Re: st: Is there a way to use or emulate the behaviour of --predict-- with --by-- groups?

[Nick Cox <njcoxstata@gmail.com>]

Good question. In fact, if you think regressing y on log x is
appropriate, then it's a matter of interpolating on log x and there is
no call for exponentiating. I didn't look carefully enough at your
regression, but my main question remains.

Nick

On Wed, Oct 31, 2012 at 6:35 PM, Aaron Kirkman <ak1795mailserv@gmail.com> wrote:
> Hi Nick,
>
> Just to clarify (for my benefit), do you mean using --ipolate-- on the
> log(y) and log(x) variables then exponentiating the new values of
> log(y)?
>
> Aaron
>
> On Wed, Oct 31, 2012 at 1:18 PM, Nick Cox <njcoxstata@gmail.com> wrote:
>> Did you think of interpolating the logs and then exponentiating?
>>
>> Regression of any flavour can't interpolate unless all the residuals
>> are zero; it can only smooth. If that's what you want, fine, but it is
>> not like -ipolate-.
>>
>> Nick
>>
>> On Wed, Oct 31, 2012 at 5:49 PM, Aaron Kirkman <ak1795mailserv@gmail.com> wrote:
>>> Thank you Maarten. The interaction term does exactly what I need. Once
>>> I added --set matsize 3000-- to my actual data set, the interpolation
>>> works well.
>>>
>>> On Wed, Oct 31, 2012 at 3:19 AM, Maarten Buis <maartenlbuis@gmail.com> wrote:
>>>> On Wed, Oct 31, 2012 at 1:40 AM, Aaron Kirkman wrote:
>>>>> I have data grouped by a variable called --group--, in this example,
>>>>> and I'm trying to use logarithmic interpolation on another variable.
>>>>> Linear interpolation using the --ipolate-- command works perfectly in
>>>>> --by-- group <snip> Unfortunately, this does not work with logarithmic
>>>>> interpolation and --regress--/--predict--.
>>>>
>>>> You can avoid the -by:- prefix by adding interactions to your linear
>>>> regression model:
>>>>
>>>> *---------- begin example ----------
>>>> clear
>>>> quietly input str1 group x y
>>>> A 1 1000
>>>> A 2 .
>>>> A 3 3000
>>>> A 4 .
>>>> B 5 45
>>>> B 6 .
>>>> B 9 20
>>>> end
>>>>
>>>> encode group, gen(groupnum)
>>>> gen lx = ln(x)
>>>>
>>>> reg y i.groupnum##c.lx
>>>>
>>>> predict y_loginterp, xb
>>>> *----------- end example -----------
>>>>
>>>> Notice that the observation A 4 is not strictly speaking an
>>>> interpolation but an extrapolation. You'll want to be more careful in
>>>> those situations.
*
*   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/