Requested Base Level Ignored in Interaction Syntax

[Mauricio Caceres]

I am trying to run an regression of this form:

clear
set seed 1729
set obs 1000
gen T = ceil(runiform() * 3)
gen W = mod(_n, 3)
gen C = 1
gen Y = C + T + runiform()
reg Y ibn.T#(C b0.W), noconstant

For some reason this excludes the last level of W instead of the first. Is that always the case? How can I know which level will be omitted? The following works as expected

reg Y ibn.T b0.W, noconstant

I don't care which level I omit as long as I can omit the same base level in both regressions. Many thanks for any suggestions!

[Andrew Musau]

reg Y ibn.T#(C b0.W), noconstant

For some reason this excludes the last level of W instead of the first. Is that always the case? How can I know which level will be omitted? The following works as expected

reg Y ibn.T b0.W, noconstant

I don't think that there is any such thing as the base level for an interaction term in the absence of a main effect and therefore I agree with the default behavior. Starting with this principle, you can ask for interactions and main effects and then omit the latter.

clear
set seed 1729
set obs 1000
gen T = ceil(runiform() * 3)
gen W = mod(_n, 3)
gen C = 1
gen Y = C + T + runiform()
reg Y ibn.T##(C b0.W) o(1/2).W, noconstant

Res.:

. reg Y ibn.T##(C b0.W) o(1/2).W, noconstant
note: 1.C omitted because of collinearity.
note: 1.T#1.C omitted because of collinearity.
note: 2.T#1.C omitted because of collinearity.
note: 3.T#1.C omitted because of collinearity.

      Source |       SS           df       MS      Number of obs   =     1,000
-------------+----------------------------------   F(9, 991)       =  16387.97
       Model |   12899.352         9  1433.26133   Prob > F        =    0.0000
    Residual |  86.6710214       991  .087458145   R-squared       =    0.9933
-------------+----------------------------------   Adj R-squared   =    0.9933
       Total |   12986.023     1,000   12.986023   Root MSE        =    .29573

------------------------------------------------------------------------------
           Y | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
           T |
          1  |   2.526311   .0281971    89.59   0.000     2.470979    2.581644
          2  |   3.458505   .0279442   123.76   0.000     3.403668    3.513341
          3  |   4.479736   .0280698   159.59   0.000     4.424653    4.534819
             |
         1.C |          0  (omitted)
             |
           W |
          1  |          0  (omitted)
          2  |          0  (omitted)
             |
         T#C |
        1 1  |          0  (omitted)
        2 1  |          0  (omitted)
        3 1  |          0  (omitted)
             |
         T#W |
        1 1  |  -.0463275   .0401552    -1.15   0.249    -.1251265    .0324715
        1 2  |  -.0416716   .0396111    -1.05   0.293    -.1194029    .0360596
        2 1  |   .0117188   .0400743     0.29   0.770    -.0669215    .0903591
        2 2  |   .0499672   .0392604     1.27   0.203    -.0270759    .1270104
        3 1  |   .0207667   .0388678     0.53   0.593    -.0555059    .0970394
        3 2  |   .0258527   .0402597     0.64   0.521    -.0531515    .1048568
------------------------------------------------------------------------------

[Mauricio Caceres]

I think this might be a matter of terminology; I just mean the level that will end up being excluded. But you do give me an idea, thanks! This seems to work:

clear
set seed 1729
set obs 1000
gen T = ceil(runiform() * 3)
gen W = mod(_n, 3)
gen C = 1
gen Y = C + T + runiform()

tempname ivars
fvexpand C b0.W
local varlist `r(varlist)'
matrix `ivars' = J(1, `:list sizeof varlist', .)
matrix colnames `ivars' = `varlist'
_ms_omit_info `ivars'
mata st_local("varlist", invtokens(select(tokens(st_local("varlist")), !st_matrix("r(omit)"))))
reg Y ibn.T#(`varlist'), noconstant