experimental-design

Use the randomization script's block_randomization function with a seed to generate a permuted-block allocation schedule as a pandas DataFrame. For balancing a prognostic variable like site or sex, use stratified_block_randomization, which block-randomizes independently within each stratum.

A full 2^k factorial estimates all main effects and interactions but needs 2^k runs, practical only up to about 5 factors. A fractional factorial runs a chosen fraction of those runs, trading some interaction estimates for efficiency, with the tradeoff summarized by the design's resolution and alias structure.

Use Plackett-Burman when screening many factors (5 to 20 or more) to find the vital few, since it needs only the next multiple of 4 above k runs. It estimates main effects only and heavily confounds two-factor interactions, so follow up with a higher-resolution design.

Pseudoreplication is treating repeated measurements of one unit as independent replicates, such as counting 100 cells from one mouse as n=100. Replicate at the level the treatment is randomized to, and analyze with a mixed model that respects the nesting structure.

No, sample size and power calculations are handled by the separate statistical-power skill once the design is chosen. This skill covers design selection, randomization, blocking, and generating the experimental layout before data collection.

The scripts require numpy, pandas, and pyDOE3, installed into the shared office-skills virtual environment with uv. pyDOE3 supplies the factorial, Plackett-Burman, central composite, Box-Behnken, and Latin hypercube generators.