How to verify NHANES variables

A concrete task with RigorMed

1. Turn a variable list into a traceable data-dictionary draft

   These are suggested task inputs and outputs to request, not a new execution, a customer outcome or a guarantee of completion.

   Materials to provide: Research definitions, cycles, candidate fields and official Codebooks, questionnaires and measurement documentation.

   Requested outputs / delivery goal: A source-linked data-dictionary draft, a cross-cycle differences table and a field-gap list.

2. An example request

   For each variable, record its official name, component, definition, eligible population, units, special codes, missingness meanings and source location in a data-dictionary draft. Leave unavailable definitions blank and explain the gap.

3. What still needs review

   Identical names do not establish comparability. Do not infer units, codes, temporal relationships or weights from column names.

What should a usable variable record contain?

Build an item-by-item mapping for the main exposure, primary outcome, and key covariates rather than retaining only a list of variable names.

Labels help discover candidates; official data documentation, codebooks, frequencies, and methods are the main evidence for meaning. If fields look similar, return to the questionnaire, examination, or laboratory component to determine what each measures.

Seven steps for checking NHANES variables

1. Locate candidate variables through official search tools

   Search by disease, measure, questionnaire concept, and possible English terms. Record the cycles and components in which candidates appear. Search results locate variables; they do not establish their final definitions.

2. Open the documentation for the relevant cycle

   Confirm whether the variable comes from demographics, questionnaires, examinations, dietary data, or laboratory files. Record the file, cycle, and participant identifier. Do not infer its current meaning from same-named documentation for another cycle.

3. Check eligibility and skip patterns

   Check applicable ages, sex or interview conditions, and answers that skip later questions. Not asked by design, refusal, 'don't know', and other missing observations must not be treated as a single state.

4. Check methods, units, and special values

   For laboratory and examination variables, verify specimen type, instruments, methods, units, detection limits, and special codes. For questionnaires, retain the exact question, response options, and respondent. Consult component methods and analytical notes where needed.

5. Compare each cycle's measurement comparability

   Create a cycle comparison table covering names, definitions, questionnaire wording, equipment, laboratory methods, units, codes, and eligible populations. Apply official bridging studies or calibration methods only under their stated conditions. Do not assume equivalence without evidence.

6. Determine merge keys and survey weights

   Participant-level components within a cycle are generally merged by SEQN, but record granularity and duplicates still require checking. Select weights based on the components actually used and the analytical population, retaining strata and PSU variables as well.

7. Freeze the dictionary and derivation rules

   Before formal statistics, fix source fields, target names, types, units, code mappings, missing-value rules, derivation formulas, cycle handling, and source links. List meanings that cannot be established from official evidence as gaps rather than filling them from experience.

Six common variable errors

1. Treating a label as a complete definition without opening the cycle-specific codebook;
2. Ignoring eligible ages and skip patterns and including people who were not asked in the denominator;
3. Treating refusals, 'don't know', values below detection limits, or other special codes as ordinary numbers;
4. Comparing column names across cycles without comparing methods, units, and codes;
5. Using multiple components while retaining a weight inappropriate for the most restricted analytical subsample;
6. Recoding in code first and documenting the dictionary and source evidence afterwards.

Checklist before a variable enters analysis

For every main variable, confirm at least the following:

• Do the official name, label, component file, and cycle agree?
• Are the target population, age range, and questionnaire eligibility clear?
• Does the exact question or measurement method support the planned definition?
• Are units, detection limits, equipment, and laboratory methods recorded?
• Is there an official explanation for every code and special missing value?
• Is comparability supported item by item rather than inferred from matching names?
• Have merge keys, record granularity, and duplicates been checked?
• Do weights, strata, and PSU variables match the components?
• Do derived variables have prespecified formulas, source fields, and handling boundaries?
• Are access dates and official documentation links retained in the dictionary?

If the main exposure or outcome cannot pass these checks, do not immediately substitute a similar variable. Revisit the research question and choose suitable cycles, variables, or sources.

Frequently asked questions