Could anyone advise me on the procedure for estimating the significance , p values, of parameter estimates in a path analyis involving mixed categorical and continuous variables using WLS and WLSMV. Other SEM packages such as EQS and Amos calculate them automatically.
The ratio of the parameter estimate to the standard error (column three of the results) is a z-statistic (standard normal variable). You can look up the absolute value in a z-table and obtain the probability for values greater than or equal to that. The probability times 2 is the p-value for this two-tailed test. For example, the z-value 1.96 has a probability of .025 or a significance level of .05.
I have a one factor model with five observed variables. Each of the observed variables represents responses to a Likert type scale coded 1 to 5. Some of the items are highly skewed. Model fit is poor [WLSMV estimator] with a chi-square of 240, df = 6, RMSEA ~ .17. When I allow two of the residuals to correlate, the residual variance listed under the R-SQUARE heading becomes negative and large. Why does this happen?
where y1-y4 are categorical outcomes in the model.
Anonymous posted on Monday, January 20, 2003 - 7:44 am
I'm trying to test a two-factor model with ten observed variables. My data type is categorical(including binary), so I use WLS estimator in estimating the solution. Furthermore, I need to allow two covariances (for convenience, call A and B) among the residuals.
The output of model results show A is significant, while B is not, although the Est(and Std) of B is larger than that of A. I can't know why does this happen? How should I interpret this result?
It is not the size of the parameter estimate that determines statistical significance but the ratio of the parameter estimate to its standard error. Even though the size of the parameter estimate may be large, it may not be measured with as much precision as a smaller estimate.
Amy posted on Tuesday, November 22, 2005 - 12:47 pm
When WLSMV estimation is used for observed categorical and observed continuous dependent variables in the same model, what is the metric of the effects of the independent variables on the dichotmous dependent variable (i.e. the estimates reported in the coefficient column)? Is it using the probit or logit models, or percent difference, etc.?
bmuthen posted on Tuesday, November 22, 2005 - 4:30 pm