假设有如下数据集：

data a;
      input Treatment y @@;
      datalines;
1  17   1  28   1  19   1 21   1  19
1  43   1  30   1  39   1 44   1  44
1  16
2  21   2  21   2  24   2 25
2  39   2  45   2  42   2 47
2  19   2  22   2  16
3  22   3  30   3  33   3 31
3  46
3  26   3  31   3  26   3 33   3  29   3  25
;

proc glm data=a;
   class Treatment ;
   model y = Treatment/ss3;
   lsmeans Treatment/cl alpha=0.05 pdiff stderr;
run;
quit;

结果：

                                                     Least Squares Means for effect Treatment
                                                             Pr > |t| for H0: LSMean(i)=LSMean(j)

                                                                    Dependent Variable: y

                                                        i/j              1             2             3

                                                           1                        0.9833        0.8017
                                                           2        0.9833                      0.8179
                                                           3        0.8017        0.8179

                                                    Treatment        y LSMEAN      95% Confidence Limits

                                                    1               29.090909       22.873802    35.308016
                                                    2               29.181818       22.964711    35.398925
                                                    3               30.181818       23.964711    36.398925

                                                            Least Squares Means for Effect Treatment

                                                                  Difference
                                                                     Between    95% Confidence Limits for
                                                      i    j           Means       LSMean(i)-LSMean(j)

                                                      1    2       -0.090909       -8.883227     8.701408
                                                      1    3       -1.090909       -9.883227     7.701408
                                                      2    3       -1.000000       -9.792317     7.792317

删去数据集中treatment 3 的数据：

data b;
     set a;
  where treatment=1 or treatment=2;
run;

proc glm data=b;
   class Treatment ;
   model y = Treatment/ss3;
   lsmeans Treatment/cl alpha=0.05 pdiff stderr;
run;
quit;

结果为：

                                                          The GLM Procedure
                                                                      Least Squares Means

                                                                                                                                        H0:LSMean1=
                                                                                         Standard    H0:LSMEAN=0      LSMean2
                                            Treatment        y LSMEAN           Error       Pr > |t|                      Pr > |t|

                                            1              29.0909091       3.4775549         <.0001                    0.9854
                                            2              29.1818182       3.4775549         <.0001

                                                    Treatment        y LSMEAN      95% Confidence Limits

                                                    1               29.090909             21.836857    36.344962
                                                    2               29.181818             21.927766    36.435871

                                                            Least Squares Means for Effect Treatment

                                                                  Difference
                                                                     Between    95% Confidence Limits for
                                                      i    j           Means       LSMean(i)-LSMean(j)

                                                      1    2       -0.090909      -10.349689    10.167870

请问： 同样是treatment 1 和 treatment 2 进行比较，为什么两个结果中的置信区间和P值却不一样？

结果1：      （-8.883227 ，  8.701408）
结果2：      （ -10.349689， 10.167870）

请问LSmeans置信区间的计算方法是什么？

十分感谢！