Re: st: "table" showing summary of ~100 ternary variables?

[Phil Schumm <pschumm@uchicago.edu>]

On Oct 30, 2010, at 12:40 PM, Michael Costello wrote:
> I have about 100 ternary variables (0=incorrect, 1=correct, 2=No  
> Response, .=Missing) and I would like to get a table of the  
> responses that looks like this:
>
> -------------------------------------------------------
> |         |     Correct   Incorrect     No Response   .
> ----------+--------------------------------------------
>      Var1 |          2          21             4     21
>      Var2 |          8          19            13     18
>      Var3 |         30          19             4     19
>      Var4 |         18          21            47     22
>      Var5 |         11          27             8     30
>
> Maybe I'd even like to add in a proportion of correct or ratio of  
> correct to incorrect into the table.
>
> Is there a function to do this or something reasonable similar?


I expect that there is probably a user-written program to do this, but  
it's not too difficult to do from first principles.  We'll start by  
generating a dataset similar to what you've described:


    set obs 100
    gen id = _n
    lab def mylab 1 "Correct" 2 "Incorrect" 3 "No response" 4 "Missing"
    set seed 123456789
    forv i = 1/5 {
        gen byte y`i' = cond(runiform()<0.95, ceil(runiform()*3), 4)
        lab val y`i' mylab
    }
    replace y5 = 2 if y5==1


This generates 5 variables y1-y5, each taking values 1, 2, 3 or 4.   
You'll notice I've used 4 for the "missing" values here, only because  
that'll give you more flexibility for where the corresponding column  
appears in the final table (i.e., if we left missing values as ".", we  
wouldn't be able to place a summary column after that one).  As you  
can see, each variable takes values 1-3 with probability 1/3 each, and  
is missing in 5% of cases.  Note that I've also modified y5 so that it  
doesn't contain any correct responses, because you want to make sure  
your code can handle such cases.

Now, the first trick is to reshape your data into long form:


    reshape long y, i(id) j(Var)


Note that we could have used -stack- here instead, and in fact, that  
would have been more convenient if our variables weren't named  
systematically as they are here.  Since we want to add a column for  
the proportion of correct responses, we'll add a corresponding  
observation for each variable whose values we'll fill in later (if you  
wanted to add multiple summary columns, you could add additional  
observations here):


    set obs `=c(N) + 1'
    replace y = 5 if _n == _N
    lab def mylab 5 "Prop. correct", add


Next, we'll generate our cell counts by using -collapse-, but first  
we'll use -fillin- to make sure that all of our cells are represented  
(even if their observed counts are zero):


    fillin Var y
    collapse (count) cnt=id, by(Var y)


Note that we are also using -fillin- here to propagate the observation  
we added above to hold the proportion of correct responses across all  
of the variables.

Now, we'll compute the proportion correct (as a proportion of values  
1-3) for each variable:


    egen correct = max((y==1)*cnt), by(Var)
    egen nonmiss = sum(inlist(y,1,2,3)*cnt), by(Var)
    bys Var (y): replace cnt = correct[1] / nonmiss[1] if y == 5


And finally, we can use -tabdisp- to create our table:


    . tabdisp Var y if !mi(Var), c(cnt) format(%9.2g)

     
--------------------------------------------------------------------
         |                             y
     Var |  Correct    Incorrect   No response    Missing  Prop.  
correct
    ----- 
+--------------------------------------------------------------
       1 |       34           37            25          4            . 
35
       2 |       26           35            33          6            . 
28
       3 |       25           32            40          3            . 
26
       4 |       32           27            35          6            . 
34
       5 |        0           62            34           
4              0
     
--------------------------------------------------------------------


where I've used my text editor to narrow some of the columns so that  
the table doesn't get wrapped by people's mailers.  Of course, there  
are several ways we might embellish this -- this merely illustrates  
one possible strategy for achieving the desired result.


-- Phil

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/