run, but two situations then look identical:

• the question did not exist in that cycle
• the question existed and the participant has no value

Example: RIDRETH3, which identifies non-Hispanic Asian participants, begins in 2011–2012. Earlier cycles cannot supply the same classification.

Check variable availability by cycle before interpreting an NA.

The age sentence leaves a real decision

The paper states:

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

The reproduction uses:

filter(age >= 18)

Because NHANES age is recorded in whole years:

• age > 18 means age 19 and older
• age >= 18 includes people aged exactly 18

The age sensitivity gives conflicting evidence

Seventy reproduced participants are exactly 18 years old.

All reproduction rows are the full reconstruction row set (not the fasting-weight domain).

The strict rule matches N and deaths. The inclusive rule matches the age distribution.

Exclusions are scientific claims

• Every filter answers “who is this study about?” — so every filter needs a stated rationale, not just a line of code
• The reproducibility contract, in one rule:

# Paper, Methods: eligibility stated as "aged > 18 years"
d <- filter(d, age >= 18)     # NOT what the sentence says — deliberate, and recorded

• A comment naming the sentence you are implementing lets an auditor trace every filter() back to the paper, and makes a divergence visible instead of invisible

Same variable, three studies

Week 2 put three age rules on one slide:

NHANES age is an integer, so one variable yields three different studies.

The discipline is not to pick the right one. It is to record which one you used, so a reader can undo it. A reproduction that quietly adopts a different rule has destroyed its own evidence.

The paper constructs MASLD from measurements

In this study, a participant must have:

1. hepatic steatosis, defined using a Fatty Liver Index (FLI) of at least 60
2. metabolic dysfunction, defined by at least one of five criteria

NHANES does not provide one MASLD diagnosis column for these six cycles. The analyst constructs the case definition.

2017–2018 includes liver transient elastography, but no comparable liver-imaging measure spans all six cycles used by this paper.

The Fatty Liver Index uses four inputs

\[ \mathrm{FLI}=100\frac{\exp(L)}{1+\exp(L)} \]

\[ L=0.953\ln(TG)+0.139(BMI)+0.718\ln(GGT)+0.053(waist)-15.745 \]

One participant can sit on either side of the FLI threshold

Hold BMI at 28.5 kg/m², GGT at 30 U/L and waist at 99 cm:

The cut occurs at approximately 112.4 mg/dL for this profile.

A continuous score becomes a hard inclusion boundary.

Fasting design and item missingness are different

Triglycerides come from a fasting laboratory subsample with its own weight, WTSAF2YR.

Assignment to the fasting subsample is design-based. Participation, fasting compliance and laboratory nonresponse can still affect who has a positive weight and complete data.

Two analytic row sets answer different questions

The 323-person difference is not an arithmetic error.

Every reported N should name its row set.

Metabolic dysfunction requires at least one of five criteria

The paper also mentions treatment targeting HDL. The supplied implementation has no separate variable for that clause.

Two participants meet none of the five criteria

Both have FLI-defined steatosis, but no operational cardiometabolic criterion:

Therefore, the MASLD step changes 6,376 to 6,374 in the full reconstruction.

An unresolved criterion contributes no positive count. Check whether each NA is a structural skip, cycle absence, refusal or genuine missingness before interpreting it.

The alcohol sentence is not executable as written

The paper excludes consumption above two drinks per day for men and one for women.

The sentence leaves three questions:

• Does “per day” mean per drinking day or an average over all days?
• Which variables supply frequency and quantity?
• What should happen when an alcohol response is missing?

ALQ130 records drinks on drinking days, not average drinks per calendar day.

Three implementations produce three cohorts

Counts are the full reconstruction row set.

Among the tested alcohol implementations, switching the exclusion off comes closest to the published cohort. It does not prove what the authors implemented.

The remaining 71-person gap is unresolved: a strict age >18 reading gives 6,301 with 585 deaths, but the published mean age matches ≥18.

Reserved codes can look like extreme measurements

For ALQ130:

qty <- ifelse(ALQ130 %in% c(777, 999), NA, ALQ130)

Leaving these codes in reclassifies nine refused or don’t-know answers as extremely heavy drinkers.

Missing-value codes are variable-specific. Always check the codebook.

The analysis switch records the decision

APPLY_ALCOHOL_EXCL <- FALSE

The pipeline still:

• constructs the alcohol variable
• saves the heavy-drinking flag
• reports the alternative cohort size
• writes the switch setting into the funnel

A visible switch preserves the unresolved question for the next analyst.

The exposure has four BMI and waist phenotypes

Obesity means BMI ≥30 kg/m², or ≥25 kg/m² for Asian participants. High central adiposity means waist-to-height ratio (WHtR) ≥0.6.

Counts are the full reconstruction row set (6,371).

The exposure shares BMI and waist inputs with FLI. M1 asks what that structural dependence does to interpretation.

Asian status is unavailable in the first two cycles

RIDRETH3 supplies a separate non-Hispanic Asian category only from 2011–2012 onward.

In 2007–2010:

• the available race variable combines “other race” and multiracial participants
• Asian participants cannot be isolated reliably
• the supplied code treats missing Asian status as “not Asian”

Seventeen early-cycle participants lie in the BMI range where Asian status could change their phenotype. Seventeen is an upper bound, not a known error count.

The paper’s BMI labels conflict with its Methods

Methods: obesity uses BMI ≥30 kg/m², with the Asian-specific cut at 25.

Table headers: groups are labelled BMI ≥25 and BMI <25.

Published non-obesity medians: 27.6 and 28.8 kg/m².

Those medians cannot belong to groups defined entirely by BMI <25. Coding the general cut at 30 reproduces all four published group medians.

The table labels appear to be wrong; the Methods and reproduced distributions agree.

Cohort construction exposes under-specified covariates

The paper says that comorbidity definitions were “previously described,” without giving the full operational rules in this article.

The reproduction must choose definitions for hypertension, diabetes, dyslipidaemia, chronic kidney disease and medication use.

Do not choose a definition because it makes Table 1 match. Record the rule first, then report the residual difference.

This problem returns in P2, when Table 1 and confounding variables are constructed.

Mortality comes from a separate linked file

The NCHS linked mortality files supply:

• mortality status
• underlying cause category
• follow-up time through 31 December 2019

The reproduction uses PERMTH_EXM, months from the mobile examination centre visit, because BMI and waist were measured at that examination.

The paper does not state whether it used examination-based or interview-based follow-up.

The fasting-weight funnel names the cause of every reduction

The full reconstruction follows a different path

This path best reproduces the paper’s ordinary analyses. It fuses fasting design and data availability at the FLI step, so it should not replace the design-aware funnel.

Why you never tune the cohort to hit the paper’s N

• The temptation: quietly flip a threshold until N matches the abstract
• That is fitting the process to the answer — you have chosen your method using the result, and the method no longer carries evidence
• One sentence in this paper moves N by 2,063. So an exact match is roughly as likely to mean you guessed their ambiguity correctly as you implemented it correctly

A documented N you can defend beats a matched N you cannot.

A short decisions ledger preserves the audit trail

A decision without a quantified consequence is not fully documented.

Thursday’s L2 builds the same object from another paper

The benzodiazepine, opioid and mortality lab repeats the same mechanics:

dat.mort   <- rbind(mortality_cycles)
dat.nhanes <- rbind(nhanes_cycles)
dat <- merge(dat.mort, dat.nhanes, by = "SEQN", all = TRUE)

Then it applies a numbered chain of eligibility rules.

Your task in L2: write N remaining, N removed and the reason for every step.

If you can only do it on MASLD, you have not learned it.

One more thing to notice on Thursday

L2’s exposure has four levels, and the reference category is “Neither”:

dat6 <- subset(dat5, benzodiazepines == "Yes" | ssris == "Yes" | opioids == "Yes")
dat6$exposure <- relevel(dat6$exposure, ref = "Neither")

• Eligibility required at least one of three drug classes. So “Neither” means neither benzodiazepines nor opioids — everyone in it is an SSRI user
• The reference group is an active comparator, not an unexposed population
• Exactly the shape of our Group I: every hazard ratio in the MASLD paper is measured against a group that is obese

A ratio without its referent means nothing — and in both studies the referent is not “healthy”.

The same cohort work returns in P1 through P5

M1 asks which threat matters most. M2 asks what reproduced and what did not. M3 asks you to defend the decisions person by person.

What to take into P1

For each cohort step, write four lines:

1. Paper: the exact statement or definition
2. Problem: the ambiguity or data limitation
3. Decision: the implemented rule and rationale
4. Consequence: the change in N, classification or estimate

Then label the consequence: eligibility, design, item missingness, target population or outcome ascertainment.

Be ready to show one person who enters or leaves because of your decision.

Appendix: component map

Appendix: cardiovascular outcome check

Our counts are the full reconstruction row set (6,371).

The saved analytic variable combines heart and cerebrovascular causes. The Table 2 script creates a separate heart-only outcome, which reproduces the published event total.

Do not reuse a derived outcome variable without checking its definition.

Appendix: And L2 writes a funnel — it just does not know it yet

dat1 <- subset(dat,  RIDAGEYR >= 20)                            # N = 43,793
dat2 <- subset(dat1, RIDSTATR == "int + mec exam")              # N = 41,659
dat3 <- subset(dat2, TAKERX_INLASTMONTH %in% c("No", "Yes"))    # N = 41,614
dat4 <- dat3[complete.cases(dat3$BPQ020), ]                     # N not recorded
dat5 <- subset(dat4, mortstat == 0 | (mortstat == 1 & permth_int > 12))  # N = 31,591
dat6 <- subset(dat5, benzodiazepines == "Yes" | ssris == "Yes" | opioids == "Yes")  # N = 4,049

• Six eligibility steps, N after each one — that is a funnel, written as comments
• Now type it, the way we typed ours:

• dat4 prints no N. The complete-case steps are the ones that go uncounted — in this lab and in most papers

Appendix: Reviewing someone else’s eligibility

• Also check the tables against the Methods: here the Table 1–2 headers read “BMI < 25” while the Methods define obesity at ≥ 30, and the non-obese groups have BMI medians of 27.6 and 28.8 — so the headers appear to be a labelling error
• Charity and rigour together: assume competence, and still require that the numbers reconcile

Appendix: course resources

• Detailed recipe: COHORT_BUILD_STEPS.md
• Variable crosswalk: VARIABLE_MAP.md
• Cohort code: 02_build_analytic.R
• Table 1 code: 03_table1.R
• Cox model code: 04_table2_cox.R
• Thursday lab: Accessing and preparing NHANES mortality data

The cohort counts in this deck come from the supplied reproduction pipeline. They are reproduction results, not numbers reported by the paper unless explicitly labelled.