Healthcare Readmission/fair.py keeps number_outpatient on the grounds it's "less racially stratified than inpatient" - the actual data shows the opposite

[yakew7]

Where: Healthcare Readmission/fair.py, the mitigated-model feature-list comment.

The gap: The mitigated-model feature list drops number_inpatient as a proxy but keeps number_outpatient, justified by:

    'number_outpatient',       # retained: outpatient visits (access to care signal,
                                #           less racially stratified than inpatient)

Using the script's own binary race split (is_minority = ~race.isin(['Caucasian','Asian'])), the opposite holds:

number_inpatient  by is_minority:  0=0.633   1=0.677   (abs. gap 0.044)
number_outpatient by is_minority:  0=0.410   1=0.249   (abs. gap 0.161)

The racial gap in number_outpatient (0.161) is roughly 3.7x the gap in number_inpatient (0.044) - number_outpatient is more racially stratified, not less, so the stated rationale for keeping it in the "fair" feature set while dropping number_inpatient is backwards.

Repro:

python3 -c "
import pandas as pd
df = pd.read_csv('Healthcare Readmission/diabetic_data.csv')
df = df[~df['race'].isin(['?'])]
df = df[df['gender'] != 'Unknown/Invalid']
df['is_minority'] = (~df['race'].isin(['Caucasian','Asian'])).astype(int)
print(df.groupby('is_minority')[['number_inpatient','number_outpatient']].mean())
"
#              number_inpatient  number_outpatient
# is_minority
# 0                    0.632968           0.410347
# 1                    0.677303           0.248594

Fix direction: Either correct the comment (number_outpatient is in fact more racially stratified by raw group means, so retaining it needs a different justification, e.g. weaker correlation in the trained model's actual feature importance), or reconsider dropping it alongside number_inpatient if the goal is minimizing racial-proxy signal.

[yakew7]

Fixed in 5491aa4: Healthcare Readmission/fair.py's number_outpatient retention comment claimed it was "less racially stratified than inpatient" - verified via df.groupby('is_minority')[['number_inpatient','number_outpatient']].mean(), which shows the opposite (0.161 vs 0.044 gap). Reworded the comment to state the real numbers and the actual justification (clinical relevance, not lower stratification).