testing independent of data assumption, before doing Poisson Regression I have got a set of data and decided to do a Poisson regression so $log(\lambda)=X\beta$. 
I wonder if there is any test available for such purpose?
The first thought is to use the Durbin-Watson test statistics to test the independence of data.
My primary concern is that since the mean and variance is not the same for each observation so if the Durbin-Watson statistic is correct?
If not, is there any suggestions?
Thank you!
 A: I believe that you are correct that the regular cut-off for the DW statistic would not be appropriate, but you could find a new cutoff using a permutation test.
If the values are independent (given the model and predictors) then the ordering of the points will not matter, so under the null hypothesis of independence you can randomly permute the order of the points (response and predictors permuted together so the model, predictions, residuals, etc. are the same).  So permute the data a bunch of times (1000 or so) and compute the DW statistic (or autocorrelation on the Pearson residuals, or other measure) for each permutation.  Now compare the statistic for the data in the original order to the distribution of permutations statistics to make your decision.
A: My comment got too long for a comment, so I've made an answer:
I made one suggestion in the comment you replied to. 
My other suggestion would have been roughly similar to @GregSnow's answer, though I might have suggested bootstrapping. 
There are poisson time series-type models you could fit; here are some references:
http://www.econ.ucdavis.edu/faculty/cameron/research/CTE01preprint.pdf
gives an overview of count data regression models, which has a section on time series data (sec. 5), and their book has a chapter (ch. 7) on it.
http://www.stat.columbia.edu/~rdavis/lectures/montreal.pdf
Gives a sequence of steps for estimating a GLM on time series of counts
Somne other potentially useful references:
M. J. Campbell (1994)
Time Series Regression for Counts: An Investigation into the Relationship between Sudden Infant Death Syndrome and Environmental Temperature
Journal of the Royal Statistical Society. Series A 
Vol. 157, No. 2, pp. 191-208
http://www.utdallas.edu/~pbrandt/pests/apsa99sc.pdf
If you fit a GLM with a simple time series component where the time component doesn't pick up much, you can argue you don't need it.
