# Prior predictive check multivariate model
**URL:** https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220
**Category:** brms
**Created:** [August 4, 2020, 2:04pm UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220 "2020-08-04T14:04:34Z")
**Posts on this page:** 8
**Page:** 1
### Author:  [@jroon](https://discourse.mc-stan.org/u/jroon)
#### Post date: [August 4, 2020, 2:04pm UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220/1 "2020-08-04T14:04:34Z")
I’m fitting a multivariate model in brms and I wanted to do a prior predictive check. But, I’m running into issues with setting priors.
Here is a toy example:
```
library(brms)
data("BTdata", package = "MCMCglmm")
fit0 <- brm(
mvbind(tarsus, back) ~ sex + hatchdate + (1|p|fosternest) + (1|q|dam),
data = BTdata, chains = 2, cores = 2, sample_prior = "only"
)
Setting 'rescor' to TRUE by default for this model
Error: Sampling from priors is not possible as some parameters have no proper priors. Error occured for parameter 'b_tarsus'
```
Ok so I put sample\_prior to “no” and fit a model and then do `prior_summary(fit0)` to get:
```
> prior_summary(fit0)
prior class coef group resp dpar nlpar bound
1 b back
2 b hatchdate back
3 b sexMale back
4 b sexUNK back
5 b tarsus
6 b hatchdate tarsus
7 b sexMale tarsus
8 b sexUNK tarsus
9 student_t(3, 0, 10) Intercept back
10 student_t(3, 0, 10) Intercept tarsus
11 lkj_corr_cholesky(1) L
12 L dam
13 L fosternest
14 lkj_corr_cholesky(1) Lrescor
15 student_t(3, 0, 10) sd back
16 student_t(3, 0, 10) sd tarsus
17 sd dam back
18 sd Intercept dam back
19 sd dam tarsus
20 sd Intercept dam tarsus
21 sd fosternest back
22 sd Intercept fosternest back
23 sd fosternest tarsus
24 sd Intercept fosternest tarsus
25 student_t(3, 0, 10) sigma back
26 student_t(3, 0, 10) sigma tarsus
```
Ok so the `b` terms don’t have priors so I attempt to set them:
```
fit1 <- brm(
mvbind(tarsus, back) ~ sex + hatchdate + (1|p|fosternest) + (1|q|dam),
data = BTdata, chains = 2, cores = 2, ,
prior = set_prior("normal(0, 5)", class = "b", coef = "",
resp = c("tarsus", "back")), sample_prior = "no")
prior_summary(fit1)
```
However now I’m still getting the error and when I check priors this time:
```
> prior_summary(fit1)
prior class coef group resp dpar nlpar bound
1 normal(0, 5) b back
2 b hatchdate back
3 b sexMale back
4 b sexUNK back
5 normal(0, 5) b tarsus
6 b hatchdate tarsus
7 b sexMale tarsus
8 b sexUNK tarsus
9 Intercept back
10 Intercept tarsus
11 lkj_corr_cholesky(1) L
12 L dam
13 L fosternest
14 lkj_corr_cholesky(1) Lrescor
15 student_t(3, 0, 10) sd back
16 student_t(3, 0, 10) sd tarsus
17 sd dam back
18 sd Intercept dam back
19 sd dam tarsus
20 sd Intercept dam tarsus
21 sd fosternest back
22 sd Intercept fosternest back
23 sd fosternest tarsus
24 sd Intercept fosternest tarsus
25 student_t(3, 0, 10) sigma back
26 student_t(3, 0, 10) sigma tarsus
```
I see it has not set priors for all the coefficients for the outcomes. I tried to do this but can’t figure it out. Also, now for some reason the priors for the Intercepts have dissappeared.
Where am I going wrong here?
---
### Author:  [@jroon](https://discourse.mc-stan.org/u/jroon)
#### Post date: [August 4, 2020, 10:39pm UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220/2 "2020-08-04T22:39:33Z")
So I eventually got this to work by specifying many of the priors manually as follows:
```
fit1 <- brm(
mvbind(tarsus, back) ~ sex + hatchdate + (1|p|fosternest) + (1|q|dam),
data = BTdata, chains = 2, cores = 2, ,
prior = c(
set_prior("normal(0, 5)", class = "b", resp = c("tarsus", "back"),
coef = "sexMale"),
set_prior("normal(0, 5)", class = "b", resp = c("tarsus", "back"),
coef = "sexUNK"),
set_prior("normal(0, 5)", class = "b", resp = c("tarsus", "back"),
coef = "hatchdate"),
set_prior("normal(0, 5)", class = "Intercept", resp = c("tarsus", "back"),
coef = "")),
sample_prior = "only")
pp_check(fit1, resp = "tarsus")
```
In my real data I have more outcomes and so the priors to set will grow - it gets clunky. @paul.buerkner - am I missing a neater way to set the priors for multiple coefficients in multivariate models ?
---
### Author:  [@paul.buerkner](https://discourse.mc-stan.org/u/paul.buerkner)
#### Post date: [August 5, 2020, 5:38am UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220/3 "2020-08-05T05:38:12Z")
Drop the coef argument and only specify the prior once.
---
### Author:  [@jroon](https://discourse.mc-stan.org/u/jroon)
#### Post date: [August 5, 2020, 8:28am UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220/4 "2020-08-05T08:28:38Z")
Ok thanks. I think I am confused by the `prior_summary()` output.
i.e. when it looked like this:
… I thought there were no priors on the hatchdate, sexMale and sexUNK covars. I take it thats wrong? Does this instead mean the prior applies to all the covars ?
---
### Author:  [@paul.buerkner](https://discourse.mc-stan.org/u/paul.buerkner)
#### Post date: [August 5, 2020, 8:35am UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220/5 "2020-08-05T08:35:51Z")
Yes. See ?set\_prior
---
### Author:  [@jroon](https://discourse.mc-stan.org/u/jroon)
#### Post date: [August 5, 2020, 10:33am UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220/6 "2020-08-05T10:33:14Z")
@paul.buerkner I have one more follow on question here. I’m trying to make my priors more informative - however changing that priors for `b` or `intercept` doens’t really make much difference.
My real data model is:
```
fit1 <- brm(
mvbind(out_fvc, out_svc, out_snip, out_peak) ~ cohort + days_from_baseline +
(days_from_baseline | uin),
data = df2, chains = 4, cores = 4,
prior = c(set_prior("normal(0, 1)", class = "b", resp = c("outfvc", "outsvc", "outsnip", "outpeak")),
set_prior("normal(0, 1)", class = "Intercept",
resp = c("outfvc", "outsvc", "outsnip", "outpeak") )),
sample_prior = "yes" )
```
My prior predictive check looks like this:
Current priors are:
```
> prior_summary(fit1)
prior class coef group resp dpar nlpar bound
1 normal(0, 1) b outfvc
2 b cohortLON outfvc
3 b cohortLVN outfvc
4 b cohortSHD outfvc
5 b cohortTUR outfvc
6 b cohortUTR outfvc
7 b days_from_baseline outfvc
8 normal(0, 1) b outpeak
9 b cohortLON outpeak
10 b cohortLVN outpeak
11 b cohortSHD outpeak
12 b cohortTUR outpeak
13 b cohortUTR outpeak
14 b days_from_baseline outpeak
15 normal(0, 1) b outsnip
16 b cohortLON outsnip
17 b cohortLVN outsnip
18 b cohortSHD outsnip
19 b cohortTUR outsnip
20 b cohortUTR outsnip
21 b days_from_baseline outsnip
22 normal(0, 1) b outsvc
23 b cohortLON outsvc
24 b cohortLVN outsvc
25 b cohortSHD outsvc
26 b cohortTUR outsvc
27 b cohortUTR outsvc
28 b days_from_baseline outsvc
29 normal(0, 1) Intercept outfvc
30 normal(0, 1) Intercept outpeak
31 normal(0, 1) Intercept outsnip
32 normal(0, 1) Intercept outsvc
33 lkj_corr_cholesky(1) L
34 L uin
35 lkj_corr_cholesky(1) Lrescor
36 student_t(3, 0, 2.5) sd outfvc
37 student_t(3, 0, 148.3) sd outpeak
38 student_t(3, 0, 31.1) sd outsnip
39 student_t(3, 0, 2.5) sd outsvc
40 sd uin outfvc
41 sd days_from_baseline uin outfvc
42 sd Intercept uin outfvc
43 sd uin outpeak
44 sd days_from_baseline uin outpeak
45 sd Intercept uin outpeak
46 sd uin outsnip
47 sd days_from_baseline uin outsnip
48 sd Intercept uin outsnip
49 sd uin outsvc
50 sd days_from_baseline uin outsvc
51 sd Intercept uin outsvc
52 student_t(3, 0, 2.5) sigma outfvc
53 student_t(3, 0, 148.3) sigma outpeak
54 student_t(3, 0, 31.1) sigma outsnip
55 student_t(3, 0, 2.5) sigma outsvc
```
Now I know that Y can never be negative and will never be higher than about 400 in value, so I’ like to constrain y to about the 0 to 2000 range or so (or I guess -1000 to 1000 range). Do I need to change the sd and sigma priors for each outcome and if so - both sd and sigma or just one?
---
### Author:  [@jroon](https://discourse.mc-stan.org/u/jroon)
#### Post date: [August 5, 2020, 12:13pm UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220/7 "2020-08-05T12:13:42Z")
Zooming in on the pp\_checks above I see this:
Is there anyway to set a different prior on the intercept for each outcome? I tried this but it gives an error:
```
> fit1 <- brm(
+ mvbind(out_fvc, out_svc, out_snip, out_peak) ~ cohort + days_from_baseline +
+ (days_from_baseline | uin),
+ data = df2, chains = 4, cores = 4,
+ prior = c(set_prior("normal(0, 1)", class = "b", resp = c("outfvc", "outsvc", "outsnip", "outpeak")),
+ set_prior("normal(0, 1)", class = "Intercept", resp = c("outfvc"),
+ "normal(0, 1)", class = "Intercept", resp = c("outsvc"),
+ "normal(50, 10)", class = "Intercept", resp = c("outsnip"),
+ "normal(250, 50)", class = "Intercept", resp = c("outpeak") )
+ ),
+ sample_prior = "only" )
Setting 'rescor' to TRUE by default for this model
In the future, 'rescor' will be set to FALSE by default for all models. It is thus recommended to explicitely set 'rescor' via 'set_recor' instead of using the default.Error in set_prior("normal(0, 1)", class = "Intercept", resp = c("outfvc"), :
formal argument "class" matched by multiple actual arguments
```
---
### Author:  [@jroon](https://discourse.mc-stan.org/u/jroon)
#### Post date: [August 5, 2020, 4:09pm UTC](https://discourse.mc-stan.org/t/prior-predictive-check-multivariate-model/17220/8 "2020-08-05T16:09:51Z")
OK that error seems to have been caused by c( ) around the resp argument in set\_prior. Without it it works perfectly:
```
fit1 <- brm(
mvbind(out_fvc, out_svc, out_snip, out_peak) ~ cohort + days_from_baseline +
(days_from_baseline | uin),
data = df2, chains = 4, cores = 4,
prior = c(set_prior("normal(0, 1)", class = "b", resp = c("outfvc", "outsvc", "outsnip", "outpeak")),
set_prior("normal(2.5, 1)", class = "Intercept", resp = "outfvc" ),
set_prior("normal(2.5, 1)", class = "Intercept", resp = "outsvc" ),
set_prior("normal(50, 10)", class = "Intercept", resp = "outsnip" ),
set_prior("normal(250, 50)", class = "Intercept", resp = "outpeak" )
),
sample_prior = "only" )
```
