Business Context
A healthcare provider is assessing a model that predicts the likelihood of patient readmission within 30 days. The dataset is imperfect, with some missing readmission statuses (target variable). Furthermore, they want to give more importance to high-risk patients (e.g., those from the ICU) during the assessment by using a weight variable.
About the Set : percentile
Precise calculation of percentiles and quantiles.
Discover all actions of percentile Data Preparation
Create a patient dataset with missing target values ('READMITTED_30D'). A 'PATIENT_WEIGHT' is assigned, giving higher weight to ICU patients. 'P_READMIT' is the model's predicted probability.
Copied!
| 1 | DATA casuser.patient_readmission; |
| 2 | call streaminit(789); |
| 3 | DO PATIENT_ID = 1 to 2000; |
| 4 | ICU_STAY = rand('bern', 0.2); |
| 5 | IF ICU_STAY = 1 THEN DO; |
| 6 | PATIENT_WEIGHT = 2.5; |
| 7 | base_prob = 0.4; |
| 8 | END; |
| 9 | ELSE DO; |
| 10 | PATIENT_WEIGHT = 1.0; |
| 11 | base_prob = 0.15; |
| 12 | END; |
| 13 | P_READMIT = base_prob + rand('uniform')*0.2; |
| 14 | READMITTED_30D = rand('binomial', P_READMIT, 1); |
| 15 | /* Introduce missing targets for 5% of records */ |
| 16 | IF rand('uniform') < 0.05 THEN call missing(READMITTED_30D); |
| 17 | OUTPUT; |
| 18 | END; |
| 19 | RUN; |
Étapes de réalisation
1
First, run assess without handling missing targets to see the default behavior. An error or warning is expected.
Copied!
| 1 | PROC CAS; |
| 2 | percentile.assess |
| 3 | TABLE={name='patient_readmission', caslib='casuser'}, |
| 4 | response='READMITTED_30D', |
| 5 | inputs={{name='P_READMIT'}}, |
| 6 | event='1', |
| 7 | weight='PATIENT_WEIGHT', |
| 8 | fitStatOut={name='readmit_fit_fail', caslib='casuser', replace=true}; |
| 9 | QUIT; |
2
Now, correctly run the assessment using `noMissingTarget=true` to exclude records with missing outcomes and apply the `weight` parameter.
Copied!
| 1 | PROC CAS; |
| 2 | percentile.assess |
| 3 | TABLE={name='patient_readmission', caslib='casuser'}, |
| 4 | response='READMITTED_30D', |
| 5 | inputs={{name='P_READMIT'}}, |
| 6 | event='1', |
| 7 | weight='PATIENT_WEIGHT', |
| 8 | noMissingTarget=true, |
| 9 | includeFitStat=true, |
| 10 | fitStatOut={name='readmit_fit_weighted', caslib='casuser', replace=true}; |
| 11 | QUIT; |
Expected Result
The first step should produce a warning or error in the log indicating that missing values were found in the response variable. The second step should run successfully. The resulting 'readmit_fit_weighted' table will contain fit statistics (AUC, etc.) calculated only on the non-missing target observations, and these calculations will be influenced by the 'PATIENT_WEIGHT' variable, giving more influence to the high-risk ICU patients.