Skip to content
No Vape

Did the vaping estimate change because the model changed or because the rows changed?

Compare the analysis rosters when an additional predictor leaves some completed surveys outside the model.

No Vape · Updated

Stack of old newspapers on a table
Illustrative image · Mr Cup / Fabien Barral / Unsplash

Find the requirements for each model column

Read the footnotes under the model table. One model may require the behavior outcome and one background field. A second might add a field that some respondents did not supply. A respondent can have an observed vaping outcome and still be outside the second analysis. Completion of the survey and inclusion in a specific model are separate events.

Sterne and colleagues explain that a complete-case approach can leave out records with incomplete analysis data. That is the starting point for this reading check. It does not tell you why particular values are missing or whether excluding them changed the estimate in a specific study.

BMJ: Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls

Sketch two rosters for a fictional model table

Imagine a table where 80 respondents supplied both the outcome and age group. A second model adds time on a particular activity. Ten of the 80 lack that field. The first model uses 80 cases and the second uses 70 if both apply the described complete-case rule independently. Even if the reported relationship changes, those two columns alone do not tell you how much change comes from each difference.

Draw two lines for each model. One lists the required fields. The other lists the reported count and the kind of missing field that affects entry, if described. If two counts match, check any further description of membership. Counts alone cannot verify that two analyses used exactly the same records.

Look for a same-roster comparison

Some papers provide a version of the simpler model restricted to the same records as the fuller model. That makes the adjustment comparison easier to describe within the shared set. It does not determine what the estimate would have been for the excluded observations or resolve assumptions about why values were unavailable.

If the paper gives no same-roster comparison, retain that uncertainty. Do not subtract one reported estimate from another and assign the difference entirely to the newly added variable. Keep the coefficient scale and the outcome definition alongside the roster as well.

Explain the exact change you verified

A careful summary can say the second estimate added a particular covariate and used a smaller complete-case set. If the report gives a shared-roster comparison, describe it as a separate analysis. This is an evidence-reading exercise about which observations contributed, without trying to repair a result or make a treatment decision.

Questions about this step

Can a completed follow-up still be excluded from one model?

Yes. A missing predictor or other variable required by that model can affect inclusion even if the follow-up outcome itself was recorded.

Sources and scope

Sources inform the guidance. The worksheets and examples are No Vape editorial suggestions, not tested treatments.

  1. BMJ — Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls

Written with AI assistance and checked against the cited sources. No clinical review is claimed. This guide covers practical planning, not individual medical treatment. How we research and write