Confidence limits in tabulate

[RichardDeVen]

I was asked to report the % of some event occurring and the 95% confidence interval about that %.

In this sample, a flag variable 'replied' is 0 or 1, and represents wether or not some survey question was answered.

'replied' is used categorically as a CLASS variable in table so that an across percentage can be specified in the table statement.

'replied' can not be reused analytically as a VAR variable, so replied_CL is created to have a something to work with in tabulate.

My question is this... is there any statistical problems with computing a LCLM UCLM from a two-valued variable ?

Thanks for listening.

Richard

data have;

do region = 'A', 'B';

  do year = 2005 to 2015;

    do _i = 1 to 100 + 50 *ranuni(123);

      id + 1;

      replied = ranuni(123) < 0.15;

      replied_CL = 100 * replied;

      output;

    end;

  end;

end;

drop _:;

run;

proc format;

  value replied 0='No Reply' 1='Reply';

options nocenter;

proc tabulate data=have;

  class region year replied;

  var replied_CL;

  table

    year

  , region

    * (replied='' * (N pctn<replied>=' % of region')

       replied_CL = '% of region Replied(*ESC*){newline}confidence interval' * ( LCLM='95% CI LB' UCLM='95% CI UB' )

      )

  ;

  format replied replied.;;

run;

* i guess this is a more 'canonical' way to get CI;

proc freq;

  by region year;

  table replied / binomial;

  table replied_CL / binomial;

run;

[Ksharp]

I was surprised by you can use LCLM ,UCLM in proc tabulate , so I quickly check the documentation . Here is :

"Use both LCLM and UCLM to compute a two-sided confidence limit for the mean."

According to documentation, LCLM ,UCLM is for T statistical estimator, in other words, it test the H0: mu=0 .

It is not the confidence limit you are talking about (univariate's binormial distribution ) . I think you should merge it back to your original dataset.

[ballardw]

Other concerns with any tabulate output involving confidence limits and a survey relates to weights and which divisor to use in calculations of variance / standard deviation. The VARDEF option on the Proc Tabulate statement indicates  which divisor to use.

Second is if your survey design involved any form of sampling other than a simple random sample then the weights aren't quite applied correctly and should use one of the Survey procedures to generate confidence limits.

And since you didn't include a weight statement at all your results cannot be applied to the population only the respondent pool unless your data is an actual census of the population of interest.