Loading field map
Biostatistics is the science of developing and applying statistical methods to questions in medicine, public health, and biology. It is a subfield of statistics, distinguished less by its mathematical core than by the kinds of data it handles and the kinds of decisions it informs. Where general statistics might study any sort of numerical information, biostatistics is organized around biological variation: the fact that living organisms differ from one another, respond differently to treatments, and are studied in settings where measurement error, missing data, and confounding are the norm rather than the exception. Its central task is to extract trustworthy conclusions from data that are incomplete, noisy, and often collected from people who cannot be randomly assigned to conditions for practical or ethical reasons.
The stakes are unusually high. Biostatistical reasoning determines whether a new drug is judged effective enough to license, whether a screening test does more harm than good, whether an environmental exposure is declared a public health hazard, and whether a genomic finding is real or a statistical artifact. Errors are not abstract: a false-positive result can expose patients to ineffective treatments, while a false-negative result can keep a useful therapy from reaching the clinic. This connection to consequential decisions shapes how biostatisticians work, what they publish, and how they evaluate evidence.
Most of biostatistics can be understood as a response to a single difficulty. Researchers almost never observe an entire population of interest—all patients with a given disease, all children in a country, all people exposed to a pollutant. They observe a sample, and they want to say something about the population from which it was drawn. The fundamental question is: how much can be learned about a population from a finite, imperfect sample, and how certain should we be?
This question has two linked parts. The first is estimation: what is the best guess for a quantity like a mean blood pressure, a risk ratio, or a survival time? The second is uncertainty quantification: how precise is that guess, and how likely is it that the observed difference between two groups arose by chance? The standard toolkit for these tasks—confidence intervals, hypothesis tests, p-values, and standard errors—was largely developed in the early twentieth century by statisticians working alongside agricultural and biological researchers. The logic they established still underlies most medical research, though its limitations have become much better understood over time.
A distinctive feature of biostatistics is that these tools are used not merely to describe data but to support causal claims. When a study reports that a treatment reduces mortality, the claim is that the treatment causes the reduction, not merely that the two are associated. The gold standard for such claims is the randomized controlled trial, in which patients are assigned to treatment or control groups by chance. Randomization works because it breaks the link between treatment assignment and any pre-existing differences among patients—known or unknown. In trials, the statistical machinery of estimation and testing provides a rigorous basis for causal inference.
But many important questions cannot be studied by randomization, either because it would be unethical (randomly assigning people to smoke) or impractical (randomly assigning populations to different health systems). In these observational settings, biostatisticians must attempt to separate cause from confounded association using statistical adjustment, matching, or more recent methods designed to mimic randomization after the fact. This is arguably the hardest and most contested part of the field: the gap between what a randomized trial can establish and what an observational study can only suggest is a constant source of methodological innovation and disciplinary debate.
Biostatistics grew out of statistics itself, and statistics grew out of the need to make sense of variation in the natural world. In the nineteenth century, the Belgian astronomer and statistician Adolphe Quetelet applied the normal distribution to human characteristics like height and chest circumference, treating individuals as imperfect instances of an underlying average "type." This idea of variation around a mean became central, but it was still largely descriptive.
The modern statistical framework arose in the early twentieth century, driven by researchers who needed to analyze agricultural field trials. Ronald Fisher, working at the Rothamsted Experimental Station in England, developed the analysis of variance, the design of experiments, and many of the concepts—sufficiency, likelihood, randomization—that remain foundational. Fisher's work was deeply biological in motivation but general in application, and it created a template for how quantitative data from designed experiments could be analyzed. A parallel tradition developed around the work of Karl Pearson, who emphasized correlation and regression, and later Jerzy Neyman and Egon Pearson, who formalized the logic of hypothesis testing in terms of error rates.
Biostatistics as a self-conscious discipline emerged as these tools were taken up by public health and medicine. Early applications included the statistical evaluation of vaccines, the analysis of disease rates across populations, and the study of occupational and environmental exposures. In the mid-twentieth century, statisticians such as Bradford Hill in Britain and the developers of the Framingham Heart Study in the United States helped establish the prospective cohort study—a design in which a group of healthy people is followed over time and their health outcomes are compared according to their measured exposures. This design, along with the case-control study (in which people with and without a disease are compared for past exposures), became the backbone of epidemiology, and biostatisticians developed the methods to analyze them.
A second major expansion came with the rise of computers and, later, of biomedical measurement technologies. From the 1970s onward, statisticians developed methods that could handle large datasets, complex sampling designs, longitudinal measurements repeated on the same individuals, and data with intricate structure such as patients clustered within hospitals or genes measured across thousands of individuals. The sequencing of the human genome and the later proliferation of high-dimensional biological assays created a need for statistical methods that could cope with situations where the number of measured variables exceeds the number of study participants—a regime that invalidates many classical approaches.
Rather than being divided into rival schools, biostatistics is characterized by complementary traditions that address different problems, sometimes overlapping and sometimes in tension. Understanding these traditions matters more than memorizing method names, because they represent different answers to the question of what it means to do statistical analysis well.
The dominant tradition in biostatistics is frequentist inference. Its core idea is that statistical statements should be evaluated in terms of what would happen if a procedure were repeated many times. A 95% confidence interval, in this framework, is an interval constructed by a method that, in repeated sampling, would contain the true value 95% of the time. A p-value is the probability of obtaining a result at least as extreme as the observed one, assuming the null hypothesis (usually "no effect") is true. These definitions sound technical but imply a specific worldview: probability is understood as long-run frequency, and statistical claims are evaluated by the error rates of the procedures that generate them.
This tradition underlies most clinical trial regulation, most medical journal publication, and most of the machinery of sample-size calculation and study design. Its strengths are rigor and clarity about uncertainty. Its limitations have become increasingly recognized. P-values are routinely misinterpreted as the probability that the null hypothesis is true; confidence intervals are misinterpreted as the probability that the true value lies within them; and the threshold of p < 0.05 has created incentives to run many analyses and report only the significant ones, a problem known as p-hacking. These are problems of misuse as much as of the framework itself, but they are built into the incentives of the system that uses it.
A smaller but influential tradition is Bayesian inference. Bayesians treat probability as a degree of belief rather than a long-run frequency. The researcher starts with a prior distribution—a quantified belief about the likely value of the quantity of interest before seeing the data. The data then update this belief, via Bayes' theorem, into a posterior distribution. The posterior directly expresses the probability that a treatment effect lies in a given range, which is the question researchers actually want answered, rather than the more indirect question that a p-value addresses.
Bayesian methods are especially useful when data are limited, when prior information is genuinely available, or when hierarchical structure must be modeled—for example, when patients are nested in hospitals and the analysis must estimate both hospital-level and patient-level effects simultaneously. Recent advances in computational methods have made Bayesian analysis practical for large and complex models. Its main challenges are the choice of prior (which introduces a subjective element that can feel uncomfortable in regulatory settings), the computational burden, and the fact that much of the medical research establishment is indexed to frequentist categories. In practice, many biostatisticians work in both traditions, using whichever is more appropriate for a given problem.
A third body of methods, often considered a subfield in its own right, is survival analysis. Many medical questions are about timing: how long until a patient dies, relapses, or develops a complication. These data have a special structure—some patients are still event-free when the study ends, and others drop out—so for them the event time is unknown. This is called censoring, and it cannot be simply ignored. Survival analysis developed methods to use all the available information: from censored patients, one knows only that the event did not occur during follow-up, but that information is still useful.
The Kaplan-Meier curve, which estimates the proportion of patients surviving over time, and the Cox proportional hazards model, which relates survival to predictors without requiring strong assumptions about the shape of the baseline risk, are the standard tools. This tradition is central to oncology (where progression-free survival is a common endpoint), cardiology, vaccine trials, and health services research on wait times and re-admission. The key intellectual move is the explicit treatment of partial information, a theme that generalizes to many settings where data are missing or incomplete.
A major organizing distinction in the field is between designed experiments and observational studies. In a randomized trial, the design itself justifies the analysis: randomization ensures, in expectation, that comparison groups are balanced, so a simple comparison of outcomes can be interpreted causally. In observational data, no such guarantee holds. The exposure of interest (say, smoking or taking a statin) is associated with many other factors that also affect the outcome, and the analysis must try to remove these confounders.
The traditional model-based approach to this problem is regression adjustment: fit a statistical model that includes the exposure and a set of candidate confounders, and estimate the exposure effect conditional on those variables. This approach is flexible and widely used, but it depends on assumptions that are hard to verify—that all confounders have been measured, that the model is correctly specified, and that effects do not vary in ways that would make the adjusted estimate misleading.
A more recent family of methods, sometimes called the causal inference tradition, makes these assumptions explicit and develops tools that are more robust to model misspecification. Propensity score methods, for example, model the probability of receiving the exposure given the measured covariates, then balance treated and untreated groups by matching or weighting on that score. Instrumental variable analysis looks for natural experiments—variables that affect the exposure but not the outcome directly—to isolate causal effects even when confounders are unmeasured. Directed acyclic graphs (DAGs) provide a formal language for stating assumptions about how variables are related and for identifying which variables must be adjusted for. These methods have transformed how observational research is evaluated, but they have not resolved the fundamental limitation: causal claims from observational data always rest on assumptions about unmeasured factors that cannot be tested from the data alone.
This is not a disagreement about mathematics; it is a disagreement about what can be concluded from data collected without randomization. The mainstream view, shared by most biostatisticians, is that observational studies can provide suggestive evidence, can support causal claims when the evidence is strong and consistent, and are sometimes the only possible source of information—but they cannot provide the certainty of a well-designed trial. The tension between these positions is productive and ongoing.
The most recent development is the convergence of biostatistics with computational biology and machine learning. Modern biological data can include measurements on tens of thousands of genes, millions of genetic variants, or enormous medical images. Classical regression models fail when there are more variables than observations, and statisticians have developed regularization methods that can nonetheless extract signal. At the same time, prediction methods from machine learning—random forests, gradient boosting, deep neural networks—have become common tools for tasks like risk stratification, disease classification from imaging, and electronic health record analysis.
This tradition is less a set of formal inferential procedures than an orientation toward prediction and pattern discovery. It has been enormously successful in certain applications, but it raises its own problems. Prediction accuracy from a model does not tell you whether you have identified causes; a model that predicts disease risk beautifully may incorporate variables that are not actionable or may produce results that do not generalize across populations. The field's methods also come with new risks of overfitting, hidden biases in training data, and opacity that conflicts with the medical requirement to understand why a decision was made. Biostatisticians working in this area see themselves as adapting the field's core concern—rigorous inference from noisy data—to a new scale of measurement.
These traditions are best understood as a division of labor rather than as competitors. A typical research program in public health will draw on several of them in sequence. A question about whether a dietary factor causes heart disease might begin with an observational cohort study analyzed using model-based adjustment, proceed to a randomized trial analyzed with frequentist methods, and end with a meta-analysis—a statistical synthesis of multiple studies—that might use either frequentist or Bayesian tools. The methods are compatible in practice because they answer different aspects of the question: initial hypothesis generation, rigorous confirmation, and quantitative synthesis.
The relationship between the traditions is also shaped by a deeper division in the field regarding the goals of statistical analysis. One view, associated with Fisher, holds that statistics is primarily a tool for drawing inferences from data via likelihood and that its proper role is to summarize evidence. Another view, associated with Neyman and Pearson, holds that statistics is about choosing procedures with known long-run error rates. A third view, associated with Bayesians, is that statistical analysis should produce a probability distribution over hypotheses. Most working biostatisticians move between these positions pragmatically, but the philosophical differences resurface whenever the field debates standards for publication, regulation, or interpretation.
The present-day field is large and institutionally embedded. Biostatisticians are employed by academic medical centers, public health agencies, pharmaceutical companies, regulatory bodies, and international health organizations. They are standard members of clinical research teams, and their approval is required for most human subjects research. In many countries, biostatisticians hold formal roles in drug regulation and health monitoring.
Several durable questions organize current work. The first is how to handle the reproducibility crisis in science—the recognition that many published findings do not reproduce when others try. Biostatisticians are central to this debate because they design the standards for analysis, pre-registration, and reporting. The field's response has included calls for greater transparency, more conservative use of statistical significance, and methodologically stronger study designs.
A second question concerns the analysis of massive and complex health data. Electronic health records, wearable sensors, genomic sequencing, and medical images are generating data at a scale and richness that strain conventional methods. The field is actively developing procedures that can handle this structure while preserving the rigor that randomized trials taught it to value. Whether the causal language of trials can be applied to such data remains a deep unresolved issue.
A third question is equity and fairness: whether statistical models that guide medical decisions perform differently across racial, ethnic, gender, or socioeconomic groups. Biostatisticians increasingly recognize that a model that is accurate overall can be systematically wrong for subgroups, and that the very definitions of disease and risk encode historical biases. This has led to new methods for measuring and correcting disparities, and to a broader rethinking of what "valid" statistical inference should mean when the populations being studied are themselves shaped by discrimination and unequal access to care.
None of these questions has a settled answer, which is typical of a field that is methodologically mature but continuously pushed into new territory by changes in medicine, computation, and society. What remains constant is the core commitment: to make uncertainty visible and manageable, to separate signal from noise in a world of highly variable biological data, and to give decision-makers an honest accounting of what is known and what is not.