Skip to content

Group Comparisons

Summary

PyFLASH group comparisons test whether numeric measurements differ between groups. The core comparison helper is used by bar plots and by the group_comparison pipeline. The pipeline adds marker-by-marker tables, optional screening corrections, effect sizes, power columns, and saved figures.

Where PyFLASH Uses It

  • plot_mean_bars uses the shared comparison helper when drawing significance annotations and writing statistics tables.
  • group_comparison runs planned comparisons across markers and can use auto, mixed, or bootstrap engines.
  • plot_group_matrix and plot_effect_forest display group-effect summaries after comparison results have been computed.
  • Structured report records can summarize the selected test, group descriptives, p-values, pairwise rows, and effect-size text.

Inputs

Typical inputs are numeric measurement columns, a grouping column, and optional comparison labels. Plot-level functions may receive groupList; pipeline runs usually use group_col, control, comparisons, factor, and row filters.

The shared comparison helper drops non-numeric and missing values before testing. Each non-empty group is counted after this filtering. When comparisons are not provided, PyFLASH uses all pairwise group combinations.

The group_comparison pipeline can run:

  • auto: marker-wise group comparisons using the shared helper.
  • mixed: ROI-level mixed modelling where ROI rows and nesting information are available.
  • bootstrap: ROI-level bootstrap summaries where ROI rows are available.

If ROI-level inputs are not available or a nested model cannot be fitted, the pipeline falls back to the auto engine for that marker and records the engine used in the output.

Outputs

Common output columns include:

  • marker, comparison, reference, and group.
  • test and engine, describing the method used.
  • n_reference, n_group, mean_reference, and mean_group.
  • percent_change, effect-size columns, and confidence interval columns.
  • p, optional q, and significant.
  • omnibus_p, icc, achieved_power, and required_n_80 where available.

The pipeline also writes descriptives, skipped-marker tables, omnibus summaries, figures, and a manifest.json file in its run folder.

Interpretation

For two groups, the shared helper uses an independent two-sample t-test when both groups have enough values. If a group is too small for that route, it uses a Mann-Whitney U test. For three or more groups, PyFLASH checks distributional support and chooses between one-way ANOVA with Tukey post-hoc comparisons or Kruskal-Wallis with Conover/Dunn-style post-hoc comparisons.

The normality check pools group values and combines several tests when enough data are available. It is a routing aid for multi-group comparisons, not a proof that assumptions are fully satisfied.

Read a significant comparison as evidence that the tested groups differ for the specified marker under the selected method and filtering. Read it alongside the effect size, confidence interval, sample sizes, skipped rows, and correction family.

Limitations

The automatic route does not establish causality and does not replace a planned statistical design. Small groups, zero variance, many missing values, repeated measurements treated as independent, and post-hoc selection can make p-values and effect sizes unstable.

For multi-group auto comparisons in the pipeline, PyFLASH uses one-way comparison logic. Use the linear-model tools when covariates, categorical terms, interactions, or adjusted means are the primary question.

See Also