Two lab results show a change, three show a trend

How much does a number have to move before the change is real?

Longetic charts each marker across every report you upload, and the first thing a chart shows is that numbers move even when nothing has happened. A result has to clear the combined noise of the assay and your own day to day biology before the difference means anything. That threshold has a name, the reference change value, and for most markers it is larger than people expect.

The reference change value combines two sources of noise. One is analytical: the assay itself does not return the same number twice on the same tube. The other is within-subject biological variation, the random drift of a healthy person around their own set point. Fraser set out the standard form as RCV equals the square root of two, multiplied by Z, multiplied by the square root of the sum of the squared analytical and squared within-subject coefficients of variation (Fraser CG, CCLM 2011). Z is 1.96 for a two-sided 95% probability and 2.58 for 99% (Vasani NM et al, 2018).

The idea is older than the formula. Harris and Yasaka described a reference change in 1983 as "that difference between two consecutive test results in an individual that is statistically significant in a given proportion of all similar persons", and noted that the same method extends to evaluating trends across three successive measurements (Clinical Chemistry, 1983).

The practical consequence is blunt. A ferritin that drops from 50 to 40 has not necessarily dropped. An HbA1c that moves from 5.6% to 6.1% almost certainly has. The difference between those two sentences is not clinical judgement, it is arithmetic about measurement noise, and it is arithmetic laboratory medicine publishes for every marker.

The real-world version is worse than the theory. In 587,023 primary care patients, about 1 in 4 people whose first HbA1c landed in the 48 to 53 mmol/mol band had a second result differing by 5.5 mmol/mol or more, which is 0.5 of a point on the DCCT % scale. That is the band where a diagnosis gets made.

This article covers adults without symptoms, which is a different question from what a doctor orders when something is wrong. Nothing here is medical advice, and no marker below replaces a conversation with a clinician who knows your history.

How many measurements does it take before a trend is real?

Two results are enough to ask whether a change is real, because the reference change value compares one result against the one before it. A trend is a different question, and it needs at least three. Beyond three the returns fall away quickly: in a model of personal reference intervals built from three or more steady-state results, adding further measurements changed little.

Coskun and colleagues modelled personalised reference intervals from a person's own history and reported that "only a few previous test results and reliable estimates of within-subject biological variation are required". In their data, "using ≥3 previous measurement results delivered robust personalized reference intervals", and adding more measurements beyond three changed little: "Increasing the number of measurements had little impact on the total variation around the true homeostatic set point" (Clinical Chemistry, 2021).

There is also a formula for the other direction, how many samples you need to pin down your own set point to a chosen tolerance. The number of tests equals Z multiplied by the square root of the sum of the squared analytical and squared within-subject coefficients of variation, divided by the deviation you are willing to accept, all squared (Choy KW et al, 2024). The reason the answer is small for some markers and large for others is mostly down to that within-subject number, which is tabulated below.

Guidelines apply the same logic without naming it. For kidney function, NICE asks for "a minimum of 3 GFR estimations over a period of not less than 90 days" to identify the rate of progression of chronic kidney disease, and separately tells clinicians to "allow for biological and analytical variability of serum creatinine (±5%) when interpreting changes in eGFR" (NICE NG203).

Large scale data points the same way. Analysing routine blood counts, Foy and colleagues found that "routine CBC indices fluctuate around stable values or setpoints, and setpoints are patient-specific", and that those set points hold for at least twenty years (Nature, 2025). A single reading tells you where you were on one morning. A short series tells you where you sit.

Is a result inside the reference range the same as a healthy result?

No, because a reference range describes a population, not a person. When someone's own values occupy only a narrow slice of that range, a number can travel a long way from where it usually sits and still read as normal. Laboratory medicine measures this with the index of individuality, and for many common markers it lands in the range where the population range is the weaker signal.

The index of individuality is the ratio of within-subject to between-subject variation. Where it is above 1.4 an individual's values cover much of the population range, so the range is informative. Where it is below 0.6 an individual's values "span only a small part of the reference interval", and the population range tells you comparatively little about that person (Vasani NM et al, 2018). The EFLM meta-analysis of thyroid measurands found that "the greater majority" had values "of less than or about 0.6", and concluded that this "reinforces the importance of RCVs" (Fernandez-Calle P et al, 2021).

Published values put TSH at 0.42 and free T4 at 0.56. Serum iron, at 1.98, is one of the few where the population range genuinely applies. Work the same ratio on the published variation figures for ferritin, apolipoprotein B, LDL-C, creatinine, cystatin C, haemoglobin and hs-CRP, and every one of them lands below 0.6.

In plain terms, a result sitting comfortably inside the normal range can still be a long way from where your own number usually sits, and a result at the edge of the range may be exactly where you always are. This is the honest case for tracking rather than screening, and the reason a first result is close to uninterpretable on its own.

MedlinePlus puts the same caution in plain words: a result outside the range "may or may not be a sign of a health problem", it is "common for healthy people to sometimes have results outside the reference range", and "some people with health problems can have results that are considered 'normal.'" (MedlinePlus).

How much does each marker move on its own?

Within-subject biological variation is published per marker, and it differs by more than a factor of ten across a routine panel. HbA1c drifts by under 2%, triglycerides by close to 20% and C-reactive protein by about 35%. That single number is what decides whether a change you are looking at is signal or noise, which is why the same move in % means very different things on different lines of the same report.

Each figure is the published within-subject coefficient of variation from the source named in its row. Half the rows come from the EFLM Biological Variation Database, which has run since May 2019 and in 2024 held "over 3060 BV data sets" with global estimates for 185 measurands (Sandberg S et al, 2024).

The C-reactive protein row needs a note. The 2024 systematic review of C-reactive protein and high-sensitivity CRP reported a median of 44% across studies and a pooled intraclass correlation of 0.62, and did not pool the CV itself (PLOS ONE, 2024). The 34.7% in the table is the EFLM database's own global estimate, which is why the two figures differ.

A few markers have a published reference change value, calculated with the study laboratory's own analytical imprecision. Those are worth quoting as they stand, including the asymmetry, where the % needed for a rise is not the % needed for a fall.

Two notes on reading these. The asymmetric form is now the one the EFLM database recommends, because biological data are log-normal rather than symmetric (Roys EA et al, 2025). And every one of these thresholds assumes the analytical side is under control: the database sets the target for analytical imprecision at half the within-subject variation, which keeps instrument noise small relative to biology (Sandberg S et al, 2024). A laboratory that misses that target widens every threshold built on these figures.

The current biomarker list in 2026

Six markers have a named body behind them and a stated interval: a lipid panel with LDL-C and non-HDL-C, lipoprotein(a) once in adulthood, HbA1c or fasting glucose, eGFR together with urine albumin-to-creatinine ratio, apolipoprotein B for residual risk, and hs-CRP with its own repeat rule. Two things moved the list this year, and both cut against the biomarker articles published in 2024 and 2025.

Three absences are deliberate. There is no vitamin D, because the Endocrine Society's 2024 guideline suggests against routine 25(OH)D screening in healthy adults and retired its own 30 ng/mL target. There is no TSH, because USPSTF has given screening asymptomatic non-pregnant adults for thyroid dysfunction a grade I, insufficient evidence, since 2015. And there is no fasting insulin or HOMA-IR, because the Association for Diagnostics and Laboratory Medicine states that HOMA-IR "is not recommended and should be reserved for research purposes only."

That is not the same as saying those tests are useless. It means nobody has shown that measuring them in a person without symptoms changes what happens next.

The 2026 ACC, AHA and multisociety guideline on the management of dyslipidemia, published 13 March 2026, states that Lp(a) should be measured at least once in adulthood. It puts high at 125 nmol/L or 50 mg/dL and above, associated with about a 1.4-fold increased long-term risk of heart attack or stroke, and 250 nmol/L with at least a two-fold increase. The National Lipid Association had already recommended measuring Lp(a) in every adult at least once in its 2024 focused update, where it notes that elevation is common, at roughly 1 in 5 individuals.

Once is usually enough because the number barely moves. The European Atherosclerosis Society reports a within-person correlation of 0.87 over a mean of 8.3 years, and states that diet and physical activity do not have a major influence on it.

Then the drug news. On 4 September 2026 Novartis announced that Lp(a)HORIZON, an 8,323-patient phase III trial of pelacarsen in people with Lp(a) of 70 mg/dL or above and established cardiovascular disease, "did not meet its primary endpoint of reducing the risk of cardiovascular events." Lp(a) levels did come down. The events did not follow. Full data are due at a medical congress that has not yet been named.

This matters for how you read your own number. An elevated Lp(a) still changes decisions, because the 2026 guideline uses it to push LDL-C lowering harder. What it does not yet buy you is a drug that treats the Lp(a) itself. Three phase III outcomes trials are still running and none has reported: OCEAN(a)-Outcomes with olpasiran from Amgen, estimated completion March 2028, ACCLAIM-Lp(a) with lepodisiran from Eli Lilly, estimated March 2029, and MOVE-Lp(a) with oral muvalaplin, also Lilly, estimated March 2031.

There is a third change worth knowing, because it is an argument rather than a settled position. In September 2025 the American College of Cardiology published a scientific statement calling for universal hs-CRP screening alongside cholesterol, on the strength of Ridker's 30-year Women's Health Study analysis published in the New England Journal of Medicine in 2024, where the top-versus-bottom quintile hazard ratio was 1.70 for hs-CRP against 1.36 for LDL-C and 1.33 for Lp(a). A scientific statement is not a guideline, and its own authors write that the time has come for guidelines to implement broad screening, which is a way of saying they have not. The 2021 European Society of Cardiology prevention guideline still carries a class III recommendation against routinely collecting additional circulating biomarkers. Both positions are current, and they disagree.

Why does running more tests find more abnormal results?

Because a reference interval is defined as the central 95% of a healthy population, so 5% of healthy people fall outside it on any single test by construction. Run twenty independent tests and at least one comes back flagged in 64.2% of healthy people. At sixty tests it is about 95%. The panel size on a marketing page is not the feature it is presented as.

The 20-test figure is quoted directly from a 2025 review in the Journal of the American Board of Family Medicine: "64.2% of 'normal' patients will have at least one 'abnormal' test result in a panel of 20 tests." The rest is the same arithmetic, 1 minus 0.95 to the power of N.

Each flag carries a repeat test, sometimes an imaging study, and always some worry. The counter-argument is real too: a flagged result that turns out to matter is worth a great deal to the person it happens to. The point is that the arithmetic belongs in the decision rather than arriving as a surprise afterwards. Whether a large paid package is worth its price is a separate question, and a money question rather than a measurement one.

Which marker do the popular lists all skip?

Urine albumin-to-creatinine ratio. It appears in none of the widely circulated biomarker listicles, and it is where the gap between what guidelines ask for and what actually gets measured is widest. KDIGO's 2024 guideline says to test filtration and albuminuria together, not one or the other, yet among US adults with hypertension and no diabetes, only 5.1% had ever been tested.

KDIGO's 2024 chronic kidney disease guideline opens with the instruction to test people at risk using both urine albumin measurement and an assessment of glomerular filtration rate. Categories are A1 below 30 mg/g, A2 from 30 to 300, A3 above 300. An abnormal random sample should be confirmed on a first morning void, and chronicity established over 3 months rather than assumed from a single result.

The AHA's 2023 cardiovascular-kidney-metabolic advisory goes further, advising annual uACR alongside eGFR from stage 2, which sweeps in anyone with hypertension, diabetes, metabolic syndrome or triglycerides at 135 mg/dL or above.

What actually happens: in a study of 192,108 US adults with diabetes or hypertension, 17.5% had albuminuria testing. Among those with diabetes it was 52.3%. Among those with hypertension and no diabetes, 5.1%. The authors estimated that only 35.2% of the albuminuria present in that population had been detected.

An honest caveat belongs here. USPSTF has no recommendation for chronic kidney disease screening in asymptomatic adults, and its review is still in progress. So uACR is strongly recommended for people at risk and unsettled for everyone else. Its other use is quieter: eGFR is a required input to the AHA's PREVENT risk equations, published in 2024 for adults aged 30 to 79, and uACR is one of the optional inputs that can be added to them.

What can make a result wrong before the lab even runs it?

A gym session, a supplement, a tourniquet left on too long, and which analyser ran the sample. One weightlifting session pushed muscle and liver enzymes above the reference limit for at least seven days. Biotin supplements bias immunoassays in both directions. Between four commercial ferritin systems, one platform read more than 60% higher than another on the same samples.

A gym session. In fifteen healthy men, a single weightlifting session pushed creatine kinase, AST, LD and myoglobin above the upper reference limit in every subject, and ALT above it in five, with elevations still present seven days later. If you train and then get a liver panel, you may be reading your workout. We looked separately at what a hard session does to five of those results.

Biotin supplements. The FDA warned in 2017 and again in November 2019 that biotin can cause falsely high or falsely low results depending on the assay, and reported a death following falsely low troponin results. The recommended daily allowance is 0.03 mg. Hair and nail supplements commonly contain 20 mg, and some pills contain up to 100 mg. Prescription drugs are a longer story, and we cover which medicines shift a lab result separately.

Whether you ate. Less than most people think, and not in the direction they expect. In the EAS and EFLM joint consensus data, the maximum mean change after a normal meal is about plus 26 mg/dL for triglycerides and about minus 8 mg/dL for total cholesterol and LDL-C. Apolipoprotein B and Lp(a) do not move at all. A fasting repeat is advised when non-fasting triglycerides reach 400 mg/dL under the US guideline, 440 mg/dL under the European consensus. If you are changing when you eat rather than skipping one breakfast before a draw, our guide to tracking health markers while fasting covers what the trials measured.

How long the tourniquet stayed on. At 30 seconds, nothing measurable. At 60 seconds, glucose, total protein, albumin, triglycerides, potassium, calcium and magnesium all rise. At 180 seconds, everything tested does.

Which analyser ran it. Across four commercial ferritin systems, the median between-assay coefficient of variation was 22.9%, and one platform read more than 60% higher than another on the same samples. For thyroid tests, a College of American Pathologists survey found a spread of 37.8% between the highest and lowest platform for free T4. In the vitamin D standardization programme's interlaboratory study, only half of the immunoassays met the precision criterion. Lp(a) has its own version of this problem: mass units in mg/dL and molar units in nmol/L cannot be reliably converted, because the mass of an Lp(a) particle varies with apo(a) isoform size, which is why the EAS asks that results be reported in the units the assay was calibrated in.

It helps to know what the number is. MedlinePlus describes ferritin as "a protein that binds to iron and stores it in your body", which is why the test is read as a window on iron stores. It is not a pure iron reading, though: the same page lists inflammation, "like from infection or surgery", among the conditions that can raise ferritin, alongside liver disease, obesity and alcohol use disorder. A ferritin result taken during or just after an infection is therefore not the same measurement as one taken when you are well.

The practical rule that falls out of all this: same laboratory, same assay, same conditions, or the trend you are looking at is partly a trend in methods.

How often should you retest each marker?

No single schedule covers the panel, so each interval below comes from the body that states it. Lipoprotein(a) once in adulthood. Lipids about every five years in low-risk adults. HbA1c or glucose every three years if normal. eGFR with urine albumin annually if you have hypertension or diabetes. Two rules hold it together: confirm an abnormal result before acting on it, and leave enough time between tests for a real change to be possible.

Which of these apply to you depends on your age and your risk factors, and our screening tool walks that part. Confirmation is built into the guidelines themselves: the ADA requires two abnormal results for a diabetes diagnosis unless symptoms are unmistakable, KDIGO says not to assume chronicity from a single abnormal eGFR or ACR, and the AHA and CDC ask for two hs-CRP measurements averaged.

A diagnosis changes the schedule again, and what a chronic condition adds to your own schedule runs on shorter intervals than the general adult one. Whichever schedule you follow, the comparison that matters is with your own previous result, which means every report has to live in one place with its date, its laboratory and its units. A consistently named folder of PDFs in cloud storage costs nothing and does that job, and a tracker app does it with a chart. As a disclosure, Longetic is our product, so weigh this accordingly: it keeps the series and draws the line, on iPhone only at $9.99 a month or $69.99 a year in the US App Store, and it does not order tests, does not compute a reference change value for you and does not tell you what a number means, so the arithmetic above is still yours to do.

Without the laboratory recorded you cannot tell a real change from a change of analyser. Without the units you cannot tell 50 mg/dL from 50 nmol/L, and for Lp(a) those are not the same number. What a clinician actually needs from that history is a shorter list than most people expect, and seven sections belong in a doctor-ready summary.

Frequently asked questions