The reference change value: the number that tells you whether a delta is real

A reference interval is the central ninety-five per cent of results from a defined reference population, which has a consequence people rarely state: one in twenty perfectly healthy people falls outside it for any given analyte, by construction and by design. On a twelve-analyte panel, assuming independence, the probability that a healthy person produces at least one flagged result is approximately forty-six per cent. On a twenty-analyte panel it is about sixty-four per cent. The flagged result on a routine panel is therefore closer to the expected outcome than to a finding.

What a reference interval actually is

A reference interval is an empirical statement about a population. A laboratory recruits a reference group meeting defined health criteria, measures the analyte, and reports the central ninety-five per cent of the resulting distribution, usually as the 2.5th to 97.5th percentiles. Everything about that construction has consequences. The interval is specific to the assay and platform used to derive it. It is specific to the reference population — its age structure, sex distribution, ethnicity and, for some analytes, its diet and altitude. And it deliberately excludes one in twenty healthy people at each end by design.

Two further points follow. The interval is not a target: for several analytes the optimal value on outcome grounds sits well inside it or below it, and low-density lipoprotein cholesterol is the standard example. And it is not a diagnostic threshold: decision limits, which are what clinical guidelines actually use, are derived from outcome data rather than from a healthy distribution, which is why the diagnostic cut-off for diabetes is not the upper limit of a reference interval.

Laboratories that report both a reference interval and a decision limit are doing the reader a service. Most report one number and one flag, and leave the distinction to be inferred.

The arithmetic of the comprehensive panel

If an analyte’s reference interval excludes five per cent of healthy people, and if analytes were independent, the probability that a healthy person produces at least one flagged result on a panel of n analytes is one minus 0.95 to the power of n. For a twelve-analyte panel that is approximately 46 per cent. For twenty analytes, approximately 64 per cent. For a thirty-analyte comprehensive panel with lipids and thyroid included, approximately 79 per cent.

Analytes are not independent — electrolytes covary, liver enzymes covary, so the true figures are somewhat lower — but the direction and rough magnitude hold. The practical implication is uncomfortable and rarely stated: on a comprehensive panel, the flagged result is the normal outcome, and treating each flag as requiring explanation is a commitment to explaining noise.

This is the strongest single argument for ordering panels against questions rather than by habit. A panel assembled because each analyte answers something the clinician wants to know produces flags that mean something. A panel assembled because it comes as a bundle produces a document in which the interesting result, if there is one, is hidden among four uninteresting ones. The Journal makes this point in a publication whose readers frequently order their own panels privately, and it applies with more force there rather than less.

The reference change value, worked

Two results in the same person differ for three reasons: the analyte genuinely changed, the assay is imprecise, and the analyte varies within the person from day to day. The last two are quantified in the biological variation literature as the analytical coefficient of variation and the within-subject coefficient of variation, and databases of the latter have been maintained for decades.¹

The reference change value combines them: approximately 2.77 times the square root of the sum of their squares, for a two-sided ninety-five per cent probability that a difference is real. The results are instructive. Sodium, with tiny biological variation, has a reference change value of around three per cent. Creatinine is about fourteen per cent. Alanine aminotransferase, with within-subject variation above twenty per cent, requires something like a sixty per cent change. Triglycerides, more variable still, require more.

Apply that to a routine monitoring situation. An ALT moving from 28 to 41 units per litre — a rise of forty-six per cent that crosses no threshold and is unlikely to be flagged — sits inside the reference change value and may be nothing at all. An ALT moving from 28 to 62 has moved. Nothing on the report distinguishes the two cases, and the distinction is the entire question.

The alanine aminotransferase interval is too wide

Most clinical laboratories report an upper limit of normal for alanine aminotransferase somewhere between about 40 and 55 units per litre, with a modest sex difference or none. Those intervals were derived from reference populations that were screened for viral hepatitis and heavy alcohol use but not, in most cases, for hepatic steatosis — which was neither commonly diagnosed nor considered when many of the intervals were established.

Work redefining the healthy range in a large population of prospective blood donors, screened for viral markers, alcohol intake and metabolic risk factors, arrived at substantially lower limits: in the region of 30 units per litre for men and around 19 for women.² Those figures have been influential in hepatology and have largely not propagated into general laboratory reporting.

The consequence for this population is direct. A person starting treatment with an ALT of 44 has a flagged result by a strict standard and an unflagged one by their laboratory interval; a fall to 31 during treatment represents normalisation by one standard and continued abnormality by the other. Neither reading is wrong. The Journal reports ALT against both where it can, and regards a laboratory report giving only the wider interval as incomplete rather than incorrect.

Against measuring too often

Frequent monitoring in a person doing well is a reliable generator of work. Each comprehensive panel carries a substantial probability of at least one flagged result; the flags are mostly noise; each requires explanation, repetition or investigation; and the cumulative effect over a year of monthly panels is several investigations and no additional information about the person.

There is also a specific problem with monitoring an analyte more frequently than its own window. HbA1c integrates three months. Measuring it monthly produces overlapping windows in which two-thirds of the data is shared between consecutive results, so the apparent trend is smoother than the underlying glycaemia and the independent information per measurement is low. The trials in this class scheduled it quarterly for exactly this reason.

The counter-argument deserves stating fairly: monitoring during escalation, when tolerability problems and their metabolic consequences are most likely, is a different proposition from monitoring during stable maintenance, and the case for closer observation in the first three months is reasonable. What the Journal has not seen is any evidence that a fixed frequent schedule during maintenance detects anything that a symptom-prompted panel would miss. Readers who know of such evidence should write to standards@compoundjournal.com.

The next instalment in this department takes up the analytical side of the same coin: not what a clinical laboratory measures in a person, but what an independent testing service measures in a vial, and why the two documents look more alike than they are. That is a different set of instruments and a different set of failure modes, and it is covered in Analytics.