Collinearity in SEM with nested data (using lavaan.survey)

[Masha van der Sande]

Dear all,

I am trying to run a SEM with nested data, for which I use the lavaan.survey function. Including this nested structure is quite straightforward:

# Running the normal SEM

fit1 <- sem(model1, fixed.x=T, data=data)

# Creating a design factor for the nested structure

design <- svydesign(ids= ~ Random.factor, nest=TRUE, data=data)

# Combining SEM and the design factor

fit.total1 <- lavaan.survey(lavaan.fit=fit1, survey.design=design)  

However, I receive a warning in the lavaan.survey function, no matter what I change in the model:

Warning message:

1: In lavaan.survey(lavaan.fit = fit.surv12, survey.design = design,  :

  Some of the standard errors may not be trustworthy.  Some of the observed covariances or means are collinear, and this has generated collinearity in your parameter estimates.  This may be a sample size issue, missing data problem, or due to having too few clusters relative to the number of parameters. Problem encountered in group(s) 1

 

I have tested for the variance inflation factor among my variables, but it is absolutely not present. I expect that missing data and the sample size are no issue. Hence, the problem should be in "having too few clusters relative to the number of parameters". I have 18 groups in my random factor, ranging between 1 and 32 observations per group:

Group.nr	Nr of replicates
1	16
2	16
3	32
4	16
5	15
6	8
7	1
8	1
9	1
10	1
11	1
12	1
13	1
14	1
15	6
16	5
17	15
18	10

Would anyone if there is a problem, what could be the problem, and how I could solve it?

Any help is welcome! Thanks a lot in advance!

[yrosseel]

On 09/21/2015 07:59 PM, Masha van der Sande wrote:
> clusters relative to the number of parameters". I have 18 groups in my
> random factor, ranging between 1 and 32 observations per group:

I am not familiar with the technical details of the lavaan.survey
package, but it might be that these methods can not handle the setting
where some 'clusters' only have 1 observation. Does it work if you
remove those 'clusters' (group.nr 7--14)?

Yves.

[Terrence Jorgensen]

I have 18 groups in my random factor, ranging between 1 and 32 observations per group:

I only recently started looking at examples using the lavaan.survey package.  Here is a link to a script co-written by the developer, as support for analyzing European Social Survey data in R (see more details here).

Pay special attention to lines 69-74 of the script:

# By default, R will crash if a primary sampling unit (psu) has a single observation.

# Set R to produce conservative standard errors instead of crashing by running this line:

options( survey.lonely.psu = "adjust" )

# this setting matches the MISSUNIT option in SUDAAN

# http://r-survey.r-forge.r-project.org/survey/exmample-lonely.html

Hope this helps,

Terry

[mett...@gmail.com]

Hi folks,

I also get this error message and I have come to the conclusion that I have too few clusters relative to the number of parameters. I have 6 clusters (regions), each with 16 observations (sites). So I have 96 sites in total, and the sites were sampled in two consecutive years. Is there a rule of thumb for the number of clusters one should have relative to the number of parameters, and how concerned should one be about borderline cases.

Thanks
Mark