stan_surv — Rearranging the data produces different results

Summary:

stan_surv produces different results after rearranging (sorting) the observations despite using the same seed.

Description:

As per the summary. With the example data below, this issue happens with basehaz = "weibull" or basehaz = "bs", but not with basehaz = "exp" or basehaz = "weibull-aft".

Reproducible example:

library(survival)
library(rstanarm)

data(mgus)

surv <- stan_surv(Surv(futime, death) ~ mspike, 
                  data = mgus, 
                  basehaz = "weibull",
                  chains = 1, 
                  seed = 1234)


mgus_order <- mgus[order(mgus$mspike), ] 
surv_order <- stan_surv(Surv(futime, death) ~ mspike, 
                        data = mgus_order, 
                        basehaz = "weibull",
                        chains = 1, 
                        seed = 1234)

all.equal(as.matrix(surv), as.matrix(surv_order))

summary(surv, digits = 5)
summary(surv_order, digits = 5)

Relevant output from the last two summary functions

Estimates:
                mean      sd        10%       50%       90%    
(Intercept)   -10.16572   0.64438 -10.98888 -10.15258  -9.35997
mspike         -0.11383   0.16630  -0.33068  -0.12206   0.09692
weibull-shape   1.19092   0.06669   1.10818   1.18821   1.27728

Estimates:
                mean      sd        10%       50%       90%    
(Intercept)   -10.16413   0.69352 -11.03953 -10.13223  -9.30872
mspike         -0.11555   0.16577  -0.32517  -0.12085   0.11076
weibull-shape   1.19101   0.07182   1.09859   1.19006   1.28189

RStanARM Version:

2.21.2

R Version:

3.6.3

Operating System:

macOS 10.14.6

Start at the stan_surv entry point and reproduce the example using the mgus data, comparing the original and sorted observations with basehaz="weibull" and "bs". Check that identical seeds produce equivalent results after sorting, while preserving the behavior reported for "exp" and "weibull-aft".