Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, statalist.org is already up and running.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: RE: Poisson regression with score/scale as DV
Nick Cox <email@example.com>
Re: st: RE: Poisson regression with score/scale as DV
Tue, 3 Apr 2012 16:44:29 +0100
On Tue, Apr 3, 2012 at 4:28 PM, Austin Nichols <firstname.lastname@example.org> wrote:
> Nick et al.--
> The ordered *it does not help combining, but could be a good way to
> learn more about the data. If you have 9 activities measured with 4
> intensities, and you find that X has a large effect on the sum, that
> is one finding of dubious import; suppose you find that X has a very
> large impact on one of the 9 activities (with -oprobit- say) and none
> at all on the other 8: the impact on the sum is still there but who
> would care when you have identified the source of the impact? Clearly
> the mvreg version of ordered *it is a good complement to the summed
> outcome; ssc desc cmp for one useful user-written contribution.
> On Tue, Apr 3, 2012 at 6:36 AM, Nick Cox <email@example.com> wrote:
>> I like to agree with people, but I don't think we agree much at all.
>> What Clifton is doing is addition -- that's mathematically well
>> defined. It is a scientific or substantive issue on whether the
>> operation is invalid -- I would rather say "dubious" -- because you
>> are adding arbitrary scores. This divides the purists and the
>> pragmatists and always will. Any measurement scale comes with caveats
>> and cautions and complications. Many of us deal with binary scales
>> such as male/female but we all know that it can be more complicated
>> than that in practice for some fraction of people.
>> I don't think ordinal *it will help with Clifton's desire to combine
>> the scales.
>> On Tue, Apr 3, 2012 at 11:23 AM, Reinhardt Jan Dietrich
>> <firstname.lastname@example.org> wrote:
>>> I certainly agree with Nick, at least to a large degree ... If you dichotomize you will of course loose information as well and just adding everything up should not be done after inquiring into the dimensionality of the data (for instance with pca if you don't have any idea or confa if you have a hypothesis on this). For sure summing up ordinal items is meaningless because of different reference standards people will apply. Most reviewers would critisize this as an operation that is mathematically not feasible either ...
>>> What you could also do is apply some ordinal probit or logit model to your problem ...
>>> -----Original Message-----
>>> From: email@example.com [mailto:firstname.lastname@example.org] On Behalf Of Nick Cox
>>> Sent: Dienstag, 3. April 2012 11:35
>>> To: email@example.com
>>> Subject: Re: st: RE: Poisson regression with score/scale as DV
>>> Jan and I are not bound to agree. I don't agree with the argument that
>>> it's clearly OK to count yes-no answers to different questions but
>>> clearly not OK to add graded answers to different questions.
>>> Scores will never satisfy measurement purists, but the job of the data
>>> analyst is to squeeze the juice out of never-ideal data, not to
>>> pontificate about perfect oranges that would yield perfect juice.
>>> In this case, one obvious difficulty is whether (e.g.) my "Often" is
>>> equivalent to anybody else's, let alone everybody else's, but
>>> dichomotimising the scale would not remove that difficulty.
>>> It's unclear whether piling up just refers to skewness or you have
>>> zero inflation too.
>>> On Tue, Apr 3, 2012 at 10:09 AM, Clinton Thompson
>>> <firstname.lastname@example.org> wrote:
>>>> Many thanks for the replies, Jan & Nick. As for the suggestion to
>>>> create a sum index based on the dichotomization of the ordinal
>>>> variables, I must admit that I'm unsure of how/why this would be
>>>> superior to the current index. In my situation, the score follows
>>>> from the summing of nine composite questions about the frequency with
>>>> which a person engages in an activity where each composite question
>>>> has four responses ("Never", "Rarely", "Sometimes", "Often"). The
>>>> corresponding values for the responses are [0,3]. Maybe I don't yet
>>>> understand the intricacies of the Poisson distribution but re-scaling
>>>> the component questions from [0,3] to [0,1] will just re-scale the
>>>> score variable from [0,27] to [0,9], which still leaves me w/ a
>>>> bounded DV with a pile-up of responses at zero. Either way (and if I
>>>> understand both of you), it sounds like Poisson is a reasonable way to
>>>> model this variable/response?
>>>> Nick -- I hadn't considered -glm, f(binomial)- but I'll look further
>>>> into it. (And thanks for correcting my reference to Austin Nichols'
>>>> presentation. My spelling implied his last name is Nichol -- not
>>>> Nichols. Embarrassing mistake.)
>>>> Thanks again,
>>>> On Tue, Apr 3, 2012 at 10:43 AM, Nick Cox <email@example.com> wrote:
>>>>> Lots of social scientists agree with you, while lots of other social
>>>>> and other scientists spend most of the time doing precisely that.
>>>>> On Tue, Apr 3, 2012 at 9:07 AM, Reinhardt Jan Dietrich
>>>>> <firstname.lastname@example.org> wrote:
>>>>> ... Ordinal items should definitely not be summed up ...
> * For searches and help try:
> * http://www.stata.com/help.cgi?search
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
* For searches and help try: