Question (Omitted for multicollinearity)

[Hugo Rocha]

Dear all,

I am running some regressions in which I want to investigate the role of foreign technological transfers in firms' performance. Nevertheless, my coefficients for the variable technological transfer (percent of foreign firms' technological transfer in industry in years 1, year 2, and year 3) are omitted for multicollinearity and I do not understand exactly why. Any hints?

 * Example generated by -dataex-. To install: ssc install dataex
clear
input float(lsales_emp_us_y1 lsales_emp_us_y2) int isic3_2d float(tech_intensity lsales_emp_us_y3 pr_foreign_ipexp_isic1 pr_foreign_ipexp_isic2 pr_foreign_ipexp_isic3)
        . 10.408776 21 1         . 0 0  .4371358
        . 10.743725 22 1         . 0 0 .28848907
        .         . 36 1         . 0 0 .01065994
        .         . 18 1         . 0 0         0
        . 11.922093 16 1         . 0 0         0
        . 10.572225 36 1         . 0 0 .01065994
13.406651 13.316334 15 1         . 0 0 .58766276
 12.03148  12.24458 15 1         . 0 0 .58766276
 9.683147  9.752334 15 1   9.81349 0 0 .58766276
        .  10.34347 22 1         . 0 0 .28848907
        .         . 17 1         . 0 0  .0270067
        .         . 22 1         . 0 0 .28848907
        .  9.850177 18 1         . 0 0         0
        .         . 15 1         . 0 0 .58766276
        . 10.718296 36 1         . 0 0 .01065994
        .         . 21 1         . 0 0  .4371358
        .  8.997816 19 1         . 0 0  .3687027
        .    10.336 18 1         . 0 0         0
        .  9.659329 18 1         . 0 0         0
        .         . 17 1         . 0 0  .0270067
        .         . 15 1         . 0 0 .58766276
        .         . 21 1         . 0 0  .4371358
        .  10.87047 15 1         . 0 0 .58766276
        . 10.536288 19 1         . 0 0  .3687027
        .  12.89594 15 1         . 0 0 .58766276
        .         . 36 1         . 0 0 .01065994
        .         . 15 1         . 0 0 .58766276
13.178212 13.200963 15 1 12.956532 0 0 .58766276
        . 12.765702 15 1         . 0 0 .58766276
        .         . 15 1         . 0 0 .58766276
   10.159 10.457202 15 1 10.867983 0 0 .58766276
        .         . 15 1         . 0 0 .58766276
        .         . 22 1         . 0 0 .28848907
 10.51813  10.61326 20 1         . 0 0  .5017243
        .         . 21 1         . 0 0  .4371358
        .         . 36 1         . 0 0 .01065994
        .  9.536554 36 1         . 0 0 .01065994
        . 11.279724 20 1         . 0 0  .5017243
10.813397  10.82667 15 1 10.928153 0 0 .58766276
        .  10.19028 36 1         . 0 0 .01065994
        .  10.67447 19 1         . 0 0  .3687027
        . 11.390025 22 1         . 0 0 .28848907
        .         . 19 1         . 0 0  .3687027
        .         . 17 1         . 0 0  .0270067
        .         . 18 1         . 0 0         0
        .         . 15 1         . 0 0 .58766276
11.265405  11.14699 20 1         . 0 0  .5017243
        .         . 15 1         . 0 0 .58766276
        .         . 18 1         . 0 0         0
11.650347 11.914528 21 1  11.76032 0 0  .4371358
        .         . 15 1         . 0 0 .58766276
11.230142 11.211963 36 1  11.08524 0 0 .01065994
11.741138 12.105224 20 1  11.77603 0 0  .5017243
  10.7231  10.88147 17 1 10.500517 0 0  .0270067
        . 10.030972 19 1         . 0 0  .3687027
        .  9.393082 21 1         . 0 0  .4371358
        . 10.769274 15 1         . 0 0 .58766276
        . 13.007306 15 1         . 0 0 .58766276
 9.575509  9.814736 20 1  10.18476 0 0  .5017243
        .         . 22 1         . 0 0 .28848907
        . 10.898132 22 1         . 0 0 .28848907
        . 11.635665 22 1         . 0 0 .28848907
        .         . 36 1         . 0 0 .01065994
        . 10.735565 15 1         . 0 0 .58766276
        .         . 15 1         . 0 0 .58766276
        .  11.45083 17 1         . 0 0  .0270067
        .         . 36 1         . 0 0 .01065994
        . 11.212776 22 1         . 0 0 .28848907
        .         . 18 1         . 0 0         0
        .         . 20 1         . 0 0  .5017243
        . 10.087008 15 1         . 0 0 .58766276
        .         . 36 1         . 0 0 .01065994
 11.03718  11.44674 20 1  11.38199 0 0  .5017243
        .         . 36 1         . 0 0 .01065994
        . 10.261843 20 1         . 0 0  .5017243
        .         . 36 1         . 0 0 .01065994
        . 11.011687 15 1         . 0 0 .58766276
        .         . 22 1         . 0 0 .28848907
        .         . 15 1         . 0 0 .58766276
12.495256  12.27831 15 1         . 0 0 .58766276
11.167412 11.386222 19 1 11.364215 0 0  .3687027
        .         . 17 1         . 0 0  .0270067
        .  9.743364 36 1         . 0 0 .01065994
        .         . 22 1         . 0 0 .28848907
        . 11.091274 15 1         . 0 0 .58766276
        .         . 36 1         . 0 0 .01065994
        .         . 15 1         . 0 0 .58766276
        . 11.900102 15 1         . 0 0 .58766276
        .         . 15 1         . 0 0 .58766276
        .         . 17 1         . 0 0  .0270067
        .         . 15 1         . 0 0 .58766276
        .  9.617929 18 1         . 0 0         0
        . 10.634501 17 1         . 0 0  .0270067
 9.092028  9.361426 18 1  9.481264 0 0         0
        .  11.13112 17 1         . 0 0  .0270067
10.383406  10.69398 20 1  10.98167 0 0  .5017243
        .         . 36 1         . 0 0 .01065994
        . 12.643037 15 1  12.65743 0 0 .58766276
        .         . 22 1         . 0 0 .28848907
        .         . 20 1         . 0 0  .5017243
end

Regressions of technological transfer's share by foreign firms in year 3. Dependent variable log of sales per employee in the same year.

  reg lsales_emp_us_y2 pr_foreign_ipexp_isic2 i.isic3_2d  i.year i.countryc1 if tech_intensity==1,  robust

("note: pr_foreign_ipexp_isic2 omitted because of collinearity"). Any explanation?

Thank you very much!

[Daniel PV]

Hi Hugo Rocha in your example, the pr_foreign_ipexp_isic2 variable is only zeros.

[Carlo Lazzaro]

Hugo:
your data excerpt does not include -i.year-.
It may also be that, in your original dataset, -pr_foreign_ipexp_isic2- is perfectly collinear with -i.year-

[Hugo Rocha]

Yes, I just saw that but I still do not understand why pr_foreign_ipexp_isic1 and pr_foreign_ipexp_isic2 give me zero. The way I constructed these variables is the following:

 forval y = 1/3 { 
       egen total_ipexp_isic`y' = total(ipexp_y`y'), by(isic3_2d) 
                egen foreign_ipexp_isic`y'= total(ipexp_y`y'* dforeign_y`y'), by(isic3_2d)
                gen pr_foreign_ipexp_isic`y' = foreign_ipexp_isic`y' /total_ipexp_isic`y'

When I summarize the variables used to construct I get that foreign_ipexp1 foreign_ipexp2 are 0. Ipexp means technological transfer. That means that foreign firms didn't do technological transfers in years 1 and 2? Thanks!

[Hugo Rocha]

Hi Hugo Rocha in your example, the pr_foreign_ipexp_isic2 variable is only zeros.

Thanks, but still I showed how I constructed the variables and cannot understand why they are zero. Also, if I pick the raw variable and plot, I can see that

 pr_foreign_ipexp_isic`y'.

for year2 and year 3 should not be zero

[Hugo Rocha]

Hugo:
your data excerpt does not include -i.year-.
It may also be that, in your original dataset, -pr_foreign_ipexp_isic2- is perfectly collinear with -i.year-

That's a good point. But, my question is why? Should just remove the i.year?

[Carlo Lazzaro]

Hugo:
yes, if you want to estimate the variables you're interested in.
That said, I would give a try at:

c.year##c.year

and see if a turning point comes alive.
If this is not the case, you can keep the linear term only in.
What above provided that no collinearity issue creeps up.
If collinearity is still a nuisance, omit -year- altogether from the right-side of your regression equation.

[Hugo Rocha]

Hugo:
yes, if you want to estimate the variables you're interested in.
That said, I would give a try at:

c.year##c.year

and see if a turning point comes alive.
If this is not the case, you can keep the linear term only in.
What above provided that no collinearity issue creeps up.
If collinearity is still a nuisance, omit -year- altogether from the right-side of your regression equation.

Thank you, I tried both approaches.

 reg lsales_emp_us_y2 pr_foreign_ipexp_isic2 i.isic3_2d  c.year##c.year i.countryc1 if tech_intensity==1,  robust

(still get 0's and multicollinearity)

 reg lsales_emp_us_y2 pr_foreign_ipexp_isic2 i.isic3_2d  i.countryc1 if tech_intensity==1,  robust

(still get 0's and multicollinearity)

I think the answer lies in what Daniel PV was saying, I just don't understand why I get zeros if the actual variable is not zero. ipexp_`y' is not zero for years 2 and 3 and, as I show, in the graph the percent of that done by foreigners is not negliglible...

[Carlo Lazzaro]

Hugo:
try again adding one predictor at a time and see when the collinearity issue creeps up.

[Hugo Rocha]

Hugo:
try again adding one predictor at a time and see when the collinearity issue creeps up.

Thank you very much, let me try this at the same time I check another data. This forum has been very helpful