Parceling method and Estimation approach

[E. Bernardo]

Dear everyone,

 

My hypothesize model can be described as follows: consists  of two latent exogenous, one latent endogenous and one dichotomous dependent (yes/no) (sorry I used the term "dependent variable" for a variable that is not affected by any variables in the model and it caused only by the exogenous and endogenous variables); the endogenous variable is caused directly by each of the exogenous variables, while the dependent is caused directly by each of the exogenous and endogenous; the two exogenous are correlated; and each of the exogenous and endogenous consist of six sub-factors with number of items (7-point scale) ranging between 5 and 15.  Thus, the model is too large and the sample size of 329, I think, is inadequate for the model.

 

My remedy is to use the parceling approach. For each subfactor, I will parcel all the items into a single parcel then each parcel will represent the subfactor; thus, each of the exogenous and endogenous will have six indicators where the indicators are the parcels.  

 

My questions: (1) Is it required that the items should be significantly loaded to the corresponding subfactor before I will do the parceling? (2) What estimation approach is appropriate for my model considering that the dependent is dichotomous?  I am using AMOS.

 

Thank you in advance for your help.

 

Eins

[Ryan]

Eins,

 

I haven't read your post very carefully, but here are a couple of thoughts at glance:

 

(1) There have been times where I've needed to reduce the number of manifest variables entered into a structural equation model. One approach I've used is to combine certain variables based on previous theory/research and evidence that the variables in question "hang together" . You might consider estimating Cronbach's alphas to measure internal consistency. Then I enter the composites of these variables as manifest variables.

 

(2) Treating a dichotomous dependent variable as continuous variable is risky, regardless if you're talking about a structural equation model. I found this website that provides examples of fitting models with binary dependent variables.

 

http://www.structuralequations.com/3.html

 

I cannot speak to the validity of this approach, but I will be certain to explore it. Sounds interesting.

 

(3) As the structural equation modeling field has grown dramatically over the past years, so has the hierarchical modeling field. There are various procedures (e.g. Nlmixed procedure in SAS) that should have little difficulty fitting your model with all appropriate speficiations (e.g. canonical link function and binary distribution for dependent variable, while incorporating latent variables).

 

So where do you start? First I think you should start by asking youserlf what the PRIMARY research questions are, and then begin to write equations [using standard notation] that best answer such a quesiton. There are many little details that will need to be addressed sooner rather than later.

 

Perhaps if I followed more closely I'd be able to provide specific advice.

 

Sorry. HTH.

 

Ryan 

On Sat, Mar 5, 2011 at 2:28 AM, Eins Bernardo <[hidden email]> wrote:

Dear everyone,

 

My hypothesize model can be described as follows: consists  of two latent exogenous, one latent endogenous and one dichotomous dependent (yes/no) (sorry I used the term "dependent variable" for a variable that is not affected by any variables in the model and it caused only by the exogenous and endogenous variables); the endogenous variable is caused directly by each of the exogenous variables, while the dependent is caused directly by each of the exogenous and endogenous; the two exogenous are correlated; and each of the exogenous and endogenous consist of six sub-factors with number of items (7-point scale) ranging between 5 and 15.  Thus, the model is too large and the sample size of 329, I think, is inadequate for the model.

 

My remedy is to use the parceling approach. For each subfactor, I will parcel all the items into a single parcel then each parcel will represent the subfactor; thus, each of the exogenous and endogenous will have six indicators where the indicators are the parcels.  

 

My questions: (1) Is it required that the items should be significantly loaded to the corresponding subfactor before I will do the parceling? (2) What estimation approach is appropriate for my model considering that the dependent is dichotomous?  I am using AMOS.

 

Thank you in advance for your help.

 

Eins

[E. Bernardo]

Thanks a lot Ryan.

--- On Sun, 3/6/11, R B <[hidden email]> wrote:

From: R B <[hidden email]>
Subject: Re: Parceling method and Estimation approach
To: "Eins Bernardo" <[hidden email]>
Cc: [hidden email]
Date: Sunday, 6 March, 2011, 3:58 AM

Eins,

 

I haven't read your post very carefully, but here are a couple of thoughts at glance:

 

(1) There have been times where I've needed to reduce the number of manifest variables entered into a structural equation model. One approach I've used is to combine certain variables based on previous theory/research and evidence that the variables in question "hang together" . You might consider estimating Cronbach's alphas to measure internal consistency. Then I enter the composites of these variables as manifest variables.

 

(2) Treating a dichotomous dependent variable as continuous variable is risky, regardless if you're talking about a structural equation model. I found this website that provides examples of fitting models with binary dependent variables.

 

http://www.structuralequations.com/3.html

 

I cannot speak to the validity of this approach, but I will be certain to explore it. Sounds interesting.

 

(3) As the structural equation modeling field has grown dramatically over the past years, so has the hierarchical modeling field. There are various procedures (e.g. Nlmixed procedure in SAS) that should have little difficulty fitting your model with all appropriate speficiations (e.g. canonical link function and binary distribution for dependent variable, while incorporating latent variables).

 

So where do you start? First I think you should start by asking youserlf what the PRIMARY research questions are, and then begin to write equations [using standard notation] that best answer such a quesiton. There are many little details that will need to be addressed sooner rather than later.

 

Perhaps if I followed more closely I'd be able to provide specific advice.

 

Sorry. HTH.

 

Ryan 

On Sat, Mar 5, 2011 at 2:28 AM, Eins Bernardo <einsbernardo@...> wrote:

Dear everyone,

 

My hypothesize model can be described as follows: consists  of two latent exogenous, one latent endogenous and one dichotomous dependent (yes/no) (sorry I used the term "dependent variable" for a variable that is not affected by any variables in the model and it caused only by the exogenous and endogenous variables); the endogenous variable is caused directly by each of the exogenous variables, while the dependent is caused directly by each of the exogenous and endogenous; the two exogenous are correlated; and each of the exogenous and endogenous consist of six sub-factors with number of items (7-point scale) ranging between 5 and 15.  Thus, the model is too large and the sample size of 329, I think, is inadequate for the model.

 

My remedy is to use the parceling approach. For each subfactor, I will parcel all the items into a single parcel then each parcel will represent the subfactor; thus, each of the exogenous and endogenous will have six indicators where the indicators are the parcels.  

 

My questions: (1) Is it required that the items should be significantly loaded to the corresponding subfactor before I will do the parceling? (2) What estimation approach is appropriate for my model considering that the dependent is dichotomous?  I am using AMOS.

 

Thank you in advance for your help.

 

Eins