BIOSTATISTICS HUB / 04
Biostatistics connects health questions, study design, data, uncertainty, and decisions. The goal is not to run the most complicated test; it is to choose methods that match the question and data, examine assumptions, quantify uncertainty, and explain what the result does—and does not—mean.
Follow the reasoning
Start in order, but do not treat research as a straight line. Later evidence often sends you back to fix an earlier decision—and that is part of doing the work properly.
Identify outcomes, exposures, predictors, confounders, identifiers, dates, and derived measures. Check coding, units, impossible values, duplicates, missingness, and whether observations are independent or clustered.
Quality check: Keep an untouched raw file and create a reproducible cleaning log or script.Use counts and percentages for categories. For numerical variables, choose mean and standard deviation when the distribution supports them, or median and interquartile range when skew or outliers make those summaries more informative.
Quality check: A table should communicate the sample, missing values, denominators, units, and distribution—not only a list of numbers.Use confidence intervals to show the range of values reasonably compatible with the data and model. Report an effect measure that fits the question: mean difference, risk difference, risk ratio, odds ratio, rate ratio, correlation, or another justified measure.
Quality check: Statistical uncertainty is not the only uncertainty; bias, measurement error, and confounding also matter.Match the test to the outcome type, groups, pairing, distribution, study design, and objective. Examples include t tests, analysis of variance, rank-based tests, chi-square or exact tests—but assumptions must be checked.
Quality check: A p-value is not an effect size, the probability that the null is true, or proof of importance.Use linear models for suitable continuous outcomes, logistic models for binary outcomes, Poisson or negative-binomial approaches for counts or rates, and other models when the outcome structure requires them. Pre-specify key confounders and assess fit.
Quality check: Do not select variables using p-values alone; subject knowledge and a causal framework should guide adjustment.Define the event, time origin, follow-up, censoring, and competing events. Summarise survival with appropriate curves and use models such as Cox regression only when their assumptions and interpretation fit the data.
Quality check: Two participants with the same outcome but different follow-up times do not contribute identical information.Base sample-size calculations on the primary objective, design, effect worth detecting, variability or baseline risk, significance level, desired power, allocation, design effect, and expected non-response or loss to follow-up.
Quality check: Report every assumption and run scenarios; a single unexplained sample-size number is not enough.Present effect estimates with uncertainty, exact denominators, missing-data handling, model specification, diagnostics, sensitivity analyses, and practical interpretation. Separate association from causation and statistical from clinical importance.
Quality check: Make tables, figures, and code traceable to the cleaned analysis dataset.See it in context
Start with a binary outcome defined using stated clinical criteria, describe the sample, estimate unadjusted associations, then use a justified multivariable model while checking missingness, influential observations, collinearity, and model performance.
Hypertension coded from a stated measurement protocol and threshold
Age, sex, body-mass index, smoking, physical activity, diabetes, and other pre-specified factors
Prevalence with a confidence interval plus distributions and missingness for every key variable
A model suited to the binary outcome, with adjusted effect estimates and uncertainty
Functional form, collinearity, influential cases, calibration, discrimination, and sensitivity checks
Association in the sampled setting—not automatic proof of causation or population-wide prediction
Supporting learning resource
Use the practical primer to review populations, variable types, power, common test choices, regression and responsible interpretation before returning to the full biostatistics workflow.
Before you move on
The analysis starts from the research objective and outcome type.
01Cleaning, missingness, coding, denominators, and exclusions are documented.
02Effect estimates and confidence intervals accompany significance tests.
03Model assumptions, fit, diagnostics, and sensitivity analyses are reported.
04The conclusion respects study design, uncertainty, bias, and practical importance.
05Biostatistics is strongest when the question, design, data, method, and interpretation tell the same story.
Follow the next decision
A good question should shape the literature search; the evidence should sharpen the design; and the methods should make the result possible to understand and challenge.