Dichotomizing a Factor Variable

[Jae Park]

Dear All,

Hi, I have a question about dichotomizing a factor variable (or a variable that contains non-integer values).


I am currently using -prchange- from spost9 after conducting a logistic regression.
One of my major independent variables is an average of six related variables (-alpha- value 0.8).
The six original variables are set variables, which were meant to measure participation frequency in six different areas from level 1 to 4.
I would dichotomize one of the original variables by recoding 1 through 2 to 0 and 3 through 4 to 1.
Yet, after I combined these six variables it became hard to dichotomize this variable.

As many of you might know already, -prchange- command would not accept factor variables for analysis.

Therefore, I have to dichotomize my averaged variable that contains non-integer values.
I have been searching for any standard or recommended method to do so, yet I failed to find any.
Does this mean dichotomizing an averaged variable is against mathematical rules?
Would it be wrong to dichotomize my average variable by recoding any values above 2 to 1?
(In other words, what should be the cut point in this case?)

Any recommendation or advice will be welcome.
Thank you for reading this post.

[Nick Cox]

You can dichotomise anything you like, and that can make sense.

(speed of car > speed limit)

keeps important information. Below, you're legal. Above, you're illegal. Watch out for police, but slow down now. If it's somebody else's car, keep out of its way.

But you could probably collect a hundred articles and books warning that you should not usually do this in statistical analysis, as in most cases it is arbitrary, increases the scope for error and throws away information.

I don't know anything much about the prchange command (not function), but if it doesn't apply, use something that does.

I don't understand enough about your problem to say more than this.

[Jae Park]

Dear Nick Cox ,
Thank you for your advice. I also believe that making a binary variable out of an averaged variable would risk losing a lot of information. I will consider using my original variables for this command, too.

[Richard Williams]

You are better off upgrading to spost13 which supports factor variables and using commands like mtable and mchange. Do -findit spost13_ado-. For more see

http://www.indiana.edu/~jslsoc/web_s..._faqspost9.htm

For Long & Freese's book, see

http://www.stata.com/bookstore/regre...ent-variables/

[Jae Park]

Richard Williams, Thank you so much! In fact I had installed spost13 but I was using the commands such as -prchange-only (I did not realize spost13 would require different commands). I will try again with spost13 commands again!

[Richard Williams]

Jae, follow Scott Long's advice about what to do if you want both spost9 and spost13 installed. Otherwise you may run into some problems.

http://www.indiana.edu/~jslsoc/web_s..._faqspost9.htm

Long and Freese's book is well worth reading:

http://www.stata.com/bookstore/regre...ent-variables/

[Marcos Almeida]

Under recent versions of Stata, and depending on what you wish, - margins - may probably do the trick.

You may also try - listcoef - under SPost packages. I wonder whether the option "percent" wouldn't present what you wish. Please, see the example below:

. use http://www.stata-press.com/data/r14/lbw.dta
(Hosmer & Lemeshow data)

. logistic low c.age i.smoke i.race

Logistic regression                             Number of obs     =        189
                                                LR chi2(4)        =      15.81
                                                Prob > chi2       =     0.0033
Log likelihood =  -109.4311                     Pseudo R2         =     0.0674

------------------------------------------------------------------------------
         low | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         age |   .9657186   .0322573    -1.04   0.296     .9045206    1.031057
             |
       smoke |
     smoker  |    3.00582   1.118001     2.96   0.003     1.449982    6.231081
             |
        race |
      black  |   2.749483   1.356659     2.05   0.040     1.045318    7.231924
      other  |   2.876948   1.167921     2.60   0.009     1.298314    6.375062
             |
       _cons |    .365111   .3146026    -1.17   0.242     .0674491    1.976395
------------------------------------------------------------------------------

. listcoef, help percent

logit (N=189): Percentage change in odds

  Odds of: 1 vs 0

------------------------------------------------------------------------
             |         b        z    P>|z|         %     %StdX     SDofX
-------------+----------------------------------------------------------
         age |   -0.0349   -1.044    0.296      -3.4     -16.9     5.299
             |
       smoke |
     smoker  |    1.1006    2.959    0.003     200.6      71.4     0.489
             |
        race |
      black  |    1.0114    2.050    0.040     174.9      41.8     0.345
      other  |    1.0567    2.603    0.009     187.7      66.0     0.480
             |
    constant |   -1.0076   -1.169    0.242         .         .         .
------------------------------------------------------------------------
       b = raw coefficient
       z = z-score for test of b=0
   P>|z| = p-value for z-test
       % = percent change in odds for unit increase in X
   %StdX = percent change in odds for SD increase in X
   SDofX = standard deviation of X

Best,

Marcos

[Jae Park]

Thank you so much for your grateful advices. Richard Williams Fortunately I have access to Long and Freese's book which I will start reading right away. Also I will try -margins- and other two options you have suggested, Marcos Almeida. I appreciate your help.

[Richard Williams]

Margins is great. But the spost13 commands are basically shells for margins and make things easier, especially with multiple outcome commands like mlogit and ologit. I also like the mcp command if you want graphs involving continuous variables after single-outcome commands like logit. For examples, see the appendices of

http://www3.nd.edu/~rwilliam/xsoc73994/Ologit01.pdf

For a discussion of mcp, see

http://www3.nd.edu/~rwilliam/xsoc73994/Margins03.pdf