Erythrocytosis and Raised Haematocrit in Testosterone Therapy

Evidence Note

What is the question?

Testosterone therapy raises haemoglobin and haematocrit — the concentration of red cells in the blood. Professional societies specify haematocrit levels that trigger clinical action: commonly above 54%, with some bodies naming 50% before starting treatment. The question this review asks is narrow and factual: what are those numbers calibrated to?

It is a question about provenance, not about practice. The review grades the evidence; it does not answer whether, at what level, or how often haematocrit should be monitored, nor what should be done when a threshold is crossed.

What does the evidence say?

The rise is real, measured directly, and reasonably precise. A meta-analysis of seven longitudinal studies — 312 people, low heterogeneity between them — puts the increase at about four and a half percentage points over the first year. A prospective 192-person cohort reports a similar figure. In the prospective evidence, the largest increment falls in the first three months. This is the strongest directly measured statement the review makes.

The average rise does not put a typical person near the action thresholds — but individuals do cross them. Among 192 prospectively followed trans men, four reached at least 52%; one reached exactly 54·0%, and no value above 54·0% was reported. In a retrospective cohort of 511 patients, twelve — 2·3% — had a recorded haematocrit of at least 54%.

Those frequencies are not comparable to each other. Some studies count anyone who ever crossed a line; others describe a single sampled value; one estimates time-specific cumulative risk with censoring. Different thresholds, different observation periods, different ways of finding an event. An earlier version of this review treated the spread between them as a substantive finding about populations. It is not, and that reading is withdrawn.

No study identified here shows that acting at a threshold prevents anything. Thirteen reports give thrombotic-event status alongside a quantitative haemoglobin or haematocrit. None has a comparator defined by haematocrit exposure; none can attribute an event to a threshold crossing. The rationales cited around the current numbers draw on general-population data, on evidence from other conditions that raise the red cell count, and on expert judgement. The numerical derivation of the individual cut-offs has not been reconstructed, and no trans-masculine outcome study identified here validates them.

The tabulations of those guidelines contain errors. Checking eight of the nine guideline rows in a published systematic review against the primary documents found discrepancies of value, of function and of attribution — and a structural problem underneath them: the table cannot represent a threshold for starting treatment separately from one for acting during it. They are different clinical objects and the format runs them together. One current guideline carries a confirmed transmission defect of its own.

What this review does not claim

Three things it would be easy to read into the findings, and which the review explicitly does not say. Each was argued in an earlier draft and withdrawn.

It does not say haematocrit is the wrong measure. Haematocrit agrees only weakly with directly measured red-cell mass — in one referral cohort it explained eight per cent of the variance. But haematocrit is itself a concentration, and concentration is a major determinant of how thick blood is. Weak agreement with red-cell mass does not by itself disqualify it. An earlier version argued that it did; that argument is gone.

It does not say the treatment effect is weak. A person’s serum testosterone level is a poor predictor of their blood count compared with other people’s — under three per cent of the between-person variance in a study of 6,670, and no relationship at all within a cohort of 79 on stable therapy. That constrains using one testosterone reading as a proxy for someone’s haematocrit. It says nothing about whether changing a dose changes the count, and nothing about causation.

It does not say testosterone makes the blood more likely to clot. The direct coagulation evidence is mixed by assay family, and the larger plasma-based analyses lean the other way: in 348 trans men, thrombin generation fell after six of seven modalities; in a nested analysis of 320 of the same people, clots formed less and broke down more readily. An earlier version summarised this evidence in one direction. It does not run in one direction.

How strong is the evidence?

Strong where it is measured directly, thin where it is inferred. The haematological change is graded robust, from randomised and longitudinal evidence in this population. The mechanism is not settled — the human erythropoietin data disagree with each other, and in animal work the best-characterised alternative pathway turned out contributory but not essential. And the link from a number on a blood test to a clinical event in this population has not been established by any study identified here.

The review states its own retrieval limits. Its search of PubMed and ClinicalTrials.gov was only partially executed to protocol, and Embase and CENTRAL were not successfully searched. Where it says something was not found, that is a statement about this review’s retrieval, not a claim that the field contains nothing. That distinction is carried on every absence claim in the review, and it is carried here.

Where are the uncertainties?

Almost everything between the blood test and the decision. Whether crossing 54% — or 52%, or 50% — changes a person’s risk of a clot or a cardiovascular event. Whether reducing the dose, switching formulation or taking blood off at that point changes anything. How the numbers in the guidelines were arrived at. Whether the relationship between haematocrit and red-cell mass measured in other populations holds in testosterone-treated trans-masculine people — it has not been measured there. And the quantitative contribution of the mechanisms in humans.

What does this mean for practice?

For clinicians. Nothing here is a reason to change what you do, and nothing here is a reason not to. The rise is real and predictable, and it is steepest early. The thresholds you are working to are expert judgement resting on evidence from other populations, not on outcome data in this one — which is worth knowing when a patient asks what a number means, and worth knowing when a borderline result prompts a decision about their dose. That the calibration is unverified does not make monitoring unwarranted; an unvalidated threshold and a useless one are not the same thing.

For patients. A rising haematocrit on testosterone is expected, not a sign that something has gone wrong. If your result crosses a line in a guideline, that line is a cautious number drawn from other groups of people rather than a measured danger point for you. It is a reason for a conversation with your clinician about dose and formulation — not a reason to stop treatment on your own, and not a reason to dismiss the result either.

For anyone reading the headlines. Two opposite misreadings are available here, and this review supports neither. That the thresholds are uncalibrated does not mean testosterone is dangerous at levels nobody has measured; and it does not mean monitoring is theatre. The honest position is narrower and less quotable than either: the change is certain, the consequence is not, and the numbers in use have not been shown to be the right ones or the wrong ones.

A note on this review’s own accuracy

The review publishes an audit of itself, and it is one quarter complete. 447 of its sentences carry a citation; 111 have been checked against the source they cite. Nine were wrong — a quantifier, a qualifier, a modal verb, a population term, a count, a denominator attached to the wrong scope — and all nine are corrected in the published version. On that rate, the remaining 336 should be expected to hold twenty-five to thirty more of the same kind. Finishing the audit is roughly ninety-four further source reads, and that will be version 2.0.

The full correction record — seven hundred and three entries, each stating what was wrong, where the error came from, who caught it and what changed — is published as a companion page →.