Q: (Different slopes in different groups):
I am fitting the same model to two samples
-- for example, a sample from the US, and a sample from Sweden.
It looks to me as though the slopes
for the US are different from those for Sweden. But how I can I
formally test this?
A: There are couple of popular approaches. These should give
roughly the same result. I'd be interested in hearing about cases where
they don't.
- One approach involves comparing the slopes directly. For a
particular variable X, suppose that b1 and se1 are the estimated slope
and standard error for the US, while b2 and se2 are the estimated
slope and standard error for Sweden. The test statistic is Z = (b1-b2)
/ (se12+se22)1/2, which is very
similar to the statistic for comparing two means. Under the null
hypothesis of equal population slopes, Z has a standard normal
distribution in large samples. So in large samples, you might reject
the null hypothesis if |Z|>2 or so. [Note that some researchers have
used a different, erroneous formula for Z (see Paternoster et al 1998
for a discussion).]
- Another approach involves interactions.Fit a model that combines
both groups. In addition to variables from
the separate models, the combined model will include a dummy variable
representing
group membership--for example, a dummy variable that is 1 in Sweden and
0 in the US.
Also include interactions between this dummy variable and all the other
variables in
the model. This allows for the possibility that the slopes are
different in the different
groups. To test whether a particular slope is different, look
at the p
values for the interaction terms.
To test whether any slopes are different, conduct a joint test
of all the interactions at
once. SAS and STATA have commands for such joint tests. In an OLS
model, a
joint test of interactions is
equivalent to the "Chow test; this general approach to testing complex
hypotheses in generalized linear models is discussed
in chapter 4 of Long (1997).
References
Long, J.Scott. 1997.
Regression Models for Categorical and Limited Dependent Variables.
Thousand Oaks, CA: Sage.
Paternoster, R.; Brame, R.; Mazerolle, P.; Piquero, A. 1998. "Using
the Correct Statistical Test for the Equality of Regression
Coefficients." Criminology
36(1), 859-866.