| name | statistics |
| description | Choose, run and report the right analysis: descriptives before inference, test selection driven by the design, assumptions checked with a statistic rather than assumed, effect sizes with intervals as the headline and p-values as support, multiplicity handled, power computed before data collection. Use whenever the task involves a test, a p-value, a sample size, a regression, or a claim that two groups differ.
|
| version | 1.0.0 |
| author | Remedy |
| tags | ["statistics","inference","effect-size","power","regression","research"] |
| requires | [] |
| tools | ["power_analysis","stats_assumptions","stats_effect_size","stats_multiplicity","data_profile","analysis_run","skill_activate","file_read","file_write"] |
| triggers | ["\\b(p[- ]?values?|confidence intervals?|effect sizes?|statistical(?:ly)?\\s+(?:significan\\w+|power))\\b","\\b(t-?test|ANOVA|chi[- ]?squared?|Mann[- ]?Whitney|Wilcoxon|Kruskal[- ]?Wallis|Fisher.?s exact)\\b","\\b(sample size|power analysis|a ?priori power|multiple comparisons?|Bonferroni|Benjamini|false discovery rate|\\bFDR\\b)\\b","\\b(mixed[- ]effects?|random effects? model|regression (?:coefficients?|assumptions?)|heteroscedastic\\w*|collinearit\\w+|multicollinearit\\w+)\\b"] |
Statistics
skill_activate(skill="research-method") first, and do not restate the
spine here — that pack owns question framing, evidence standards,
preregistration, citation honesty and how to say "we do not know". This
pack owns only the analysis itself.
Order of work (do not skip a step to get to the p-value)
- Write the estimand in one sentence before touching a test: what
quantity, in what population, contrasted against what? "Is there a
difference?" is not an estimand. If the owner cannot answer it, ask —
the test follows from it, not the other way round.
- Describe.
data_profile(path, target=<outcome>) → n, missingness,
duplicates, cell sizes, class balance, oddities. Read the warnings and
leakage_suspects before any model. A distribution you have not looked
at will surprise you later.
- Select by design, not by habit —
references/test-selection.md.
Between/within/paired, outcome type (continuous, binary, count, ordinal,
time-to-event), independence structure. stats_assumptions(data_path, outcome=, group=, design=) returns recommended[] with why and
when_it_fails, plus fallbacks[].
- Check assumptions with a number. Same call reports normality of
residuals (D'Agostino-Pearson, refused below n=20 on purpose), variance
homogeneity (Brown-Forsythe), outliers, balance, zero-inflation. When a
check fails,
references/assumptions-and-diagnostics.md says what to do
about it — the answer is rarely "transform and carry on".
- Estimate before you test.
stats_effect_size(kind=..., ...) →
estimate with CI. That pair is the result. Report the p as support.
- Handle multiplicity. If the analysis implies more than one test,
stats_multiplicity(pvalues=..., labels=..., method="holm"|"bh") and
say what the family was.
- Power belongs before the data.
power_analysis(test=, solve="n", effect_size=<the smallest effect worth detecting>). Post-hoc power is
arithmetic on the p-value and carries no information — if the owner
asks for it, give the confidence interval or the minimum detectable
effect at the achieved n instead, and say why.
- Report the whole recipe: n per cell, exclusions and why, the test,
the estimate with its interval, the exact p, the software and versions.
Hard rules
- Never pick or switch the test after seeing the p-value; never drop a
covariate, an outlier or a subgroup because it moved significance. If
you did try alternatives, all of them go in the report.
