Different signs of my explanatory variables

[Ammanna Azwai]

Hello, I'm running some regressions using Fixed effects estimations. In fact, I'm trying to study the impact of increasing and decreasing the governmental rules on my dependent variable. Moreover, I have classified all the countries who did increase their rule in last 20 years and used a binary which gives 1 for any country increased its activity and zero otherwise (country increased its rules on bank activities). Similarly, all countries that did decrease their rule are classified and multiplied by a binary (decrease) which take value of 1 in the case that country relaxed its activity, and zero otherwise. However, after running my regression , it turns out that both inc*governmental rules and decrease*governmental rules have the sign (both positive) . is there any suggestions to help with this as we are expecting both interactions to have the opposite sign. Thanks a lot for your help in advance.

[Ammanna Azwai]

xtreg A B C D E F J H  inc_govrules dec_govrules , fe cluster(Country)
Fixed-effects (within) regression               Number of obs      =     44433
Group variable: Bvdidnum                        Number of groups   =      8566
 
R-sq:  Within  = 0.0636                         Obs per group: min =         1
       Between = 0.0920                                        avg =       5.2
       Overall = 0.0794                                        max =        13
 
                                                F(9,21)            =  9.67e+06
corr(u_i, Xb)  = -0.2268                        Prob > F           =    0.0000
 
                                     (Std. err. adjusted for 22 clusters in Country)
------------------------------------------------------------------------------------
                   |               Robust
                  A | Coefficient  std. err.      t    P>|t|     [95% conf. interval]
-------------------+----------------------------------------------------------------
                   B |   .3077334   .0449607     6.84   0.000     .2142324    .4012344
                   C |  -1.191879   .1244949    -9.57   0.000    -1.450781   -.9329778
                    D|  -.0247429   .0143517    -1.72   0.099    -.0545889    .0051031
                   E |    .056238    .118463     0.47   0.640    -.1901193    .3025953
                   F |  -5.146442   6.618746    -0.78   0.446    -18.91088    8.617994
                   J |   .0227529   .0186228     1.22   0.235    -.0159754    .0614812
                   H |   .4109422   .5995536     0.69   0.501    -.8358977    1.657782
        inc_govrules |   .1144277   .0109234    10.48   0.000     .0917112    .1371441
        dec_govrules |   .1752997   .0091983    19.06   0.000     .1561708    .1944285
               _cons |   16.00984   .6063505    26.40   0.000     14.74886    17.27081
-------------------+----------------------------------------------------------------
           sigma_u |  3.6134625
           sigma_e |  1.3005847
               rho |  .88531001   (fraction of variance due to u_i)
------------------------------------------------------------------------------------

[Clyde Schechter]

Your command appears to be correctly coded.

There are many possibilities here. For a start, your expectation could simply be wrong. Perhaps changes in rules, in either direction, have effects in the same direction and they contrast with continuation of the status quo ante.

But let's also consider some common ways in which the model could be wrong:

1. There could be errors in the data. Are you sure that you have correctly identified which countries increased, and which decreased rules? Are the values of the other variables correct?
2. Are the measurements used to ascertain the values of the variables really valid measurements of what you think they represent conceptually?
3. Have you drawn and inspected a directed acyclic graph showing the presumed causal relationships among all the variables in your model? Perhaps you have mistakenly included a collider variable, which is giving you spurious results.
4. Are the relationships between each of the explanatory variables and outcome variable A really linear? If they are not, it is possible that the effect estimates for the rules variables are being distorted because they are "correcting" for the non-linearity of other variables' relationships to A. You may need to transform some of these variables, or include interactions to properly specify the relationships.
5. Is there something unusual about the countries or Bvdidnum's included in the sample, or the years in which you sampled them, that would cause the unexpected findings to apply to them, even though in a more general sample the findings would be the way you expect them to be?

These are, I think, the commonest modeling errors to consider. Applying these considerations to your problem requires knowledge of what all the variables are and how they relate to each other, as well as substantive knowledge of the field. I don't meet any of those qualifications. If you do not feel comfortable with your own ability to answer these questions, you should seek advice on them from somebody in your own field for further help.