Chapter 1 — Why Biostatistics Matters at the Bedside

The monitor and the manuscript: same problem

You walk into the ICU at 03:00. The arterial line shows MAP 58. The pulse oximeter reads 88%. The bedside glucose says 2.4 mmol/L. Three numbers. Three alarms. You don’t treat the numbers — you treat what they represent, and you question whether each sensor is giving you the truth.

A clinical paper is the same thing. The abstract gives you a number — “mortality reduced from 42% to 31%, p = 0.03.” Your job, before you change practice, is to ask the same three questions you ask about that monitor:

1. Is the number real, or is it artifact?
2. Does it apply to this patient, in this ICU, with these resources?
3. What’s the cost of acting on a false signal?

Biostatistics is the set of tools that answers those questions. It is not a separate subject you study in year two and forget. It is the language in which medical evidence is written, and if you cannot read that language, you are prescribing by reflex, not by evidence.

What clinical studies actually do

Every clinical study tries to answer one of three questions:

• Descriptive: “What happened?” — incidence of VAP in our ICU last year, median SOFA score on admission, proportion of patients with AKI stage 3.
• Analytical: “Is X associated with Y?” — does early dialysis associate with mortality? Is fluid overload linked to longer ventilation?
• Interventional (causal): “Does X cause Y?” — does prone positioning reduce mortality, or do sicker patients just spend more time prone?

The statistics you need depends on which question is being asked. Descriptive questions need summaries (means, medians, proportions). Analytical questions need association measures (odds ratios, hazard ratios, correlation coefficients). Interventional questions need designs that isolate causality — and that is where randomization, blinding, and control groups enter.

Most papers you read mix these three without being explicit about which question they are actually answering. Your first checkpoint is always: what is the study for? If the authors designed an analytical study but write the discussion as if it were interventional, the statistics will mislead you.

The population you care about vs the one studied

Every study draws a sample from a population. The population is the group you want to generalize to (e.g., all adults with septic shock in ICUs). The sample is who actually enrolled (e.g., patients at 12 centers in France, excluding immunocompromised, excluding those on ECMO).

The gap between the two is called external validity or generalizability. A statistically significant result in a highly selected population may not apply to the septic patient with cirrhosis and renal replacement therapy who sits in your bed. Ask:

• Who was excluded and why?
• How does the enrolled population differ from yours?
• Was the setting (community hospital vs academic ICU) comparable?

In critical care, exclusion criteria are often broader than in other specialties — exactly because sicker patients are harder to manage and their outcomes are messier. This means ICU literature frequently underestimates the heterogeneity you see on the ground.

Bias: the enemy that statistics cannot fully fix

Statistics can adjust for measured confounders. It cannot adjust for what was not measured, or for systematic error in how data were collected.

Three biases you will meet daily:

• Selection bias: the enrolled group differs from the target population in a way related to the outcome. Example: a trial of early enteral nutrition that excludes patients with gastroparesis — those patients may be the ones who benefit most.
• Detection bias: one arm is monitored more aggressively, so outcomes are detected more often. Example: the ICU protocol group gets daily blood cultures; the standard group only gets cultures when febrile. The protocol group will “find” more infections — not because the intervention caused them, but because surveillance was different.
• Attrition bias: patients drop out differentially between arms. Example: the intervention arm has 15% loss to follow-up because patients transferred out; the control arm has 2% loss. The surviving patients in the intervention arm may be the healthier ones — and the analysis becomes biased.

No statistical test can rescue a study with severe bias. The statistics describe the data as collected; if the collection process is flawed, the description is of flawed data.

Parameters, statistics, and the sampling error you live with

The parameter is the true value in the population (e.g., the true mortality rate of septic shock in all ICUs). The statistic is what you observe in your sample (e.g., 28% mortality in your 80-patient cohort). They are never identical — that difference is sampling error.

This is why confidence intervals exist. A 95% CI of 22–34% mortality says: if we repeated this study 100 times, 95 of those intervals would contain the true parameter. It does not mean there is a 95% probability the true rate is between 22 and 34% in any one study — that is a common misinterpretation.

The width of the CI tells you precision; the sample size, the event rate, and the variability all determine it. A study of 40 patients reporting “mortality 15% (95% CI 5–25%)” gives you a very different clinical picture than “mortality 15% (95% CI 12–18%).”

Why the ICU is statistically harder than most specialties

• Heterogeneous populations: sepsis, trauma, post-surgical, medical — each with different baseline risks.
• Time-varying exposures: vasopressor doses, fluid balances, ventilator settings change hourly; a single baseline measurement misses this.
• Competing risks: death from multiorgan failure competes with death from hemorrhage; standard survival analysis may misestimate if not handled with cause-specific or subdistribution hazards.
• Cluster effects: patients nested within nurses, within attending physicians, within ICUs — outcomes cluster, violating independence assumptions.
• Small samples, large effects: ICU trials often enroll hundreds, not thousands, so they are powered for large effect sizes; modest but clinically meaningful effects go undetected.

If you understand these ICU-specific challenges, you will read ICU literature more skeptically — and more usefully.

Your bedside checklist for any paper

Before you change practice based on one study, verify:

1. Question match: does the statistical method match the clinical question (descriptive, analytical, interventional)?
2. Population: does the enrolled group resemble your patients? Who was excluded?
3. Design safeguards: randomization, allocation concealment, blinding — present and adequate?
4. Bias assessment: attrition, detection, selection — how handled?
5. Effect size with precision: not just p-value, but the point estimate and its CI.
6. Clinical significance: is the magnitude of effect big enough to matter at the bedside?
7. Harm and cost: what are the downsides of the intervention, and were they reported?

Go deeper

• StatPearls — “Understanding Biostatistics Interpretation” (NCBI NBK431056): free, peer-reviewed, covers the purpose of biostatistics in clinical research, types of data, and basic inference. https://www.ncbi.nlm.nih.gov/books/NBK431056/
• BMJ Statistics Notes (Altman & Bland, free PDF): short, practical pieces on interpreting common statistical reports in medical literature. https://www.bmj.com/content/bmj_stats_notes
• “Biostatistics for Medical and Biomedical Practitioners” (2nd ed., Academic Press / ScienceDirect): free chapter available on ResearchGate, written explicitly for clinicians who need to critically read the literature. https://www.sciencedirect.com/book/9780128170847

Next: Chapter 2 — From question to design: RCT vs observational