Durbin Watson critical values

[bbmra]

Hi,

Is there an API that can calculate the upper and lower critical values of Durbin Watson test?  I am using smt.durbin_watson(residuals) to compute Durbin Watson test value. With that value alone, I can not reject/accept or be inconlusive about that test. It seems that people use Durbin-Watson significance table to look up the critical values. But in my case number of sample points and number of regressors varies greatly.  Manually looking up a table is not practical. Does anyone have some insights to share?

Thank you in advance.

[josef...@gmail.com]

No, we don't have any code for critical values of Durbin Watson.

There are some tables available which could be used, but I've never seen them in python.

However, I would recommend the other autocorrelation tests which are more robust and those have critical values.

http://statsmodels.sourceforge.net/devel/stats.html#residual-diagnostics-and-specification-tests

ljungbox which is a test on the sum of autocorrelations and breusch_godfrey which is a Lagrange multiplier test.

Note, I misspelled `breusch` in the names of two tests (missing "c").

My impression is that DW is used mostly because of historical familiarity, and I was thinking of dropping it from the summary table.

Josef

[bbmra]

Thank you Josef. 

How about heteroskedasticity? Is goldfeld quant a good test for that?

[josef...@gmail.com]

I don't remember what the recommendations where, but I think breusch-godfrey is the best and most robust for different cases, assumptions

for example google found this

http://stats.stackexchange.com/questions/148004/testing-for-autocorrelation-ljung-box-versus-breusch-godfrey

AFAIR, breusch godfrey is also more robust to heteroscedasticity.

I didn't find a comparison with DW in a quick search.

Josef

 

Josef

[josef...@gmail.com]

On Mon, Jun 15, 2015 at 5:31 PM, bbmra <chang...@gmail.com> wrote:

Thank you Josef. 

How about heteroskedasticity? Is goldfeld quant a good test for that?

goldfeld quant is splitting the sample, so it is useful if there is a ordering and you want to test whether heteroscadisticiy changed across subsamples.

het_breushpagan and het_white test for heteroscedasticity that is related to explanatory variables

Both are LM tests, and AFAIR, mainly differ in which explanatory variables might affect the variance.

Also some of those tests also reject on more general misspecification. Since it has been some time, since I looked at those, I don't remember all the details.

Josef

[bbmra]

Thanks so much once again.

On Monday, June 15, 2015 at 4:49:18 PM UTC-4, bbmra wrote:

[chrims]

I had the very same question and problem just two days ago. I used the table in http://www.jstor.org/stable/1914122
Since I know the number of observations and regressands I was lucky and could define some constants.

I wonder whether there are tables for more than ten regressands.

[chrims]

regressands

regressors, or course.

[josef...@gmail.com]

this goes to 21, http://web.stanford.edu/~clint/bench/dwcrit.htm  for 3 significance levels

IIRC, for R they produced their own extended tables.

Josef