longitudinal measurement invariance doesnt work with target="stan"

[maugavilla]

When testing longitudinal measurement invariance, target="stan" doesnt work, while "stanclassic" does work. Does work when I add the factor loadings or intercept equality labels across time

The error is gives is
SAMPLING FOR MODEL 'stanmarg' NOW (CHAIN 1).
Chain 1: Unrecoverable error evaluating the log probability at the initial value.
Chain 1: Exception: normal_lpdf: Random variable has dimension = 3, expecting dimension = 4; a function was called with arguments of different scalar, array, vector, or matrix types, and they were not consistently sized; all arguments must be scalars or multidimensional values of the same shape. (in 'model_stanmarg' at line 689)

[1] "Error in sampler$call_sampler(args_list[[i]]) : "
[2] " Exception: normal_lpdf: Random variable has dimension = 3, expecting dimension = 4; a function was called with arguments of different scalar, array, vector, or matrix types, and they were not consistently sized; all arguments must be scalars or multidimensional values of the same shape. (in 'model_stanmarg' at line 689)"
error occurred during calling the sampler; sampling not done
Stan model 'stanmarg' does not contain samples.
Error in rowvec[tmpw[, 1] == 0] <- grep(stanvec[j], names(b.est)) :
replacement has length zero

Here is an example that reproduce the error with simulated data

library(blavaan)


mod_sim <- '
f1 =~ 0.7?y1_t1 + 0.7?y2_t1 + 0.7?y3_t1
f2 =~ 0.7?y1_t2 + 0.7?y2_t2 + 0.7?y3_t2

f1 ~~ 1?f1
f2 ~~ 1.5?f2

f1 ~0?1
f2 ~0.7?1

f1 ~~ 0.4?f2

y1_t1 ~~ 0.3?y1_t2
y2_t1 ~~ 0.3?y2_t2
y3_t1 ~~ 0.3?y3_t2
'


dat_sim <- simulateData(mod_sim, sample.nobs = 300)
head(dat_sim)

mod_conf <- '
f1 =~ y1_t1 + y2_t1 + y3_t1
f2 =~ y1_t2 + y2_t2 + y3_t2

y1_t1 ~~ y1_t2
y2_t1 ~~ y2_t2
y3_t1 ~~ y3_t2
'


fit_conf <- bcfa(mod_conf, data=dat_sim, std.lv=T)
summary(fit_conf)


mod_weak <- '
f1 =~ l1*y1_t1 + l2*y2_t1 + l3*y3_t1
f2 =~ l1*y1_t2 + l2*y2_t2 + l3*y3_t2

f1 ~~ 1*f1
f2 ~~ NA*f2

y1_t1 ~~ y1_t2
y2_t1 ~~ y2_t2
y3_t1 ~~ y3_t2
'

fit_weak <- bcfa(mod_weak, data=dat_sim, std.lv=T)
summary(fit_weak)


mod_strong <- '
f1 =~ l1*y1_t1 + l2*y2_t1 + l3*y3_t1
f2 =~ l1*y1_t2 + l2*y2_t2 + l3*y3_t2

f1 ~~ 1*f1
f2 ~~ NA*f2

f1 ~0*1
f2 ~NA*1

y1_t1 ~~ y1_t2
y2_t1 ~~ y2_t2
y3_t1 ~~ y3_t2

y1_t1 ~i1*1
y2_t1 ~i2*1
y3_t1 ~i3*1
y1_t2 ~i1*1
y2_t2 ~i2*1
y3_t2 ~i3*1
'

fit_strong <- bcfa(mod_strong, data=dat_sim, std.lv=T)
summary(fit_strong)

[ecmerkle]

Thanks for the report. This is a conflict between std.lv=TRUE and freely estimating an lv variance. I will have to dig to see how to fix it. In the meantime, it is possible to estimate the model by using blavaan() and avoiding std.lv=TRUE:

fit_weak <- blavaan(mod_weak, data=dat_sim, auto.fix.first=F, auto.var=T,
                    int.ov.free=T, auto.cov.lv.x=T)

The model also appears to work with the other targets (jags, stanclassic, and there is a new "stancond" on github that is not fully finished).

[maugavilla]

Good to see there is a way around it with blavaan(). I did tested with the other jags and stanclassic targets and worked fine.
Did see the stancond in the github changes. What is this target for?

Thanks

[ecmerkle]

It treats latent variables as parameters, like stanclassic, but the code is simpler and it is often more efficient sampling.

[ecmerkle]

I think it is now fixed, but let me know if not.

[maugavilla]

yes, I just run the test, it works. Thanks