If you have taken an elementary probability course, you have been exposed to the binomial distribution. Remember that the binomial distribution finds the probability of x success occurring in n trials. In other words, if we flip a coin 10 times (n = 10), what is the probability that we would observe 6 heads (x = 6)? My mother gave birth to 7 children, 4 of which were boys. We could apply the binomial distribution to discover the probability that my mom could have 4 boys (x = 4) out of 7 children (n = 7).
ORR and DCR can be related to the binomial distribution because either a patient experiences an ORR (or a DCR) or she doesn't. Success or failure. Heads or tails. Boy or girl. Recall that ORR and DCR are derivations from RECIST standards that show if a patient's tumor is increasing (progressing), decreasing (responding), or stable (no getting better and not getting worse). Here is a general table to help define ORR and DCR:
|
Term
|
Abbreviation
|
Definition
|
|
Complete Response
|
CR
|
A tumor completely disappears
|
|
Partial Response
|
PR
|
A tumor decreases in size by at
least 30% from baseline but does not fully disappear
|
|
Stable Disease
|
SD
|
A tumor does not decrease enough to
be a PR or increase enough to be a PD
|
|
Progressive Disease
|
PD
|
A tumor increases by 20% or more
from baseline
|
|
|
|
|
|
Overall Response Rate
|
ORR
|
Number of patients experiencing a CR
or PR out of all patients of interest
|
|
Disease Control Rate
|
DCR
|
Number of patients experiencing a CR,
PR, or SD out of all patients of interest
|
Now that we have cleared up the background information, let's look at how to get the confidence intervals.
There are many forms of the binomial proportion confidence interval, but the most common used among statisticians is the Clopper-Pearson confidence interval. The formula for the confidence interval is as follows:
Where:
- x is the number of successes
- n is the number of trials
- p is the proportion of successes (x/n)
- α is the level of significance
- F() is the specified percentile of the F distribution
Now that the hard part is out of the way, we get to do the fun part—SAS programming!
The Macro
This macro truly only requires 2 parameters. The other 5 are for more specification and output preferences. The 2 required parameters are any two of x, n, or p (i.e., x and n, x and p, or n and p). If only one of these parameters is specified, an error message is displayed, and the macro stops execution. x, n, and p have the same meanings as above (number of successes, number of trials, and proportion of successes).alpha is defaulted to 0.05 for a 95% confidence interval, but it can be specified to 0.01 or 0.10 (for a 99% or 90% confidence interval, respectively) or any other level of desired significance.
sided defaults to 2, meaning a two-sided confidence interval. Replace this with sided=1 for a one-sided confidence interval.
print defaults to Y (or yes), meaning that the user wants the results to print to the output screen in SAS. If printed output is not needed, this can be changed to N (or no).
Finally, dsout allows the user to specify an output dataset. If the results are not printed, it would be a good idea to specify an output dataset; otherwise, no results will be retained. This is also handy if the results are needed for use in another program.
%macro cp_ci(x=,n=,p=,alpha=0.05,sided=2,print=Y,dsout=);
%if
(%sysfunc(upcase(&print.))=Y
| %sysfunc(upcase(&print.))=YES)
& &dsout.= %then
%do;
%put
WARNING: No print option or output dataset has been specified. No output will be displayed or saved.;
%end;
%let
missing=0;
%if
&x.= %then
%let missing=%eval(&missing.+1);
%if
&n.= %then
%let missing=%eval(&missing.+1);
%if
&p.= %then
%let missing=%eval(&missing.+1);
/*** Print ERROR message and quit execution if
insufficient number of entries ***/
%if
&missing.>1
%then %do;
%put
ERROR: The CP_CI macro requires at least 2 of the 3 parameters - x, n, p;
%if
&x.= %then
%put WARNING: x is empty.;
%if
&n.= %then
%put WARNING: n is empty.;
%if
&p.= %then
%put WARNING: p is empty.;
%return;
%end;
/*** Calculate Confidence Limits ***/
data cp;
/* calculate missing
parameter */
%if
&x.= %then
%do;
n=&n.;
p=&p.;
x=round(n*p);
%end;
%else
%if &n.=
%then %do;
x=&x.;
p=&p.;
n=round(x/p);
%end;
%else
%if &p.=
%then %do;
x=&x.;
n=&n.;
p=x/n;
%end;
%else
%do; *
if all 3 parameters are given, run a check to overwrite p based on x and n;
x=&x.;
n=&n.;
p=x/n;
%end;
/* determine alpha
based on 1-sided or 2-sided test */
%if
&sided.=1
%then %do;
a=1-(&alpha.);
%end;
%else
%if &sided.=2
%then %do;
a=1-(&alpha./2);
%end;
/* calculate the lower
limit */
v11=2*(n-x+1);
v12=2*x;
fscore1=finv(a,v11,v12);
coef1=(n-x+1)/x;
cp_lcl=1/(1+coef1*fscore1);
/* calculate the upper
limit */
v21=2*(x+1);
v22=2*(n-x);
fscore2=finv(a,v21,v22);
coef2=(x+1)/(n-x);
cp_ucl=(coef2*fscore2)/(1+coef2*fscore2);
/* combine lower and
upper limits into a single string and add variable labels */
label x='Successes'
n='Trials'
p='Proportion'
cp_lcl='Clopper-Pearson Lower
Limit'
cp_ucl='Clopper-Pearson Upper
Limit';
keep x n p cp_lcl cp_ucl;
run;
%let
ci=%sysevalf(100*(1-&alpha.)); * for use in the title;
/*** Set up title information ***/
%if
&sided.=1
%then %do;
title "1-sided
&ci.% Clopper-Pearson Confidence Interval";
%end;
%else
%if &sided.=2
%then %do;
title "2-sided
&ci.% Clopper-Pearson Confidence Interval";
%end;
/*** Print to Output screen if requested ***/
%if
%sysfunc(upcase(&print.))=Y
| %sysfunc(upcase(&print.))=YES
%then %do;
proc print data=cp label noobs;
var x n p cp_lcl cp_ucl;
format p cp_lcl cp_ucl 6.4;
run;
%end;
/*** Save to output dataset if requested ***/
%if
&dsout.^= %then
%do;
data &dsout.;
set cp;
run;
%end;
/*** Clean up
***/
%if
&dsout.= | %sysfunc(upcase(&dsout.))^=CP
%then %do;
proc datasets lib=work;
delete cp;
run;
quit;
%end;
title;
%mend
cp_ci;
An Example
An investigational drug is given to 45 patients. The following RECIST results are collected:|
|
% (n)
|
|
CR
|
0% (0)
|
|
PR
|
4% (2)
|
|
SD
|
36% (16)
|
|
PD
|
60% (27)
|
|
ORR (CR+PR)
|
4% (2)
|
|
DCR (CR+PR+SD)
|
40% (18)
|
You now know the ORR and DCR (4% and 40%, respectively), but your manager is interested in the 95% confidence intervals for each result. You should be able to see this as a binomial problem (a success/failure situation). A patient is either counted in the ORR or not. A patient is either counted in the DCR or not.
Simply running your new macro, you can now code:
%cp_ci(x=2,n=45);
%cp_ci(x=18,n=45);
and the following results will magically appear in your SAS output:
%cp_ci(x=2,n=45,dsout=orr_cp);
%cp_ci(x=18,n=45,dsout=dcr_cp);
where orr_cp is a SAS dataset containing the Clopper-Pearson confidence limits for the ORR and dcr_cp is a SAS dataset containing the Clopper-Pearson confidence limits for the DCR, both datasets saved in the WORK library.
Pretty simple, huh?