Daniel posted on Monday, February 14, 2005 - 12:10 pm
Hi, Is use of bootstrap standard errors for calculating indirect effects better than WLSMV-derived stanadard errors when working with categorical outcomes? When I calculate indirect effects without bootstrap, I get significant indirect effects. However, when I bootstrap, I get non-significant effects.
I don't know. I think it would depend on many factors.
Daniel posted on Friday, February 18, 2005 - 9:53 am
Linda, one of my indirect effects (a*b) in this mediation analysis is not significant. However, each individual path (path a and path b) from the exogenous variable to the slope factor is significant. Do you have any idea how I would report this finding? Essentially, the pieces are significant but not the whole indirect effect.
BMuthen posted on Friday, February 18, 2005 - 11:04 am
This can happen. For example, if the a and b parameter estimates are positively correlated, this increases the standard error yielding non-significance.
What is the sample size in the analysis where the z was sig. and the percentile was not? We found this to occur rarely in simulation data with continuous variables. In most cases the significant z is a Type I error though. When sample size is 200+, the z is never signifcant if the percentile is not.
bmuthen posted on Saturday, February 26, 2005 - 5:55 pm
That typically says that your model is difficult to estimate for your data, so that for different subsets of the data the estimation does not converge properly. For instance, the sample size may be small or the model ill-fitting. You can send your files to support if you want guidance on that.
Amelia Rock posted on Tuesday, September 17, 2019 - 11:28 am
Hello, I am conducting conventional mediation analysis using data from two time points (T1 and T2). I have an ordinal mediator and binary outcome, exposures are continuous. I am estimating half-longitudinal indirect effects (HLIEs) by computing the product of the parameter estimate for each ‘a’ path (the effect of each T1 exposure variable on the T2 mediator controlling for the T1 mediator) with the ‘b’ path (the effect of the T1 mediator on the T2 outcome controlling for the T1 outcome). I am using MODEL CONSTRAINT. I want to report bootstrapped HLIEs and confidence intervals. However, when I request bootstrapped estimates, the standard errors increase greatly and/or become unavailable (i.e. “*******” – perhaps because they are extremely large??) and all pvalues for direct effects and new/additional paramters (i.e. the HLIEs) become equal to 1. When I run the model without bootstrapping, standard errors and pvalues appear reasonable. I understand that bootstrapping can lead to large SEs, but is there anything you would recommend to ameliorate this issue in my case such that I can obtain meaningful estimates and pvalues? Many thanks.