[R] Kolmogorov Smirnov Test
Greg Snow
Greg.Snow at imail.org
Thu Nov 11 18:05:18 CET 2010
Consider the following simulations (also fixing the pnorm instead of dnorm that Ted pointed out and I missed):
out1 <- replicate(10000, {
x <- rnorm(1000, 100, 3);
ks.test( x, pnorm, mean=100, sd=3 )$p.value
} )
out2 <- replicate(10000, {
x <- rnorm(1000, 100, 3);
ks.test( x, pnorm, mean=mean(x), sd=sd(x) )$p.value
} )
par(mfrow=c(2,1))
hist(out1)
hist(out2)
mean(out1 <= 0.05 )
mean(out2 <= 0.05 )
In both cases the null hypothesis is true (or at least a meaningful approximation to true) so the p-values should follow a uniform distribution. In the case of out1 where the mean and sd are specified as part of the null the p-values are reasonably uniform and the rejection rate is close to alpha (should asymptotically approach alpha as the number of simulations increases). However looking at out2, where the parameters are set not by outside knowledge or tests, but rather from the observed data, the p-values are clearly not uniform and the rejection rate is far from alpha.
--
Gregory (Greg) L. Snow Ph.D.
Statistical Data Center
Intermountain Healthcare
greg.snow at imail.org
801.408.8111
> -----Original Message-----
> From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-
> project.org] On Behalf Of Kerry
> Sent: Thursday, November 11, 2010 12:02 AM
> To: r-help at r-project.org
> Subject: Re: [R] Kolmogorov Smirnov Test
>
> Thanks for the feedback. My goal is to run a simple test to show that
> the data cannot be rejected as either normally or uniformally
> distributed (depening on the variable), which is what a previous K-S
> test run using SPSS had shown. The actual distribution I compare to my
> sample only matters that it would be rejected were my data multi-
> modal. This way I can suggest the data is from the same population. I
> later run PCA and cluster analyses to confirm this but I want an easy
> stat to start with for the individual variables.
>
> I didn't think I was comparing my data against itself, but rather
> again a normal distribution with the same mean and standard deviation.
> Using the mean seems necessary, so is it incorrect to have the same
> standard deviation too? I need to go back and read on the K-S test to
> see what the appropriate constraints are before bothering anyone for
> more help. Sorry, I thought I had it.
>
> Thanks again,
> kbrownk
>
> On Nov 11, 12:40 am, Greg Snow <Greg.S... at imail.org> wrote:
> > The way you are running the test the null hypothesis is that the data
> comes from a normal distribution with mean=0 and standard deviation =
> 1. If your minimum data value is 0, then it seems very unlikely that
> the mean is 0. So the test is being strongly influenced by the mean
> and standard deviation not just the shape of the distribution.
> >
> > Note that the KS test was not designed to test against a distribution
> with parameters estimated from the same data (you can do the test, but
> it makes the p-value inaccurate). You can do a little better by
> simulating the process and comparing the KS statistic to the
> simulations rather than looking at the computed p-value.
> >
> > However you should ask yourself why you are doing the normality tests
> in the first place. The common reasons that people do this don't match
> with what the tests actually test (see the fortunes on normality).
> >
> > --
> > Gregory (Greg) L. Snow Ph.D.
> > Statistical Data Center
> > Intermountain Healthcare
> > greg.s... at imail.org
> > 801.408.8111
> >
> >
> >
> > > -----Original Message-----
> > > From: r-help-boun... at r-project.org [mailto:r-help-bounces at r-
> > > project.org] On Behalf Of Kerry
> > > Sent: Wednesday, November 10, 2010 9:23 PM
> > > To: r-h... at r-project.org
> > > Subject: [R] Kolmogorov Smirnov Test
> >
> > > I'm using ks.test (mydata, dnorm) on my data. I know some of my
> > > different variable samples (mydata1, mydata2, etc) must be normally
> > > distributed but the p value is always < 2.0^-16 (the 2.0 can change
> > > but not the exponent).
> >
> > > I want to test mydata against a normal distribution. What could I
> be
> > > doing wrong?
> >
> > > I tried instead using rnorm to create a normal distribution: y =
> rnorm
> > > (68,mean=mydata, sd=mydata), where N= the sample size from mydata.
> > > Then I ran the k-s: ks.test (mydata,y). Should this work?
> >
> > > One issue I had was that some of my data has a minimum value of 0,
> but
> > > rnorm ran as I have it above will potentially create negative
> numbers.
> >
> > > Also some of my variables will likely be better tested against non-
> > > normal distributions (uniform etc.), but if I figure I should learn
> > > how to even use ks.test first.
> >
> > > I used to use SPSS but am really trying to jump into R instead, but
> I
> > > find the help to assume too heavy of statistical knowledge.
> >
> > > I'm guessing I have a long road before I get this, so any bits of
> > > information that may help me get a bit further will be appreciated!
> >
> > > Thanks,
> > > kbrownk
> >
> > > ______________________________________________
> > > R-h... at r-project.org mailing list
> > >https://stat.ethz.ch/mailman/listinfo/r-help
> > > PLEASE do read the posting guidehttp://www.R-project.org/posting-
> > > guide.html
> > > and provide commented, minimal, self-contained, reproducible code.
> >
> > ______________________________________________
> > R-h... at r-project.org mailing
> listhttps://stat.ethz.ch/mailman/listinfo/r-help
> > PLEASE do read the posting guidehttp://www.R-project.org/posting-
> guide.html
> > and provide commented, minimal, self-contained, reproducible code.
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-
> guide.html
> and provide commented, minimal, self-contained, reproducible code.
More information about the R-help
mailing list