Issue with Running PPML for Gravity Model

Sahal Hossain

Join Date: Aug 2025
Posts: 2

Issue with Running PPML for Gravity Model

Yesterday, 10:38

Good afternoon, I am an undergraduate student. I am trying to estimate the effects of natural disasters on agricultural product exports for the period 2000-2020 and I have 27 countries in my dataset and which form around 259 trading pairs between them. I have taken aggregated product export values from the BIMTS dataset provided by OECD, and the yearly disaster metrics are count variables from EMDAT. Aside from this, I also have some of the usual variables that are kept in gravity models, such as GDP of the countries, distance, RTAs etc.

I am using PPML model as it is recommended for trade data. For my model, I am mainly interested in the outcome for the natural disaster variables for exporters and importers. I have also filtered a number of major natural disaster variables based on criteria in relevant literature. I am mainly having issues with the coefficients that Stata is returning for the disaster variables in the models. I first ran the following command:

Code:

ppmlhdfe ln_export_value num_of_dis_exp l.num_of_dis_exp l2.num_of_dis_exp num_of_major_dis_exp l.num_of_major_dis_exp l2.num_of_major_dis_exp l3.num_of_major_dis_exp num_of_major_dis_imp num_of_dis_imp ln_gdp_exp ln_gdp_imp ln_pop_exp ln_pop_imp ln_distcap outdegree indegree rta wto comlang_off comcol hegemony_relationship

which returned the following result:

Code:

PPML regression                                   No. of obs      =      4,921
                                                  Residual df     =      4,899
                                                  Wald chi2(21)   =    3020.22
Deviance             =  11212.80631               Prob > chi2     =     0.0000
Log pseudolikelihood =  -15929.8341               Pseudo R2       =     0.1383
---------------------------------------------------------------------------------------
                      |               Robust
      ln_export_value | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
----------------------+----------------------------------------------------------------
       num_of_dis_exp |
                  --. |   .0045309   .0013589     3.33   0.001     .0018675    .0071943
                  L1. |   .0043782   .0014044     3.12   0.002     .0016256    .0071308
                  L2. |   .0041736   .0013842     3.02   0.003     .0014606    .0068866
                      |
 num_of_major_dis_exp |
                  --. |  -.0018753   .0021146    -0.89   0.375    -.0060199    .0022693
                  L1. |   -.004591   .0023506    -1.95   0.051     -.009198    .0000161
                  L2. |  -.0057954   .0023962    -2.42   0.016    -.0104918    -.001099
                  L3. |   -.004473   .0018278    -2.45   0.014    -.0080555   -.0008905
                      |
 num_of_major_dis_imp |  -.0008275   .0016083    -0.51   0.607    -.0039798    .0023247
       num_of_dis_imp |  -.0014102   .0009266    -1.52   0.128    -.0032263    .0004058
           ln_gdp_exp |   .0806123   .0043619    18.48   0.000     .0720632    .0891614
           ln_gdp_imp |   .0961636   .0038321    25.09   0.000     .0886527    .1036744
           ln_pop_exp |  -.0369827   .0056349    -6.56   0.000    -.0480269   -.0259385
           ln_pop_imp |   .0332621   .0053604     6.21   0.000     .0227559    .0437682
           ln_distcap |  -.1153456   .0068734   -16.78   0.000    -.1288173   -.1018739
            outdegree |   .0061965   .0003727    16.63   0.000      .005466     .006927
             indegree |  -.0015845   .0004009    -3.95   0.000    -.0023702   -.0007988
                  rta |   .0170505   .0085625     1.99   0.046     .0002684    .0338327
                  wto |   .1001379   .0196778     5.09   0.000     .0615701    .1387056
          comlang_off |   .1015086    .010186     9.97   0.000     .0815444    .1214728
               comcol |   .0344936   .0237892     1.45   0.147    -.0121324    .0811195
hegemony_relationship |  -.0189402   .0176511    -1.07   0.283    -.0535358    .0156554
                _cons |  -1.195602   .1773664    -6.74   0.000    -1.543233   -.8479699
---------------------------------------------------------------------------------------

what is confusing me is the positive coefficients for number of disasters as it does not make any intuitive sense. Afterwards, I took importer and exporter fixed effects and pair-wise fixed effects as well using the following command:

Code:

ppmlhdfe ln_export_value num_of_dis_exp l.num_of_dis_exp l2.num_of_dis_exp num_of_major_dis_exp l.num_of_major_dis_exp l2.num_of_major_dis_exp l3.num_of_major_dis_exp num_of_major_dis_imp num_of_dis_imp ln_gdp_exp ln_gdp_imp ln_pop_exp ln_pop_imp ln_distcap outdegree indegree rta wto comlang_off comcol hegemony_relationship, absorb(exporter_id importer_id pair_id) vce(cluster dist)

Which had the following results

Code:

HDFE PPML regression                              No. of obs      =      4,921
Absorbing 3 HDFE groups                           Residual df     =        300
Statistics robust to heteroskedasticity           Wald chi2(17)   =     103.21
Deviance             =  6669.246421               Prob > chi2     =     0.0000
Log pseudolikelihood = -13658.05416               Pseudo R2       =     0.2612

Number of clusters (dist)   =        301
                                          (Std. err. adjusted for 301 clusters in dist)
---------------------------------------------------------------------------------------
                      |               Robust
      ln_export_value | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
----------------------+----------------------------------------------------------------
       num_of_dis_exp |
                  --. |  -.0005491   .0008931    -0.61   0.539    -.0022994    .0012013
                  L1. |   .0004038   .0008164     0.49   0.621    -.0011962    .0020038
                  L2. |   .0001799    .000816     0.22   0.826    -.0014195    .0017792
                      |
 num_of_major_dis_exp |
                  --. |   .0002221   .0013716     0.16   0.871    -.0024661    .0029103
                  L1. |   .0000312   .0014169     0.02   0.982    -.0027459    .0028083
                  L2. |  -.0001937   .0014434    -0.13   0.893    -.0030228    .0026353
                  L3. |   .0003116   .0010321     0.30   0.763    -.0017113    .0023345
                      |
 num_of_major_dis_imp |  -.0034234   .0019005    -1.80   0.072    -.0071484    .0003016
       num_of_dis_imp |    .000843   .0005096     1.65   0.098    -.0001558    .0018417
           ln_gdp_exp |   -.073666   .0595342    -1.24   0.216    -.1903509    .0430189
           ln_gdp_imp |  -.2679774   .1121216    -2.39   0.017    -.4877317   -.0482232
           ln_pop_exp |   .7049349   .3080901     2.29   0.022     .1010893     1.30878
           ln_pop_imp |   1.122893   .3021418     3.72   0.000     .5307059     1.71508
           ln_distcap |          0  (omitted)
            outdegree |    .003352   .0015725     2.13   0.033     .0002699     .006434
             indegree |   .0005123   .0012268     0.42   0.676    -.0018921    .0029168
                  rta |  -.0274514   .0169421    -1.62   0.105    -.0606574    .0057546
                  wto |   .1950682   .0623619     3.13   0.002     .0728412    .3172952
          comlang_off |          0  (omitted)
               comcol |          0  (omitted)
hegemony_relationship |          0  (omitted)
                _cons |  -21.95647   5.045151    -4.35   0.000    -31.84479   -12.06816

In this case, the coefficients for GDP became negative, and there were also some sign changes in the disaster variables. I also tried running the model with just the number of major disaster variables (excluding the overall number of disaster variables), and the results for that variable came out positive and insignificant. As a result, I am somewhat confused about what the underlying issue in my model, command or data could be. I would appreciate any sort of insight on this matter from experts. Thank you very much for your time. I am also sharing a sample of my dataset below:

[CODE]----------------------- copy starting from the next line -----------------------

Code:

* Example generated by -dataex-. For more info, type help dataex
clear
input int year str3(iso3_o iso3_d) byte(comlang_off comcol hegemony_relationship num_of_major_dis_exp num_of_major_dis_imp) float(outdegree indegree ln_gdp_exp ln_gdp_imp ln_pop_exp ln_pop_imp ln_distcap ln_export_value)
2000 "ARG" "AUS" 0 0 0 0  0  4.161259   .467287  26.70468   27.4952 17.432196 16.761465 9.372289  12.50696
2001 "ARG" "AUS" 0 0 0 1  0  4.954784    .36824  26.65816  27.51449 17.443174 16.774303 9.372289 12.782493
2002 "ARG" "AUS" 0 0 0 0  1   4.37543   .431357  26.54335  27.55414 17.453869 16.785679 9.372289  15.13579
2003 "ARG" "AUS" 0 0 0 1  0  5.614964   .664061  26.62863  27.58393 17.464201 16.797182 9.372289   16.6069
2004 "ARG" "AUS" 0 0 0 0  0  5.949921   .585038  26.71473  27.62601  17.47434 16.807873 9.372289  14.39857
2005 "ARG" "AUS" 0 0 0 0  0  6.934795   .662563  26.79861  27.66058 17.484615 16.820045 9.372289 14.071252
2006 "ARG" "AUS" 0 0 0 0  1  6.278434   .769001    26.876  27.67981 17.494898  16.83354 9.372289 14.274624
2007 "ARG" "AUS" 0 0 0 0  1 10.481373   .927276  26.96354   27.7172 17.504808 16.851791 9.372289 16.364223
2008 "ARG" "AUS" 0 0 0 0  1  15.12983  1.062732  27.00179  27.75324 17.514938  16.87183 9.372289 15.582834
2009 "ARG" "AUS" 0 0 0 0  2  7.046975   .957076 26.941866  27.77078 17.525536 16.892439 9.372289  15.48392
2010 "ARG" "AUS" 0 0 0 0  3  12.42004   1.12896 27.038624 27.796535 17.536098 16.907995 9.372289 15.802192
2011 "ARG" "AUS" 0 0 0 0  1  16.01985  1.407277 27.096586  27.82164 17.546747  16.92189 9.372289 15.753852
2012 "ARG" "AUS" 0 0 0 0  0 15.847264  1.399465 27.086294 27.862133 17.557024 16.939348 9.372289  15.90519
2013 "ARG" "AUS" 0 0 0 1  1  15.52329  1.335947 27.110146 27.885664 17.566954  16.95656 9.372289 15.526612
2014 "ARG" "AUS" 0 0 0 0  1  11.49686  1.497158  27.08457 27.908653  17.57727 16.971476 9.372289 15.776915
2015 "ARG" "AUS" 0 0 0 0  1 11.272637   1.39542  27.11183 27.931126 17.587744 16.985868 9.372289 15.958762
2016 "ARG" "AUS" 0 0 0 1  1 12.623044  1.427795 27.089737 27.960325 17.597431 17.001488 9.372289 15.413207
2017 "ARG" "AUS" 0 0 0 0  1 11.991758   1.63659  27.11853  27.98168 17.606245 17.017956 9.372289 15.631765
2018 "ARG" "AUS" 0 0 0 1  1 11.104562  1.588034 27.091454 28.009457 17.614473 17.032915 9.372289 15.154465
2019 "ARG" "AUS" 0 0 0 0  2 15.508174   1.69718 27.070656 28.029797 17.621584 17.047691 9.372289 14.659573
2020 "ARG" "AUS" 0 0 0 0  3 14.567545  1.746344  26.96743 28.029797  17.62643 17.060024 9.372289 15.162526
2021 "ARG" "AUS" 0 0 0 0  1 19.196016  1.728934 27.067146  28.04973  17.62909 17.061434 9.372289 15.329987
2000 "ARG" "BGD" 0 0 0 0  2  4.161259   .667196  26.70468  25.14803 17.432196 18.717403 9.727705  16.23662
2001 "ARG" "BGD" 0 0 0 1  4  4.954784   .844089  26.65816 25.197554 17.443174 18.734074 9.727705 15.488054
2002 "ARG" "BGD" 0 0 0 0  2   4.37543   .771186  26.54335  25.23517 17.453869 18.749508 9.727705   12.1258
2003 "ARG" "BGD" 0 0 0 1  1  5.614964  1.123751  26.62863 25.281475 17.464201 18.764061 9.727705  9.103163
2004 "ARG" "BGD" 0 0 0 0  4  5.949921  1.292155  26.71473 25.328436  17.47434 18.777676 9.727705 16.769659
2005 "ARG" "BGD" 0 0 0 0  3  6.934795  1.214314  26.79861 25.396095 17.484615 18.790283 9.727705 18.097454
2006 "ARG" "BGD" 0 0 0 0  1  6.278434  1.624914    26.876 25.459465 17.494898 18.801893 9.727705 17.489979
2007 "ARG" "BGD" 0 0 0 0  4 10.481373  2.106431  26.96354 25.527287 17.504808  18.81265 9.727705 18.025837
2008 "ARG" "BGD" 0 0 0 0  1  15.12983  2.246869  27.00179   25.5908 17.514938 18.822636 9.727705 17.927904
2009 "ARG" "BGD" 0 0 0 0  2  7.046975  2.747937 26.941866  25.63592 17.525536 18.831953 9.727705  16.32062
2010 "ARG" "BGD" 0 0 0 0  2  12.42004  3.618653 27.038624  25.69308 17.536098  18.84072 9.727705  14.31912
2011 "ARG" "BGD" 0 0 0 0  2  16.01985  4.364283 27.096586 25.753704 17.546747 18.849804 9.727705 18.058853
2012 "ARG" "BGD" 0 0 0 0  3 15.847264   3.57821 27.086294 25.817017 17.557024 18.859388 9.727705 17.637295
2013 "ARG" "BGD" 0 0 0 1  1  15.52329  4.259815 27.110146  25.87656 17.566954 18.868773 9.727705 18.533525
2014 "ARG" "BGD" 0 0 0 0  2  11.49686  4.262775  27.08457  25.93275  17.57727 18.877884 9.727705 18.241816
2015 "ARG" "BGD" 0 0 0 0  2 11.272637  4.614795  27.11183 25.996265 17.587744 18.886822 9.727705 17.787113
2016 "ARG" "BGD" 0 0 0 1  2 12.623044  2.859848 27.089737   26.0656 17.597431 18.895746 9.727705  17.72174
2017 "ARG" "BGD" 0 0 0 0  2 11.991758   4.98801  27.11853  26.13044 17.606245 18.904255 9.727705 18.670645
2018 "ARG" "BGD" 0 0 0 1  0 11.104562  5.250591 27.091454  26.19973 17.614473 18.912464 9.727705 18.152012
2019 "ARG" "BGD" 0 0 0 0  3 15.508174  6.251867 27.070656 26.276226 17.621584 18.920929 9.727705 18.316277
2020 "ARG" "BGD" 0 0 0 0  2 14.567545   6.97281  26.96743 26.310514  17.62643 18.929293 9.727705 18.734407
2021 "ARG" "BGD" 0 0 0 0  2 19.196016  8.909783 27.067146 26.375755  17.62909 18.937443 9.727705 18.540874
2000 "ARG" "BRA" 0 0 0 0  1  4.161259  2.160713  26.70468 27.804974 17.432196  18.97467 7.755339 20.933893
2001 "ARG" "BRA" 0 0 0 1  1  4.954784  1.752217  26.65816 27.813343 17.443174 18.987705 7.755339  20.84301
2002 "ARG" "BRA" 0 0 0 0  1   4.37543  1.770926  26.54335  27.84613 17.453869 19.000118 7.755339 20.583223
2003 "ARG" "BRA" 0 0 0 1  1  5.614964  2.160287  26.62863 27.854164 17.464201  19.01192 7.755339   20.7335
2004 "ARG" "BRA" 0 0 0 0  2  5.949921  1.720203  26.71473   27.9162  17.47434  19.02322 7.755339 20.599745
2005 "ARG" "BRA" 0 0 0 0  0  6.934795  1.689548  26.79861  27.94583 17.484615  19.03418 7.755339  20.54539
2006 "ARG" "BRA" 0 0 0 0  1  6.278434  2.299382    26.876  27.98168 17.494898 19.044762 7.755339 20.881567
2007 "ARG" "BRA" 0 0 0 0  2 10.481373  3.137951  26.96354  28.04313 17.504808 19.054886 7.755339  21.12944
2008 "ARG" "BRA" 0 0 0 0  3  15.12983  4.126668  27.00179 28.094755 17.514938 19.064466 7.755339  21.29766
2009 "ARG" "BRA" 0 0 0 0  3  7.046975  3.038094 26.941866 28.088446 17.525536 19.073421 7.755339  20.85564
2010 "ARG" "BRA" 0 0 0 0  1  12.42004  4.226881 27.038624  28.16165 17.536098  19.08183 7.755339  21.08443
2011 "ARG" "BRA" 0 0 0 0  3  16.01985  5.293038 27.096586    28.202 17.546747  19.08997 7.755339  21.49753
2012 "ARG" "BRA" 0 0 0 0  2 15.847264  4.753578 27.086294  28.21881 17.557024 19.098085 7.755339  21.41978
2013 "ARG" "BRA" 0 0 0 1  1  15.52329  5.585747 27.110146   28.2516 17.566954  19.10619 7.755339  21.11085
2014 "ARG" "BRA" 0 0 0 0  2  11.49686  4.802975  27.08457  28.25696  17.57727 19.114254 7.755339  20.86071
2015 "ARG" "BRA" 0 0 0 0  1 11.272637   3.54652  27.11183  28.21881 17.587744  19.12217 7.755339 21.039684
2016 "ARG" "BRA" 0 0 0 1  0 12.623044  4.861575 27.089737 28.184906 17.597431  19.12979 7.755339 21.262035
2017 "ARG" "BRA" 0 0 0 0  1 11.991758  4.042827  27.11853    28.202 17.606245 19.137074 7.755339  21.16939
2018 "ARG" "BRA" 0 0 0 1  0 11.104562  4.115983 27.091454  28.21881 17.614473  19.14391 7.755339 21.327425
2019 "ARG" "BRA" 0 0 0 0  1 15.508174  4.196014 27.070656  28.22986 17.621584 19.150427 7.755339  21.30669
2020 "ARG" "BRA" 0 0 0 0  1 14.567545  4.017918  26.96743 28.196335  17.62643  19.15622 7.755339 21.134523
2021 "ARG" "BRA" 0 0 0 0  4 19.196016  5.230016 27.067146  28.24079  17.62909 19.160475 7.755339  21.53823
2000 "ARG" "CHN" 0 0 0 0  8  4.161259  6.428595  26.70468 28.671297 17.432196 20.956474 9.865993  20.09719
2001 "ARG" "CHN" 0 0 0 1 12  4.954784  6.952546  26.65816 28.749435 17.443174  20.96374 9.865993  20.56557
2002 "ARG" "CHN" 0 0 0 0 14   4.37543  7.131054  26.54335  28.83699 17.453869  20.97044 9.865993 20.071836
2003 "ARG" "CHN" 0 0 0 1 15  5.614964  11.70507  26.62863 28.933933 17.464201  20.97667 9.865993 20.968706
2004 "ARG" "CHN" 0 0 0 0 10  5.949921  16.70713  26.71473  29.02974  17.47434 20.982607 9.865993  20.96224
2005 "ARG" "CHN" 0 0 0 0 19  6.934795 18.242695  26.79861 29.139534 17.484615  20.98849 9.865993 21.331245

Tags: None

Joao Santos Silva

Join Date: Apr 2014

Posts: 3025
#2

Today, 00:49

Dear Sahal Hossain,

The major problem with what you are doing is that you are using a dependent variable in logs. The major advantage of PPML is that is uses a dependent variable in levels; using a dependent variable in logs may lead to invalid estimates and PPML avoids that. See if things improve if you make that change.

Beat wishes,

Joao
1 like
Comment

Sahal Hossain

Join Date: Aug 2025
Posts: 2

Today, 11:22

Dear @Joao Santos Silva Sir,

Thank you very much for your suggestion and for helping me understand this. I reran the regression using the following code this time:

Code:

ppmlhdfe export_value num_of_dis_exp l.num_of_dis_exp l2.num_of_dis_exp num_of_major_dis_exp l.num_of_major_dis_exp l2.num_of_major_dis_exp l3.num_of_major_dis_exp num_of_major_dis_imp num_of_dis_imp ln_gdp_exp ln_gdp_imp ln_pop_exp ln_pop_imp ln_distcap outdegree indegree rta wto comlang_off comcol hegemony_relationship, absorb(pair_id) vce(cluster dist)

Which yielded a result that makes much more sense intuitively.

Code:

HDFE PPML regression                              No. of obs      =      4,921
Absorbing 1 HDFE group                            Residual df     =        300
Statistics robust to heteroskedasticity           Wald chi2(17)   =    2362.55
Deviance             =  1.10466e+11               Prob > chi2     =     0.0000
Log pseudolikelihood = -5.52330e+10               Pseudo R2       =     0.9799

Number of clusters (dist)   =        301
                                    (Std. err. adjusted for 301 clusters in dist)
---------------------------------------------------------------------------------
                |               Robust
   export_value | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
----------------+----------------------------------------------------------------
 num_of_dis_exp |
            --. |  -.0084032   .0039961    -2.10   0.035    -.0162355    -.000571
            L1. |  -.0067222   .0031367    -2.14   0.032      -.01287   -.0005744
            L2. |  -.0089766   .0035183    -2.55   0.011    -.0158723   -.0020809
                |
num_of_major~xp |
            --. |  -.0043574   .0065979    -0.66   0.509    -.0172889    .0085742
            L1. |   .0112812   .0057236     1.97   0.049     .0000631    .0224992
            L2. |   .0181635    .006141     2.96   0.003     .0061272    .0301997
            L3. |    .012584   .0042424     2.97   0.003      .004269    .0208989
                |
num_of_major~mp |   .0083644   .0037056     2.26   0.024     .0011016    .0156272
 num_of_dis_imp |   .0020358   .0020987     0.97   0.332    -.0020776    .0061491
     ln_gdp_exp |   .0373858   .1749167     0.21   0.831    -.3054446    .3802162
     ln_gdp_imp |   .6493967   .2616866     2.48   0.013     .1365004    1.162293
     ln_pop_exp |  -1.196168   .8267145    -1.45   0.148    -2.816499    .4241621
     ln_pop_imp |   3.074979   .8722083     3.53   0.000     1.365482    4.784476
     ln_distcap |          0  (omitted)
      outdegree |   .0327526   .0070118     4.67   0.000     .0190098    .0464954
       indegree |   .0006305   .0035959     0.18   0.861    -.0064172    .0076783
            rta |   .0838937   .1124331     0.75   0.456     -.136471    .3042585
            wto |   1.215673   .2234146     5.44   0.000     .7777882    1.653557
    comlang_off |          0  (omitted)
         comcol |          0  (omitted)
hegemony_rela~p |          0  (omitted)
          _cons |   -38.2628   16.15396    -2.37   0.018    -69.92399   -6.601611
---------------------------------------------------------------------------------

Absorbed degrees of freedom:
-----------------------------------------------------+
 Absorbed FE | Categories  - Redundant  = Num. Coefs |
-------------+---------------------------------------|
     pair_id |       259           0         259     |
-----------------------------------------------------+

I have another query though, I received a warning message on Stata during the command

Code:

dependent variable takes very low values after standardizing (2.5189e-11)

I do believe that this should not cause any significant issues, but I would really appreciate it if you could kindly share your insight regarding this as well.

And thank you very much for your kind assitance and time regarding this matter.

Announcement

Issue with Running PPML for Gravity Model

Comment

Comment