How to choose NHANES survey weights

A concrete task with RigorMed

1. Organise a weighting question into a pre-analysis decision record

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

   Materials to provide: Proposed cycles, the components containing the variables, official Analytic Notes, the target population and intended estimands.

   Requested outputs / delivery goal: A table of weighting evidence, a complex-sampling design draft and items for statistical review.

2. An example request

   Using my variables and the official guidance, prepare an evidence table for candidate weights, strata/PSUs and cross-cycle exceptions. List unresolved component combinations separately and draft an analysis-plan paragraph for statistical review.

3. What still needs review

   This does not promise automatic correct weight selection or executed complex-survey analysis. Special releases such as 2017–2020 and 2021–2023 must not be handled by mechanically dividing by the number of cycles.

Identify the weight level from variable sources

This table narrows the candidates; it is not an automatic selector. The official analytical notes for the actual cycle and components determine the final choice.

Weights reflect sampling, nonresponse, and coverage adjustments, not simply 'multiplying the sample size'. With several components, weights are usually chosen for the most restrictive eligible sample. However, official guidance specifically addresses combinations of 2021–2023 blood draw and dietary variables; do not mechanically choose whichever group has fewer observations.

Six checks before analysis

1. Specify the target estimand and population

   Distinguish population means or proportions, prespecified subgroup estimates, between-cycle trends, and overall pooled estimates. State ages, respondent conditions, and exclusions. Different questions cannot share an unverified weighting plan.

2. Bind every variable to its cycle and component

   Record variable names, component files, collection setting, eligible populations, and official Analytic Notes. Questionnaire variables may be collected in the MEC; the 'Questionnaire' category alone does not justify interview weights.

3. Select applicable weights and check missingness

   Choose interview, MEC, blood draw, dietary, or other subsample weights covering the main analytical variables. Distinguish sampling or component nonparticipation, item nonresponse, and complete-case losses. Existing weights do not automatically adjust for all item missingness.

4. Retain complex-survey design variables

   SDMVSTRA and SDMVPSU in public files are masked strata and pseudo-PSUs used for variance estimation. Use survey-capable software with applicable weights, strata, and PSUs. Do not treat weighted records as independent observations from a simple random sample.

5. Check comparability and special releases before pooling

   Compare questionnaires, equipment, laboratory methods, units, codes, target populations, and design changes. Determine whether the files are regular two-year releases, the 1999–2002 four-year base period, the 2017–2020 pre-pandemic 3.2-year release, or the 2021–2023 post-pandemic release. Read the corresponding pooling guidance.

6. Freeze the plan and check estimate reliability

   Prespecify weight variables, construction, strata, PSUs, cycles, subgroup definitions, effective samples, and reliability criteria. Use domain/subpopulation methods for subgroup analysis rather than deleting non-subgroup records before estimating variance.

When does 'divide each of two cycles' weights by 2' not apply?

1. Official four-year weights are provided for 1999–2002; longer periods crossing this base period require specific construction rules;
2. The 2017–March 2020 pre-pandemic files represent approximately 3.2 years, not a regular two-year cycle;
3. A gap without data collection separates August 2021–August 2023 from earlier periods; pooling requires assessment of assumptions about the unobserved period;
4. The 2021–2023 blood analytes introduce blood draw weights; combinations with dietary and other components require that release's special guidance;
5. Fasting, dietary, and other dedicated subsamples have their own weights; also reading demographics does not justify replacing them with MEC or interview weights;
6. Do not first delete everyone outside a subgroup and then calculate standard errors using ordinary sample methods.

Include these details in the analysis plan

Retain a reviewable record of every item before formal statistics:

• Is the goal between-cycle comparison, a pooled overall estimate, or a subgroup estimate?
• Are each core variable's cycle, component, eligible population, and Analytic Notes confirmed?
• Does the weight cover the most restrictive applicable component, and are there release-specific exceptions?
• Do SDMVSTRA and SDMVPSU correspond to the official design guidance for these cycles?
• Have definitions, measurements, units, codes, and periods without data collection been assessed across cycles?
• Are weight sources and construction, subgroup methods, effective samples, and unstable-estimate rules frozen?

If a key component has no applicable weight or cross-period comparability cannot be confirmed, narrow the cycles, revise the question, or stop the population inference. A generic formula cannot fill an evidence gap.

Frequently asked questions