[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: testparm
Richard Williams <Richard.A.Williams.5@ND.edu>
Re: st: testparm
Tue, 04 Mar 2008 09:52:02 -0500
At 04:02 AM 3/4/2008, Chiara Mussida wrote:
By way of analogy, suppose you had gender in your model, and its
effect was significant. Would you immediately leap to estimating
totally separate models for men and women? Probably not. The
significant effect of gender would tell you that the intercepts
differ for men and women, but it would not tell you whether the slope
coefficients for the independent variables differ by gender.
sorry for my previous formatted messages, but the other e-mail address
didn't give me the opportunity to write in palin text and I didn't
know how it processes my e-mail...sorry again (Thanks Marteen).
my question today is: if the null of testparm is rejected by the data
(my 4 categories related to geographical regions of residence) can I
quietly proceed with the estimation of (4 in my case) separated models
for the duration of unemployment? or are there other procedures/test
all the best,
As Maarten notes, estimating separate models means you estimate a bunch of parameters, perhaps unnecessarily so. This reduces your likelihood of rejecting various null hypotheses even when you should reject.
Keep in mind, too, that in many cases people have all sorts of group characteristics they could focus on, e.g. race, gender, region... If you start estimating separate models for every group characteristic of interest (and separate models for every combination of those characteristics) you are going to get overwhelmed with parameters pretty quickly. Generally, you want as much parsimony as possible in your models.
In short, you have to break down and use theory to guide you at some point. If you have good reasons for believing the models will be very different across your groups, then estimating separate models may be a good idea. But, you wouldn't base such a decision simply on the finding that the intercepts differ across groups.
I'm a little more positive than Maarten about conducting tests to help you make these decisions. A series of nested models, guided by theory, can be very helpful in deciding how to proceed (as opposed to, say, a mindless stepwise regression approach.) They can be particularly helpful if you need to justify the decision NOT to estimate separate models by groups, e.g. you can show that, say, other than the intercepts, model coefficients do not significantly differ by gender; or that you maybe you only need one or two interaction terms (which your brilliant theory had previously identified) as opposed to zillions of interactions. If there are going to be dueling perspectives on how to proceed (e.g. a reviewer says you should have estimated separate models for each group, or conversely, that you shouldn't have) it is nice to have both theoretical arguments and empirical evidence to defend what you did.
I outline some procedures for testing group differences in
Richard Williams, Notre Dame Dept of Sociology
OFFICE: (574)631-6668, (574)631-6463
* For searches and help try: