Biostatistics & evidence interpretation

1. Describe the data and its uncertainty

• Mean and SD summarize the center and spread of suitable continuous data; median and interquartile range are often more informative for markedly skewed data. State the unit and distribution.
• SD describes variation among observations. SE describes uncertainty in an estimate; for an independent sample mean, SE = SD/√n. Increasing n can reduce SE without reducing individual variability.
• For a mean under a suitable normal approximation, 95% CI = mean ± 1.96 × SE. Example: 12.0 ± 1.96 × 1.5 = 9.06 to 14.94. When population variability is estimated in smaller samples, an appropriate t critical value is commonly needed.
• In repeated sampling under the model, a 95% confidence-interval method captures the fixed population parameter about 95% of the time. One observed CI is not the range containing 95% of patient outcomes.

Pause & recall
Does a smaller SE mean that individual patients are less variable?

Sources: CDC Principles of Epidemiology: Statistical Inference · FDA/ICH E9: Statistical Principles for Clinical Trials

2. Keep risk, rate and effect measures distinct

Swipe or scroll to read the full table.

• ARR and benefit NNT here assume treatment reduces an adverse outcome. If the absolute difference is zero, there is no finite NNT. If treatment increases the adverse outcome, interpret the increase as harm rather than reporting a negative benefit NNT.

Pause & recall
Can a 50% relative reduction have an ARR of only 1 percentage point?

Sources: CDC Principles of Epidemiology: Measures of Association · CDC Principles of Epidemiology: Morbidity Frequency Measures · Cochrane Handbook chapter 15: Interpreting results and drawing conclusions · Cochrane Handbook chapter 6: Choosing effect measures and computing estimates

3. Build the diagnostic two-by-two table

Swipe or scroll to read the full table.

Pause & recall
Which denominator answers: “Among positive tests, who has disease?”

Sources: FDA: Statistical Guidance on Reporting Results from Studies Evaluating Diagnostic Tests

4. Interpret tests without overstating certainty

• A p value describes the extremeness of data under a specified null model. It is not the probability the null is true, the chance a patient benefits, or an effect-size measure.
• Type I error rejects a true null; alpha is its planned probability. Type II error fails to reject a false null; beta describes it for a specified alternative. Power = 1 − beta. Sample size, variability, effect size and design affect power.
• For conventional two-sided tests, a ratio CI is compared with 1; a difference CI with 0. An interval crossing the null does not prove no effect. Evaluate its precision and whether it allows clinically important benefit or harm.
• Noninferiority requires a defensible prespecified margin and suitable trial conduct. For failure risk (new minus control), a CI of −1 to +3 percentage points lies below a +5-point harm margin: noninferiority, but not superiority. This does not prove equality.
• Multiple outcomes and repeated analyses require a prespecified error-control strategy for confirmatory claims. A significant result in one subgroup and a nonsignificant result in another do not establish an interaction.

Pause & recall
Why does p > 0.05 fail to prove equivalence?

Sources: CDC Field Epidemiology Manual: Analyzing and Interpreting Data · FDA/ICH E9: Statistical Principles for Clinical Trials · FDA: Non-Inferiority Clinical Trials to Establish Effectiveness · FDA: Multiple Endpoints in Clinical Trials

5. Ask how the comparison was created

Swipe or scroll to read the full table.

Pause & recall
Why can rescue therapy appear harmful in an observational study?

Sources: FDA/ICH E9: Statistical Principles for Clinical Trials · FDA/ICH E9(R1): Estimands and Sensitivity Analysis · CDC Field Epidemiology Manual: Analyzing and Interpreting Data

6. Match the analysis to the question

Swipe or scroll to read the full table.

Pause & recall
What additional information is needed before choosing paired versus independent analysis?

Sources: NIST/SEMATECH: Two-sample t test for equal means · NIST/SEMATECH: ANOVA table and tests of means · CDC Field Epidemiology Manual: Analyzing and Interpreting Data

7. Translate the result into a clinical decision

• Define the population, intervention, comparator, outcome and time frame. Assess whether the trial population and setting resemble the patient or service in question.
• Report absolute effects alongside relative effects and uncertainty. NNT depends on baseline risk, outcome definition and duration; it is not a universal property of a drug.
• Inspect individual components of composite endpoints. A common laboratory or hospitalization component may drive a result without proving a separate mortality benefit.
• Look for prespecified analyses, complete outcome reporting, losses to follow-up, important harms and practical treatment burden. Statistical significance alone does not establish a worthwhile net benefit.
• For diagnostic studies, examine patient spectrum and the reference standard. Performance in obvious disease and healthy controls may not transfer to patients with uncertain symptoms.

Pause & recall
Does a significant composite outcome establish a mortality benefit?

Sources: FDA: Multiple Endpoints in Clinical Trials · FDA: Statistical Guidance on Reporting Results from Studies Evaluating Diagnostic Tests · Cochrane Handbook chapter 15: Interpreting results and drawing conclusions

8. Read the units in health economics

Swipe or scroll to read the full table.

Pause & recall
Does an ICER below a single dollar figure prove that a treatment is appropriate for every patient?

Sources: CDC: Economic Evaluation — Part V, Cost-Effectiveness Analysis