Reading your statistics

Start here

Almost every result in the Data Hub reports a p-value, a number that tells you how surprising your data would be if nothing were really going on. A small p-value is a useful flag, but it does not tell you how big an effect is or whether it matters. For that you need the effect size and its confidence interval. That foundation page is worth reading before anything else, because every test below reports a version of the same three things, a direction, a size, and a range of uncertainty.

• Effect sizes and confidence intervals, why the size of a difference matters more than a bare p-value, and how to read a 95% confidence interval.

Comparing groups

These tests answer "are these groups really different, and by how much?" Reach for them when you have a measured outcome (a fluorescence reading, a growth rate, a concentration) split across two or more conditions.

• ANOVA, post-hoc tests, and two-way ANOVA, comparing three or more groups at once, then finding which specific pairs differ.
• Repeated measures, mixed models, and nested designs, for when the same subjects are measured more than once, or your replicates are cells within mice rather than independent samples.

Relationships

These answer "do these two things move together, and can I predict one from the other?" Reach for them when both of your variables are measured numbers rather than groups.

• Correlation and regression, measuring how tightly two variables track each other, and fitting a line you can read a slope off of.

Curves

Some biology is not a straight line. A drug saturates, a binding curve plateaus, an enzyme runs out of substrate. These pages cover fitting a shaped curve and reading the parameters that summarize it.

• Dose-response curves, covering EC50 and IC50, the 4PL and 5PL sigmoid fits, the Hill slope, and comparing or sharing fits across datasets.

Survival and time-to-event

When your outcome is "how long until something happens," and some subjects have not had it happen yet, you need methods built for that. Reach for these for time to relapse, time to death, or time to any defined event.

• Survival curves and hazard ratios, covering Kaplan-Meier curves, the log-rank test, and the Cox hazard ratio.

Counts and categories

When your data are counts in categories (responded or did not, mutant or wild type) rather than measured numbers, you compare proportions instead of means.

• Contingency tables, odds ratios, and relative risk, covering chi-square and Fisher exact tests, plus the odds ratio that logistic regression also reports.

Screening and data quality

Two practical tools that sit alongside the rest, one for judging how well a measurement separates two groups, one for the honest question of whether a stray point is really an error.

• ROC curves and AUC, how well a continuous measurement tells two groups apart, with sensitivity, specificity, and a cut point.
• Outlier tests, the Grubbs test, and the honest caution that removing a data point needs a real reason.

Planning and assumption checking

Before you run a test, the power and sample-size planner lets you work out three questions in any direction: given an effect size, alpha, and a target power, how large does your sample need to be? Given the sample you already have, what power do you achieve? Given your sample and power target, what is the smallest effect you could detect? The planner covers the two-sample t-test, the paired t-test, one-way ANOVA, and Pearson correlation.

After you have data, the assumption Report Card runs the checks your chosen test relies on and surfaces them in plain language. For parametric tests it runs a Shapiro-Wilk normality test on each group and a Levene or Brown-Forsythe equal-variance test across groups, and tells you whether each assumption passes at your chosen alpha. A failing check is a prompt to look at a nonparametric alternative, not a verdict that the parametric result is wrong.