Hi,

I would like to measure the performance of two different Groups A and B. Group A works directly at the project sponsor's location. Group B works remotely. For Group B only I have further details, such as the training status (1= no training) and full IT access (1 = restricted IT).

Code:
regress performance group_b

----------------------------------------------------------------------------------
performance |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-----------------+----------------------------------------------------------------
group_b     |   .0326976   .0028966    11.29   0.000     .0270203    .0383749
      _cons |   .3165059   .0017716   178.65   0.000     .3130336    .3199783
----------------------------------------------------------------------------------
Group A is the omitted case. Now, I would like to analyze the effect of "training" and "IT". I could do this by subsampling on group B:

Code:
regress performance group_b training IT if group_b == 1

---------------------------------------------------------------------------------------
performance |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
----------------------+----------------------------------------------------------------
group_b     |          0  (omitted)
training    |   .0169437   .0050673     3.34   0.001      .007012    .0268755
IT          |   .0405026    .004614     8.78   0.000     .0314592    .0495459
       _cons|   .3140697   .0042685    73.58   0.000     .3057035    .3224359
---------------------------------------------------------------------------------------
But then I can't compare the effect of group B against group A. So I would like to include everything into a single model and breakdown the "Group B" variable. My goal is to measure following effects:
  • Effect of Group A as reference (omitted case)
  • Effect of Group B (with full IT access and/or training)
  • Effect of no training for Group B
  • Effect of restricted IT for Group B
For the last two, I introduce an "interaction term" with no main effect. So the interpretation is different than a real interaction term. I get following model:

Code:
regress performance group_b group_b#training group_b#IT


-----------------------------------------------------------------------------------------------------
performance      |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
------------------------------------+----------------------------------------------------------------
group_b          |   .0149012   .0045581     3.27   0.001     .0059675    .0238349
                                    |
group_b#training |
            1 1  |   .0114282   .0044497     2.57   0.010     .0027068    .0201495
            1 0  |          0  (omitted)
                 |
group_b#IT       |
            1 1  |    .039985   .0042443     9.42   0.000     .0316663    .0483037
            1 0  |          0  (omitted)
                 |
           _cons |    .316844   .0017725   178.75   0.000     .3133699     .320318
-----------------------------------------------------------------------------------------------------
Does the model's interpretation match my goal?

Best regards
Olaf