Structural Equation Modelling (SEM)using STATA

[Ainun Nisa]

Hi,
I want to analyse my data with SEM using STATA. But i still confuse because there is no unobserve variable in my model. This is what my model looks like:

Can i still using SEM?

[Marcos Almeida]

In the Stata Manual) we read that structural equation models are ‘a way of thinking ‘. A frame of mind, so to speak. Therefore, path analysis, which matches with the diagram shown in #1, is part of other SEM resources, such as latent variables, growth models, MIMIC, etc.

Structural equation modeling encompasses a broad array of models from linear regression to measurement models to simultaneous equations, including along the way confirmatory factor analysis (CFA), correlated uniqueness models, latent growth models, multiple indicators and multiple causes (MIMIC) models, and item-response theory (IRT) models.
Structural equation modeling is not just an estimation method for a particular model in the way that Stata’s regress and probit commands are, or even in the way that stcox and mixed are. Structural equation modeling is a way of thinking, a way of writing, and a way of estimating

[Joseph Coveney]

. . .there is no unobserve variable in my model.

Can i still using SEM?

Yes.

You don't need unobserved variables (latent factors) in your SEM model in order to use sem. As Marcos mentions, you have what is called a "path analysis" model, and these can be fitted using sem as well.

When I re-create your path diagram using Stata's sembuilder thingy, I get the following.

Which, when I execute it against an artificial dataset, produces the following line of code.

sem (income -> health, ) (asset -> health, ) (distance -> income, ) (distance -> asset, ), nocapslatent

This can be rendered equivalently—but a little more readable—manually, as in the output below. (Begin at the "Begin here" comment; the first part is just to create the phony data.)

.ÿ
.ÿversionÿ16.1

.ÿ
.ÿclearÿ*

.ÿ
.ÿsetÿseedÿ`=strreverse("1557304")'

.ÿ
.ÿtempnameÿCorr

.ÿmatrixÿdefineÿ`Corr'ÿ=ÿJ(4,ÿ4,ÿ0.5)ÿ+ÿI(4)ÿ*ÿ0.5

.ÿ
.ÿquietlyÿdrawnormÿdistanceÿassetÿincomeÿhealth,ÿdoubleÿcorr(`Corr')ÿn(250)

.ÿ
.ÿ*
.ÿ*ÿBeginÿhere
.ÿ*
.ÿsemÿ(healthÿ<-ÿincomeÿasset)ÿ(assetÿincomeÿ<-ÿdistance),ÿ///
>ÿÿÿÿÿÿÿÿÿnocnsreportÿnodescribeÿnofootnoteÿnolog

StructuralÿequationÿmodelÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿNumberÿofÿobsÿÿÿÿÿ=ÿÿÿÿÿÿÿÿ250
Estimationÿmethodÿÿ=ÿml
Logÿlikelihoodÿÿÿÿÿ=ÿ-1293.1549

------------------------------------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿÿÿÿ|ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿOIM
ÿÿÿÿÿÿÿÿÿÿÿÿÿ|ÿÿÿÿÿÿCoef.ÿÿÿStd.ÿErr.ÿÿÿÿÿÿzÿÿÿÿP>|z|ÿÿÿÿÿ[95%ÿConf.ÿInterval]
-------------+----------------------------------------------------------------
Structuralÿÿÿ|
ÿÿhealthÿÿÿÿÿ|
ÿÿÿÿÿÿincomeÿ|ÿÿÿ.2690603ÿÿÿ.0560947ÿÿÿÿÿ4.80ÿÿÿ0.000ÿÿÿÿÿ.1591167ÿÿÿÿ.3790039
ÿÿÿÿÿÿÿassetÿ|ÿÿÿ.3759859ÿÿÿ.0554518ÿÿÿÿÿ6.78ÿÿÿ0.000ÿÿÿÿÿ.2673023ÿÿÿÿ.4846694
ÿÿÿÿÿÿÿ_consÿ|ÿÿ-.0382719ÿÿÿ.0497898ÿÿÿÿ-0.77ÿÿÿ0.442ÿÿÿÿ-.1358581ÿÿÿÿ.0593144
ÿÿ-----------+----------------------------------------------------------------
ÿÿincomeÿÿÿÿÿ|
ÿÿÿÿdistanceÿ|ÿÿÿ.4885884ÿÿÿ.0575273ÿÿÿÿÿ8.49ÿÿÿ0.000ÿÿÿÿÿ.3758369ÿÿÿÿ.6013399
ÿÿÿÿÿÿÿ_consÿ|ÿÿ-.0079931ÿÿÿ.0553355ÿÿÿÿ-0.14ÿÿÿ0.885ÿÿÿÿ-.1164487ÿÿÿÿ.1004624
ÿÿ-----------+----------------------------------------------------------------
ÿÿassetÿÿÿÿÿÿ|
ÿÿÿÿdistanceÿ|ÿÿÿ.4259697ÿÿÿ.0603151ÿÿÿÿÿ7.06ÿÿÿ0.000ÿÿÿÿÿ.3077542ÿÿÿÿ.5441851
ÿÿÿÿÿÿÿ_consÿ|ÿÿÿ.0438545ÿÿÿÿ.058017ÿÿÿÿÿ0.76ÿÿÿ0.450ÿÿÿÿ-.0698568ÿÿÿÿ.1575659
-------------+----------------------------------------------------------------
var(e.health)|ÿÿÿ.6180408ÿÿÿ.0552793ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ.5186608ÿÿÿÿÿ.736463
var(e.income)|ÿÿÿ.7635723ÿÿÿÿ.068296ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ.640791ÿÿÿÿ.9098796
ÿvar(e.asset)|ÿÿÿ.8393711ÿÿÿ.0750756ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ.7044014ÿÿÿÿ1.000202
------------------------------------------------------------------------------

.ÿ
.ÿexit

endÿofÿdo-file


.

In general, I recommend using lower case for names of observed variables, especially in the context of sem and gsem.

[Christopher Bratt]

Leaving your orignal question aside, your model claims that there is no direct effect from Distance to Health, all the association between these two variables is mediated by Income and Asset. I'm not sure that's reasonable. But you can test that assumption against a model with that path added. (Or, in this case, you can simply test the model you propose and have a look at model fit, that will be sufficient here). Even though path analysis is not SEM in a strict sense, you test path models with SEM software and if you leave out one or several paths (as you do), the model becomes falsifiable and you need to check model fit.

[Richard Williams]

If it helps any, you can think of the observed variables as being perfectly measured. You therefore don't need latent variables for them.

[Ainun Nisa]

Thank you all for all the answer. Its really helpful.
But I have this question. If variable health is chategorical and its value is 1-4 (1 for very bad health condition and 4 for very good health condition). Can i still use sem? or should i use Gsem?