如何在SAS中计算测试组与对照组的Uplift及Uplift_1？

针对按risk_level分组，计算测试组/对照组Uplift及Uplift_1的需求，以下是两种可靠的实现方法（假设数据集名为app_data，包含变量：risk_level、group（取值'test'/'control'）、loans、approved_apps）：

方法一：Proc SQL实现 #

通过子查询分别计算两组的Uplift，再关联计算Uplift_1，逻辑清晰且易调试：

proc sql;
    create table uplift_calc as
    select 
        a.risk_level,
        a.uplift as test_uplift,
        b.uplift as control_uplift,
        -- 处理对照组Uplift为0的情况，避免除以0错误
        case when b.uplift = 0 then . 
             else (a.uplift / b.uplift) - 1 end as uplift_1
    from 
        -- 计算测试组各risk_level的Uplift
        (select 
             risk_level,
             sum(loans)/sum(approved_apps) as uplift
         from app_data
         where group = 'test'
         group by risk_level) as a
    inner join
        -- 计算对照组各risk_level的Uplift
        (select 
             risk_level,
             sum(loans)/sum(approved_apps) as uplift
         from app_data
         where group = 'control'
         group by risk_level) as b
    on a.risk_level = b.risk_level;
quit;

方法二：Proc Summary + Data步实现 #

先通过聚合工具生成分组统计值，再用Data步合并计算，适合后续需要对聚合数据做更多扩展处理的场景：

步骤1：生成分组聚合数据 #

proc summary data=app_data nway;
    class risk_level group;
    var loans approved_apps;
    output out=group_agg sum(loans)=total_loans sum(approved_apps)=total_approved;
run;

步骤2：合并并计算Uplift指标 #

data uplift_calc;
    merge group_agg(where=(group='test') 
                   rename=(total_loans=test_loans total_approved=test_approved))
          group_agg(where=(group='control') 
                   rename=(total_loans=control_loans total_approved=control_approved));
    by risk_level;
    
    test_uplift = test_loans / test_approved;
    control_uplift = control_loans / control_approved;
    
    if control_uplift ne 0 then uplift_1 = (test_uplift / control_uplift) - 1;
    else uplift_1 = .; -- 对照组无有效数据时标记为缺失
    
    keep risk_level test_uplift control_uplift uplift_1;
run;

关键说明 #

• Uplift的计算逻辑为组内总放款数/组内总批准数，符合应用级数据的聚合要求
• 两种方法均处理了approved_apps总和为0或对照组Uplift为0的异常情况，避免运行错误
• 若你的loans或approved_apps是比例而非计数变量，只需将sum()改为mean()即可适配

内容的提问来源于stack exchange，提问作者Swati Kanchan