Below are my questions: Goal is to estimate comp and sales_1 as instrument variable method ? How would I state these two variables in the instrument statement ? Can I use both fdone and nolevels ?, since comp and sales_1 are correlated with differenced error term, is it correct to mention in instrument statement under correlated ?  Would fdone or nolevels in model take care of first difference estimation? or should I difference it ahead of time and supply the data in Proc Panel ? Is GMM1 or GMM2 appropriate? How to choose appropriate GMM? how to write the equation (1) using SAS proc panel?    To answer your questions in order:   1. To get instruments for sales_1, just use the keyword DEPVAR.  To get instruments for the lagged values of comp you would include those in one of the other options for the INSTRUMENTS statement, for example    instruments depvar correlated = (comp);   2. You do not need the FDONE option.  Once you specify GMM you are getting first-differenced equations.  When combined with the NOLEVELS options, you are getting first-differenced equations exclusively.   3. The differencing is done for you once you specify GMM.   4. GMM1 assumes no serial correlation in the differenced residuals.  If that's true then you gain some efficiency,  but it's usually safer to use GMM2.   GMM2 uses a data-based variance estimator for the differenced residuals.   5. If you are wanting to do standard fixed effects estimation then use  Proc panel data = cust_data; id cust time; model Sales = promo promo_1 promo_2 comp comp_1 comp_2 sales_1 sales_2/ fixone; run; However, if you are wanting to do dynamic panel estimation using only the level equations then use Proc panel data = cust_data; id cust time; instrument leveleq = (comp); model Sales = promo promo_1 promo_2 comp comp_1 comp_2 sales_1 sales_2/ gmm2 nodiffs; run; Note that I changed CORRELATED to LEVELEQ in the INSTRUMENTS statement.  That's because correlated variables are not instruments in the level equations, since they are correlated with the individual effects.   Please email me at Bobby.Gutierrez@sas.com (or post here) if you have any further questions. ... View more

Because you would like a Hausman test, you could try the following to get PROC PANEL to work.   1. Increase the memory footprint so that the dataset may fit.  See the following link for the MEMSIZE system option:   http://support.sas.com/documentation/cdl/en/hostwin/69955/HTML/default/viewer.htm#n0srij4mcdpzv2n1vtphltb00xjw.htm   2. Consider using the high-performance version of PROC PANEL: PROC HPPANEL.  See   http://support.sas.com/documentation/cdl/en/etshpug/68149/HTML/default/viewer.htm#etshpug_hppanel_toc.htm   Finally, we continuously make performance improvements to our panel-data procedures in SAS/ETS.  If you would like to email me at Bobby.Gutierrez@sas.com and let me know the version of SAS/ETS you are using, and how big of a dataset we're dealing with, I'd be glad to provide more specific advice.   --bobby ... View more

Try the following code, which    1.  Runs the model as pooled regression with dummy variables for firm.  Because you specified the PRINTFIXED option I figured the number of firms as small to moderate, and thus the dummy variable approach would not be a problem.   2.  Switches the order of the id variables so that quarter is now picked up as the main cluster.   3.  Makes a copy of the firm variable so that firm can be both an ID variable and a CLASS variable (That you need to do this is a quirk of current PROC PANEL, something we intend to change).   data have; set have; cfirm = firm; run; proc sort data = have; by quarter firm; run; proc panel data=have plot=all; id quarter firm; class cfirm; model Y = x1 x2 cfirm / pooled noint HCCME=3 cluster; run; Please email me at Bobby.Gutierrez@sas.com if you have any questions.  In the future, we hope to add general clustering ability to PROC PANEL, which would not require this workaround.   --Bobby ... View more

  The most likely cause of the error is a failed auxiliary regression that is used to calculate HCCME standard errors.  If you can email me the data at Bobby.Gutierrez@sas.com, I'll be glad to take a closer look at the root cause.   --Bobby ... View more

To answer your questions,   1. This is correct.   By default (NOLEVELS not specified), both difference and level equations are used.   2. This is correct.   3. This is not correct.  PROC PANEL assumes an AR(1) model in developing the instruments.  Making this work with higher-order AR models would require some coercion of the INSTRUMENTS statement.  In my experience, it would be difficult to get the full instrument matrix just right because PROC PANEL won't innately recognize the higher orders.  This is an area where could improve PROC PANEL in the future.   4. This is not correct. See 3.   5. Correct.  When you specify NOLEVELS, the CORRELATED option plays no role for the reason you described.   ... View more

  1. As far as getting p-values and other inference for combinations of parameter estimates (e.g. Beta_2 + Beta_3); you can issue a TEST statement, such as             TEST X1 + X1_1 = 0   which will give you the p-value for the combined effect Beta_2 + Beta_3.   2. For testing across two sub-samples, there is nothing currently in PROC PANEL to do this directly. As an alternative, I would fit it as one regression, placing an indicator variable for small vs. big firm any place that will take it.  That is, you would have a main effect for this indicator, and then interact that indicator with all the regressors in the model.     ... View more

  AbuChowdhury emailed me his data, and we were able to ascertain that his missing AR test was caused by gaps in the time series for one or more of the firms in his data.        ... View more