Dear Mark,
Thanks very much for your answer!
Thanks to Anderson's addition too.
Cheers,
Yuwen
On 30 October 2017 at 05:19, Mark Jenkinson <[log in to unmask]>
wrote:
> Hi,
>
> You can do something like this but you need a separate EV for each
> outlier. For example, if you have 5 outliers then you need 5 separate
> EVs, each of which is zero for all entries except one as each EV can
> only remove the effect of a single outlier (trying to put any more
> into one EV makes an assumption that the values of the outliers are
> exactly equal to each other, which is clearly not true).
>
> All the best,
> Mark
>
>
> > On 27 Oct 2017, at 15:22, Yuwen Hung <[log in to unmask]> wrote:
> >
> > Dear Anderson,
> > If I want to control for the effect of outlier data before running
> > glm,
> I was thinking whether I could, in a new EV, put the outlier subjects
> as 1 and all others as 0, and this EV in the contrast matrix would be 0 effect.
> Please let me know? Thank you very much for any help!
> >
> > Best,
> > Yuwen
> >
Date: Wed, 1 Nov 2017 17:35:10 -0400
From: "Anderson M. Winkler" <[log in to unmask]>
Subject: Re: modeling outlier in glm
Hi Yuwen,
Just adding a bit of extra information to MJ's answer: need also one exchangeability block for the outliers, and another for the non-outliers, so that residuals of the outliers (all identical to zero) aren't shuffled with the residuals of the non-outliers, as these have different distributions.
All the best,
Anderson
|