Covariance between exogenous binary variables in SEM

[Dai Duong]

Hello experts and fans of lavaan,

I want to include co-variances between exogenous binary variables into my SEM model. I have two questions:

1. In lavaan syntax, to allow such co-variances, I use two options: "fixed.x = FALSE,conditional.x = FALSE" 

so the command will be: sem(fit, data=data,std.lv = TRUE,fixed.x = FALSE,conditional.x = FALSE)

Is that correct?

2. The above command only work with exogenous variables that have at least three values. When I add an exogenous binary variable, errors show up 

"Error in eigen(VarCov, symmetric = TRUE, only.values = TRUE) : 

  infinite or missing values in 'x'

In addition: Warning message:

In sqrt(A1[[g]]) : NaNs produced"

I don't understand why this happens and how to overcome it?

Here is my model

ctrl.l <- "

ctrl_l =~ income+wh+skill+mwh21+njob #in which njob and skill are categorical indicators.

ctrl_l~tedu+m1ac2+ahtedu+hleadera+ titw+urbanrate+exp #all exogenous variables are continuous or ordered

wh ~~  mwh21

mwh21 ~~   njob

income ~~  skill

income ~~ wh

income ~~ mwh21

wh ~~  njob

"

fit.l <- sem(ctrl.l, data=data,std.lv = TRUE,fixed.x = FALSE,conditional.x = FALSE)

==> Regress successfully!

But if I add binary variables as gender, urban/rural, or other dummy variables, errors occur:

  

  ctrl.l <- "

ctrl_l =~ income+wh+skill+mwh21+njob    # njob and skill are categorical indicators.

ctrl_l~tedu+m1ac2+ahtedu+hleadera+ titw+urbanrate+exp+gender+urban    #all exogenous variables are continuous or ordered while gender and urban are binary

wh ~~  mwh21

mwh21 ~~   njob

income ~~  skill

income ~~ wh

income ~~ mwh21

wh ~~  njob

"

fit.l <- sem(ctrl.l, data=data,std.lv = TRUE,fixed.x = FALSE,conditional.x = FALSE)

Error in eigen(VarCov, symmetric = TRUE, only.values = TRUE) : 

  infinite or missing values in 'x'

In addition: Warning message:

  In sqrt(A1[[g]]) : NaNs produced

Please help me, 

Thank you very much,

Dai Duong

[Dai Duong]

Hi all,

I have found a way to make lavaan running with co-variances of exogenous binary variables: set binaries as ordered variables.

However, another issue occurs:

when I add an exogenous binary variable, residual variances of two indicators (njob, skill) become negative

$cov

          inc001 wh0001 skill  mwh21  njob   ahtedu hleadr migrtd

income001  0.000                                                 

wh0001    -0.015  0.000                                          

skill      0.059 -0.133 -0.016                                   

mwh21      0.001 -0.016  0.012  0.000                            

njob       0.031  0.164  0.000  0.149 -0.021                     

ahtedu     0.022 -0.025  0.097  0.005  0.016  0.000              

hleadera   0.002 -0.006  0.003  0.003  0.008  0.000  0.000       

migrated   0.022 -0.033 -0.025  0.008 -0.003  0.000  0.002  0.000

In the above residual covariance matrix, residual variances of skill and njob become negative when migrated (binary) is added.

 

If I add non-binary variable, for example tedu01 (schooling years) while keeping njob and skill in model then residual variances of skill and njob are 0

$cov

          inc001 wh0001 skill  mwh21  njob   ahtedu hleadr tedu01

income001  0.000                                                 

wh0001    -0.019  0.000                                          

skill      0.036 -0.143  0.000                                   

mwh21      0.001 -0.021  0.012  0.000                            

njob       0.029  0.148  0.000  0.170  0.000                     

ahtedu     0.018 -0.027  0.087  0.005  0.014  0.000              

hleadera   0.001 -0.006  0.003  0.003  0.007  0.000  0.000       

tedu01     0.015 -0.053  0.115  0.005  0.044  0.109  0.008  0.000

 

I try to remove both njob and skill from measurement part, then no issue occurs even when a binary variable is added

$cov

          inc001 wh0001 mwh21  ahtedu hleadr prisct

income001  0.000                                   

wh0001    -0.009  0.000                            

mwh21      0.000 -0.004  0.000                     

ahtedu     0.032 -0.020  0.003  0.000              

hleadera   0.002 -0.007  0.004  0.000  0.000       

prisect   -0.015  0.157 -0.011  0.000 -0.016  0.000

 

It seems that, the inclusion of exogenous binary variables makes the model invalid if covariances between exogenous variables are allowed.

 

How should I solve this issue?

Thank you very much,

Dai Duong

[Terrence Jorgensen]

1. In lavaan syntax, to allow such co-variances, I use two options: "fixed.x = FALSE,conditional.x = FALSE" 

so the command will be: sem(fit, data=data,std.lv = TRUE,fixed.x = FALSE,conditional.x = FALSE)

Is that correct?

Yes.

2. The above command only work with exogenous variables that have at least three values. When I add an exogenous binary variable, errors show up 

"Error in eigen(VarCov, symmetric = TRUE, only.values = TRUE) : 

  infinite or missing values in 'x'

In addition: Warning message:

In sqrt(A1[[g]]) : NaNs produced"

I don't understand why this happens and how to overcome it?

Can you post enough data to reproduce this error?

Terrence D. Jorgensen

Postdoctoral Researcher, Methods and Statistics

Research Institute for Child Development and Education, the University of Amsterdam

http://www.uva.nl/profile/t.d.jorgensen

[Dai Duong]

Dear Dr.Terrence

For the problem:

2. The above command only work with exogenous variables that have at least three values. When I add an exogenous binary variable, errors show up 

"Error in eigen(VarCov, symmetric = TRUE, only.values = TRUE) : 

  infinite or missing values in 'x'

In addition: Warning message:

In sqrt(A1[[g]]) : NaNs produced"

Can you post enough data to reproduce this error?

--> My sample contains 5856 obs.

I fixed the issue by setting exogenous binary variables as ordered variables. Then fitting runs successfully. It is ok, isn't it? 

If I don't specify as ordered ones, there are other errors like this:

> ctrl.labor <- "

+ control_labor =~ income001+wh0001+skill+mwh21+njob

+ control_labor~ahtedu+hleadera+tedu01+exp01+ttnt

+ wh0001 ~~  mwh21

+ mwh21 ~~   njob

+ income001 ~~  skill

+ income001 ~~ wh0001

+ income001 ~~ mwh21

+ wh0001 ~~  njob

+ "

> fit.l <- sem(ctrl.labor, data=x180108_lkh,std.lv = TRUE,fixed.x = FALSE,conditional.x = FALSE,ordered=c("njob","skill")) # only two indicators is set "ordered", none of predictors is set "ordered".

Warning messages:

1: In muthen1984(Data = X[[g]], ov.names = ov.names[[g]], ov.types = ov.types,  :

  lavaan WARNING: trouble constructing W matrix; used generalized inverse for A11 submatrix

2: In lav_model_vcov(lavmodel = lavmodel, lavsamplestats = lavsamplestats,  :

  lavaan WARNING:

    The variance-covariance matrix of the estimated parameters (vcov)

    does not appear to be positive definite! The smallest eigenvalue

    (= -4.591183e-20) is smaller than zero. This may be a symptom that

    the model is not identified.

Thank you very much

Dai

On Wednesday, 26 September 2018 00:21:04 UTC-5, Dai Duong wrote: