RMSEA returns NA

2 replies [Last post]
Mrkwht's picture
Offline
Joined: 03/27/2012

I fit a basic latent growth curve model (using covariance data and means) and everything seems to work fine, but the RMSEA is returning NA in the summary function. Looking at the behavior of the summary function, it appears that this is because the squaredRMSEA is less than 0 (model chi-sq=3 with df=4). I am just now in the process of learning SEM, so forgive the basic question. But I thought that RMSEA should be given as 0 if this happens, not NA. I'm sure that I'm just missing something, but what is the reasoning behind returning NA?

Thanks in advance!
Mark

Ryne's picture
Offline
Joined: 07/31/2009
Thanks for pointing this out.

Thanks for pointing this out. I've made a change to the summary function that is available to users building from source and should be included in the next binary release unless someone points out a problem. When squaredRMSEA is less than zero, we'll now return a comment that reads "0.000 (Non-centrality parameter is negative)."

First, does anyone see a problem with this? Second, what should we do for negative values of RMSEA confidence intervals should we ever include them (i.e., should we allow the lower bound of the CI to go below zero)?

mspiegel's picture
Offline
Joined: 07/31/2009
Oh wait, I wasn't paying

Oh wait, I wasn't paying attention during this discussion. I understand the need for informing the user when the non-centrality parameter is negative, but I don't think we should be reporting that information by placing a character string in the 'RMSEA' field of the summary output. The RMSEA should return either a number or NA (which is number-ish) but not a string. Can we add a new boolean field to the summary object that indicates when the non-centrality parameter is negative?

The fancy option is to store RMSEA as a number and have a new boolean field, and when the summary object is printed we combine this information in the output to highlight the non-centrality parameter is negative. We use this technique to highlight when a free parameter is resting on an upper or lower bound (see MxSummary.R:323-330).