Message/Author |
|
Paul Silvia posted on Wednesday, April 28, 2010 - 5:50 am
|
|
|
Hi: Congrats on a rich and mature update. I was wondering what books you might recommend for those of us new to the Bayesian/MCMC approach to SEM estimation. All the best, Paul |
|
|
Thanks from the whole Mplus team. Funny you should ask about Bayes books - I have a about a dozen sitting on the floor. Many are good. Two stand out so far: Gelman, Carlin, Stern & Rubin (2004). Bayesian data analysis. - My favorite among "advanced" texts. Lynch (2010). Introduction to applied Bayesian statistics and estimation for social scientists. - A somewhat more applied book Lots of good articles as well. I will reference some of them in my writing "A brief introdution to using Bayes in Mplus", which will be posted before too long. |
|
|
Hi, I was wondering whether it makes sense to interpret p-values in BSEM the same way than in frequentist approach (rejection over a certain threshold). Similarly, standardized coefficients in BSEM are interpreted in the same way than in ML? thank you Davide |
|
|
No, the p-values are not interpreted in the same way as in the frequentist approach. With Bayes, you should interpret the credibility intervals. |
|
|
Thanks for your reply, I still have a doubt: can i make an interpretation about the proporion of cases on each side of the zero (that is the lower bound of th CI divided by the full range)? Or the credibility intervals give only range of the estimate, like confidence intervals? In other words, do the credibility intervals refer to number of observations or the possible values that the parameter can assume? |
|
|
Credibility intervals are just like confidence intervals. |
|
|
For an multiple group analysis of an ALT Model where the estimator = bayes, I understand that we must use TYPE=MIXTURE and the KNOWNCLASS option. If I have 2 groups (say gender with 0 and 1) and I do not want to explore classes within gender (only looking at gender differences) then do I set it as follows? CLASSES = cgender (2) c (1); KNOWNCLASS = cgender (gender = 0 gender = 1); Thank you, |
|
|
You can use only a KNOWNCLASS variable: CLASSES = cgender (2); KNOWNCLASS = cgender (gender = 0 gender = 1); |
|
|
Dear Mplus team I am fitting an endogenous treatment model. All variables continuous, T binary. Sample size=6500 Define: ONE=1; Model: L1 BY s1@1 s2@1;L2 BY d1@1 d2@1; [L1 L2](m n); [s1 s2 d1 d2](is1 is2 id1 id2); s1 ON T (t1);s2 ON T (t1);d1 ON T (t2);d2 ON T (t2); T ON L1 L2 ONE (a b c);[T$1@0]; Model priors: t1~N(-17,3.2);t2~N(-15,2.25); is1~N(0,25);is2~N(0,25); cov(is1,is2)=-24.99999; id1~N(0,25);id2~N(0,25); cov(id1,id2)=-24.99999; m~N(110,100);n~N(90,100); I used bayes to tackle underidentification using prior knowledge. Because of slow convergence, I fixed threshold at 0 and simulated an intercept with an auxiliary variable. The convergence improved (still slow). 1) Do you think that fitting an underidentified model using external information embedded in priors is a reasonable approach? 2) When I fit the model, I get an error message for singular sample covariance matrix. I understand why, but is there a way to obtain posterior p-values in this case? 3) Any suggestion to improve convergence? Thank you |
|
|
I don't recognize this as an endogenous treatment model. T is regressed on 2 factors at the same time as the factor indicators are regressed on T - what does that mean? |
|
|
The idea is that the observed values of the continuous indicators are affected by treatment status T, which, in turn, is not independent by the values of the underlying factors. In the model, L1 and L2 represent blood pressure values (systolic and diastolic) in absence of treatment. With increasing L1 and/or L2, the probability of being treated increases, and this affects the observed values s1-d2. Treatment is endogenous in this sense. Do you think that this reasoning makes sense? Thanks |
|
|
That is like a reciprocal interaction model (in econometrics). It makes substantive sense, but seems hard to identify. Usually with reciprocal interaction you need exogenous variables to identify. Or you need longitudinal data. |
|
|
I have various exogenous covariates (omitted from the model I posted, for simplicity). But I suppose you are thinking to instrumental variables, correlated with T but not with s1-d2. Am I interpreting correctly your comment? |
|
|
I am thinking of an econometric reciprocal model like y1 on y2 x1; y2 on y1 x2; Here, the fact that y1 and y2 both have a unique predictor makes the model identified, whereas this model is not: y1 on y2 x1 x2; y2 on y1 x1 x2; |
|
Lynne W posted on Wednesday, March 07, 2018 - 5:08 am
|
|
|
Hello, I am rather new to Bayesian estimation and am unsure about the priors. This is the model I want to estimate: ANALYSIS: ESTIMATOR=BAYES; POINT=MEAN; BITERATIONS=10000; MODEL PRIORS: y ~ N(2.066,0.001); m ~ N(0,3); COVARIANCE: (y,m)=2,21 MODEL: y ON m x; m ON x; The mean for y is a population value, hence I set the variance rather low. The variable m is binary, and according to Gelman 2008, I use a normal distribution and set the mean to 0 and variance to 3. The covariance is also a population value. Is this a correct specification? Also, do all variables have to be standardized in MPlus like Gelman 2008 suggests? thanks for your help |
|
|
Priors should not be put on variables but on parameters. You label the parameters in the Model command, like y on m (b); and then refer to b in Model prior. The variables don't have to be standardized. |
|
|
Hello, We are running a bifactor model with 56 subjects. I know our N is pretty small, so we are using estimator = bayes. When we set fbiter to 20,000, we get a negative estimated number of parameters, but when we set fbiter to 1,000, the number of parameters is positive. To clarify, a negative value for number of parameters is an indication that the model fit is poor, correct? Also, is there a default number of iterations (fbiter) or thinning value that we should be using? Any insight would be really helpful! Thank you! |
|
|
Please send your output to Support. |
|
Back to top |