Multiple Group Path Analysis Final Jan 2021

You might also like

Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 27

Multigroup path analysis using AMOS

Mike Crowson, Ph.D.


University of Oklahoma
January 2021

Video link: https://youtu.be/t7zdpwFQLwo


The example in this presentation is based on that which is provided in the AMOS 24 User’s Guide (
IBM_SPSS_Amos_User_Guide (7).pdf), Example 11. The example provided in the User’s Guide involves a multisample path
analysis based on a sample of 209 girls & 207 boys from a study by Felson and Bohrnstedt’s (1979). Although that
presentation focused on testing for invariant path coefficients, the example I provide involves testing of invariant path
coefficients and covariances among exogenous variables in the model

Citation: Felson, R. B., & Bohrnstedt, G. W. (1979). "Are the good beautiful or the beautiful good?" The relationship between
children's perceptions of ability and perceptions of physical attractiveness. Social Psychology Quarterly, 42(4), 386–392.
For our multigroup path analysis, we begin by testing the model for fit in each group separately. Assuming the model fits
equally in each group, then we proceed to testing a fully unconstrained model, where there are no equality constraints
imposed. Next, we test a constrained model with those coefficients we wish to test for evidence of invariance are held
constant across groups against the fully unconstrained model. If we have evidence of non-invariant parameters, we then go
through a series of steps where test each parameter by freeing one at a time and testing the difference in fit between the
fully constrained model and the partially constrained models. Generally, the more restrictive model will fit the data worse
than a partially constrained model (where a single parameter is freed). Each chi-square difference test associated with each
parameter is a test of whether the fit of the model fit improves significantly as a result of relaxing a constraint. If significance
is found, we assume the parameter is non-invariant (i.e., unequal) across groups. If significance is not achieved, then we
assume the parameter is invariant (i.e., equal) across groups. After identifying non-invariant parameters, we can test a final
model where invariant parameters are constrained equal and non-invariant parameters are freely estimated.
Step 1a: We will test the fit of our model among the sample comprised of girls. Link:
https://drive.google.com/file/d/10EttAf0o9DRdK7XfEmbWSfGd75tfKQMg/view?usp=sharing

Standardized coefficients for the model based on girls’ data.

The overall fit of the model was quite good based on review of the
fit indices.

For a review of criteria for model fit using these indices see:
https://drive.google.com/file/d/1U1ExQrBr_EOXioaYij85SKT-lLTXM
iYr/view?usp=sharing
In the girls sample, the unstandardized path coefficient from
‘gpa’ to ‘academic’ was significant (p<.001), as was the path
from ‘rating’ to ‘attract’ (p<.001) and the path from
‘academic’ to ‘attract’ (p<.001).
Step 1b: We will test the fit of our model among the sample comprised of boys

Standardized coefficients for the model based on boys’ data.

The overall fit of the model was quite good based on review of the
fit indices.
In the boys sample, all unstandardized path coefficients were
statistically significant (at p<.05), except for the path from
‘attract’ to ‘academic’.

https://drive.google.com/file/d/1JlCY6r_XBAMY-bBQ7NgkFZHJB52f7bLa/view?usp=sharing
Step 2: Given that the fit of the model (based on the fit indices) for each group was acceptable, we progress to the
multisample path analysis itself using the model specification we arrived at based on our previous analyses.

We begin by using the Manage Groups drop-down to establish our two groups…
Next, we select the data file that will be used
in the analysis of each group.

After specifying our analysis properties, we run


the model.
The model being tested first (here) is an
unconstrained model – i.e., no parameters
were constrained across groups. The fit
statistics that are printed in AMOS will allow
us to use this model as a baseline model
against which we carry out part of our
invariance testing. Specifically, we will
compare the fit of this model where the
parameters are freely estimated in each
group against a model where the
parameters of interest are all constrained
equal.

The unstandardized (shown here) and


standardized coefficients for each group are
the same as before…

Link to AMOS file containing unconstrained model:


https://drive.google.com/file/d/10qcoYjkxvtAWJ1x-efeXG5kgM2VHzTdx/view?usp=sharing
In general, the fit of this unconstrained model exhibited a strong fit to
the data. [The chi-square value printed here is approximately equal to
the sum of the chi-square values associated with the analyses from
each group].
Now, let’s test a model where we constrain all of the coefficients of interest to be equal across groups. To accomplish this,
we will assign the same label to each parameter that is being tested for invariance. Below, I have assigned the label ‘p1’
to the path from ‘gpa’ to ‘academic’. Additionally, the ‘All groups’ box remains clicked (which assigns the same label to
both groups).
For this demonstration, we will constrain all path coefficients and the covariances among the exogenous variables to
equality…

https://drive.google.com/file/d/1KwVaGDbqtI6nyE3ku-9bDUthxjb4nnKI/view?usp=sharing
Fit of fully constrained model Previous unconstrained model

Notice that the fit of the current constrained model is worse than that of the fit of the unconstrained model. This occurs when
equality constraints are placed on parameters. The chi-square value of the current model is χ² = 37.235, whereas the previous
chi-square was χ² = 3.183. Since a chi-square of 0 indicates exact fit, then larger of the two chi-square values indicates the
worse-fitting model.
Fit of fully constrained model Previous unconstrained model

We can test whether the current model exhibits a significantly worse fit than the previous unconstrained model by performing
a chi-square difference test. The difference in chi-square [37.235 – 3.183 = 34.052] is compared to the tabled chi-square,
assuming α=.05 and df = 16 – 4 = 12.
Here is an online calculator to compute the p-value: https://www.socscistatistics.com/pvalues/chidistribution.aspx

The chi-square test indicates that the change in


fit from the unconstrained model to the
constrained model is statistically significant
(p<.001). This indicates that at least some of
the parameters we have constrained to
equality are non-invariant (i.e., not equal).

Given this result, we will want to identify the non-invariant parameters. We can do this through a series of chi-square
difference tests comparing the fit of the fully constrained model against the fit of a series of models where one parameter
constraint is relaxed (one at a time). The assumption behind these tests if the fit of the model improves significantly with the
relaxation of a constraint, then that parameter is non-invariant across groups (which justifies freeing that parameter in the final
model). [Note: If our chi-square difference test (comparing the constrained and unconstrained models) had been non-
significant, then there would be no need to test each of the parameters.]
Fit of fully constrained model Previous unconstrained model

Side note: Although I am emphasizing the use of chi-square difference tests in this demonstration, it is also possible to use
other descriptive indices when evaluating whether the difference in nested models is substantial or of importance. One
possible threshold is a difference in CFI of > .01 (see Cheung & Rensvold, 2002) as a basis for assuming non-invariant
parameters.
Since each of our tests will result in df=1, then we can use 3.84 (assuming α=.05) as the chi-square critical value for a given
test. To make things even simpler, we could subtract 3.84 from the critical value of our fully constrained model to arrive at a
chi-square threshold at or below which would indicate statistical significance for the test. The chi-square value for the fully
constrained model is 37.235. Thus, we could use 37.235 – 3.84 ≤ 33.395 as the chi-square value for any particular test that
would represent a statistically significant improvement in fit.

As an example, let’s start by relaxing the equality constraint on ‘p1’. This requires us to remove the label for that parameter
and then to run the analysis and note the chi-square value and degrees of freedom for the model…
The chi-square value for the model where the
path from ‘gpa’ to ‘academic’ is 37.214 (which is
> 33.395). Parameter can be treated as
invariant.
Next, we re-constrain ‘p1’ and then proceed to testing the next parameter (‘p2’).

The chi-square value for the model where the


path from ‘height’ to ‘attract’ is 35.993 (which is
> 33.395). Parameter can be treated as
invariant.
Next, we re-constrain ‘p2’ and then proceed to testing the next parameter (‘p3’).

The chi-square value for the model where the


path from ‘weight’ to ‘attract’ is 37.181 (which
is > 33.395). Parameter can be treated as
invariant.
Next, we re-constrain ‘p3’ and then proceed to testing the next parameter (‘p4’).

The chi-square value for the model where the


path from ‘rating’ to ‘attract’ is 33.455 (which is
> 33.395). Although close, we may consider this
parameter invariant across groups.
The parameters identified using these tests as non-invariant include: c1, c2, c5.

Download file with computations here:


https://drive.google.com/file/d/1rEgCVqrJqkWofS_eVTAodPOFW6TorlxF/view?usp=sharing

Let’s re-specify our model with all the previous constraints minus these three…
Here is the final model and the final fit statistics (download .amw file:
https://drive.google.com/file/d/16fLGU5EhLQYukzOYkcZ0OJ4f68byBQIy/view?
usp=sharing
) The difference in chi-square between this model containing partial
constraints and our previous model with a full set of constraints is:

37.235 – 11.130 = 26.105.

The chi-square difference test (where df = 16-13 = 3) is statistically significant


(see next slide).
The partially constrained model represents a significant improvement in fit relative to the model containing our full set of
constrained parameters.
Most important, we want to let’s test the fit of our partially constrained model and the full unconstrained model. The
expectation is that this model does not represent a significant decrease in fit relative to the unconstrained model.

The chi-square difference value is 11.130 - 3.183 = 7.947. The degrees of freedom for the comparison is 15-4 = 11…

The results indicate that our final model did not the fit data significantly worse than the unconstrained model.
See the following presentation that covers invariance testing in the context of confirmatory factor analysis.

Video: https://youtu.be/PSK5dhEwJ98
Powerpoint: https://drive.google.com/file/d/1GVi5dqRiScVdxJdJ27LGpP_x_Aixm2o0/view?usp=sharing

View another video demonstration on how to test for differences in indirect effects between groups using this same
data can be found here: https://youtu.be/3r1llTC0YIk

You might also like