GraphHelix @GraphHelix

When the rule is violated, switch to Fisher's exact. Modern algorithms handle large tables fine — the old Fisher's only works for 2×2 restriction is gone. And: turn off Yates's continuity correction in R unless you have a reason.

Cochran (1954): no expected count below 1, and no more than 20% of cells with expected counts below 5. For 2×2, that means zero cells below 5. For 3×3, at most one cell below 5. Always check expected counts, not observed.

Chi-square test of independence: looks easy in software, but the footnote X cells have expected count less than 5 trips up most practitioners. The fix is Cochran's rule and Fisher's exact — not ignore the warning.

Skip the effect size and reviewers will catch it. 2×2: phi or odds ratio with 95% CI. Larger: Cramér's V. p-values scale with N; effect size answers how strong is the association — the practical question. graphhelix.ai/blog/chi-squar…

Chi-square test of independence: the assumption that bites is expected cell counts. Cochran's rule: no expected count <1, and no more than 20% of cells with expected <5. For a 2×2 table, that means ZERO cells below 5. Below that → Fisher's exact. graphhelix.ai/blog/chi-squar…

Power loss vs. Student's when variances really ARE equal: typically <1%. Power loss isn't the worry. Type I error inflation under heterogeneity is. Full comparison and verdict by researcher type: graphhelix.ai/blog/welch-vs-…

The Levene's-then-pick workflow doesn't help either. Levene's lacks power at small n, so p > .05 usually means we couldn't detect unequal variance, not variances are equal. Skip the test, default to Welch's. Delacre, Lakens & Leys (2017) lays out the case.

You have two groups, continuous data, and you reach for a t-test. The software gives you a choice: Student's (equal variance) or Welch's (unequal). Most introductory training points the wrong way.

SPSS labels them equal variances assumed vs not assumed. R defaults to Welch's; Prism defaults to Student's. Use Welch's by default. It controls Type I error correctly under both equal AND unequal variances. Details: graphhelix.ai/blog/welch-vs-… #stats

And: if your data are cross-sectional, mediation is still informative but the causal direction can't be established from the data alone. Say so in limitations. Full worked example: graphhelix.ai/blog/mediation…

Standard now: 5,000 bias-corrected and accelerated (BCa) bootstrap resamples. R's mediation package: boot=TRUE, boot.ci.type=bca, sims=5000. Report a×b, its 95% CI, the resample count, and the package version used.

Mediation analysis: testing X → M → Y. The interesting quantity is a×b, the indirect effect carried through M. Baron & Kenny's 1986 causal-steps method never tested a×b directly — modern practice is to bootstrap it with a CI.

Color: red-green fails for ~8% of male readers. Blue-white-red (RdBu) for row-scaled data, viridis for raw log2 expression. Code with pheatmap + when to switch to ComplexHeatmap: graphhelix.ai/blog/heatmap-g…

Clustering: 1−Pearson distance treats genes that go up/down together as similar regardless of expression level. Ward.D2 linkage gives readable cluster blocks. pheatmap defaults give neither.

Default heatmap in base R: red-green palette, raw values, Euclidean distance, complete linkage. Four bad defaults you have to override deliberately for a publication-quality gene expression heatmap.

Reviewer asks for a bootstrapped CI on the indirect effect and you reach for Sobel. Don't. Sobel assumes the product a×b is normally distributed — it isn't. Use 5,000 BCa bootstrap resamples instead. Worked example: graphhelix.ai/blog/mediation… #stats #rstats

Row-scaling raw RNA-seq counts in your heatmap? A few highly expressed genes dominate the SD and compress everything biological toward zero. log2(x+1) FIRST, then row-scale. Plus: Pearson distance + Ward.D2 linkage beats the defaults. graphhelix.ai/blog/heatmap-g… #RNAseq