Hot answers tagged

8

Many statistical tests can be thought of as structural equation models, and one of those is the paired samples t-test. As you say, the advantage of the SEM approach is that you can use FIML estimation - which is asymptotically equivalent to multiple imputation, but can be easier. You estimate a parameter which represents the difference between the means, ...


7

What you want is in the lavaan package, the function is named sem. Try writing an argument estimator = "WLSMV". For more information read this.


3

Neither normal nor beta ;-) Let's say that there are 6 items (presented to the children as scenarios) and that each response may be 1 (may be depressed) or 0. Thus you have seven possible outcomes: 0 (0/6), 0.167 (1/6), 0.333 (2/6), 0.5 (3/6), 0.667 (4/6), 0.833 (5/6), and 1 (6/6). You can think of these outcomoes as count data (how many ones?), thus you ...


2

Background The different fit indices tend to be sensitive to different forms of misspecification. Looking at the formulae, where $T$ is the model statistic, $df$ the degrees of freedom, and subscripts indicate baseline versus target: $$ TLI = \frac{(T_b / df_b) - (T_t / df_t)}{(T_b / df_b) - 1} $$ $$ SRMR = \sqrt{\left( \frac{2\Sigma_{i=1}^{p}\Sigma_{j=1}...


2

You are likely attempting to extract too many factors from your item pool. If you have reason to believe that certain items should load to certain factors, then switch to a CFA approach. However, if you don't have a reasonable belief as to which items the factors should load, then stick with the CFA. How many factors to extract and what criteria to use is ...


2

This question is very broad. It first of all really depends on the model you want to test, in which a higher complexity would decrease the validity of an ML-SEM model (but probably also of a BSEM model). I would say, as a starter, try both and experience/see which difference you get. To give you a gross insight in the debate between both you could read the ...


2

This is one of a class of more general questions in SEM about how fit indices are calculated - it's not just relevant to Mplus. The incremental fit indices (CFI, etc) all work by comparing the fitted model chi-square with the null model chi-square. They are not hard to work out. For example: $$ CFI = (\chi^2_0 - \chi^2_m) / \chi^2_0 $$ TLI/NNFI includes ...


1

If you end up with the [0,1] averaged distribution, you could use a zero-one-inflated beta model as well. Flexible, easy to implement.


1

For those who do not know what the "sample-adjusted meta-analytic deviancy (SAMD)" statistic is -- it's just externally studentized residuals (see Wikipedia) for meta-analytic data (also called 'studentized deleted residuals'). See: Huffcutt, A. I., & Arthur, W., Jr. (1995). Development of a new outlier statistic for meta-analytic data. Journal of ...


1

For exogenous categorical predictors, maximum likelihood estimation is fine. As long as you treat the exogenous variables as fixed, you are not modeling their distribution, so you can be agnostic about their distribution and treat them as if they were any continuous variable. If you decide to treat them as random, you can assume the binary variable is a ...


1

This would be pretty unusual, but not unheard of, and it would be possible to rationalize such a finding. However, to take a more agnostic stance, it might make more sense to simply correlate the errors of those two indicators, which should yield the same improvement in fit without requiring the causal assumptions that are laden in a directed path. If you ...


1

You can create new parameters in lavaan with the := symbol. First, label your parameters: model <- ' # measurement model opp =~ FTP1 + FTP2 + FTP3 + FTP9 ext =~ FTP4 + FTP5 + FTP6 + FTP7 const =~ FTP8 + FTP9 + FTP10 ses =~ SES1 + SES2 + SES3 + SES4 + SES5 + SES6 # regressions ses ~ a * opp + b * ext + c * const ' You want to ...


1

To obtain valid inference in an SEM model, the residuals for each of X, Y, M1, and M2 must be conditionally independent of the confounder variable(s) to eliminate its/their confounding effects. The easiest and most typical way to do this is to add a path from the confounding variables to each of these variables. However, this will not work in all scenarios. ...


1

Yes, it's possible. You will lose information about variances, which you might care about. You should not do this if you have a multiple group model. If you use standardized variables, you can obviously only report standardized estimates.


1

Here are a couple of blog entries by Andrew Gelman about this issue: http://andrewgelman.com/2007/11/28/clustered_stand/ , http://andrewgelman.com/2009/08/21/clustered_stand_1/. If you add complex (and don't have weights) then all that Mplus will do is correct your standard errors - the estimates should stay the same. If you use multilevel models, you model ...


1

If you care about the organizations (the clusters), then use type=twolevel and estimate a within and a between part of your model. If you were not concerned about the clustering level but wanted to control for within-cluster correlation (you should do that), you would use type=complex to make Mplus use a sandwich estimator. These two solutions are ...


1

I'd echo the comments that asking for code examples is off-topic, and yet I'd agree this is an--albeit complicated--statistical question. I've been dealing with similar issues, mediation from panel data. Intuitively I understand that mediation has a longitudinal aspect to it in the sense that the mediator must be caused by the primary exposure and, in turn, ...


1

It's important to distinguish causal or conceptual assumptions from statistical assumptions here. Reverse causality, for instance, has no test, since statistically $A \rightarrow B$ leads to a probability model for $P(A, B)$ that is identical to $B \rightarrow A$ without control for additional causal moderators. Unlike most regression techniques, SEM ...


1

In Lavaan you can calculate new parameters, using the := operator. Use this to calculate the sum of indirect effects. Here's an example. We have 4 predictors (x1-x4), one mediator (m) and one outcome (y). m ~ x1 (mx1) m ~ x2 (mx2) m ~ x3 (mx3) m ~ x4 (mx4) y ~ x1 (yx1) y ~ x2 (yx2) y ~ x3 (yx3) y ~ x4 (yx4) y ~ m totIndEffect := yx1 + yx2 + yx3 + yx4


1

Use the mvrnorm function in the MASS package to get a multivariate normal distribution (it comes with base R). Here's some example code: library(MASS) library(ggplot2) # used for plotting, otherwise, it's optional Sigma <- matrix(c(10,3,3,2),2,2) var(mvrnorm(n = 1000, rep(0, 2), Sigma)) example <- mvrnorm(n = 1000, rep(0, 2), Sigma) examplePlot <-...


1

Estimated sample statistics in Mplus is the covariances of the variables. In SPSS you have asked for the covariances of the parameters. In Mplus you need to request tech3 on the output line, i.e. OUTPUT: tech1 tech3 tech4 samp stand mod(4); The estimators are not the same, you're using ML in Mplus and OLS in SPSS, but the results should be extremely ...


1

@ Erik Actually, Mplus offers a way to restrict correaltions between latent factors. f1 WITH f2@1 should restrict the correlation between f1 and f2 to 1, if variances within both factors are fixed to 1 (by f1@1; f2@1). Otherwise, you restrict the covariance. See e.g. the question by Boliang Guo on http://www.statmodel.com/discussion/messages/9/380.html?...


1

If the 7-item scale measures some meaningful construct, then I would recommend the first approach. If each item itself is informative or relevant to your research question, then I would go with the latter.


1

I would have to agree with the previous post, R is a good alternative to these. I've used both HLM and M-PLUS and find that data manipulation is difficult to impossible, and the scripting language of R is much better for actually using the results from the models. my 2 cents.


1

Assuming that your model is correctly specified, then it looks like the likelihood is highly non linear or unreliable at neighborhood of the last iteration which seemed to be maxima (likely a local one). My strategy would be to try several starting values by using the STARTING option which allows you to set the number of starting value sets. I suppose that ...


1

When complicated things are going wrong, look to simple things :-). Things to look at What does the correlation matrix look like? How severe is the left censoring? How different are your data from those in the FA you are trying to confirm? Then you could try an exploratory FA, to see if your data have 4 factors (much less the same four).


Only top voted, non community-wiki answers of a minimum length are eligible