Compare levelsof to original var in if condition

[Henry Strawforrd]

I am looping over a variable, running a regression, and extracting the coefficient for each level (to plot on a twoway graph). I created a new variable to collapse later (maybe not most efficient but easy) and store the values in there. But I am struggling with the fact that the value in the local of levelsof is different to the level of the variable.

This MWE does not work but should illustrate what I am attempting to do

sysuse auto.dta, replace

gen corr = .

levelsof foreign, local(foreign)
foreach f in `foreign' {
    reg price weight
    mat result = e(b)
    replace corr = result[1,1] if foreign == `f'
}

[Nick Cox]

Your regression is for all available data. Being inside a loop doesn't affect that. I don't know if that answers your question, which I find hard to follow.

I first wrote levelsof against some scepticism that it was needed or useful. Now the joke's on me as I see many examples in which I wouldn't use it. For your token example I would do this:

. sysuse auto, clear
(1978 automobile data)

. rangestat (reg) price weight, int(foreign 0 0)

. tabdisp foreign, c(b_weight)

--------------------------------
Car       |
origin    |             b_weight
----------+---------------------
 Domestic |            2.9948135
  Foreign |            5.3620402
--------------------------------

I note that

1. To be fair, rangestat from SSC wasn't written until after levelsof

2. What you are picking up after the regression is not the correlation, as your variable name might be thought to imply, but the coefficient of weight.

[Henry Strawforrd]

Thank you for the reply. In the real example the variable over which I want to loop is year and there are years. A table might be okay for that, but a twoway graph of the coefficient on y and the year on the x axis will be much better I think. But to graph this I need to store the coefficients somewhere. That's why I wanted to create this variable, call it b_weight, and store the regression coefficient for each year in there. Then I will collapse and plot

collapse (mean) b_weight, by(year)
twoway connected b_weight year

On the correlation. I thought the was just a scaling issue but indeed even for rescaled variables the two are different. Interesting

. sysuse auto.dta, replace
(1978 automobile data)

. 
. foreach var of varlist price mpg {
  2.     qui sum `var'
  3.         replace `var' = (`var' - `r(min)') / (`r(max)'-`r(min)')
  4. }
variable price was int now float
(74 real changes made)
variable mpg was int now float
(74 real changes made)

. 
. corr price mpg
(obs=74)

             |    price      mpg
-------------+------------------
       price |   1.0000
         mpg |  -0.4686   1.0000


. reg price mpg

      Source |       SS           df       MS      Number of obs   =        74
-------------+----------------------------------   F(1, 72)        =     20.26
       Model |  .876278986         1  .876278986   Prob > F        =    0.0000
    Residual |  3.11437388        72  .043255193   R-squared       =    0.2196
-------------+----------------------------------   Adj R-squared   =    0.2087
       Total |  3.99065287        73  .054666478   Root MSE        =    .20798

------------------------------------------------------------------------------
       price | Coefficient  Std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
         mpg |  -.5491824   .1220154    -4.50   0.000    -.7924156   -.3059492
       _cons |   .4039103   .0459861     8.78   0.000     .3122386     .495582
------------------------------------------------------------------------------

[Nick Cox]

This appears to miss a main point of #2, which is that the code I give also puts the regression results into new variables directly, so that you could graph them straight away. I couldn't have used tabdisp otherwise.

You don't need to collapse as something akin to

egen tag = tag(foreign)
scatter b_weight foreign if tag

will suffice.

Your procedure implicitly assumes that range is proportional to SD, which is likely to be roughly correct, but highly unlikely to be exactly correct. If you standardize each variable to (value MINUS mean) / SD you will then find that the correlation and regression coefficients are identical.