Re: st: Relative Importance of predictors in regression

[Lucas <lucaselastic@gmail.com>]

Agreed!

On Wed, Nov 6, 2013 at 12:47 PM, Marcello Pagano
<pagano@hsph.harvard.edu> wrote:
>
> I think we have belabored this point sufficiently.
> I suspect that further discussion will not elicit any more light.
> Thanks to all who partook in the discussion.
> Let us agree to disagree.
>
> m.p.
>
>
> On 11/6/2013 2:52 PM, Lucas wrote:
>>
>> On Rich's point, of course if we estimate:
>>
>> 0. y=b1*YrsSchl + b2*Male
>> 1. y=g1*YrsSchl + g2*Male + g3*white
>> 2. y=h1*YrsSchl + h2*Male + h3*white + h4*age
>>
>> we would not expect b1=g1=h1 necessarily.  This has nothing to do with
>> whether we have "held constant" the variables that are in model 0 when
>> we are interpreting b1.
>>
>> On William's point, yes, the data has men and women in the example,
>> else no expected values could be obtained for Y3 and Y4 in the
>> example. So, paraphrasing William, he says, "You have adjusted your
>> estimates for gender." Given that claim, what is the correct
>> interpretation of b1 in model 0 above?  Sounds like you'd say "b1 is
>> the difference in Y associated with a one year difference in YrsSchl,
>> once the association between Y and sex has been accounted for." So,
>> basically, this phrasing reduces to "b1 tells us the association once
>> we hold constant all the other variables in the model, i.e.,
>> differences in those variables DO NOT EFFECT our estimate of b1."
>> [Note: if we interacted sex and education this interpretation would be
>> inappropriate]. This is what I people mean when they say "held
>> constant."
>>
>> It is interesting that there are varying interpretations of David H.'s
>> point, which suggests his point escapes some and perhaps many.  I
>> wonder if the formula he mentioned would clarify everything.
>>
>> Sam
>>
>> On Wed, Nov 6, 2013 at 11:37 AM, Richard Goldstein
>> <richgold@ix.netcom.com> wrote:
>>>
>>> Hi Sam,
>>>
>>> a little more seriously, consider the following two models:
>>>
>>> 1. y=b0 + b1*age + b2*female + b3*white
>>> 2. y=b0 + b1*age + b2*female
>>>
>>> so, there is no reason to expect that either b1 or b2 would be the same
>>> in these two models -- that I think is (part of) David's point
>>>
>>> I don't understand the "hold constant" part and how it might apply here,
>>> or, really elsewhere when talking about the "effect" of a
>>> right-hand-side variable; but I don't think that is what you are talking
>>> about; so, I think that at least part of this discussion has people
>>> talking past each other. Further, I don't think that this discussion is
>>> related to the subject line either.
>>>
>>> Rich
>>>
>>> On 11/6/13, 2:22 PM, Lucas wrote:
>>>>
>>>> Hi Rich,
>>>>
>>>> Depends on which of us you ask.  I'd say if you compare a male w/ 9
>>>> YrsSchl and a male w/ 8YrsSchl you've held sex constant and b1 is the
>>>> difference in Y associated with that one year difference in schooling.
>>>>   I think David H. would say that you've held nothing constant.  Is
>>>> that a correct interpretation of your claim, David H.?
>>>>
>>>> Sam
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/