How to recover stored results from e() class?

[Santosh Dash]

Hello,
I am working on mwitch command of Stata. I work in Stata 15.
I ran the following:

. mswitch dr gcf, switch(gds) varswitch states(2) nolog vsquish

Performing EM optimization:

Performing gradient-based optimization:


Markov-switching dynamic regression

Sample: 1970 - 2016 No. of obs = 47
Number of states = 2 AIC = 4.4781
Unconditional probabilities: transition HQIC = 4.5966
SBIC = 4.7930
Log likelihood = -97.23547


gcf Coef. Std. Err. z P>z [95% Conf. Interval]

State1
gds .0924856 .1473673 0.63 0.530 -.1963491 .3813203
_cons 15.42887 3.11739 4.95 0.000 9.318895 21.53884

State2
gds .736845 .0485931 15.16 0.000 .6416043 .8320856
_cons 6.129055 1.193445 5.14 0.000 3.789946 8.468164

sigma1 2.250846 .3264654 1.693895 2.990922

sigma2 1.059222 .2000414 .7315295 1.533707

p11 .9288938 .0537234 .7262418 .9846927

p21 .0772527 .0579521 .0167304 .2917522



. matrix list e(b)

e(b)[1,8]
State1: State1: State2: State2: lnsigma1: lnsigma2: p11: p21:
gds _cons gds _cons _cons _cons _cons _cons
y1 .09248559 15.428868 .73684496 6.1290551 .8113061 .05753501 -2.5698196 2.4802733

. matrix list e(uncprob)
e(uncprob)[1,2]
State1 State2
r1 .52071497 .47928503

I want to generate transition probabilities p11 and p21 from the above table. When i try to recover p11 and p21 by running the following code:
gen p11=[p11]_cons

It is generating the value as -2.5698196, NOT .9288938, as it is reported in Stata's output
Similarly, I want to export results into excel using outreg2. When I'm exporting, p11 and p21 are stored as (-2.5698196 2.4802733), not as (.9288938, .0772527). Do not know why.

Observations:
1. Transition probabilities reported in the result (shaded in blue) do not match with probabilities stored in matrix (shaded in red). Why is the difference?
2. Further, the unconditional probabilities do not match with the results of matrix list e(uncprob) (shaded in purple)


Any help would be greatly appreciated.

[Jesse Wursten]

Please enclose any code and output in code tags, that makes them much easier to read.

To extract data from the e-results, you should first copy the e-matrix of interest.

mat results = e(b)

Then you can extract the elements you need using standard matrix syntax, see

help matrix extraction

[Weiwen Ng]

Actually, in this case, the correct probabilities are returned in the r(table) matrix. That is the matrix that Stata usually outputs to your console. A demonstration:

use https://www.stata-press.com/data/r16/usmacro
mswitch dr fedfunds

Performing EM optimization:

Performing gradient-based optimization:

Iteration 0:   log likelihood = -508.66031  
Iteration 1:   log likelihood =  -508.6382  
Iteration 2:   log likelihood = -508.63592  
Iteration 3:   log likelihood = -508.63592  

Markov-switching dynamic regression

Sample: 1954q3 - 2010q4                         No. of obs        =        226
Number of states =    2                         AIC               =     4.5455
Unconditional probabilities: transition         HQIC              =     4.5760
                                                SBIC              =     4.6211
Log likelihood = -508.63592

------------------------------------------------------------------------------
    fedfunds |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
State1       |
       _cons |    3.70877   .1767083    20.99   0.000     3.362428    4.055112
-------------+----------------------------------------------------------------
State2       |
       _cons |   9.556793   .2999889    31.86   0.000     8.968826    10.14476
-------------+----------------------------------------------------------------
       sigma |   2.107562   .1008692                      1.918851    2.314831
-------------+----------------------------------------------------------------
         p11 |   .9820939   .0104002                      .9450805    .9943119
-------------+----------------------------------------------------------------
         p21 |   .0503587   .0268434                      .0173432    .1374344
------------------------------------------------------------------------------


mat t = r(table)
mat list t

t[9,8]
            State1:     State2:    lnsigma:        p11:        p21:   _diparm1:   _diparm1:   _diparm1:
             _cons       _cons       _cons       _cons       _cons       sigma         p11         p21
     b     3.70877   9.5567934   .74553174  -4.0045427   2.9369133   2.1075618   .98209385   .05035868
    se   .17670834   .29998894    .0478606    .5914089   .56131057   .10086917   .01040023   .02684338
     z   20.988087   31.857152   15.577151   -6.771191   5.2322429          .b          .b          .b
pvalue   8.427e-98   1.05e-222   1.041e-54   1.277e-11   1.675e-07          .b          .b          .b
    ll    3.362428   8.9688259   .65172669  -5.1636828   1.8367648   1.9188512   .94508055   .01734316
    ul    4.055112   10.144761   .83933678  -2.8454025   4.0370618   2.3148312   .99431195   .13743437
    df           .           .           .           .           .           .           .           .
  crit    1.959964    1.959964    1.959964    1.959964    1.959964    1.959964    1.959964    1.959964
 eform           0           0           0           0           0           0           0           0

The bolded values should look familiar.

Basically, it appears that the raw coefficients for the transitions (e.g. p11:_cons) are not probabilities. They're some transformation of the probability. I can't quite tell what the transformation is; it's not an exponent (I checked), and it's not a natural log (ln of a negative number is undefined). e(b) only appears to contain the raw coefficients, whatever scale they were estimated on. It's moot, since the actual probabilities are returned in the matrix r(table), which the code above saves to a matrix called t.

I have no idea how you would export this to outreg2, so you're on your own there.

[Santosh Dash]

@Jesse, Thank you for your reply. I had read and tried but it was not of much help.

[Santosh Dash]

Weiwen Ng Thank you for your reply. It is difficult to export matrix table to an excel. However, last night I used another way to recover the probabilities and their SE. That worked well.

* Extracting prob
estat transition
matrix b = r(prob)
gen p11_coef= b[1,1]
* Extracting SE of prob
matrix b = r(se)
gen p11_se = b[1,1]

That worked well. Now, I am struggling to extract the standard deviation of the sigma.
I could extract the sigma using the following code and then take the exponent of lnsigma1 to get the value which is reported in the Stata table.

di _b[lnsigma:_cons]
di exp(_b[lnsigma:_cons])

Using your example, I have extracted the sigma as 0.74553174
exp(0.74553174)= 2.1075618 which is the same as sigma reported in your table.

However, I did the same to export the standard deviation of the sigma. It is not working.

di _se[lnsigma:_cons]

The output is .0478606
Then I tried taking exponent of it

di exp(_se[lnsigma:_cons])

I am getting it as 1.0490244 whereas it is .1008692 in the Stata output table.

Do not know how to extract it.
Any help in this regard will be greatly appreciated.

[Santosh Dash]

Weiwen Ng Finally, I could nail it. Your code of recovering the matrix table helped it.

mat t = r(table)
mat list t
di t[2,9]

That would give output .0478606 which is reported in your illustration table.

Thank you for your reply