[R] How to fit a line through the "Mountain crest", i.e., through the highest density of points - in a "loess-like" fashion.

[Emmanuel Levy]

Hi,

Thanks a lot for your reply - I posted a second message where I
provide a "dummy" example, entitled
"How to improve the robustness of "loess"? - example included".

I need to fit a curve which makes it a bit difficult to work with kde2d only.

I'm actually trying to use kde2d in combination with loess - basically
I give the output density of kde2d as weights in the "loess" function.
It seems to give nice results :)

In my second post I wrote that the "weight" option did not work but
that's because I was writing "weigth" - not sure why I did not get an
error message.

I'll post the lines of code as a reply to the second post.

All the best,

Emmanuel




On 10 March 2012 19:46, David Winsemius <dwinsemius at comcast.net> wrote:
>
> On Mar 10, 2012, at 3:55 PM, Emmanuel Levy wrote:
>
>> Hi,
>>
>> I'm trying to normalize data by fitting a line through the highest density
>> of points (in a 2D plot).
>> In other words, if you visualize the data as a density plot, the fit I'm
>> trying to achieve is the line that goes through the "crest" of the
>> mountain.
>
>
> Are you familiar with the kde2d  of bkde2D functions in various packages? If
> you then collected the max density for each X and Y you might want to see
> whether that 2-d function would follow a sufficiently regular path that
> would represent the projection of the ridge on the z=0 plane.
>
>
>>
>> This is similar yet different to what LOESS does.
>
>
> Do you want a curve or a line?
>
>
>> I've been using loess
>> before, but it does not exactly that as it takes into account all points.
>> Although points farther from the fit have a smaller weight, they result in
>> the fit being a bit off the crest.
>>
>> Do you know a package or maybe even an option in loess that would allow me
>> achieve this?
>
>
> I don't. I happen to have a dataset where I could test it. But you are
> likely to get better responses if you provide a test case.
>
>> Any advice or idea appreciated.
>>
>> Emmanuel
>>
>>        [[alternative HTML version deleted]]
>
>
> Plain text is preferred.
>
> --
>
> David Winsemius, MD
> West Hartford, CT
>