qinyan-nature-statistics

Start from the design, not the data: define the estimand, the independent experimental unit, and the replication hierarchy first. Then select a model matching the outcome type and structure, such as mixed models for repeated measures or generalized models for counts, rather than relying on normality tests alone.

Run the included reporting_audit.py script on your Methods, Results, or legend text with the matching --context flag. It checks for required elements like sample size, named tests, effect estimates, p values, and software versions, and flags risky phrases such as p = 0 or claims of no difference.

The independent unit is the smallest entity independently assigned to a condition, such as an animal, an independent cell culture, or a participant. Technical replicates, fields of view, or repeated measurements of the same unit do not increase n and should be summarized or modeled hierarchically.

No. A non-significant p value does not establish equivalence or absence of a difference. Equivalence claims require a design that supports them, such as a pre-specified equivalence margin; otherwise report the effect estimate and confidence interval instead.

The script uses pattern matching to detect the presence of reporting elements, not their statistical validity. It cannot verify that a chosen test is appropriate, that numbers are correct, or that assumptions hold; those judgments require the design-first review workflow.