Power¶

Summary¶

Power tools estimate the probability of detecting an effect under a specified test model, effect size, alpha, and sample size. PyFLASH uses power calculations as planning and diagnostic summaries, not as proof that an observed result is true or false.

Where PyFLASH Uses It¶

plot_power_curve draws expected power across sample sizes and effect sizes.
group_comparison can report achieved power and estimated sample size required for 80 percent power.
data_overview can report minimum detectable effect summaries for marker contrasts.

Inputs¶

Power calculations need an effect size, alpha, sample size, and test family. The main helper functions use two-sample t-test assumptions for standardized differences. Plot-level power curves can also draw ANOVA-style curves depending on the requested configuration.

For group-comparison outputs, observed data are used to estimate an observed effect size and current sample sizes. Required sample size estimates are reported per group under the configured target power.

Outputs¶

Power-related outputs can include:

achieved_power: estimated power for the observed sample size and effect.
required_n_80: estimated per-group sample size for 80 percent power.
effect_size, n_per_group, and power rows from power-curve tables.
minimum detectable effect summaries in data-overview outputs.
power-curve figures and optional CSV tables.

Interpretation¶

Power is a model-based planning quantity. A higher value means the specified analysis would more often detect the specified effect if the assumptions held. Required sample size estimates are useful for planning similarly structured future experiments.

Post-hoc achieved power is descriptive. It is largely a transformation of the observed effect and sample size, so it should not be used as independent evidence that a non-significant result proves no difference.

Limitations¶

Power estimates depend strongly on assumptions about variance, independence, effect size, alpha, balance, and the selected test. Estimates from small pilot datasets can be noisy. ROI-level rows should not be treated as independent subjects unless the analysis design supports that.

Data-overview minimum detectable effect summaries are screening aids. They are not substitutes for a prospective power analysis matched to the final design.