Gamma and Inverse Gaussian models don't replicate Stata results exactly

[nathanielolin]

Working on #28. They're close, but not exact.

> data(margex)
>   margex$ycn[margex$ycn == 0] <- 1
>   mod <- glm(ycn ~ treatment + age + distance, data = margex, family = inverse.gaussian)
>   
>   z1 <- marg(mod, var_interest = 'age', at_var_interest = seq(20, 60, 5), 
+             type = 'levels')[[1]]
>   
>   z2 <- marg(mod, var_interest = 'treatment', type = 'effects')[[1]]
>   
>   stata <- aiEstimation::mod_marg(
+     'glm ycn i.treatment c.age c.distance, family("igaussian")',
+     margs = list(age = sprintf("margins, at(age = (%s))", 
+                                paste(seq(20, 60, 5), collapse = ", ")),
+                  treat = sprintf("margins, dydx(treat)")),
+     df = margex
+   )
>   
>   z1[, c("Label", "Margin", "Standard.Error", "P.Value")]
     Label   Margin Standard.Error P.Value
1 age = 20 73.68419      0.8351090       0
2 age = 25 73.86944      0.6894061       0
3 age = 30 74.05614      0.5564396       0
4 age = 35 74.24433      0.4508355       0
5 age = 40 74.43401      0.3988748       0
6 age = 45 74.62522      0.4245450       0
7 age = 50 74.81797      0.5194893       0
8 age = 55 75.01229      0.6567655       0
9 age = 60 75.20819      0.8174053       0
>   stata$margins$age
  variable margins_b margins_se p
1    1._at  73.68441  0.8330135 0
2    2._at  73.86966  0.6876766 0
3    3._at  74.05638  0.5550439 0
4    4._at  74.24458  0.4497050 0
5    5._at  74.43428  0.3978749 0
6    6._at  74.62550  0.4234810 0
7    7._at  74.81827  0.5181878 0
8    8._at  75.01260  0.6551204 0
9    9._at  75.20852  0.8153583 0
>   
>   z2[, c("Label", "Margin", "Standard.Error", "P.Value")]
          Label   Margin Standard.Error      P.Value
1 treatment = 0  0.00000      0.0000000          NaN
2 treatment = 1 15.42396      0.8370207 7.946186e-76
>   stata$margins$treat
      variable margins_b margins_se        p
1 0b.treatment   0.00000         NA       NA
2  1.treatment  15.42372  0.8349224 3.39e-76

Similar issue with Gamma models