如何通过SPSS宏在GENLIN中执行AIC逐步变量选择？

需求背景 #

仅可使用SPSS，无法借助R或Python工具，需在GENLIN模块中通过宏实现基于AIC的逐步变量选择流程，针对二项逻辑回归（广义模型）完成变量缩减，自动筛选出AIC值最低的最优模型。现有全变量GENLIN语法如下：

GENLIN SITIOINYECGLUT2DELT1 (REFERENCE=LAST) WITH DOSIS ARI DHA TOTAL
  /MODEL DOSIS ARI DHA TOTAL INTERCEPT=YES
 DISTRIBUTION=BINOMIAL LINK=LOGIT
  /CRITERIA METHOD=FISHER(1) SCALE=1 COVB=MODEL MAXITERATIONS=100 MAXSTEPHALVING=5 
    PCONVERGE=1E-006(ABSOLUTE) SINGULAR=1E-012 ANALYSISTYPE=3(WALD) CILEVEL=95 CITYPE=WALD 
    LIKELIHOOD=FULL
  /MISSING CLASSMISSING=EXCLUDE
  /PRINT CPS DESCRIPTIVES MODELINFO FIT SUMMARY SOLUTION .

手动逐一移除变量计算AIC效率极低，需实现自动化流程：先构建全变量模型，再逐一移除单个变量计算对应AIC，保留最优模型，重复此过程直至无法得到更低AIC的模型。

实现宏代码 #

以下SPSS宏可自动执行上述逐步变量选择流程：

DEFINE !StepwiseAIC(DV=!TOKENS(1), IVs=!CMDEND)
* 初始化变量列表与最优AIC存储.
!LET !CurrentIVs = !IVs
!LET !BestAIC = 999999
!LET !BestModelIVs = !CurrentIVs

* 循环迭代直至无更优模型.
!DO !WHILE (!TRUE)
  * 运行当前变量集的GENLIN模型并提取AIC.
  GENLIN !DV (REFERENCE=LAST) WITH !CurrentIVs
    /MODEL !CurrentIVs INTERCEPT=YES
    DISTRIBUTION=BINOMIAL LINK=LOGIT
    /CRITERIA METHOD=FISHER(1) SCALE=1 COVB=MODEL MAXITERATIONS=100 MAXSTEPHALVING=5 
      PCONVERGE=1E-006(ABSOLUTE) SINGULAR=1E-012 ANALYSISTYPE=3(WALD) CILEVEL=95 CITYPE=WALD 
      LIKELIHOOD=FULL
    /MISSING CLASSMISSING=EXCLUDE
    /PRINT FIT SUMMARY.
  * 从输出中提取AIC值到临时数据集.
  OMS SELECT TABLES
    /IF SUBTYPES='Model Fit'
    /DESTINATION FORMAT=SAV OUTFILE='TempAIC.sav'
    /TAG='ExtractAIC'.
  EXECUTE.
  OMSEND TAG='ExtractAIC'.

  * 读取当前模型的AIC值.
  GET FILE='TempAIC.sav'.
  SELECT IF INDEX(VARNAME, 'AIC')>0.
  COMPUTE CurrentAIC = VALUE.
  N OF CASES 1.
  SAVE OUTFILE='CurrentAIC.sav' /KEEP CurrentAIC.
  GET FILE='CurrentAIC.sav'.
  !LET !CurrentAIC = !QUOTE(!STRING(CurrentAIC, F8.2))

  * 首次运行时更新最优AIC.
  !IF (!BestAIC = 999999) !THEN
    !LET !BestAIC = !CurrentAIC
    !LET !BestModelIVs = !CurrentIVs
  !ENDIF

  * 生成所有移除单个变量后的变量集.
  !LET !IVList = !CurrentIVs
  !LET !TempIVs = !NULL
  !LET !MinAIC = !CurrentAIC
  !LET !MinIVs = !CurrentIVs

  !DO !IV !IN !IVList
    !LET !TempIVs = !REMOVE(!IVList, !IV)
    * 运行移除单个变量后的模型.
    GENLIN !DV (REFERENCE=LAST) WITH !TempIVs
      /MODEL !TempIVs INTERCEPT=YES
      DISTRIBUTION=BINOMIAL LINK=LOGIT
      /CRITERIA METHOD=FISHER(1) SCALE=1 COVB=MODEL MAXITERATIONS=100 MAXSTEPHALVING=5 
        PCONVERGE=1E-006(ABSOLUTE) SINGULAR=1E-012 ANALYSISTYPE=3(WALD) CILEVEL=95 CITYPE=WALD 
        LIKELIHOOD=FULL
      /MISSING CLASSMISSING=EXCLUDE
      /PRINT FIT SUMMARY.
    * 提取该模型的AIC.
    OMS SELECT TABLES
      /IF SUBTYPES='Model Fit'
      /DESTINATION FORMAT=SAV OUTFILE='TempAIC.sav'
      /TAG='ExtractTempAIC'.
    EXECUTE.
    OMSEND TAG='ExtractTempAIC'.

    GET FILE='TempAIC.sav'.
    SELECT IF INDEX(VARNAME, 'AIC')>0.
    COMPUTE TempAIC = VALUE.
    N OF CASES 1.
    SAVE OUTFILE='TempAICValue.sav' /KEEP TempAIC.
    GET FILE='TempAICValue.sav'.
    !LET !TempAICVal = !QUOTE(!STRING(TempAIC, F8.2))

    * 比较AIC，记录最小值对应的变量集.
    !IF (!TempAICVal < !MinAIC) !THEN
      !LET !MinAIC = !TempAICVal
      !LET !MinIVs = !TempIVs
    !ENDIF
  !DOEND

  * 判断是否找到更优模型.
  !IF (!MinAIC < !BestAIC) !THEN
    !LET !BestAIC = !MinAIC
    !LET !BestModelIVs = !MinIVs
    !LET !CurrentIVs = !MinIVs
  !ELSE
    * 无更优模型，终止循环.
    !BREAK
  !ENDIF
!DOEND

* 输出最优模型结果.
PRINT / '最优模型变量集: ' !BestModelIVs.
PRINT / '最优模型AIC值: ' !BestAIC.
EXECUTE.

* 运行最优模型并输出完整结果.
GENLIN !DV (REFERENCE=LAST) WITH !BestModelIVs
  /MODEL !BestModelIVs INTERCEPT=YES
  DISTRIBUTION=BINOMIAL LINK=LOGIT
  /CRITERIA METHOD=FISHER(1) SCALE=1 COVB=MODEL MAXITERATIONS=100 MAXSTEPHALVING=5 
    PCONVERGE=1E-006(ABSOLUTE) SINGULAR=1E-012 ANALYSISTYPE=3(WALD) CILEVEL=95 CITYPE=WALD 
    LIKELIHOOD=FULL
  /MISSING CLASSMISSING=EXCLUDE
  /PRINT CPS DESCRIPTIVES MODELINFO FIT SUMMARY SOLUTION .

* 清理临时文件.
DELETE FILE='TempAIC.sav'.
DELETE FILE='CurrentAIC.sav'.
DELETE FILE='TempAICValue.sav'.
!ENDDEFINE.

使用方法 #

1. 将上述宏代码粘贴到SPSS语法编辑器中；
2. 调用宏时替换因变量和自变量列表，示例调用：

!StepwiseAIC(DV=SITIOINYECGLUT2DELT1, IVs=DOSIS ARI DHA TOTAL)

3. 运行语法，宏会自动执行逐步变量选择流程，最终输出最优模型的完整结果。

注意事项 #

• 宏依赖OMS功能提取AIC值，确保SPSS的OMS功能正常启用；
• 若变量较多，宏运行时间会相应增加，耐心等待即可；
• 宏默认针对二项逻辑回归（DISTRIBUTION=BINOMIAL LINK=LOGIT），如需适配其他广义模型，修改GENLIN语句中的分布和链接函数即可。

内容的提问来源于stack exchange，提问作者Jorge A