Reading focus
Syllabus: What Can Statistical Evidence Tell Us?
Distinguish model-based uncertainty from doubts about measurement, missing observations and selective reporting.
Why you’re reading it. Interpret a statistical summary without giving it more authority than its assumptions support.
A question to consider. A school reports p = .08 and calls the intervention ineffective. What additional results and design information would you request before judging that claim?
Why this reading
Greenland and colleagues connect familiar statistical summaries to the assumptions that make them interpretable. Their examples show why a calculation cannot, by itself, establish whether a scientific claim is true.
Before you read
A statistical model includes assumptions about sampling, measurement, analysis and reporting. A p value concerns a test statistic under those assumptions; it is not the probability that the tested hypothesis is true. Confidence-interval coverage describes a procedure across repeated valid applications.
Read for
- For each interpretation, identify what is conditional on the model and what would require additional information.
- Distinguish the estimate’s magnitude and precision from its position relative to a significance threshold.
Consider afterwards
- A narrow confidence interval comes from a selectively reported analysis with substantial missing data. Which uncertainties does its width capture, and which might it miss?
Editorial context
The paper addresses both tests and intervals; switching between them does not remove their shared assumptions. A large p value does not establish equivalence, and interval endpoints are not sharp boundaries between possible and impossible effects. Its cautions also apply to the conditions under which reported Bayesian probabilities are meaningful.