How big is this really?
Step 7Live-room allocation: 2:00
Open in Colab — 02_act2.ipynb, opens in a new tab
NHANES is designed to move from a sample to the US population.
A result can look small when you only count the people directly in front of you.
In this cohort, 70 Black adults cross into CKD when the race coefficient is removed.
But NHANES is not an ordinary convenience sample. It is a complex probability sample designed to represent the US population. Each participant carries a survey weight reflecting how many people they represent nationally, while strata and primary sampling units preserve the structure of the survey design.
So the question is not only:
How many people did we observe?
It is also:
How many people could this represent across the country?
From 70 people to a national estimate
Section titled “From 70 people to a national estimate”After applying the NHANES survey design, 70 observed Black adults correspond to an estimated
427,489 Black adults nationally
who were kept on the healthier-looking side of the CKD threshold by the race coefficient.
That is approximately 2.08% of Black adults, with a design-based 95% confidence interval of
277,359– 577,620.
That is not just a large number.
It is a reportable national estimate under the reliability rule used here.
The estimate is supported by at least 30 unweighted observations, and its relative standard error is no greater than 30%.
In other words, the national projection survives the reliability check.
The projection is large because the sample was designed to represent more than itself
Section titled “The projection is large because the sample was designed to represent more than itself”Survey weights do not simply make a number bigger.
They encode how each sampled participant contributes to an estimate of the US population. The confidence interval then reflects the uncertainty introduced by sampling that population rather than observing everyone directly.
That is why the interval matters:
427,489 is the estimate.
277,359–
577,620 is the uncertainty around it.
Not every finding scales with the same confidence
Section titled “Not every finding scales with the same confidence”Now return to the metformin finding from the previous step.
| Finding | Observed n | National estimate | Status |
|---|---|---|---|
| Black adults reclassified into CKD | 70 | 427,489 | RELIABLE |
| Black diabetics with a masked metformin tier | 13 | 68,455 | UNRELIABLE FOR NATIONAL PROJECTION |
The two findings are useful at different levels.
The CKD result is supported by enough observed cases to estimate the national burden with acceptable reliability.
The metformin result is still clinically informative in this cohort: 13 Black diabetic adults changed prescribing tier, including 2 who crossed into contraindication.
What does not hold up is the next step — turning those 13 observed patients into a precise national count.
There are simply too few unweighted observations to support that projection under the reliability rule used here.
So the limitation belongs to the scale of the claim, not to the finding itself.
We can say:
In this cohort, the equation change altered metformin eligibility for observed patients.
We cannot say with the same confidence:
This corresponds to exactly 68,455 people nationally.
Scale and certainty are different questions
Section titled “Scale and certainty are different questions”The CKD result tells us something about national scale.
The metformin result tells us something about clinical mechanism and consequence.
Both are useful, they simply support different levels of inference.