wwu-mmll / confound_corrected_cpm / 35736801717

88%

main: 88%

Build Type

push

github

Committed by NilsWinter

Commit Message

Phase 10: make the report say which analysis produced it

Phase 8 made confound control two run-level choices, which means one report is
one cell of the 2x2 and nothing can be compared side by side inside it. That
makes it the report's job to say which cell it is -- and it did not. A naive run
and a fully controlled one opened with the identical sentence:

  The connectome model predicted target: r = 0.31, p = .020 (5-fold CV).
  The connectome model predicted target: r = 0.34, p = .020 (5-fold CV).

`selection_input` and `model_input` were recorded only in the configuration
table at the bottom of the Data & Methods appendix. Anyone handed the HTML could
not tell the analyses apart.

Now: `run_config.json` sits alongside `task_type.txt` (a file rather than a
parse of cpm_log.txt, so a run with logging turned down still produces a report
that knows its own configuration); the headline and a stat chip name the cell --
none / edge selection only / features only / edge selection and features -- and
the Model Comparison section explains what that means for the numbers below it.
An older results directory has no run_config.json and says nothing rather than
guessing: claiming "none" for a run that was in fact controlled would be worse
than silence.

Also here, all things a reader would otherwise misread:

  - the glossary describes the models *this* run produced. Under
    model_input='residualized', `connectome` is the deconfounded-strength model
    while carrying the same row label as the naive one.
  - the em dash in the increment's Pearson column is explained (Fisher z /
    Steiger, not subtraction), so it does not read as a failed computation.
  - parse_config_block folds wrapped lines into the previous value. A multi-line
    estimator repr was producing rows like "threshold=[0.05])]," with an empty
    value in the configuration table.

Two test-quality notes worth keeping:

  - the headline test originally read the flattened page text, so it passed... (continued)

Coverage Stats

28 of 29 new or added lines in 5 files covered. (96.55%)

3 existing lines in 1 file now uncovered.

2504 of 2869 relevant lines covered (87.28%)

0.87 hits per line