Analyses built to order

Which employee attributes actually predict attrition? (asked in plain English of the public IBM HR attrition dataset)

The data. card_data.json, the numbers file published with the delivered report IBM HR Employee Attrition Drivers. The underlying data is the public IBM HR attrition dataset.

The calculation. Python, standard library only. This exact code ran against the file when this page was built, and printed the output below it.

import json
from math import exp, sqrt
card = json.load(open("card_data.json"))
table = card["rf_data"]["summary_table"]["tables"]["attrition_by_segment"]
ot = {r["segment_value"]: r for r in table if r["segment"] == "Overtime"}
yes, no = ot["Yes"], ot["No"]
p1 = yes["attrited"] / yes["headcount"]
p0 = no["attrited"] / no["headcount"]
rr = p1 / p0
se = sqrt(1 / yes["attrited"] - 1 / yes["headcount"]
          + 1 / no["attrited"] - 1 / no["headcount"])  # log risk ratio
print("overtime:", yes["attrited"], "of", yes["headcount"], f"left ({p1:.1%})")
print("no overtime:", no["attrited"], "of", no["headcount"], f"left ({p0:.1%})")
print("risk ratio", round(rr, 2), "95% interval",
      round(rr * exp(-1.96 * se), 2), "to", round(rr * exp(1.96 * se), 2))

overtime: 127 of 416 left (30.5%)
no overtime: 110 of 1054 left (10.4%)
risk ratio 2.93 95% interval 2.33 to 3.68

What it says. The delivered report's own table shows 30.5% of employees who worked overtime left, against 10.4% who did not: overtime workers left about 2.9 times as often, with a 95% interval of 2.33 to 3.68. Recomputing a headline figure from the published table is a check anyone can repeat on this report.

What it does not say. This is a raw comparison of two groups. Overtime goes with job role and level, which the report's logistic regression adjusts for; there the odds ratio for overtime is 5.97 (95% interval 4.20 to 8.55). An odds ratio is not a risk ratio and runs larger when the outcome is common, so the two figures are not meant to match. Neither is proof that overtime causes attrition.

Open the full report to see the methods the build chose and the diagnostics behind them.