[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: RE: RE: Residuals in svy:intreg
Nick Cox <firstname.lastname@example.org>
Re: st: RE: RE: Residuals in svy:intreg
Fri, 1 Jun 2012 09:55:08 +0100
I can't add much to my previous answer. Clearly you have in mind that
censoring complicates things. But what does "OK" mean here? That it is
technically correct? That you cannot be misled? That someone in
authority might object or assert that you are wrong? I think it
depends what you are doing, in what circumstances, and who is judging
on what criteria.
For example, I think it can be helpful to look at residual plots.
Frequently they don't help, but when they do, it is in an important
Some economists (in particular) seem to object to anything that is not
a formal test as arbitrary, subjective and lacking in rigour. So, they
don't seem to do things like that.
But I can see that any formal procedure that treats residuals from an
interval regression and ignores their origin is likely to be suspect.
On Fri, Jun 1, 2012 at 9:43 AM, Ángel Rodríguez Laso
> Then, would it be OK for individuals with censored variables to
> consider that their observed values are those where censoring took
> place? Shouldn't there be some 'allowance' for error in the observed
> values, as I suppose there is when calculating the interval
> Thank you very much.
> Angel Rodriguez-Laso
> 2012/6/1 Nick Cox <email@example.com>:
>> I think there are two levels to this.
>> For informal (e.g. graphical) analysis, nothing is fundamentally
>> different, so that absence of pattern is good news and presence of
>> pattern may make you think about whether the model can be improved.
>> For formal analysis, I don't know of any published procedures either,
>> but if you invented your own, simulation may be the most practical way
>> to establish their properties.
>> On Fri, Jun 1, 2012 at 8:49 AM, Ángel Rodríguez Laso
>> <firstname.lastname@example.org> wrote:
>>> Dear Nick,
>>> Thank you for your answer.
>>> Unfortunately, after intensive search in the web I haven´t been able
>>> to find any document on the use of residuals in interval regression or
>>> the checking of assumptions of interval regression in the survey
>>> setting. Of the two references that Stata manual v11 gives for an
>>> introduction to interval regression (Wooldridge, J. M. 2009.
>>> Introductory Econometrics: A Modern Approach. 4th ed. Cincinnati, OH:
>>> South-Western. Davidson, R., and J. G. MacKinnon. 2004. Econometric
>>> Theory and Methods. New York: Oxford University Press.), I've only had
>>> access to Wooldridge's and it does not say anything on how to use
>>> residuals in this context.
>>> I suppose the difficult part in calculating (observed-predicted
>>> values) is assigning values from which censoring takes place as
>>> observed values for individuals with censored data.
>>> Angel Rodriguez-Laso
>>> 2012/5/31 Nick Cox <email@example.com>:
>>>> I doubt that this was the question, but I am assuming here that if you want residuals, then it's just an extra line calculating the residuals as difference between observed and predicted.
>>>> -----Original Message-----
>>>> From: firstname.lastname@example.org [mailto:email@example.com] On Behalf Of Nick Cox
>>>> Sent: 31 May 2012 11:09
>>>> To: 'firstname.lastname@example.org'
>>>> Subject: st: RE: Residuals in svy:intreg
>>>> Still true in 12.1.
>>>> I would guess rather that as residuals need some careful interpretation with -intreg-, StataCorp lets users make their own decisions about working with them.
>>>> If you can refer us to literature defining and using residuals carefully for -intreg-, that would strengthen the case for adding them to official Stata.
>>>> Ángel Rodríguez Laso
>>>> I'm working with Stata 9.2 for Windows.
>>>> I have to carry out an interval regression with survey data, because
>>>> there are top and bottom censored values. I've noticed Stata version
>>>> 9.2 does not provide residuals for this model. It calculates predicted
>>>> values, but if it does not provide (observed-predicted values), there
>>>> must be a good reason.
>>>> I understand that, because I'm in a survey environment, I do not have
>>>> to check for homoskedasticity of residuals and that they are not
>>>> expected to be independent. But residuals would still be useful to
>>>> check for model lack of fit (nonlinearity and presence for influential
>>>> points and outliers). Do you know of any alternatives?
>> * 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:
> * 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: