Dear Paul and Radstats, re NPR (Non-parametric regression)
I am very intrested to find out how this strand evolved. Paul wrote to say that the paper he was reading didn't actually use the kernel regression type of non-parametric regression after all. Instead it was probit regression.
This was just what I expected was going to be the outcome, because at the moment the use of kernel regression is still pretty rare. Instead, what some of us might casually have called non-parametric regression (but which I might like to call qualitative regression and certainly technically can be called non-linear regression) is where the dependent variable is of the binary, ordinal or probit scale type. In other words the regression isn't assumed to meet all the usual parametric assumptions, but only some of them.
Whereas in true non-parametric regression, hardly any parametric assumptions are made.
What I find interesting is that there are still numerous assumptions to be made even in NPR, e.g. whether to fit a curve or a line; what causes to allow for; how to handle endogeneity; and whether ordinality or cardinality is to be claimed for the resulting predictions. It would be so interesting to use NPR because of non-cardinal measurement, and then drop us back into cardinal interpretations at the end. It might be incoherent because of using conflicting assumptions at different stages.
Maybe in the coming year or so we can study an exemplar of NPR to get a better feel for the new methods.
Yours
Wendy
Wendy Olsen
Senior Lecturer in Socio-Economic Research
Cathie Marsh Centre for Census & Survey Research
and Inst. for Development Policy and Management
Univ. of Manchester
Manchester M13 9PL
tel 0044-161-275-3043
web www.ccsr.ac.uk/staff/wo.htm
See also
www.humanities.manchester.ac.uk/socialchange
-----Original Message-----
From: Paul Spicker <[log in to unmask]> [mailto:[log in to unmask]]
Sent: 28 June 2007 12:41
To: Wendy Olsen
Subject: Re: Non-parametric regression
Thanks. Stephen McKay's reference was helpful for getting to grips with the range of methods covered. I found another paper at http://www.maths.bris.ac.uk/~maxak/vorau.pdf which suggests that outliers and discontinuities might be a problem, but I can't judge it very effectively, and I can't tell how big the effects are likely to be.
As it's turned out, though, I don't know whether I need to plunge much further in. I was reviewing a paper for a project which claimed to be performing a multivariate nonparamateric regression. I couldn't make sense of the figures at all, and I was wondering where the residuals were. Now that I've got a clearer idea what I'm looking for, I realise that I've been led down the garden path; the paper isn't multivariate at all, it's a series of bivariate probits that's been mis-labelled, and that much I can cope with.
Paul Spicker
******************************************************
Please note that if you press the 'Reply' button your
message will go only to the sender of this message.
If you want to reply to the whole list, use your mailer's
'Reply-to-All' button to send your message automatically
to [log in to unmask]
Disclaimer: The messages sent to this list are the views of the sender and cannot be assumed to be representative of the range of views held by subscribers to the Radical Statistics Group. To find out more about Radical Statistics and its aims and activities and read current and past issues of our newsletter you are invited to visit our web site www.radstats.org.uk.
*******************************************************
|