s*******y
发帖数: 2977 | 1 Can we test if a count data set follows negative-binomial distribution in
SAS? It seems we can check the over-dispersion of count data using deviance/
df, but this is for Poisson distribution only.
Any methods for negative-binomial distribution?
Thanks a lot in advance! |
A*******s
发帖数: 3942 | 2 i would also like to know the answer. seems there is no one-sample KS test,
or other distribution tests in SAS (except for normality test).
Do u need to compare NB with other alternative distributions? Then we can
use likelihood as a criterion.
I also think such tests, if exist, lack of sense in reality. How can u tell
the prob of a sample follows NB but not Poisson, zero-inflated Poisson or
zero-inflated NB? They look quite similar...
deviance/
【在 s*******y 的大作中提到】
: Can we test if a count data set follows negative-binomial distribution in
: SAS? It seems we can check the over-dispersion of count data using deviance/
: df, but this is for Poisson distribution only.
: Any methods for negative-binomial distribution?
: Thanks a lot in advance!
|
a*****3
发帖数: 601 | 3 did you mean zero-truncated/modified distribuation ?
, : or other distribution tests in SAS (except for normality test).
sense in reality. How can u tell: the prob of a sample follows NB but not
Poisson, zero-inflated Poisson or: zero-inflated NB? They look quite similar
...
【在 A*******s 的大作中提到】
: i would also like to know the answer. seems there is no one-sample KS test,
: or other distribution tests in SAS (except for normality test).
: Do u need to compare NB with other alternative distributions? Then we can
: use likelihood as a criterion.
: I also think such tests, if exist, lack of sense in reality. How can u tell
: the prob of a sample follows NB but not Poisson, zero-inflated Poisson or
: zero-inflated NB? They look quite similar...
:
: deviance/
|
A*******s
发帖数: 3942 | 4 different...
zero truncated means zero value is not valid in your assumption.
zero inflation means you have more zero values than theoretical models
people use ZT distribution plus delta distribution at zero to build a
mixture model, while zero inflated models contain count distribution (Poi/NB
) and delta.
similar
【在 a*****3 的大作中提到】
: did you mean zero-truncated/modified distribuation ?
:
: , : or other distribution tests in SAS (except for normality test).
: sense in reality. How can u tell: the prob of a sample follows NB but not
: Poisson, zero-inflated Poisson or: zero-inflated NB? They look quite similar
: ...
|
a*****3
发帖数: 601 | 5 受教了
NB
【在 A*******s 的大作中提到】
: different...
: zero truncated means zero value is not valid in your assumption.
: zero inflation means you have more zero values than theoretical models
: people use ZT distribution plus delta distribution at zero to build a
: mixture model, while zero inflated models contain count distribution (Poi/NB
: ) and delta.
:
: similar
|
F****3
发帖数: 1504 | |
y*****w
发帖数: 1350 | 7 If your count data does not include zero values, you can use PROC UNIVARIATE
with the HISTOGRAM statement to get goodness-of-fit test statistics for
several distributions, however those distributions do not include NB.
http://support.sas.com/documentation/cdl/en/procstat/63104/HTML
I have recently run a project in which I had to choose a best distribution
for my non-zero historical count data for the purpose of sample size
calculation. I applied the HISTOGRAM statement to compare GAMMA, LOGNORMAL,
and NORMAL distributions based on their goodness-of-fit test statistics as
well as the density curves. |
w*******9
发帖数: 1433 | 8 2.Do you mean likelihood ratio test? Are the models embeded?
3.You just give priority to the model in the null hypothesis.
,
tell
【在 A*******s 的大作中提到】
: i would also like to know the answer. seems there is no one-sample KS test,
: or other distribution tests in SAS (except for normality test).
: Do u need to compare NB with other alternative distributions? Then we can
: use likelihood as a criterion.
: I also think such tests, if exist, lack of sense in reality. How can u tell
: the prob of a sample follows NB but not Poisson, zero-inflated Poisson or
: zero-inflated NB? They look quite similar...
:
: deviance/
|
w*******9
发帖数: 1433 | 9 It's a good idea but the chisq stuff is quite sensitive to the bins you use.
UNIVARIATE
,
【在 y*****w 的大作中提到】
: If your count data does not include zero values, you can use PROC UNIVARIATE
: with the HISTOGRAM statement to get goodness-of-fit test statistics for
: several distributions, however those distributions do not include NB.
: http://support.sas.com/documentation/cdl/en/procstat/63104/HTML
: I have recently run a project in which I had to choose a best distribution
: for my non-zero historical count data for the purpose of sample size
: calculation. I applied the HISTOGRAM statement to compare GAMMA, LOGNORMAL,
: and NORMAL distributions based on their goodness-of-fit test statistics as
: well as the density curves.
|
w*******9
发帖数: 1433 | 10 the alternative of KS test is everything other than the null, so you can
imagine that the power is very low. However, if you do like it, you can
always construct the
sampling distribution of the KS statistics under NB by simulation. It should
be very quick with nowadays computer. Based on the sampling distribution,
you can get a valid test, but again, don't have high expectation with your
power.
deviance/
【在 s*******y 的大作中提到】
: Can we test if a count data set follows negative-binomial distribution in
: SAS? It seems we can check the over-dispersion of count data using deviance/
: df, but this is for Poisson distribution only.
: Any methods for negative-binomial distribution?
: Thanks a lot in advance!
|
s*********e
发帖数: 1051 | 11 theoretically speaking, if you want to compare 2 alternative but non-nested
models, you might use vuong test or its nonparametric alternative clarke
test.
of course, everything mentioned above also makes sense from empirical
standpoint.
deviance/
【在 s*******y 的大作中提到】
: Can we test if a count data set follows negative-binomial distribution in
: SAS? It seems we can check the over-dispersion of count data using deviance/
: df, but this is for Poisson distribution only.
: Any methods for negative-binomial distribution?
: Thanks a lot in advance!
|
s*********e
发帖数: 1051 | 12 here is a macro i wrote for vuong test. |
