Different results when using splines vs interaction variables

[Scott Rick]

Hi. I am attempting to manually run a single-group ITSA for my study that has two interventions at week 78 and 107 (the in-built stat command doesn't work for my purpose). I tried two methods (from here: https://rpubs.com/mbounthavong/itsa_stata):

1. using spline (mkspline) to generate new variables that take into consideration the time after the intervention
2. the more conventional method using interaction variables

Both methods should yield the same results. However, as seen in the stata output below the results for the "period" variable differ, while the results for the other variables are the same. Could someone please help me understand why there is this difference?

*1. Using splines

 *creating knots
mkspline knot1 78 knot2 107 knot3 = week_num, marginal


// Create a variable to separate the treatment periods
gen period = .
    replace period = 0 if week_num < 78                                      /* Baseline phase */ 
    replace period = 1 if week_num >= 78 & week_num<107 & !missing(week_num) /* Mandate phase */
    replace period = 2 if week_num >= 107 & !missing(week_num)                 /* Platform phase */
tab period, m 



xtset con_id week_num
xtreg week_docalltele i.period c.knot1 c.knot2 c.knot3 
predict linear_spline


*Output 1

------------------------------------------------------------------------------
week_docal~e | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
      period |
          1  |   1.379151   .2534505     5.44   0.000     .8823975    1.875905
          2  |   .8693821   .4655648     1.87   0.062    -.0431081    1.781872
             |
       knot1 |  -.0583853   .0030684   -19.03   0.000    -.0643993   -.0523713
       knot2 |    .053078   .0136945     3.88   0.000     .0262372    .0799188
       knot3 |   .0147004   .0188625     0.78   0.436    -.0222693    .0516702
       _cons |    7.62364   .3468154    21.98   0.000     6.943895    8.303386
-------------+----------------------------------------------------------------
     sigma_u |   2.663207
     sigma_e |  3.3603617
         rho |  .38579202   (fraction of variance due to u_i)
------------------------------------------------------------------------------

. predict linear_spline
(option xb assumed; fitted values)



2. Using interaction variables


// Create a variable to separate the treatment periods
gen period = .
    replace period = 0 if week_num < 78                                      /* Baseline phase */ 
    replace period = 1 if week_num >= 78 & week_num<107 & !missing(week_num) /* Mandate phase */
    replace period = 2 if week_num >= 107 & !missing(week_num)                 /* Platform phase */
tab period, m 


xtset con_id week_num
xtreg week_docalltele i.period c.week_num i.period#c.week_num
predict linear_spline


*Output 2

-----------------------------------------------------------------------------------
  week_docalltele | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
------------------+----------------------------------------------------------------
           period |
               1  |  -2.760932   1.231048    -2.24   0.025    -5.173741   -.3481229
               2  |  -4.843649   1.624719    -2.98   0.003    -8.028039   -1.659259
                  |
         week_num |  -.0583853   .0030684   -19.03   0.000    -.0643993   -.0523713
                  |
period#c.week_num |
               1  |    .053078   .0136945     3.88   0.000     .0262372    .0799188
               2  |   .0677784   .0136487     4.97   0.000     .0410274    .0945295
                  |
            _cons |    7.62364   .3468154    21.98   0.000     6.943895    8.303386
------------------+----------------------------------------------------------------
          sigma_u |   2.663207
          sigma_e |  3.3603617
              rho |  .38579202   (fraction of variance due to u_i)
-----------------------------------------------------------------------------------

. predict linear_spline


***Data sample


input float week_num long con_id float week_docalltele
1  4 12
1 11  5
1 12 10
1 13  2
1 17  8
1 22  7
1 23 19
1 24  4
1 25  3
1 26  3
1 30  7
1 31  7
1 33  2
1 37 12
1 38 11
1 40 16
1 41 14
1 42  1
1 43  3
1 45  8
1 46  2
1 51  4
1 54 22
1 55  2
1 56  9
1 57 17
1 58 21
1 59 20
1 61  2
1 63  6
1 65  2
1 67  9
1 69  8
1 71 24
2 12 17
2 13  6
2 16 12
2 17 12
2 23 14
2 25  2
2 26  5
2 28  2
2 31  7
2 32  3
2 33  1
2 35 10
2 37 14
2 38  6
2 39  5
2 41  1
2 42  6
2 43  6
2 45 12
2 46  2
2 48  1
2 49  9
2 50  4
2 51  6
2 54 19
2 55  1
2 56 10
2 57 20
2 58 25
2 59 16
2 61  8
2 62  6
2 65  2
2 67 15
2 69  6
2 70 15
2 71 17
3  4  6
3 11 15
3 12 15
3 16 10
3 17  9
3 22 14
3 24 14
3 25  2
3 27 10
3 28  3
3 29  3
3 31 12
3 33  5
3 37  4
3 38 10
3 39 10
3 40 20
3 41 14
3 42  5
3 43  7
3 45 21
3 46  1
3 48  1
3 49  5
3 50  5
3 51  8
3 56  1
3 57 27
3 58 19

[Clyde Schechter]

The difference is because the period variable takes on different meanings in these two representations of the model. Both models are algebraically equivalent: the predicted values they result in are exactly the same.

In the interaction model, each period variable represents the expected outcome value in that period when week_num == 0. Since week_num is never 0, even in the first period, this is a hypothetical, or, you could say, meaningless output.

In the spline model, each period variable represents the "jump" in the expected value at knot points.

So there is no reason to expect these to come out the same.

[George Ford]

In the second model, (2.period#c.week_num - 1.period#c.week_num) equals the knot3 coefficient from the first. The others match up.