2 min read

A SAS Implementation of Confidence Intervals for Single Proportion: Eleven Methods

The following two papers by Professor Robert Newcombe, in my limit observation, are the most frequently cited papers in the industry for CI calculation:

This post serves as an implementation using SAS accompany with the first paper on confidence intervals for single proportion(method 1-7). Additional 4 methods also provided:

1.  Simple asymptotic, Without CC                       
2.  Simple asymptotic, With CC                          
3.  Score method, Without CC                            
4.  Score method, With CC                               
5.  Binomial-based, ‘Exact’                             
6.  Binomial-based, Mid-p                               
7.  Likelihood-based                                    
8.  Jeffreys                                            
9.  Agresti-Coull,z^22 successes                       
10. Agresti-Coull,2 successes and 2 fail                
11. Logit

The codes are available at

http://jiangtanghu.com/docs/en/CI_Single_Proportion.sas 

It is a purely SAS/Base(data step and SQL) approach. I prefer data steps because it can be seamlessly incorporate into production work which looks like:

/r is the observed number of events/responders in n observations;/

%macro _Wald(r,n,alpha=0.05);

      p=&r/&n;
      z = probit (1-&alpha/2);
      sd=sqrt(&np(1-p)); *Standard Deviation ;
      se=sd/&n;            *standard error;
      me=z*se;             *margin of error;
      p_CI_low = p-me;
      p_CI_up  = p+me;
     p_CI=compress(catx(“”,”[“,put(round(p_CI_low,0.0001),6.4),“,”,put(round(p_CI_up,0.0001),6.4),“]”));
%mend _Wald;

data test;
      input r n;
     %_Wald(r,n);
      keep r n p p_CI;
datalines;
81 263
15 148
0 20
1 29
29 29
;  
run;

proc print data=test;run;

But when simulation required (in methods 6-7), some SQL clauses pop up. Admit that It makes the codes less elegant. In this situation, matrix operation language such as SAS/IML or R will do a better job.

Note that in SAS 9.2, PROC FREQ can produce the following five intervals(method 1, 3, 9 ,8, 5):

Wald (also with method 2)                  
Wilson                
Agresti-Coull         
Jeffreys              
Clopper-Pearson (Exact)

For the five intervals in SAS 9.2, a good reference is Confidence Intervals for the Binomial Proportion with Zero Frequency by Xiaomin He and Shwu-Jen Wu:

www.pharmasug.org/download/papers/SP10.pdf

Also, an equivalent R version of the eleven methods by Abteilung Kriminologie is also available at

http://www2.jura.uni-hamburg.de/instkrim/kriminologie/Mitarbeiter/Enzmann/Software/prop.CI.r

Such R repository is also a good resource for SAS users. I almost “translate” R implementations directly to SAS for this post.