Business Context
A factory wants to determine if a 'HighSpeed' machine setting causes more defects compared to 'Normal'. The outcome is binary (Defect: Yes/No). They use Targeted Maximum Likelihood Estimation (TMLE) for precise bias reduction on this critical binary metric.
About the Set : causalanalysis
Causal inference analysis and effect estimation.
Discover all actions of causalanalysis Data Preparation
Small dataset with a Binary Outcome (Defect 1/0) and a categorical treatment.
Copied!
| 1 | |
| 2 | DATA mycas.factory_qc; |
| 3 | call streaminit(777); |
| 4 | DO batch = 1 to 200; |
| 5 | temp = rand('NORMAL', 100, 5); |
| 6 | humidity = rand('UNIFORM', 30, 60); |
| 7 | IF rand('UNIFORM') > 0.5 THEN setting = 'HighSpeed'; |
| 8 | ELSE setting = 'Normal'; |
| 9 | logit_defect = -5 + 0.05*temp + 2*(setting='HighSpeed'); |
| 10 | prob_defect = 1 / (1 + exp(-logit_defect)); |
| 11 | IF rand('UNIFORM') < prob_defect THEN defect_flag = 1; |
| 12 | ELSE defect_flag = 0; |
| 13 | OUTPUT; |
| 14 | END; |
| 15 | |
| 16 | RUN; |
| 17 |
Étapes de réalisation
1
Fitting a Gradient Boosting model to predict the binary defect (simulating a complex outcome model).
Copied!
| 1 | |
| 2 | PROC CAS; |
| 3 | decisionTree.gbtreeTrain TABLE={name='factory_qc'}, target='defect_flag', inputs={'setting', 'temp', 'humidity'}, casout={name='gb_model', replace=true}; |
| 4 | decisionTree.gbtreeScore TABLE={name='factory_qc'}, modelTable={name='gb_model'}, casout={name='qc_scored', replace=true}, copyVars='ALL', encodename=true; |
| 5 | |
| 6 | RUN; |
| 7 |
2
Execution of caEffect using TMLE for a CATEGORICAL outcome.
Copied!
| 1 | |
| 2 | PROC CAS; |
| 3 | causalanalysis.caEffect TABLE={name='qc_scored'}, method='TMLE', treatVar={name='setting'}, outcomeVar={name='defect_flag', type='CATEGORICAL', event='1'}, outcomeModel={predName='P_defect_flag1'}, pom={{trtLev='HighSpeed'}, {trtLev='Normal'}}, difference={{evtLev='HighSpeed', refLev='Normal'}}, inference=true; |
| 4 | |
| 5 | RUN; |
| 6 |
Expected Result
The action correctly handles the 'CATEGORICAL' outcome type. It estimates the probability of Defect (event='1') for both machine settings using TMLE and reports the risk difference (causal effect) with valid confidence intervals.