Macro for Co - PDF document

1 1 Macro for Co mputing Complier - Aver
1 Macro for Co mputing Complier - Average Causal Effect of Treatment (CACE) U sing Instrumental Variable Method Patricia R Houck, MSH, Sati Mazumdar, PhD, Ben oit H Mulsant, MD University of Pittsburgh ABSTRACT It is common in clinical research to experience mis sing observations in a longitudinal trial due to dropout and nonadherence . The treatment effect from these trials may be biased. The complier - average causal effect of treatment (CACE) is considered to be a measure that corrects for this bias. Instrument al variables from a latent class model can be used to improve CACE ( Dunn, Maracy, Tomenson, 2005 ). A SAS® macro is developed and applied to a randomized clinical trial for the treatment of depression in the elderly ( Mulsant et al, 2001 ). INTRODUCTION No ncompliance is a common problem in randomized clinical trials (RCT) the subject receives which has an effect on the outcome. However, there are confounding variables that influence both the treatment and the out c ome. We examine here the confound of noncomplia nce defined as missing at least o ne dose of medication in a randomized clinical trial in the treatment of depr ession in an elderly sample and compute CACE. METHODOLOGY A RCT is comprised of 4 latent classes. Compliers (a) are subjects who always take the treatment to which they are rand omized. Defiers (d) are subjects who always take the opposite treatment from the one to which they are randomized. Then there are t hose who always take a particular treatment (tx=1) regardless of assignment (b) , and finally those who Randomized n11 =n1a+n1b n10 =n1c N1 Completed m11 =m1a+m1b m10 =m1c M1 Responded s11 =s1a+s1b s10 = s1c S1 2 s1b/m1b=s0b/m0b and s1c/m1c=s0c/m0c 4) Given treatment and compliance status, outcome is independent of wh ether it is actually observed. Computation of CACE=s1a/m1a - s0a/m0a w here : s1a=S1 - (N1/N0 )*s0 0 - s10 m1a=M10 - (N1/N0)*m0 0 - m10 s 0a=S0 - (N0/N1)*s10 - s0 0 m0a=M10 - (N0/N1)*m10 - m0 0 MOTIVATING EXAMPLE The latent class method in computing CACE was applied to a 12 - week clinical trial on the outcome of elderly depressed patients randomized to treatme nt with nortriptyline or paroxetine (Mulsant, 2001). The trial failed to find a clinical or statistical significant difference in the r esponse rate (no lon

2 ger having depression) between the t wo
ger having depression) between the t wo treatments . H owev er patients were twice as likely to disconti nue nortriptyline because of side effects. The use of IV in a latent class model is able to incorporate the noncompliance and dropout information to provide a better estimate of the true effect. Noncompliance information was available on 99 of the 116 su bjects of which 47 were randomized to nortriptyline (tx=1) and 52 randomized to paroxetine (tx=0). Of the 47 nortriptyline randomized subjects, 24 completed the study and 18 responded. In the paroxetine randomized arm, 32 completed the study and 27 respo nded. The macro contains the computations for calculating CACE in a two arm randomized clinical trial. PROC MEANS is a powerful procedure that makes the computations easy. The macro allows one to readily calculate CACE in a RCT . SAS® MACRO %macro cace (data=_last_, tx=tx, complete=complete, compliant=compliant, resp=resp); proc sort data= &data ; *cell information; by & tx & compliant; proc means data= &data noprint; by & tx & compliant; var & complete & resp; output out=cell sum=cell_complete cell_re sp; proc print; run; proc means data= &data noprint; *margin information; by & tx; var & complete & resp; output out=margin sum=margin_complete margin_resp; proc print; run; proc means data= &data noprint; *total information; by & tx; var & tx; output out=total n=num; data total2; set total; if tx=0 then N 0=num; if tx=1 then N 1=num; proc means noprint data=total2; var N0 N 1; output out=total3 max=; data margin2; set margin; if & tx=0 then do; M0=margin_complete; S 0=margin_resp; end ; if & tx=1 then do; M1=margin_complete; S 1=margin_resp; end; proc means noprint data=margin2; var M0 S0 M1 S 1; output out=margin3 max=; data cell2; set cell; i f & tx=0 then do; if & compliant=0 then do; m00=cell_complete; s00=cell_resp; end; 3 if & compliant=1 then do; m01=cell_complete; s01=cell_resp; end; end; if & tx=1 then do; if & compliant=0 then do; m10=cell_complete; s10=cell_resp; end; if & compliant=1 then do; m11=cell_complete; s11=cell_resp; end; end; proc means data =cell2 noprint; var m00 m01 m10 m11 s00 s01 s10 s1

3 1; output out=cell3 max=; data fin
1; output out=cell3 max=; data final; *merge files and compute CACE; merge total3 margin3 cell3; s1a= S1 - (N1/N 0)*s0 0 - s10; m1a=M1 - (N1/N 0)*m 0 0 - m 1 0; s0a=S0 - (N0/N 1)*s10 - s0 0 ; m0a=M0 - (N0/N 1)*m10 - m0 0 ; C ACE =s1a/m1a - s0a/m0a; run; %mend; %cace(data=_last_, tx=tx, complete=complete, compliant=compliant, resp=resp); RESULT The CACE method which included information on noncompliance and dropout provided a larger effect of the randomized trial compared to th e completer result (CR) . The effect of 0.26 is the difference between compliers to nortriptyline compared to compliers to paroxetine. The completer result (CR) calculated at only - 0.09 , does not include the noncompliance information from the trial and doe s not incorporate the original randomized sample size . C ompleter R esult (CR) =18/24 ± 27/32=0.75 - 0.84= - 0.09 Output: CACE=0.26 n0 n1 m0 s0 m1 s1 m00 m01 m10 m11 s00 s01 s10 s11 52 47 32 27 24 18 2 1 11 18 6 17 10 14 4 s1a m1a s0a m0a cace - 11.3654 - 12.9808 - 5.48936 - 8.91489 0.25980 COMMENT Both noncompliance and early termination are common deviat ions in a randomized clinica l trial. The calculation of CACE using latent class is one approach that incorporates both noncompliance and early termination information into estimating the effect of the trial. REFERENCES Dunn G, Maracy M, Tomenson B. (2005) Estimating treatment eff ects from randomized clinical trials with noncompliance and loss to follow - up: the role of instrumental variable methods. Statistical Methods in Medical Research , 14:369 - 395. Mulsant BH, Pollock BG, Nebes R, Miller M, Sweet R, Stack J, Houck PR, Bensasi S, Mazumdar S, and Reynolds CF. (2001) A twelve - week double - blind randomized comparison on nortriptyline and paroxetine in older depressed inpatients and outpatients. American Journal of Geriatric Psychiatry . 9(4):406 - 414 . ACKNOWLEDGEMENT Supported in part by MH069430 CONTACT INFORMATION Patricia R. Houck, MSH University of Pittsburgh houckpr@upmc.edu SAS® SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc. in the USA and other count ries. ® indicates USA registration