| Message/Author |
|
|
| Marie Alice posted on Tuesday, September 06, 2016 - 7:51 pm
|
|
|
I am running a multi-group mediation analysis (Y=math achievement, X=single parent, M=parental involvement, Group=urban). (1) Through comparison of various equality constraints, when I find that: direct effects of Y on X, and Y on M differ across groups, but direct effect of M on X do not differ significantly across groups, should I conclude that indirect effects do not differ significantly across groups?
(2) How can I test the above indirect effect in multi-group analyses?
Currently, my syntax is as below, where I set equality constraints (A) on the direct effect one by one.
grouping=urban
analysis:
type=missing;
iterations=50000;
model:
math on single pinvolv female sibling peduc;
pinvolv on single (A);
pinvolv on peduc;
single on peduc;
(where female, sibling, peduc are control variables) |
|
|
|
|
To test equality of indirect effects across groups you create the indirect effects in Model Constraint as
ind1 = a1*b1;
ind2 = a2*b2:
diff = ind1-ind2;
where ind1, ind2, and diff are NEW parameters and the a's and b's are labels given to the parameters in the MODEL command. A z-test is then produced for diff1 being different from zero. |
|
|
| Marie Alice posted on Wednesday, September 07, 2016 - 9:02 pm
|
|
|
Thank you for your quick response.
I was able to run the above to test whether the indirect effects differed significantly. Thanks much. |
|
|
|
|
ind1 = a1*b1;
ind2 = a2*b2:
diff = ind1-ind2;
Could you clarify for the "diff" means?
How should we interpret the diff? |
|
|
|
|
| Diff is the name we give the difference between the 2 ind estimates. If diff is significant, you can reject that 2 ind estimates are the same. |
|
|
|
|
Great thank you!
Additionally, in a multipe group analysis, if I constrain direct paths between the two groups to be equal to zero one at a time to check for significant difference... how do I do chi-square diff test if my unconstrained model is over identified and therefore does not report goodness-of-fit estimates? |
|
|
|
|
Additional question regarding the model constraints. How would we interpret it for categorical data, such as male and female students?
IND_W_SI 0.005 0.005 0.913 0.361
IND_O_SI -0.052 0.028 -1.843 0.065
DIFF_SIG 0.056 0.028 1.979 0.048
The difference is significant, but the indirect estimates are not. |
|
|
|
|
I don't understand the question in your first post.
For your second post, think of the distance between each estimate and zero versus the distance between the estimates. The former distance is smaller than the latter. |
|
|
|
|
Thank you for your reply.
I'll rephrase my first question. The model fit of my unconstrained multiple group model cannot be evaluated since it's over-identified. chi-square statistic, df, p value are all 0.000. so is RMSEA and so on.
When I constrain a direct path to equality in the two groups, male and female, I get estimates of model fit.
How do I compare the two models, the unconstrained and the constrained, if I don't have model fit estimates for the unconstrained?
Regarding the second question, I should have clarified, these are not confidence interval values.
IND_W_SI 0.005 0.005 0.913 0.361
IND_O_SI -0.052 0.028 -1.843 0.065
DIFF_SIG 0.056 0.028 1.979 0.048
These are parameter estimates, so the last column are the p-values, the first column is the estimates, the second is S.E.
So the difference is significant, but the indirect estimates for both groups are not. |
|
|
|
|
The two models are compared by that model fit measure you get when constraining.
Regarding the second question, I did not think they were CIs. I just pointed out that the distance between each estimate and zero is smaller than the distance between the estimates, so it is natural that the former is insignificant and the latter significant. |
|
|
|
|
So based on your clarifications:
1. if ind1 is not significant, and ind2 and the diff are significant, we can say there is the influence of the moderator.
2. if ind1 and ind2 are not significant, but diff is significant, we cannot say there is an influence of the moderator.
3. Similarly, if ind1 and ind2 are significant, but diff is not significant, there is no influence of the moderator. |
|
|
|
|
I have one final query!
Supposing this is my syntax (multiple group moderated mediation):
Y ON M;
Y ON X;
M ON X;
MODEL BOY:
M ON X (BOYA);
Y ON X (BOYB);
MODEL GIRL:
M ON X (GIRLA);
Y ON X (GIRLB);
MODEL IND:
Y ON X1;
MODEL CONSTRAINT:
NEW (ind_b_xy ind_g_xy)
ind_b_xy = boya*boyb;
ind_g_xy = girla*girlb;
diff_xy = ind_b_xy - ind_g_xy;
MODEL TEST:
0 = boya - girla;
0 = boyb - girlb;
In this, the "diff" parameter obtained indicates whether the mediation is different between the two groups.
I would run the Walt test one by one and check the level of significance.
And the Wald test can tell us whether the difference between groups is in the first stage (X->M) or second stage (M-> Y).
Would this be a correct interpretation? |
|
|
|
|
Regarding your first post, I don't know how the moderation is modeled, so can't answer - better to send full output to Support along with your license number.
Regarding your second question, you say Y ON X (BOYB) but I think you want Y ON M (BOYB). But otherwise, the answer is yes. |
|
|
|
|
Great! Thanks.
How is the above input syntax different from the syntax given in your book (Example 3.29):
model:
liking on respappr (k1)
sexism;! (k2);
respappr on sexism;! (p1);
model zero:
liking on sexism (k20);
respappr on sexism (p10);
model one:
liking on sexism (k21);
respappr on sexism (p11);
model constraint:
new(k2diff p1diff);
k2diff = k21-k20;
p1diff = p11-p10;
I get different results from the syntax in the previous post vs this particular syntax. |
|
|
| Back to top |
