MEDICAL STATISTICS FOR NON-STATISTICIANS

Course Content

​

Statistics is a key method in medical research and evidence-based medicine (EbM). A solid foundation in statistics is therefore essential for understanding medical literature. In particular, the critical appraisal of original research articles in journals on experimental, clinical, epidemiological, and health economic studies requires sound statistical knowledge.

This two-day course provides a comprehensive introduction to, or thorough review of, the relevant statistical fundamentals. By the end, participants will be able to critically assess and evaluate the literature on evidence-based medicine and the underlying study results.

Seminar Structure
The seminar begins with an overview and classification of the statistical methods used in medicine. Measurement scales as defined by Stevens (1946) are discussed, as they provide robust guidance in selecting appropriate statistical methods. Using the EQ-5D-5L as an example, we examine quality criteria (objectivity, reliability, and validity) of measurement instruments.

Descriptive Statistics include measures of central tendency, measures of variability, and higher-order parameters (skewness and kurtosis). Each of these is discussed using practical data (e.g., length of hospital stay), with comparisons of results (differences between mean and median, impact of outliers, efficiency of estimators, etc.). Relevant measures of variability include range, percentile-based methods (interquartile range), variance, and standard deviation. The influence of outliers is considered throughout. Interpretation skills are practiced using examples from the medical literature and dossiers (early benefit assessment under §35a SGB).

Hypothesis Testing is introduced using the 2x2 contingency table and the chi-square test of independence, worked through in full with a practical example. The 2x2 table is a commonly used tool for analyzing two medications (or medication vs. placebo) in relation to a binary categorical endpoint (e.g., survival yes/no). We work through the entire process together: formulating hypotheses, setting the alpha significance level, and discussing the beta error and the resulting statistical power. Both the critical value (classical approach) and the p-value (computer-based method) are calculated. The results are then interpreted with respect to the null and alternative hypotheses. Standardized effect sizes (e.g., Cramer’s V, Cohen’s w) are calculated to aid interpretation.

The 2x2 table is also used to calculate risk parameters frequently applied in medicine, in particular risk ratio and odds ratio. Both are calculated using a practical example, with the opportunity to explore how changes in the table values affect the parameters. A confidence interval is also calculated for the odds ratio. The IQWiG methods report serves as an example to discuss thresholds (e.g., categorization of added benefit) based on confidence intervals. We then interpret results of early benefit assessments, using publications in the Federal Gazette, which provide a compact overview.

The first course day concludes with an overview of systematic reviews and meta-analyses (quantitative evidence synthesis). Practical examples from the research literature and dossiers are examined.

The second day begins with an overview of key diagnostic concepts based on false-positive and false-negative rates, especially sensitivity and specificity. The ROC curve is introduced as a commonly used tool.

We then turn to multivariate methods. Again, Stevens’ (1946) measurement scales provide useful guidance for systematic classification. We look in more detail at correlation analysis before addressing logistic regression analysis. Logistic regression is a robust method for estimating categorical outcomes (e.g., treatment effect yes/no) and also allows assessment of the influence of multiple covariates (e.g., comorbidities of patients included in the study). Knowledge of odds ratios is applied here, since logistic regression coefficients correspond to odds ratios. Interpretation is therefore straightforward but is nevertheless practiced intensively.

The final block focuses on survival analysis (also known as time-to-event analysis), an important area of statistics in medicine. Here, the key outcome is the time to the event of interest. If the event does not occur during the study period, this results in censored data. Traditionally, such data were analyzed using life tables—we review the official life tables published by the German Federal Statistical Office in Wiesbaden as an example. Because such tables take up substantial space, modern medical journals typically use graphical representations: the Kaplan-Meier estimator and the corresponding plots. These illustrate event occurrences over time, allow representation of censored observations, and can display separate curves for covariates (e.g., treatment A vs. treatment B), enabling direct comparison. To test curve differences, the log-rank test and derived robust methods are applied. The concepts learned in the section on hypothesis testing (2x2 tables) provide a foundation for understanding this test.

Finally, Cox regression is introduced. This method estimates the impact of covariates on survival curves by calculating hazard ratios for different covariates. Cox regression is particularly useful in studies where patient randomization to treatment regimens is not feasible but multiple covariates are available.

A core element of the course is the intensive practice of procedures using practical examples, as this is the only way to ensure the highest possible learning success.