RE: st: RE: Re: Kenel density function

["Nick Cox" <n.j.cox@durham.ac.uk>]

OK. I'd call this a matter of coarse resolution of the data rather than
binning, which I tend to regard as grouping of original measurements
after data production. If this were my problem I think I would look at
quantile or cumulative distribution functions directly. 

Nick 
n.j.cox@durham.ac.uk 

Janet Hill

Sorry for my ineptitude explaining the data. The master and cast are
scanned and the scanner has a fixed increment of 0.25 mm (the smallest
that can be obtained) so I have binned data at these increments and I
wanted to see how 'reproducible' the scans were. Eventually, so I have
just found out, the experiment will extend to comparing 3 different
methods of of making the cast. So my final analysis is going to compare
3 methods when the outcome is a frequency distribution - if that is
possible.

Many thanks for your time.

--- On Tue, 29/6/10, Nick Cox <n.j.cox@durham.ac.uk> wrote:

> You can smooth binned data: the
> problem, if any, is in over-interpreting
> results. 
> 
> Where does the binning come from? Why aren't unrounded data
> available? 
 
> On Behalf Of Janet Hill
 
> Thank you for that. There is a single master and 6 separate
> impressions
> have been taken of this so they are uncoupled. The
> difference between
> each impression and the master has been measured and this
> gives the
> binned data.Ideally this should have a value of zero for
> perfect fit. I
> thought that I could use the kernel density on binned data
> but I
> obviously need to do some more reading.
> 
> Any advice would be gratefully received.
> 
> Janet
> --- On Tue, 29/6/10, Nick Cox <n.j.cox@durham.ac.uk>
> wrote:
> 
> > From: Nick Cox <n.j.cox@durham.ac.uk>

> > You should tell us more about the
> > design. Do these come in groups of six, or are the
> six
> > distributions entirely uncoupled from each other? It
> sounds
> > as if that information has been omitted from what you
> have;
> > if so, retrieving it is the first priority. 
> > 
> > Either way, trying to reduce all this to
> Kolmogorov-Smirnov
> > looks severely problematic to me. As the data are
> already
> > binned, kernel estimation also looks suspect. 
> > 
> > Nick 
> > n.j.cox@durham.ac.uk
> > 
> > 
> > Janet Hill
> > 
> > I would be grateful for some advice. I have some data
> from
> > an experiment in which the difference between a cast
> and the
> > master model has been measured. The data consists of
> the
> > frequency of difference from -8 to + 8 mm in steps of
> 0.25,
> > and 6 different casts were made. Kernel density plots
> show
> > obvious differences between the casts but is there any
> way I
> > can formally test this - or am I missing an obvious
> way to
> > compare the frequency distributions/ I had thought of
> using
> > Kolmogorov-Smirnov followed by multiple comparisons
> but I
> > was not sure how applicable it would be.
> > 
> > I am using Stata 11.1

*
*   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/