Probability of a Greater Chi-Square Value vs Chi-p

43 views
Skip to first unread message

Murray Lantner

unread,
Nov 26, 2004, 4:29:09 PM11/26/04
to distance...@googlegroups.com
Hello:

I had a question with the interpretation of the results pertaining to the
Probability of a Greater Chi-Square Value and the Chi-p.

My understanding is that if the Probability of a Greater Chi-Square Value
(P) is small then the model is a poor fit to the data . Conversely if
the Probability of a Greater Chi-Square Value is high then the model is a
good fit to the data. So P=0.11990 is a better fit to the data than
P=0.00839. (I know neither is a good fit)

However at the end of the results printout in Distance there is a number
for the probability for chi-square goodness of fit test Chi-p = 0.00839.
Why does the DISTANCE summary say that the Chi-p = 0.00839 and not Chi-p
= 0.11990.

Below are the results from the runs in the attached document:

Total Chi-square value = 7.3203 Degrees of Freedom = 4.00
Probability of a greater chi-square value, P = 0.11990

Total Chi-square value = 16.2965 Degrees of Freedom = 7.00
Probability of a greater chi-square value, P = 0.02254

Total Chi-square value = 15.2357 Degrees of Freedom = 6.00
Probability of a greater chi-square value, P = 0.01850

Total Chi-square value = 27.7228 Degrees of Freedom = 13.00
Probability of a greater chi-square value, P = 0.00989


Total Chi-square value = 25.2454 Degrees of Freedom = 11.00
Probability of a greater chi-square value, P = 0.00839


The Chi-p- probability for chi-square goodness-of-fit test seen in the
results was Chi- p=0.00839


Please let me know
Thanks
Murray

--
distance-sampling list home page:
http://www.jiscmail.ac.uk/lists/distance-sampling.html
To unsubscribe, go to the home page and click "Join or leave the list",
or email jisc...@jiscmail.ac.uk with the following message body:
leave distance-sampling
Before posting to the list, read the list guidelines at
http://www.jiscmail.ac.uk/files/distance-sampling/
distance-sampling.welcome
Any questions, email the list owner at
distance-sam...@jiscmail.ac.uk
--

Len Thomas

unread,
Nov 29, 2004, 4:57:45 AM11/29/04
to distance...@googlegroups.com
Hi Murray, list,

Firstly, you can't send attachments to the list, which is why yours didn't
appear. This is to prevent viruses being spread by the list.

Nevertheless, it's still possible to make a good guess at the answer to your
question. Distance often does several chi-sq tests on the same set of data,
and the one it reports in the summary page is the last test it did.

Why does it do more than one test?

- Firstly, if you don't specify your own cutpoints for exact data, Distance
produces 3 plots and 3 sets of chi-sq tests, with a small, medium and large
number of intervals. (The exact rule for choosing number of intervals is n^
0.5, 2/3n^0.5 and 3/2n^0.5 where n is the number of observations.)

- Secondly, the method of calculating a chi-sq value is known to give a poor
approximation of the chi-sq distribution when the expected cell counts are
small (see any stats book). So, if any of the intervals has an expected
value less than 2, Distance will do a second test where it pools adjacent
intervals until all intervals have values greater than 2. It starts pooling
from the 1st interval downwards (only useful for point transect data, where
the expected count in the first interval can be small), and from the last
interval upwards. These pooled intervals are then used to perform a second
test, which is shown after the first one in the Chi-sq GOF test page of
results.

Note that, for exact data, Distance now provides Kolmogorov-Smirnov and
Craemer-von Mises tests. These are better as they avoid the need to pool
the data. Other things being equal, the uniform weighting CvM test should
be more powerful to detect departures from the fitted model than the KS
test, so that's the one I'd quote. The cosine weighting CvM should be
better at detecting departures close to 0 distance (see online help for more
on this) - but we haven't evaluated it by simulation yet, and I've found the
two to give very similar results.

Hope this helps,

- Len



--
Len Thomas l...@mcs.st-and.ac.uk http://www.creem.st-and.ac.uk/len/
Centre for Research into Ecological and Environmental Modelling
The Observatory, University of St Andrews, Scotland KY16 9LZ
Tel. (0)1334-461801 Fax. (0)1334-461800 Secretary (0)1334-461829
Reply all
Reply to author
Forward
0 new messages