Biostatistics and Epidemiological Analytics

Introduction and Context

Health-care managers make decisions under uncertainty. They must determine whether a screening programme detects disease early, whether a service intervention reduces waiting time, whether a hospital’s mortality is unusual, and whether an observed difference reflects a real effect or random variation. Biostatistics provides the methods for describing health data, quantifying uncertainty, comparing groups, and modelling outcomes. Epidemiological analytics applies these methods to populations, exposures, interventions, diseases, and health systems.

The purpose of statistical analysis is not simply to produce a small p-value. A useful analysis connects a clearly defined question with an appropriate design, valid measurement, transparent assumptions, and an interpretation that considers both statistical and clinical or managerial significance. Poorly chosen statistics can lead to inappropriate resource allocation, inequitable policies, and avoidable harm.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Distinguish populations, samples, parameters, statistics, variables, and data types.
2. Select and interpret descriptive statistics and suitable graphical displays.
3. Explain probability, sampling distributions, confidence intervals, and hypothesis tests.
4. Interpret effect measures used in epidemiology, including risk ratios, odds ratios, and rate ratios.
5. Distinguish statistical significance from clinical, public-health, and operational significance.
6. Apply and interpret linear, logistic, and Cox regression models.
7. Explain censoring, hazard, survival functions, and Kaplan–Meier analysis.
8. Recognise trends, seasonality, autocorrelation, and control limits in time-series data.
9. Design a useful health-care dashboard and communicate uncertainty responsibly.
10. Evaluate ethical, methodological, and governance risks in epidemiological analytics.

1. The Statistical Language of Health Care

A population is the complete set of units relevant to a question—for example, all patients admitted to a hospital during a year. A sample is a subset observed or measured. A parameter describes a population, whereas a statistic describes a sample. Statistical inference uses sample information to estimate or test claims about a population.

Variables may be:

• Categorical: sex, diagnosis, ward, vaccination status.
• Ordinal: disease severity rated mild, moderate, or severe.
• Discrete quantitative: number of admissions.
• Continuous quantitative: blood pressure, length of stay, or serum glucose.

The measurement scale matters. Nominal categories have no inherent order; ordinal categories have order but not necessarily equal intervals; interval measures have equal intervals but no meaningful zero; ratio measures have equal intervals and a meaningful zero. Treating an ordinal score as a continuous variable may be reasonable in some analyses but must be justified.

An epidemiological question should specify the population, exposure or intervention, outcome, comparison, time period, and estimand. For example: “Among adults discharged after heart failure, what is the difference in 30-day readmission risk between patients receiving and not receiving a structured follow-up programme?” This is more analytically useful than asking whether the programme “works.”

2. Descriptive Statistics and Data Exploration

Descriptive statistics summarise observed data without, by themselves, generalising beyond the study data.

For observations x₁, x₂, ..., x_n, the arithmetic mean is:

Here x̄ is the sample mean, x_i is observation i, and n is the number of observations. The mean is appropriate for approximately symmetric quantitative data, but it is sensitive to extreme values.

The sample variance and standard deviation are:

The denominator n−1 provides an unbiased estimator of the population variance under the usual independent random-sampling assumptions. Standard deviation describes dispersion in the original units; variance is in squared units.

The median is the middle ordered observation, and the interquartile range (IQR) is:

where Q₁ and Q₃ are the 25th and 75th percentiles. Median and IQR are generally preferable for skewed variables such as length of stay or cost.

For a binary outcome, the sample proportion is:

where x is the number with the characteristic and n is the total number observed. A rate differs from a proportion because its denominator includes person-time, such as 18 infections per 1,000 person-years.

Worked example: descriptive interpretation

Suppose five patients have lengths of stay of 2, 3, 3, 4, and 13 days. The mean is 5 days, while the median is 3 days. The mean is pulled upward by the 13-day stay. A manager planning routine bed capacity should inspect the distribution and perhaps report median and IQR, while a financial model based on total bed-days may appropriately use the mean. The “best” summary depends on the decision.

Useful displays include histograms for distributions, box plots for group comparisons, run charts for observations over time, scatterplots for two quantitative variables, and stacked or clustered bar charts for categorical comparisons. Graphs should identify denominators, units, dates, missing values, and any transformations.

3. Probability and Epidemiological Measures

Probability ranges from 0 to 1 and represents uncertainty about an event. If events A and B are mutually exclusive, then:

For general events:

Conditional probability is:

Independence means that P(A|B) = P(A), or equivalently P(A ∩ B) = P(A)P(B). Independence should not be assumed merely because two variables are measured separately.

In a cohort study, risk is the proportion developing an outcome during a specified period. If a exposed individuals develop disease among n₁ exposed individuals, exposed risk is a/n₁. The risk ratio (RR) compares risks:

An RR of 1 indicates equal risks; above 1 indicates higher risk in the exposed group; below 1 indicates lower risk. A risk difference is:

Risk difference is especially useful for managerial planning because it estimates the absolute change in events.

In case–control studies, the usual measure is the odds ratio (OR). For a 2 × 2 table with exposed cases a, unexposed cases c, exposed controls b, and unexposed controls d:

The odds ratio approximates the risk ratio when the outcome is rare and the design is appropriate. It can substantially exaggerate the apparent magnitude of common outcomes and should not automatically be described as a risk ratio.

For events occurring over person-time, an incidence rate is:

Rate ratios compare incidence rates. All effect measures require a defined numerator, denominator, time frame, and population.

4. Sampling, Bias, and Representativeness

Sampling methods include simple random sampling, systematic sampling, stratified sampling, cluster sampling, and convenience sampling. Stratification can ensure adequate representation of small but important groups, such as rural patients. Cluster sampling may reduce cost when individuals are naturally grouped in hospitals or communities, but observations within clusters are correlated and the effective sample size is smaller than the nominal sample size.

Sampling error is random variation caused by observing a sample rather than the entire population. Selection bias occurs when inclusion or participation is related to exposure and outcome. Information bias results from inaccurate measurement, such as differential recall or inconsistent diagnosis. Confounding occurs when a third variable is associated with both exposure and outcome and is not on the causal pathway.

A confounder is not simply any associated variable. Adjustment should be based on substantive knowledge, a causal diagram, and the temporal relationship among variables. Adjusting for a mediator may remove part of the effect of interest; adjusting for a collider can create bias.

The standard error of a sample mean, under independent observations with finite variance, is:

Increasing sample size reduces standard error approximately in proportion to 1/√n, not in direct proportion to n. Clustered, weighted, or correlated data require modified variance estimation.

5. Confidence Intervals and Hypothesis Testing

A confidence interval communicates estimation uncertainty. For a mean with a large sample or an approximately normal sampling distribution, a 95% confidence interval is commonly written:

where t_0.975,n−1 is the 97.5th percentile of the t distribution with n−1 degrees of freedom. For a known population standard deviation, a z critical value may be used instead. For proportions, especially near 0 or 1 or with small samples, Wilson or exact methods are generally preferable to an uncritical normal approximation.

A 95% confidence interval is produced by a procedure that would contain the true parameter in approximately 95% of repeated samples under the model and assumptions. It does not mean that there is a 95% probability that a fixed parameter lies in this particular interval under the frequentist interpretation.

Hypothesis testing begins with a null hypothesis H₀ and an alternative hypothesis H₁. A p-value is the probability, assuming H₀ and the model are correct, of obtaining a result at least as extreme as the observed result. It is not the probability that the null hypothesis is true.

A significance level α, often 0.05, defines a decision rule before analysis. A type I error rejects a true null hypothesis; a type II error fails to reject a false null hypothesis. Power is 1−β, the probability of detecting a specified effect under the alternative. Sample-size planning should specify the minimum clinically important difference, outcome variability or baseline risk, significance level, desired power, allocation, and expected loss to follow-up.

Statistical significance is not clinical significance. A very large sample may produce a statistically significant 0.2-day reduction in length of stay that is operationally trivial. Conversely, a clinically important reduction may be imprecisely estimated in a small study. Decision-makers should examine effect size, confidence interval, harms, costs, feasibility, and equity.

Multiple testing increases the chance of false-positive findings. Pre-specified primary outcomes, transparent reporting, correction or control of false discovery where appropriate, and replication are preferable to selective emphasis on favourable results.

6. Regression and Adjustment

Regression models describe associations between an outcome and one or more predictors, while allowing adjustment for measured covariates.

Linear regression

For a continuous outcome:

Here y_i is the outcome for person i, x_i is a predictor, β₀ is the intercept, β₁ is the expected change in outcome per one-unit increase in x, and ε_i is the residual. Multiple regression adds predictors as a sum of terms.

Classical inference assumes correctly specified mean structure, independent observations, constant residual variance, and approximately normal residuals for small-sample tests and intervals. Normality of the outcome itself is not required for unbiased estimation, but extreme skew, nonlinearity, heteroscedasticity, and influential observations require attention. Robust standard errors, transformations, splines, or generalized models may be appropriate.

Logistic regression

For a binary outcome with probability p:

The exponentiated coefficient e^β_j is an adjusted odds ratio for a one-unit increase in predictor xⱼ, holding other model variables constant. Odds ratios are not risk ratios, particularly when outcomes are common. Predicted probabilities and marginal effects are often easier for managers to interpret.

Causal interpretation and model quality

Regression adjustment does not automatically establish causation. It cannot remove unmeasured confounding, systematic measurement error, selection bias, or reverse causation. Analysts should define the target estimand, use a causal diagram where appropriate, assess missingness, examine interactions, validate predictive performance, and avoid overfitting.

Discrimination describes how well a model separates patients with and without an outcome; calibration describes agreement between predicted and observed risk. A model can discriminate well but be poorly calibrated in a new hospital. External validation and monitoring for performance drift are essential before deployment.

7. Survival Analysis

Survival analysis concerns time to an event, such as death, readmission, treatment failure, or discharge. It handles right censoring, where the event has not occurred before follow-up ends or the individual is lost to follow-up.

The survival function is:

where T is event time. The hazard function describes the instantaneous event rate among those still event-free at time t:

Hazard has units of inverse time and is not the same as risk. The cumulative hazard is the accumulated hazard over time. The Kaplan–Meier estimator estimates survival as successive conditional survival probabilities:

where t_j are event times, d_j is the number of events at time t_j, and n_j is the number at risk immediately before that time. Censored observations reduce the risk set after their censoring time but are not counted as events.

The Cox proportional-hazards model is:

where h₀(t) is the baseline hazard. e^β_j is a hazard ratio under the proportional-hazards assumption. A hazard ratio of 0.75 suggests a 25% lower instantaneous hazard, not necessarily a 25% lower probability of ever experiencing the event. Proportionality should be assessed; if it fails, time-varying effects, stratification, restricted mean survival time, or alternative models may be more suitable.

8. Time-Series and Quality Analytics

Health services generate observations indexed by time: daily emergency visits, weekly infections, monthly expenditure, or hourly waiting times. A time series may contain level, trend, seasonality, cycles, intervention effects, and random noise.

A simple additive representation is:

where Y_t is the observation at time t, L_t is level, T_t is trend, S_t is seasonal effect, and E_t is irregular error. In multiplicative settings, components combine by multiplication rather than addition.

Autocorrelation means observations close in time are correlated. Ignoring it can produce standard errors that are too small and false confidence. Run charts and statistical process-control charts are useful for operational monitoring. A Shewhart chart commonly uses a centre line and limits such as:

where CL is the process centre line and σ is the process standard deviation, assuming a stable process and an appropriate distributional approximation. Control limits describe process variation; they are not the same as confidence intervals or clinical thresholds.

Interrupted time-series analysis can estimate a change in level or slope after a policy intervention. The design is stronger when there are many pre-intervention observations, a clearly timed intervention, stable measurement, and consideration of concurrent events. Difference-in-differences compares changes over time between intervention and comparison groups, relying heavily on the parallel-trends assumption.

Forecasting methods should be evaluated using time-ordered validation rather than random splitting. Performance metrics include mean absolute error and root mean squared error:

MAE is in the outcome’s units; RMSE gives greater weight to large errors. Forecast intervals should be reported, especially when staffing or capacity decisions depend on them.

9. Dashboards and Communication of Uncertainty

A dashboard is a decision-support instrument, not merely a collection of attractive charts. It should begin with the decision: who must act, how often, and what threshold triggers action? Each indicator needs a precise definition, numerator, denominator, population, time window, data source, owner, update frequency, and limitations.

Effective dashboards:

• display rates or proportions with denominators, not counts alone;
• distinguish observed, estimated, and forecast values;
• show confidence or prediction intervals where meaningful;
• use run charts or control charts for process monitoring;
• allow stratification by clinically and ethically relevant groups;
• avoid unnecessary colour, three-dimensional graphics, and truncated axes that exaggerate differences;
• identify missingness, revisions, and data latency;
• preserve privacy through aggregation and access control.

A manager should receive an interpretation such as: “Readmission declined from 14.2% to 12.8%; adjusted risk difference −1.4 percentage points, 95% confidence interval −2.5 to −0.3. The result is statistically compatible with a modest benefit, but implementation cost and effects across equity groups require review.” This is more informative than “readmissions improved significantly.”

10. Illustrative Case: Reducing Emergency Department Waiting Time

Background

A regional hospital introduces a senior clinician rapid-assessment service at triage. The management question is whether the intervention reduces time to first clinical assessment without worsening safety.

Problem and design

The hospital collects 12 months of pre-intervention and 12 months of post-intervention data. Outcomes are median waiting time, the proportion assessed within 30 minutes, left-without-being-seen rate, and 72-hour return visits. Patient volume, acuity, day of week, and staffing are recorded.

A naïve comparison of monthly means could be misleading if winter demand is greater after implementation. Analysts therefore plot monthly distributions, report medians and IQRs, and use an interrupted time-series model with intervention level and slope terms. They inspect autocorrelation and adjust standard errors appropriately. A comparison emergency department without the service provides a useful sensitivity analysis.

Analysis

Before implementation, median waiting time is 78 minutes; after implementation it is 61 minutes. The unadjusted difference is −17 minutes. The model estimates an immediate level reduction of 12 minutes, with a further monthly decline of 1.1 minutes, while the comparison site shows no corresponding change. The 95% confidence interval for the immediate reduction is −20 to −4 minutes. The proportion assessed within 30 minutes increases from 41% to 53%.

However, the left-without-being-seen rate rises from 3.0% to 3.6%, and the increase is concentrated among patients from a high-deprivation area. The overall improvement therefore cannot be declared an unqualified success. Stratified analysis reveals that the rapid-assessment process improves speed but may have altered queue priorities or access patterns.

Approach and managerial response

Management retains the intervention but adds an equity and safety review, expands evening staffing, monitors clinical deterioration, and publishes monthly results with uncertainty intervals. A subsequent process evaluation identifies that registration delays disproportionately affect patients requiring language assistance. The hospital introduces interpreter support and a simplified registration pathway.

Learning points

1. Averages alone conceal distributional and equity effects.
2. Time-series methods are more appropriate than treating monthly observations as independent.
3. Operational indicators must be balanced with safety and access outcomes.
4. Statistical evidence informs, but does not replace, clinical and managerial judgement.
5. Dashboards should support corrective action rather than reward a single favourable metric.

11. Ethics, Governance, and Contemporary Analytics

Health data are sensitive. Governance requires lawful collection and use, data minimisation, secure storage, role-based access, audit trails, retention policies, and transparent data-sharing arrangements. De-identification reduces risk but does not guarantee anonymity, particularly when datasets can be linked.

Analysts should pre-specify important outcomes where feasible, preserve analysis code and data dictionaries, document exclusions and missing-data decisions, and report both favourable and unfavourable findings. Missing data should be described by group and variable. Complete-case analysis may be biased; multiple imputation can be useful when its assumptions are plausible, but it is not a remedy for systematically absent or invalid measurements.

Artificial intelligence and machine learning can improve prediction, image interpretation, demand forecasting, and anomaly detection. Their use introduces risks of dataset shift, automation bias, opaque decision rules, unequal performance, and feedback loops. A model trained in one population may be unsafe in another. Prospective evaluation, subgroup calibration, human oversight, impact assessment, and continuous post-deployment monitoring are necessary. Predictive accuracy does not establish that an intervention will improve outcomes, and a prediction model should not be interpreted as a causal model without suitable design.

Best-Practice Checklist

Before approving an analysis, ask:

1. Is the decision and target population clearly defined?
2. Are exposure, outcome, time origin, comparison, and estimand explicit?
3. Is the design appropriate for the question?
4. Are denominators, missingness, measurement validity, and selection addressed?
5. Are effect sizes and uncertainty reported, not only p-values?
6. Have assumptions, confounding, interactions, clustering, and sensitivity analyses been considered?
7. Are clinical, operational, equity, privacy, and financial consequences assessed?
8. Can a non-statistical decision-maker understand the result without being misled?

Chapter Summary

Biostatistics enables health-care professionals to describe data, quantify uncertainty, compare groups, and construct models for decisions. Descriptive statistics must match the distribution and purpose of the data. Probability underpins risk, rates, conditional reasoning, and epidemiological effect measures. Sampling error differs from bias, and representativeness cannot be inferred from sample size alone.

Confidence intervals express estimation uncertainty, while hypothesis tests provide evidence against specified null models. Statistical significance must be interpreted alongside effect magnitude, clinical importance, cost, feasibility, and equity. Linear and logistic regression allow adjusted association analysis, but adjustment does not automatically establish causation. Survival methods address time-to-event outcomes and censoring; time-series methods address trend, seasonality, autocorrelation, intervention effects, and forecasting.

Dashboards should be designed around decisions, display denominators and uncertainty, and balance speed or efficiency indicators with safety and equity measures. Ethical analytics requires secure governance, transparent methods, fair model evaluation, and responsible communication. The central managerial skill is not merely calculating a statistic but judging whether the evidence is valid, relevant, and actionable.

Key Terms

Biostatistics: Application of statistical methods to biological, medical, and health-service data.

Population: Complete set of units to which a research question refers.

Sample: Subset of a population that is observed or measured.

Parameter: Numerical characteristic of a population.

Statistic: Numerical characteristic calculated from a sample.

Standard deviation: Measure of dispersion in the original units of a quantitative variable.

Interquartile range: Difference between the 75th and 25th percentiles.

Risk: Probability of an outcome during a specified period.

Incidence rate: New events divided by total person-time at risk.

Risk ratio: Ratio of risks in exposed and unexposed groups.

Odds ratio: Ratio of odds, commonly used in case–control studies and logistic regression.

Confounding: Distortion of an exposure–outcome association by a third variable.

Confidence interval: Interval produced by a procedure designed to contain a parameter with specified long-run frequency.

p-value: Probability, under a null model, of a result at least as extreme as the observed result.

Regression: Model relating an outcome to one or more predictors.

Censoring: Incomplete knowledge of an event time, commonly because follow-up ends before the event.

Hazard: Instantaneous event rate among individuals still at risk.

Autocorrelation: Correlation among observations at different times in a series.

Calibration: Agreement between predicted and observed risks.

Clinical significance: Importance of an effect for patients, care, or outcomes, independent of statistical significance.

Review Questions

Short Answer Questions

1. Distinguish a parameter from a statistic and provide one health-care example of each.
2. Why might the median be preferred to the mean for length of stay?
3. Define confounding and explain why random sampling error is not the same as confounding.
4. What does a 95% confidence interval communicate under the frequentist interpretation?
5. Explain why a hazard ratio is not automatically a risk ratio.

Long Answer Questions

1. Explain the relationship among sampling, standard error, confidence intervals, statistical significance, and clinical significance.
2. Compare risk ratios, odds ratios, risk differences, and incidence rate ratios, including suitable study contexts.
3. Describe the assumptions, interpretation, and limitations of logistic regression in hospital outcome analysis.
4. Explain how Kaplan–Meier methods and Cox regression handle time-to-event data and censoring.
5. Design a dashboard for monitoring emergency department performance, specifying indicators, denominators, uncertainty displays, and governance controls.

Analytical / Application Questions

1. A programme reduces readmission from 16% to 13% in 2,000 patients but costs substantially more. Identify the absolute risk reduction, the managerial information still required, and the limitations of relying on a p-value alone.
2. A hospital reports a large improvement after an intervention, but the post-intervention period is winter and patient acuity increased. Propose an analytical strategy that addresses seasonality, trend, confounding, and safety outcomes.
3. A machine-learning model predicts deterioration accurately overall but performs poorly for older adults and a minority ethnic group. Explain the statistical, ethical, and managerial steps required before deployment.

Multiple Choice Questions with Answers

Question 1

Which summary is generally most appropriate for a highly right-skewed length-of-stay distribution?

A. Mean and standard deviation only

B. Median and interquartile range

C. Mode and variance only

D. Minimum and maximum only

Correct Answer: B. Median and interquartile range

Question 2

An intervention reduces risk from 10% to 7%. What is the risk difference?

A. 0.3 percentage points

B. 3 percentage points

C. 30 percentage points

D. 70 percentage points

Correct Answer: B. 3 percentage points

Question 3

Which statement best describes a p-value?

A. The probability that the null hypothesis is true

B. The probability that the alternative hypothesis is true

C. The probability, assuming the null model, of a result at least as extreme as observed

D. The probability that the study result will replicate

Correct Answer: C. The probability, assuming the null model, of a result at least as extreme as observed

Question 4

In logistic regression, exponentiating a coefficient usually gives which measure for a one-unit predictor increase?

A. Risk difference

B. Adjusted odds ratio

C. Mean difference

D. Hazard probability

Correct Answer: B. Adjusted odds ratio

Question 5

What is the principal purpose of a confidence interval?

A. To prove that a hypothesis is true

B. To show the range of plausible estimates under a specified procedure and assumptions

C. To eliminate confounding

D. To measure data quality directly

Correct Answer: B. To show the range of plausible estimates under a specified procedure and assumptions

Question 6

In survival analysis, right censoring occurs when:

A. The event occurs before follow-up begins

B. The event time is unknown because observation ends before the event is observed

C. All participants experience the event simultaneously

D. The outcome is measured on a continuous scale

Correct Answer: B. The event time is unknown because observation ends before the event is observed

Question 7

Which feature of time-series data can make ordinary independent-observation methods misleading?

A. Autocorrelation

B. Randomisation

C. Blinding

D. Stratification

Correct Answer: A. Autocorrelation

Question 8

If 20 events occur during 500 person-years of observation, the incidence rate is:

A. 0.004 events per person-year

B. 0.04 events per person-year

C. 25 events per person-year

D. 480 events per person-year

Correct Answer: B. 0.04 events per person-year

Question 9

Which statement about statistical significance is correct?

A. A statistically significant result is always clinically important

B. Clinical importance depends only on sample size

C. Statistical significance should be interpreted with effect size, uncertainty, and practical consequences

D. A non-significant result proves there is no effect

Correct Answer: C. Statistical significance should be interpreted with effect size, uncertainty, and practical consequences

Question 10

Which dashboard practice most improves interpretation of a hospital rate?

A. Display the rate without its denominator

B. Use a truncated axis to emphasise small changes

C. Show the numerator, denominator, time period, and uncertainty where appropriate

D. Combine unrelated outcomes into one score without definitions

Correct Answer: C. Show the numerator, denominator, time period, and uncertainty where appropriate