Five minutes with dropped_out in clinical-trial

Someone asked what is in dropped_out on clinical-trial, and the honest answer took one query.

900 rows, no nulls, 2 distinct values. Whether the subject left before week 12.

It is true in 9.9% of rows — 89 of 900.

That imbalance is the single most important fact about the column. It sets the baseline any model has to beat, it decides whether a per-segment breakdown will have enough rows in the minority class to say anything, and it determines how wide the confidence interval on any rate computed from it will be.

select
  count(*)                      as rows,
  count(dropped_out)            as present,
  count(*) - count(dropped_out) as nulls,
  count(distinct dropped_out)   as distinct_values
from clinical_trial;

A base rate this low is the reason accuracy is the wrong metric on anything predicting this column: always predicting false scores 90.1% and learns nothing.

Where this bites: slicing by a dimension with 4 levels leaves roughly 22 true rows per slice on average. That is thin enough that the noisiest segment will look like the most extreme one, every time, and someone will read the ranking as a finding.

The grain is one row per enrolled subject, which is the context every one of those numbers depends on. None of them survive a change of grain, which is why "profile the column" and "profile the table" are the same job. Full schema, and the CSV, on the dataset page.