|
|
Covariance between two random logv's |
|
Message/Author |
|
Jon Heron posted on Monday, September 28, 2020 - 2:48 am
|
|
|
Hi everyone this is something I have been grappling with recently. When fitting a bivariate DSEM we can incorporate random effects for the level-1 variation in both processes (logv1 logv2) and also (the log of) a random covariance which I have seen referred to as LogNCov. This additional random covariance term is not easy to interpret and furthermore it is causing me no end of estimation problems. So how important is it? It feels to me that it's strange to permit individuals to vary in their variability however NOT to allow their covariance to vary, ergo we should always include this additional random effect. Would you agree? (I'm secretly hoping you say no) all the best, Jon |
|
|
Tables 3 and 4 http://www.statmodel.com/download/DSEM.pdf show that ignoring subject specific resdiual variance can affect the structural parameters in the model. To date there is no such simulation to demonstrate the same for the residual covariance and we have tried to find such even with large samples. Model given in equation (10) in https://www.tandfonline.com/doi/pdf/10.1080/00273171.2018.1446819 can be used to estimate the subject specific covariance under the assumption that it has the same sign across individuals. Without that assumption you can use something like that f1 by y1@1; Y1@0.1; f2 by y2@1; Y2@0.1; s | f1 on f2; v1 | f1; v2 | f2; to estimate subject specific variance covariance matrix. |
|
|
Just a little correction. This concerns "residual variance and covariance", not actual variance and covariance, so the above post should have "residual" inserted everywhere. This distinction is fairly important because if you already have subject specific AR parameters you already have subject specific variance covariance (even though you might not have subject specific residual variance covariance). Typically, including the subject specific residual parameters would require longer time-series for each subject. |
|
Jon Heron posted on Monday, September 28, 2020 - 11:33 pm
|
|
|
many thanks Tihomir |
|
Back to top |
|
|