likelihood

67 views
Skip to first unread message

Xinya Liang

unread,
Oct 3, 2018, 2:05:15 PM10/3/18
to blavaan
Hi Ed,

Are those information criteria computed by blavaan based on conditional or marginal likelihood? Thanks for the clarification!

Best,
Xinya

Ed Merkle

unread,
Oct 3, 2018, 5:05:03 PM10/3/18
to Xinya Liang, blavaan
Xinya,

The information criteria are generally marginal, but you can get conditional by adding the argument "save.lvs=TRUE" to the model estimation command. Then some new metrics will appear in fitMeasures(), with "_cond" in their name. (And note that marginal are generally recommended.)

Ed



--
You received this message because you are subscribed to the Google Groups "blavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blavaan+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/blavaan/7e7d39a5-980e-4474-846f-7f5808adbea5%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Xinya Liang

unread,
Oct 3, 2018, 5:20:45 PM10/3/18
to Ed Merkle, blavaan
Thanks Ed,

If I understood correctly, there should be prior(s) used to marginalize the latent variables. What are the default prior distributions for the latent variables in the marginal likelihood in blavaan?

Best,
Xinya


Ed Merkle

unread,
Oct 4, 2018, 12:53:06 PM10/4/18
to Xinya Liang, blavaan
Xinya,

The latent variables are assumed to come from a normal distribution, which is the same as non-Bayesian SEM. We might call the latent variable distribution a "hyperdistribution". Then, in a Bayesian model, the free parameters of the hyperdistribution (mean or variance of the normal distribution) would receive prior distributions. Those prior distributions can be found in dpriors().

Ed

Xinya Liang

unread,
Mar 24, 2019, 1:11:15 AM3/24/19
to Ed Merkle, blavaan
Hi Ed,

I have a few follow-up questions about the likelihoods printed from blavaan. For a sample output below:

(n=150)
npar1  logl    ppp     bic     dic      p_dic    waic     p_waic   loo       p_loo    mlogl 
72 -2267.698 0.49825 4895.195 4676.975 70.78994 4681.485 71.05709 4682.442 71.53554 -2576.311 

(1) Is the logl the same as the standard maximum likelihood when default priors are used ?
(2) Is the BIC reported equivalent to the BIC from ML estimation when default priors are used?
(3) What are the difference between logl and mlogl? BIC seems to be computed from the marginal likelihood logl as: BIC = -2*(-2267.698) + 72*ln(150) = 4896.162. Then how should I interpret mlog1?

Thank you so much!
Xinya




Xinya Liang

Ed Merkle

unread,
Mar 24, 2019, 3:58:35 PM3/24/19
to Xinya Liang, blavaan
Xinya,

The logl is the model log-likelihood calculated at posterior means.
The posterior means sometimes are close to the ML estimates, so these
log-likelihoods (and BICs) are sometimes similar. But there is no
guarantee here.

mlogl is the log-likelihood, marginalized over all parameters (via
Laplace approximation). This is the quantity involved in Bayes
factors.

Ed

On 3/24/19, Xinya Liang <xin...@gmail.com> wrote:
> Hi Ed,
>
> I have a few follow-up questions about the likelihoods printed from
> blavaan. For a sample output below:
>
> (n=150)
>
> npar1 *logl *ppp bic dic p_dic waic p_waic
> loo p_loo *mlogl *
>>>>> <https://groups.google.com/d/msgid/blavaan/7e7d39a5-980e-4474-846f-7f5808adbea5%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>>> .
>>>>> For more options, visit https://groups.google.com/d/optout.
>>>>>
>>>> --
>>>> You received this message because you are subscribed to the Google
>>>> Groups "blavaan" group.
>>>> To unsubscribe from this group and stop receiving emails from it, send
>>>> an email to blavaan+u...@googlegroups.com.
>>>> To view this discussion on the web visit
>>>> https://groups.google.com/d/msgid/blavaan/CAJZU7FKqNigudYSi4DWGVjL2xVpZj79wGLj-L8DZJ6ZbMadnFg%40mail.gmail.com
>>>> <https://groups.google.com/d/msgid/blavaan/CAJZU7FKqNigudYSi4DWGVjL2xVpZj79wGLj-L8DZJ6ZbMadnFg%40mail.gmail.com?utm_medium=email&utm_source=footer>
>>>> .
>>>> For more options, visit https://groups.google.com/d/optout.
>>>>
>>>
>
> --
> You received this message because you are subscribed to the Google Groups
> "blavaan" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to blavaan+u...@googlegroups.com.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/blavaan/CADLks4%3D2Lnte_guVtbn6QtxVOxoaHyx57ogeq1hh4PMwzOhoKw%40mail.gmail.com.
Reply all
Reply to author
Forward
0 new messages