[Marcos Almeida]

Following Carlo's track, I wonder whether a survey design - svy + weights - would'nt fit with your needs.

[Ridvan Peshkopia]

Carlo, thank you! I did, but I am not finding the specific answer. Let me be more specific: In my dataset the male frequency is 61.19% and the female frequency is 38.81%. Instead, the INSTAT offers male and female as 51.52% and 49.48% of the population respectively. How can I create a new variable that would represent the male and female frequencies for the population? Any response would help...

[Ridvan Peshkopia]

Marco, I have a simple random sample collected via a cellphone RDD. Apparently, Albanian females use less cellphone, and we have less females than males in our sample. The problem is that the discrepancy between my female/male ratios and that of the population is problematic to my dependent variables. For instance, I am finding that Albanian female feeling temperatures toward Greeks as significantly lower than those of Albanian males, and that impacts the means unless I find a way to weight this variable in a way that it would reflect the population weights.

Also, I am not sure Ii understood your suggestion: could you be a bit more specific with it?

Best,
Ridvan

[Carlo Lazzaro]

Ridvan:
Marcos wisely pointed you to the -svy- design (although I cannot actually get whether your research is based on a survey or other sample design).
Be as it may, I think -weights- can give you some clues to elaborate on.
In the following toy-example, I arbitarily assume that the population from which the sample of car reported in -auto.dta- is equally composed of -domestic- and -foreign- cars (whereas the sample is 70.27% in favour of -domestic- ones):

. sysuse auto.dta
(1978 Automobile Data)

. tab foreign

   Car type |      Freq.     Percent        Cum.
------------+-----------------------------------
   Domestic |         52       70.27       70.27
    Foreign |         22       29.73      100.00
------------+-----------------------------------
      Total |         74      100.00

. g prob=0.50 if foreign==0
(22 missing values generated)

. replace prob=0.50 if foreign==1
(22 real changes made)

. regress price mpg i.rep78 [pw=1/prob]
(sum of wgt is   1.3800e+02)

Linear regression                               Number of obs     =         69
                                                F(5, 63)          =       5.77
                                                Prob > F          =     0.0002
                                                R-squared         =     0.2584
                                                Root MSE          =     2605.8

------------------------------------------------------------------------------
             |               Robust
       price |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         mpg |  -280.2615   61.23892    -4.58   0.000    -402.6377   -157.8853
             |
       rep78 |
          2  |   877.6347   1061.865     0.83   0.412    -1244.334    2999.603
          3  |   1425.657   649.6622     2.19   0.032     127.4107    2723.903
          4  |   1693.841   583.9808     2.90   0.005     526.8484    2860.834
          5  |   3131.982   840.6517     3.73   0.000     1452.074    4811.891
             |
       _cons |   10449.99   1332.468     7.84   0.000     7787.266    13112.72
------------------------------------------------------------------------------

PS: Ridvan: I've just read your reply about a simple random sample. Hence, no survey design. I do hope that what above can contribute to your point-to-point reply to reviewers.

[Ridvan Peshkopia]

Carlo, thank you so very much. I sense that this is exactly what I needed. However, I would humbly beg for a bit more of your time. Actually I am running tobit models, but I am not sure I want to use weights in the regression analysis. I would rather use them only in the descriptive statistics. However, in the case I decide to use weights in the models, do I need to add the [pw=1/prob] expression, or is it specific to the mock dataset that you used? Any response would help...
Best,
Ridvan

[Carlo Lazzaro]

Ridvan:
-[pw=1/prob]- notation was retrieved from an example reported in -help weights-.
Obvioulsy, you can rename -prob- as you prefer.

[Ridvan Peshkopia]

Carlo, again, I got myself the meaning of the [pw=1/prob]. However, when I use it to the sum command, it does not recognize it. How can I use the new prob variable to measure the weighted mean? Sorry for all this fuss...

[Ridvan Peshkopia]

Again, I need the mean of another variable (feeling temperatures toward Greeks, say) that has been affected by the lopsided number of males in my Sex variable. Now, how can I use the prob variable to assess that mean?
Humbly,
Ridvan

[Carlo Lazzaro]

Ridvan:
the only Stata official command suitable for descriptive statistics that supports -pweights- is -mean-.
However, it does not return a standard deviation, but a standard error.

[Ridvan Peshkopia]

Carlo, okay, great, Ii was able to perform what I needed. I am sooooooo grateful of your help. Thank you, thank you, thank you!

[Carlo Lazzaro]

Ridvan:
my pleasure.
All the best for your re-submission.

[Ridvan Peshkopia]

Carlo, I do have one more question. I did generate two weighed variables, one of them that weights the age groups (wgtage) and the other sex (wgtsex). Then I estimated the mean for another variable, FTGr (feeling temperatures toward Greeks).
. mean FTGr [pw = 1/wgtage & 1/wgtsex]

However, the estimated means that I received is the same with what I receive with mean FTGR. I do receive different estimated means though if I condition the command with only one of those variable weights (either [pw = 1/wgtage] or [1/wgtsex]). Am I missing something with the command?

[Ridvan Peshkopia]

Dear Carlo and/or Marco (and/or anybody else),
After I spent my entire day looking for answers beyond what Carlo gave me earlier today, I am back with one more question. I am using a simple random sample, and want to estimate the mean of a variable (FTGr - Feeling temperature toward Greeks) using quite a few weights (generated following some frequency estimation of population strata). Following Carlo's advise, I generated a couple weights. I am able to estimate a weighted mean for the FTGr variable using only one of those weights (for instance, I receive weighed FTGr if I run the command mean FTGr [pw = 1/wgtage], but am not able to acquire any weighted FTGr mean if I add a second weight. In this case, the command mean FTGr [pw = 1/wgtage & 1/wgtsex] only gives me the unweighted FTGr. What I got wrong? Can I use more than one weight in estimating the mean? If so, what command should I apply?
Best,
Ridvan

[Marcos Almeida]

Ridvan, if I understand correctly, you wish to apply ‘several’ weights in the same model. Unless this is a multilevel model - and it seems not to be - or a survey - and it also seems not to be - , the OLS regression won’t probably cope with ‘weighting’ in a row.

On a risky note, perhaps you may try weighted least-squares linear regression. I have no practical experience with it but, in short: the command is - regress - with the - noconst - option; yvar as well as xvars can be ‘weighted ‘ by generating new vars (example: gen yvar2 = weightvar*yvar1; gen xvar2 =xvar1*weightvar; finally: reg yvar2 xvar2 weightvar, noconst).

I strongly suggest to take a close look at the literature.

I’m afraid I’m not the best person to help you further, for at least two reasons: first, I guess the ‘selection’ of weights here is fundamentally a ‘post-hoc’ strategy to make, say, ‘ends meet’, not necessarily the search for the most appropriate model; second, because I haven’t yet used WLS in my ‘real life’ models, that would be daredevilness of mine, which is against my nature. Hence, hopefully that helped you!

[Ridvan Peshkopia]

Thank you very much, Marcos. That was extremely helpful.
Best,
Ridvan