Biostatistics
The statistics that health analytics actually needs — and the handful of errors that produce most of the confidently wrong conclusions in the literature and on your dashboard.
In one line
You don't need to be a biostatistician. You need to know when you've left your competence — and to recognise the five or six errors that produce most of the wrong conclusions in health data, including the ones already published.
The p-value, said properly
A p-value is: the probability of seeing data at least this extreme, if the null hypothesis were true.
It is not the probability the hypothesis is true. It is not the probability you're wrong. It is not a measure of effect size. p = 0.049 and p = 0.051 are not different findings, and the 0.05 line is a convention, not a law of nature.
The ASA felt compelled to publish a formal statement saying this, which tells you how routinely it's misused. Their headline: scientific conclusions should not be based only on whether a p-value passes a threshold.
What to report instead: the effect size and its confidence interval. "HbA1c fell by 0.8% (95% CI 0.3 to 1.3)" tells a clinician something. "p < 0.05" tells them almost nothing — not how much, not how certain, not whether it matters.
The errors that do the damage
Absence of evidence is not evidence of absence. Altman and Bland's classic. A non-significant result frequently means your study was too small to see it, not there is no effect. Underpowered studies produce "no difference" conclusions that are just fog.
Multiple comparisons. Test twenty things at p < 0.05 and you expect one false positive by chance. Health datasets have hundreds of columns, so "we explored the data and found that X is associated with Y" is usually a lottery win presented as a discovery. Correct for it, or say plainly that it's hypothesis-generating.
Confounding. Ice cream sales predict drowning. Both are summer. In health: sicker patients get more treatment, so naive analysis makes treatment look harmful. Every observational finding needs the question what else differs between these groups?
Regression to the mean. Take the worst-performing wards, intervene, re-measure — they improve. They would have improved anyway, because extreme measurements are partly noise and noise doesn't repeat. This single effect has "validated" an enormous number of health interventions that did nothing.
Immortal time bias. If patients must survive long enough to receive the treatment, the treatment looks life-saving. Everyone in the treated group was, by construction, alive.
Survivorship / the missing denominator. Analysing only the patients who came back. As everywhere in this domain, the ones who left are the finding.
What you'll actually use
- Descriptive statistics first. Median and IQR for skewed data — and health data is almost always skewed. Length of stay has a long tail; its mean is a fiction nobody experiences.
- Confidence intervals on every rate you report.
- Survival analysis (Kaplan-Meier, Cox) for anything time-to-event, because it handles censoring — patients who haven't had the event yet. Ignoring censoring is how you get nonsense.
- Logistic regression for binary outcomes, and know that an odds ratio is not a risk ratio — they diverge badly when the outcome is common, and people quote them interchangeably.
- Standardisation when comparing populations of different ages. Raw mortality comparisons between a retirement town and a university town are meaningless.
The number that matters clinically
NNT — number needed to treat. "Reduces relative risk by 50%" sounds enormous. If it moves absolute risk from 2 in 1,000 to 1 in 1,000, the NNT is 1,000 — treat a thousand people to help one.
Relative risk is how you sell something. Absolute risk is how you decide. Any health claim quoting only relative risk is withholding the number you need, and it does so reliably enough that you should treat it as a signal.
The professional stance
The most valuable habit isn't a technique, it's a sentence — the same one that matters in interviews and at the bedside:
"I can show you that, but here's why I wouldn't trust it yet."
Know when to fetch a statistician. Wanting the answer is not the same as having it, and in health the cost of the confident wrong answer is borne by someone who never saw your notebook.