The BCHOICE Procedure

(View the complete code for this example.)

Consider an example of allocation choice data in the MAMR study ^[25]. In this study, researchers aimed to learn about oncologists’ preferences on different scenarios of pharmaceutical therapies and benefits of these therapies in relation to the survival of patients with metastatic breast cancer.

data MAMR;
   input ID Set None TrtScheme RespRate Time Status TrtLine Risk Cost Alloc @@;
   datalines;
1     1     1     3     3     1     2     1     6     1     0.2
1     1     1     5     1     1     1     3     4     2     0
1     1     1     7     2     2     2     3     6     2     0
1     1     0     0     0     0     0     0     0     0     0.8
1     2     1     1     1     2     2     1     1     1     0.12
1     2     1     3     3     1     1     2     2     1     0.4
1     2     1     5     1     2     1     3     4     3     0
1     2     0     0     0     0     0     0     0     0     0.48
1     3     1     4     3     1     2     1     4     4     0.08

   ... more lines ...   

146   10    0     0     0     0     0     0     0     0     0
146   11    1     1     2     1     1     3     6     1     0.34
146   11    1     2     2     2     1     3     3     1     0.28
146   11    1     1     2     2     2     1     1     1     0.38
146   11    0     0     0     0     0     0     0     0     0
;

proc print data=MAMR(obs=20);
run;

The data for the first several choice sets are shown in Output 27.9.1.

The variable ID represents the IDs for the 133 oncologists in the study. Each oncologist was presented with 11 choice tasks, each consisting of four scenarios of therapies (including the NONE option), and was asked to state the number of times he or she would pick each alternative. These numbers of times were then transformed into percentages, which sum to 100% within each choice task. The variable Alloc is the response variable that gives a percentage of preference for each alternative, and the sum of percentages in each choice set is 100%. Table 27.13 lists all the attributes and their levels.

The following statements fit a logit model with random effects for these allocation choice data by specifying CHOICETYPE=ALLOCATION:

proc bchoice data=MAMR seed=124 nthreads=8 nmc=20000 nthin=2 plots=none;
   class  TrtScheme(ref=first) RespRate Time Status TrtLine Risk Cost ID Set
          / ref=last;
   model  Alloc = / choiceset=(ID Set) choicetype=allocation;
   random TrtScheme RespRate Time Status TrtLine Risk Cost
          / subject=ID remean;
run;

The NMC= option requests to run the MCMC chain for 20,000 iterations, and the THIN=2 option keeps one of every two samples for analysis. PLOTS=NONE is specified, so the plots will not be produced to save time. The REMEAN option in the RANDOM statement requests estimation of the mean of the random effects.

When CHOICETYPE=ALLOCATION, you can specify a positive integer in the WEIGHT= suboption to indicate how many choice sets the allocated percentages should apply to. For example, if the respondent were asked to imagine the next 20 occasions and to allocate a percentage of those 20 occasions for each alternative, you would specify WEIGHT=20. By default, WEIGHT=10. In this example, the default (WEIGHT=10) is close to the nature of the original data collection.

The "Model Information" table, as shown in Output 27.9.2, shows that the type of choice response is allocation.

As shown in the "Choice Sets Summary" table in Output 27.9.3, there are a total of 1,463 choice sets, all of which display the same pattern: they have a total of four alternatives, and the sum of percentages is 100%. If the sum of percentages within one choice set is not equal to 100%, the choice set is invalid. PROC BCHOICE deletes all invalid choice sets and produces a warning message.

As you see in Output 27.9.4, TrtScheme, Status, TrtLine, and Risk have impact on the choices. It seems that treatments (TrtScheme) with shorter periods of SG and SLP follow-up are more preferred than those with longer follow-up times. The oncologists in the study tend to choose the therapies that would have the best performance status (Status); the part-worth for level 1 (better) of Status is positive compared to the referent level 2 of Status, which is set at 0. More important, risk factors (Risk) have a large influence on the choice of various treatments: the estimates of the part-worths for all risk factors are all negative and have relatively large values, reflecting that doctors try to avoid them as much as possible, especially treatments with serious side effects. One surprising result is that therapy cost does not play a role in the decision-making process. But when you realize that the respondents are doctors, not their patients, and that the study deals with metastatic cancer, it might not be hard to figure out why.