Measuring Disparities¶
measure_disparity.py takes as input a pandas Dataframe containing the following observations: - model predictions and probabilities - binary outcomes - demographics
run measure_disparity.py on model output¶
To run measure_disparity.py on the dataframe, you may use the command line interface, as shown below. It is also possible to import the underlying function like so:
from measure_disparity import measure_disparity
And then call the function. See the API for a full specification of options and requirements.
Below, we demonstrate how to run measure_disparity.py from the command line using a model trained to predict risk of admission to the emergency department using the freely available MIMIC-IV repository.
[1]:
%run ../measure_disparity.py --dataset ../base_model_mimic4_admission.csv
reading in ../base_model_mimic4_admission.csv
demographic columns: ['insurance', 'ethnicity', 'gender']
========================================
Overall Performance
========================================
Measures of predictive bias on the whole population.
╭─────────┬─────────┬───────────────────┬───────┬───────┬────────────╮
│ AUROC │ AUPRC │ Positivity Rate │ FPR │ FNR │ Accuracy │
├─────────┼─────────┼───────────────────┼───────┼───────┼────────────┤
│ 0.881 │ 0.77 │ 0.299 │ 0.08 │ 0.409 │ 0.823 │
╰─────────┴─────────┴───────────────────┴───────┴───────┴────────────╯
========================================
Subgroup Fairness Violations
========================================
Measures the deviation in performance for marginal and intersectional groups.
Note that these deviation are weighted by group prevalence to produce stable estimates when sample sizes are small.
╭─────────────┬───────────────────────────────┬──────────┬─────────────────────┬─────────┬─────────┬───────────────────╮
│ insurance │ ethnicity │ gender │ Brier Score (MSE) │ FNR │ FPR │ Positivity Rate │
├─────────────┼───────────────────────────────┼──────────┼─────────────────────┼─────────┼─────────┼───────────────────┤
│ any │ any │ F │ -0.006 │ 0.007 │ -0.015 │ -0.022 │
│ any │ any │ M │ 0.006 │ -0.01 │ 0.02 │ 0.022 │
│ any │ AMERICAN INDIAN/ALASKA NATIVE │ any │ 0.0 │ 0.024 │ -0.036 │ -0.0 │
│ any │ ASIAN │ any │ 0.001 │ 0.022 │ 0.008 │ 0.0 │
│ any │ BLACK/AFRICAN AMERICAN │ any │ -0.012 │ 0.015 │ -0.058 │ **-0.035 │
│ any │ HISPANIC/LATINO │ any │ -0.004 │ 0.015 │ -0.044 │ -0.012 │
│ any │ WHITE │ any │ **0.015 │ -0.011 │ 0.033 │ 0.047 │
│ Medicaid │ any │ any │ -0.005 │ 0.008 │ -0.048 │ -0.013 │
│ Medicaid │ AMERICAN INDIAN/ALASKA NATIVE │ F │ -0.0 │ 0.021 │ -0.154 │ -0.0 │
│ Medicaid │ AMERICAN INDIAN/ALASKA NATIVE │ M │ -0.0 │ -0.0 │ -0.034 │ -0.0 │
│ Medicaid │ ASIAN │ F │ 0.0 │ 0.036 │ 0.005 │ -0.0 │
│ Medicaid │ ASIAN │ M │ 0.0 │ -0.004 │ 0.056 │ 0.0 │
│ Medicaid │ BLACK/AFRICAN AMERICAN │ F │ -0.002 │ 0.01 │ -0.085 │ -0.006 │
│ Medicaid │ BLACK/AFRICAN AMERICAN │ M │ -0.001 │ 0.009 │ -0.084 │ -0.003 │
│ Medicaid │ HISPANIC/LATINO │ F │ -0.001 │ 0.013 │ -0.07 │ -0.003 │
│ Medicaid │ HISPANIC/LATINO │ M │ -0.0 │ 0.013 │ -0.027 │ -0.001 │
│ Medicaid │ WHITE │ F │ -0.0 │ 0.007 │ -0.018 │ -0.001 │
│ Medicaid │ WHITE │ M │ -0.0 │ -0.006 │ -0.006 │ -0.0 │
│ Medicare │ any │ any │ 0.005 │ -0.024 │ 0.03 │ 0.028 │
│ Medicare │ AMERICAN INDIAN/ALASKA NATIVE │ F │ 0.0 │ **0.047 │ **0.074 │ 0.0 │
│ Medicare │ AMERICAN INDIAN/ALASKA NATIVE │ M │ 0.0 │ 0.045 │ 0.058 │ 0.0 │
│ Medicare │ ASIAN │ F │ 0.0 │ 0.019 │ 0.007 │ 0.0 │
│ Medicare │ ASIAN │ M │ 0.0 │ -0.03 │ 0.046 │ 0.0 │
│ Medicare │ BLACK/AFRICAN AMERICAN │ F │ -0.001 │ 0.013 │ -0.035 │ -0.003 │
│ Medicare │ BLACK/AFRICAN AMERICAN │ M │ -0.0 │ 0.009 │ -0.028 │ -0.002 │
│ Medicare │ HISPANIC/LATINO │ F │ -0.0 │ 0.001 │ -0.039 │ -0.001 │
│ Medicare │ HISPANIC/LATINO │ M │ -0.0 │ 0.001 │ 0.002 │ -0.0 │
│ Medicare │ WHITE │ F │ 0.003 │ -0.026 │ 0.041 │ 0.015 │
│ Medicare │ WHITE │ M │ 0.003 │ -0.046 │ 0.06 │ 0.019 │
│ Other │ any │ any │ 0.001 │ 0.012 │ -0.007 │ -0.014 │
│ Other │ AMERICAN INDIAN/ALASKA NATIVE │ F │ -0.0 │ 0.028 │ -0.076 │ -0.0 │
│ Other │ AMERICAN INDIAN/ALASKA NATIVE │ M │ 0.0 │ -0.001 │ -0.025 │ 0.0 │
│ Other │ ASIAN │ F │ 0.0 │ 0.043 │ -0.012 │ -0.001 │
│ Other │ ASIAN │ M │ 0.0 │ 0.002 │ 0.023 │ 0.001 │
│ Other │ BLACK/AFRICAN AMERICAN │ F │ -0.006 │ 0.017 │ -0.07 │ -0.017 │
│ Other │ BLACK/AFRICAN AMERICAN │ M │ -0.001 │ 0.021 │ -0.041 │ -0.005 │
│ Other │ HISPANIC/LATINO │ F │ -0.002 │ 0.018 │ -0.059 │ -0.006 │
│ Other │ HISPANIC/LATINO │ M │ -0.0 │ 0.023 │ -0.017 │ -0.001 │
│ Other │ WHITE │ F │ 0.004 │ 0.017 │ 0.012 │ 0.001 │
│ Other │ WHITE │ M │ 0.005 │ -0.007 │ 0.041 │ 0.014 │
╰─────────────┴───────────────────────────────┴──────────┴─────────────────────┴─────────┴─────────┴───────────────────╯
Subgroups with Largest Deviations
--------------------
Brier Score (MSE)
----------
- Subgroup: ethnicity=WHITE
- Brier Score (MSE) is 19.9 % higher among this group than the population.
FNR
----------
- Subgroup: insurance=Medicare,ethnicity=AMERICAN INDIAN/ALASKA NATIVE,gender=F
- FNR is 20.4 % higher among this group than the population.
FPR
----------
- Subgroup: insurance=Medicare,ethnicity=AMERICAN INDIAN/ALASKA NATIVE,gender=F
- FPR is 86.0 % higher among this group than the population.
Positivity Rate
----------
- Subgroup: ethnicity=BLACK/AFRICAN AMERICAN
- Positivity Rate is 44.9 % lower among this group than the population.
saving results to df_fairness.csv
Visualizing results¶
Running measure_disparity.py produces a df_fairness.csv file containing the resultant fairness metrics. Below, we show how to use this to generate additional figures for assessing model performance and bias.
[2]:
df_fairness = pd.read_csv('df_fairness.csv')
df_plt = (df_fairness
.melt(
id_vars = ['insurance','ethnicity','gender']
)
)
df_plt = df_plt.loc[~df_plt.variable.str.contains('Brier')]
df_fairness
import seaborn as sns
sns.set_theme(style='whitegrid',font_scale=1.3)
g = sns.catplot(
kind='bar',
edgecolor="0.8",
data=df_plt,
row='variable',
col='ethnicity',
x='gender',
hue='insurance',
y='value',
sharey='row',
aspect=0.75,
palette='Set3'
)
g.refline(y=0)
# make titles nicer
nice_cols = {
'AMERICAN INDIAN/ALASKA NATIVE':'AI/AN',
'BLACK/AFRICAN AMERICAN':'BLACK',
'HISPANIC/LATINO':'HISP/LTN',
'WHITE':'WHITE',
' any ':'ANY',
'ASIAN':'ASIAN'
}
for (row,col),ax in g.axes_dict.items():
if 'any' in col:
ax.set_ylabel(row)
if row == 'FNR':
ax.set_title(nice_cols[col])
else:
ax.set_title('')
