Differences between -xtpqml and -poi2hdfe and questions on post-estimation analysis for poi2hdfe

Cezary Baraniecki

Join Date: Jun 2018

Posts: 4
#1

Differences between -xtpqml and -poi2hdfe and questions on post-estimation analysis for poi2hdfe

27 Jun 2018, 12:08

Hello,

I'm trying to run a PPML gravity model with country-pair and time fixed effects on a 10 year panel. The number of countries I am using (175/190) prevents me from using the -ppml command with my manually created dummies (matsize is limited to 11,000).

I discovered -poi2hdfe which I believe should function the same as -ppml if identically specified. I likewise thought -xtpqml should return identical results if the regression converges.

However I'm getting some strange results with -poi2hdfe which I can't figure out My code is the following. PT* are dummy variables. My "pairid" is my exporter*importer pairs. I am limiting my regression to just China (exporter == CHN)

Code:

poi2hdfe exports lgdp_exporter lgdp_importer PTgood_gov_exp PTgood_gov_imp PTbad_gov_exp PTbad_gov_imp if exporter == "CHN" , id1(pairid) id2(year)

In my mind, this should return a value for lgdp_exporter, as China's gdp fluctuated within my 10 year panel. However I get the following:

Code:

Dropping 1 groups for pairid with a single observation Total Number of observations used in the regression -> 1855 Starting Estimation of coefficients 1 dif is -> 6244061 2 dif is -> 1.6021985 3 dif is -> 1.1138891 4 dif is -> 1.3918957 5 dif is -> .42605837 6 dif is -> .23716472 7 dif is -> .13974912 8 dif is -> .0732805 9 dif is -> .02254347 10 dif is -> .01082417 11 dif is -> .04080683 12 dif is -> .00028351 13 dif is -> 1.106e-06 14 dif is -> 2.070e-11 Coefficients converged after 14 reghdfe calls ******* Poisson Regression with Two High-Dimensional Fixed Effects ********** Number of obs = 1855 -------------------------------------------------------------------------------- | Robust exports | Coef. Std. Err. z P>|z| [95% Conf. Interval] ---------------+---------------------------------------------------------------- lgdp_exporter | 0 (omitted) lgdp_importer | .6324759 .0501047 12.62 0.000 .5342725 .7306793 PTgood_gov_exp | 0 (omitted) PTgood_gov_imp | .068497 .0281072 2.44 0.015 .013408 .1235861 PTbad_gov_exp | 0 (omitted) PTbad_gov_imp | -.2380663 .0470217 -5.06 0.000 -.3302272 -.1459055 --------------------------------------------------------------------------------

Why is lgdp_exporter omitted? This is the log gdp of my exporter, China in this case, but it IS time varying. I don't get this result when I use -xtpqml though the other regression coefficients are nearly identical. I can post the output of that command if it interests anyone.

Lastly,

Code:

estat

doesn't appear to work for poi2hdfe. Is there any way to save residuals etc. in order to do a RESET, or Hausman test?

Edit: Solved picture issue below
SOLVED-PS: I checked the guide to see the best way to post output and couldn't find a better way than posting a picture. Am I missing something or is this standard?
Attached Files

Last edited by Cezary Baraniecki; 27 Jun 2018, 12:26.
Tags: None
Joao Santos Silva

Join Date: Apr 2014

Posts: 3000
#2

27 Jun 2018, 15:14

Dear Cezary Baraniecki

I believe that the variables for China are dropped because they are collinear with the time FE; the puzzle is why xtpqml does not drop it. Can you please post those results?

Best wishes,

Joao
Comment

Cezary Baraniecki

Join Date: Jun 2018
Posts: 4

28 Jun 2018, 01:50

Hmm... it seems that I may have just forgotten to add time dummies to xtpqml as gdpexporter now drops (i've edited original post to add the output at the very bottom) . But the problem still exists for other runs done, with what I believe, should be the same specification of -xtpqml and -poi2hdfe.

Consider the following, done on the same database but with other dependent variables (note i've only left robust results for clarity/space. I can post the full regression if it helps):

Again, this is done on the same database. China's ("CHN") gdp is time-varying. "gov_imp" / "gov_exp" are governance scores (a component made from World Governance Indicators) which are also time-varying and country specific. My econometrics knowledge is not terribly advanced, but I don't think have a time dummy and a country-pair dummy should make time-varying country specific data to be perfectly colinear. But here "gov_exp" drops whereas lgdp_exporter stays. The only other odd thing I can note is that two (2) time dummies are dropped due to collinearity. In my mind, only one (1) should drop.

Code:

xtpqml exports lgdp_exporter lgdp_importer gov_imp gov_exp year_* if exporter == "CHN"  , fe i(pairid) cluster(pairid)

Code:

.
note: you are responsible for interpretation of non-count dep. variable
note: 1 group (1 obs) dropped because of only one obs per group
note: gov_exp dropped because of collinearity
note: year_1 dropped because of collinearity
note: year_10 dropped because of collinearity

Conditional fixed-effects Poisson regression    Number of obs      =      1842
Group variable: pairid                          Number of groups   =       188

                                                Obs per group: min =         4
                                                               avg =       9.8
                                                               max =        10

                                                Wald chi2(11)      =  9.94e+08
Log likelihood  = -1.402e+08                    Prob > chi2        =    0.0000

-------------------------------------------------------------------------------

-------------------------------------------------------------------------------
Calculating Robust Standard Errors...
-------------------------------------------------------------------------------
      exports |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
--------------+----------------------------------------------------------------
exports       |
lgdp_exporter |   .3134428   .0323356     9.69   0.000     .2500662    .3768195
lgdp_importer |   .6606657   .1001429     6.60   0.000     .4643892    .8569421
      gov_imp |   .1052808   .0468818     2.25   0.025     .0133941    .1971675
       year_2 |   .0272476   .0193626     1.41   0.159    -.0107024    .0651976
       year_3 |  -.1355452   .0155219    -8.73   0.000    -.1659676   -.1051228
       year_4 |   .0119779   .0162285     0.74   0.460    -.0198293    .0437852
       year_5 |   .0628282   .0231184     2.72   0.007     .0175171    .1081394
       year_6 |   .0784617   .0358336     2.19   0.029     .0082291    .1486943
       year_7 |   .1027708   .0526216     1.95   0.051    -.0003656    .2059073
       year_8 |   .1083588    .030498     3.55   0.000     .0485838    .1681337
       year_9 |   .0927698   .0099916     9.28   0.000     .0731865    .1123531
-------------------------------------------------------------------------------
Wald chi2(11) =  2046.44                                Prob > chi2 =   0.0000

Now the poi2hdfe run that I thought should be the same. Note both lgdp_exporter and gov_exp drop:

Code:

poi2hdfe exports lgdp_exporter lgdp_importer gov_exp gov_imp if exporter == "CHN", id1(pairid) id2(year)  cluster(pairid)

Code:

Dropping 1 groups for pairid with a single observation

Total Number of observations used in the regression -> 1842

Starting Estimation of coefficients

1 dif is -> 6308359.1
2 dif is -> 1.4168923
3 dif is -> 2.129642
4 dif is -> 1.0574683
5 dif is -> .44567978
6 dif is -> .29660272
7 dif is -> .12546297
8 dif is -> .0521958
9 dif is -> .0184021
10 dif is -> .01087103
11 dif is -> .00908004
12 dif is -> .00072241
13 dif is -> 4.953e-06
14 dif is -> 6.201e-10

Coefficients converged after 14 reghdfe calls


 ******* Poisson Regression with Two High-Dimensional Fixed Effects **********

                                                  Number of obs   =       1842
                                (Std. Err. adjusted for 188 clusters in pairid)
-------------------------------------------------------------------------------
              |               Robust
      exports |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
--------------+----------------------------------------------------------------
lgdp_exporter |          0  (omitted)
lgdp_importer |   .6606657   .1004376     6.58   0.000     .4638116    .8575198
      gov_exp |          0  (omitted)
      gov_imp |   .1052808   .0470198     2.24   0.025     .0131237    .1974379
-------------------------------------------------------------------------------

.

These are the results of xtpqml from the original post. Note that though lgdp_exporter drops, PTbad_gov_exp does not. This is a dummy that equal 1 if an exporter's governance score is 0.5 Standard deviations above the mean for the particular year. Therefore it is time varying. For the 10 year panel, China had 7 year of PT_bad_gov =1, and 3 with the score = 0. :

Code:

 tab year if exporter == "CHN" & PTbad_gov_exp == 0

   Year |      Freq.     Percent        Cum.
------------+-----------------------------------
       2014 |        194       33.33       33.33
       2015 |        194       33.33       66.67
       2016 |        194       33.33      100.00
------------+-----------------------------------
      Total |        582      100.00

Code:

 xtpqml exports lgdp_exporter lgdp_importer PTgood_gov_imp PTgood_gov_exp PTbad_gov_imp PTbad_gov_exp year_* if exporter == "CHN"  , fe i(pairid) cluster(pairid)

where year_* are my 10 manually created time dummies.

Code:

note: you are responsible for interpretation of non-count dep. variable
note: 1 group (1 obs) dropped because of only one obs per group
note: lgdp_exporter dropped because of collinearity
note: PTgood_gov_exp dropped because of collinearity
note: year_7 dropped because of collinearity
note: year_10 dropped because of collinearity

Iteration 0:   log likelihood = -6.730e+08  
Iteration 1:   log likelihood = -1.415e+08  
Iteration 2:   log likelihood = -1.366e+08  
Iteration 3:   log likelihood = -1.366e+08  
Iteration 4:   log likelihood = -1.366e+08  

Conditional fixed-effects Poisson regression    Number of obs      =      1855
Group variable: pairid                          Number of groups   =       189

                                                Obs per group: min =         4
                                                               avg =       9.8
                                                               max =        10

                                                Wald chi2(12)      =  9.99e+08
Log likelihood  = -1.366e+08                    Prob > chi2        =    0.0000


--------------------------------------------------------------------------------
Calculating Robust Standard Errors...
--------------------------------------------------------------------------------
       exports |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
---------------+----------------------------------------------------------------
exports        |
 lgdp_importer |   .6324759   .0822836     7.69   0.000     .4712031    .7937487
PTgood_gov_imp |    .068497    .022149     3.09   0.002     .0250857    .1119083
 PTbad_gov_imp |  -.2380663   .0930333    -2.56   0.010    -.4204082   -.0557244
 PTbad_gov_exp |   .0589282   .0546755     1.08   0.281    -.0482338    .1660902
        year_1 |  -.4184588    .045753    -9.15   0.000    -.5081331   -.3287845
        year_2 |  -.3075528   .0530627    -5.80   0.000    -.4115538   -.2035517
        year_3 |  -.4483256   .0528852    -8.48   0.000    -.5519787   -.3446725
        year_4 |    -.23516   .0462987    -5.08   0.000    -.3259037   -.1444162
        year_5 |  -.1189985   .0357842    -3.33   0.001    -.1891342   -.0488628
        year_6 |  -.0596375   .0193573    -3.08   0.002    -.0975772   -.0216979
        year_8 |   .0923141   .0343583     2.69   0.007     .0249731    .1596552
        year_9 |   .0928407   .0123281     7.53   0.000      .068678    .1170034
--------------------------------------------------------------------------------
Wald chi2(12) =  2989.31                                Prob > chi2 =   0.0000

Last edited by Cezary Baraniecki; 28 Jun 2018, 01:53.

Comment

Cezary Baraniecki

Join Date: Jun 2018

Posts: 4
#4

28 Jun 2018, 01:58

If there is any easier way I can post this information to make it legible or easier to compare, please let me know. I can post a picture of my out2feg file comparing the regressions. It's odd that the coefficients and standard errors are nearly identical save for the different in poi2hdfe dropping more variables than xtpqml. I suspect that is based on the random collinear time and country pair dummies that are dropped but that is outside of my econometrics knowledge.
Comment
Joao Santos Silva

Join Date: Apr 2014

Posts: 3000
#5

28 Jun 2018, 14:51

Dear Cezary Baraniecki,

With just one exporter, the variables from the exporter are collinear with the time dummies and you can either drop them (as poi2hdfe does) or the equivalent number of time dummies (as in xtpqml). In any case, the coefficients on the exporter are not interpretable.

Best wishes,

Joao
Comment

Announcement