Trying to understand the Wald test

[Bill MacIntosh]

Hi all,

I am struggling to interpret the results of my Wald test.

I am running four logit models, all with the same independent variables but they differ by the variable SplitH. My code looks like this:

Logit MiGroup3_dv generations edu eco religion if SplitH==1
est store h1
Logit MiGroup3_dv generations edu eco religion if SplitH==2
est store h2
Logit MiGroup3_dv generations edu eco religion if SplitH==3
est store h3

MiGroup3_dv represents feelings towards immigration, with 1 denoting negative feelings and 0 denoting positive feelings.
SplitH represents regions of origin for immigrants, with 1=west Europe, 2=east Europe, and 3=Africa
I’ve then produced a suest table:

Suest h1 h2 h3

In this table, the coefficient for edu when SplitH=1 is statistically significant, but it is not when SplitH = 2 or 3.

I have used a Wald test to investigate this further:
test[h1_MiGroup3_dv]edu = test[h2_MiGroup3_dv]edu
test[h1_MiGroup3_dv]edu = test[h3_MiGroup3_dv]edu
test[h2_MiGroup3_dv]edu = test[h3_MiGroup3_dv]edu

however, none of the chi2 values achieved statistical significance. Does this mean the value of SplitH (and therefore the region of origin of the immigrant) has no effect on my DV?

Also, apologies for the horrible layout, I am yet to fully understand dataex sufficiently.

Many thanks.

[Joseph Coveney]

. . . the coefficient for edu when SplitH=1 is statistically significant, but it is not when SplitH = 2 or 3.

I have used a Wald test to investigate this further:

test [h1_MiGroup3_dv]edu = [h2_MiGroup3_dv]edu
. . .

however, none of the chi2 values achieved statistical significance. Does this mean the value of SplitH (and therefore the region of origin of the immigrant) has no effect on my DV?

Google the difference between significant and not significant is not itself statistically significant and take a look at the first hit.

Have you considered including interactions terms in an omnibus model in lieu of suest?

[George Ford]

Agreed. I'd try to get it all in one equation using interactions.

Your results suggest that edu has a well estimated coefficient for Split1, but not so much for the others. So, you might say edu is important for split 1 (if the coef is of meaningful size), but not the others. Education affects feelings for Western Europeans but not the others.

The test across models is for equality of the coefficients. Most likely, those do not reject because the Split2 and 3 coefficients have large standard errors. (A well estimated coef with a conf interval of 0.1 to 0.15 is no different than a coefficient with a conf interval of -10 to +10).