Data Analysis
Choosing the analysis
Match method to question. Descriptive statistics (counts, means, medians, rates, distributions) characterize the population; watch for skew, where the median is the better summary. Comparative questions call for stratified rates, confidence intervals, and hypothesis tests. Causal questions demand more: a defined exposure, an outcome, confounding control (stratification, regression, matching), and attention to selection bias. Prediction questions need train/test splits and performance measures such as sensitivity, specificity, AUC, and calibration.
Rates, denominators, and risk adjustment
Most healthcare measures are proportions. Get the denominator right - eligible population, exclusions, attribution, and measurement period - because a mismatched denominator is one of the most common analytic errors. When comparing providers, risk-adjust for case mix; without it, caring for sicker patients looks like poor performance.
Bias and missing data
Know the main threats: selection bias, information bias, confounding, and measurement error. Missing data is rarely random; state the mechanism, describe the handling approach (complete case, imputation), and test whether conclusions are sensitive to it.
Traps and worked example
A p-value is not the probability that the null is true, nor a measure of effect size; with large samples, trivial differences become "significant." Correlation is not causation, and statistical significance is not practical significance. Two clinics show 12% and 8% readmission rates. Before declaring the second better, check case mix, measure specification, and sample size. If the first serves older, more comorbid patients, risk adjustment may erase the gap; if the second's denominator is 40 patients, the confidence interval is wide. Adjust, test, then interpret.