What "Normal" Means

Build a panel

Every checkbox is a real analyte NHANES measured in one blood draw from every adult in the 2017-2018 wave. The three colored strips are the fourteen of the comprehensive metabolic panel (CPT 80053), the fifteen indices of the complete blood count with differential (CPT 85025), and ten extras a "wellness" or "executive physical" screen tends to add.

One participant, screened healthy, flagged on three tests

NHANES sequence number 93878. A 59-year-old woman, non-pregnant, with a body-mass index of 24.1 and a glycohaemoglobin of 5.2% (both squarely non-diabetic and non-obese by any threshold anyone uses). She answered "no" to every one of the thirteen questionnaire items the reference sample requires: diabetes, congestive heart failure, coronary heart disease, angina, heart attack, stroke, liver condition, thyroid problem, COPD, cancer, kidney trouble, high blood pressure, high cholesterol. She is, by every questionnaire criterion, well. On the CMP+CBC her white cell count, red cell count and mean cell volume all fall outside the middle 95% of women exactly like her, and the flags are real, not paperwork.

She is one of 965 women in the reference sample. If she had walked into a clinic for the panel above with no complaint, the printout would carry three red flags and the questions "what did we miss?" that follow. The honest answer is: nothing. The panel was built to send 57% of women exactly like her home with at least one red flag, and it did.

How the reference sample is built

Because a reference range is defined by the people it is built from, the whole page turns on how the sample is selected. The screen used here is the one CLSI EP28-A3c recommends and no clinic actually performs: every person is drawn from a national household survey, ages twenty and up, not pregnant, with complete data on every one of the thirty-nine analytes, and answering no to every one of thirteen questionnaire items about heart, kidney, liver, thyroid, lung, blood pressure, cholesterol, cancer, and diabetes. Fifteen weeks of interviews and lab draws narrow to this shape:

Why the joint rate is lower than independence

Independence predicts 81.1% of the reference sample would flag on at least one of the thirty-nine, and the measured rate is 69.4%. The eleven-point gap is not noise. Several analytes measure the same thing twice. Haemoglobin and haematocrit correlate at r = 0.97 in this sample, so a person high on one is almost always high on the other, and the two together flag less often than two independent draws would. Sodium and chloride correlate at r = 0.60 for related physiological reasons. The three red-cell indices (MCV, MCH, MCHC) are arithmetic derivatives of the other three (MCHC = HGB ÷ HCT × 100 holds cell-for-cell to within the file's own rounding), so they add almost no independent chances to flag.

Add in the leukocyte differentials, whose five components sum to (100 − ε)% by definition, and roughly a dozen of the thirty-nine tests are not new information but reshaped old information. Independence overestimates; the shape of that overestimate is what a real laboratory's redundancy looks like when you count it honestly.

What this page is not saying

It is not saying that a flagged test is meaningless. A red flag raises the odds a real problem is present relative to the general population; it does not raise them to certainty, and on any single healthy adult the prior probability of the flag being spurious is not small. On a CMP with 38.5% of healthy adults flagging on at least one item, roughly two of every five flags seen by a general practitioner on a healthy patient are the interval's own 5%, not a signal.

It is also not saying that reference intervals are wrong. They are the answer to a specific question ("what value does 95% of the reference sample show?"), and they answer it exactly. They are not, and never were, the answer to the question a person receiving a printout hears them as ("am I well?"), and no arithmetic on a wellness panel can reconcile the two. The number this page prints is what the mismatch costs.

Three edges of the showing this page did not attempt. It uses NHANES's unweighted sample rather than reweighting by WTMEC2YR; it does not remove statistical outliers before deriving intervals (CLSI EP28 contemplates a Tukey step, we skipped it, so a handful of very extreme values pull each interval a fraction wider than the standard method); and it does not split intervals by age, though several analytes drift with age enough that a pediatric or geriatric panel would deserve its own. Each choice is defensible and each is a choice; the numbers change by fractions of a percentage point under either alternative, not by percentages.

Notes

The framing goes back at least as far as the 1980 Journal of Chronic Diseases paper by Bradwell, Carmalt and Whitehead, which pointed out the arithmetic on Screening for Disease panels of the era, and to the 2016 CLSI EP28-A3c guideline that codifies the reference-interval construction the analysis here reproduces. The observation is old; the honest version of it, computed on the people actually sampled, is the part a machine with the room to think can do live.

A note on the vignette. Sequence number 93878 is a real NHANES respondent; the vignette shows the CMP+CBC values she supplied to the CDC in 2017-2018. Age, sex and glycohaemoglobin are attached in the same file. She was not identified beyond the sequence number in any published record and is not identified here; NHANES ships every participant with the household and clinic identifiers stripped precisely so this kind of small honest use is possible.