Ksharp is trying to get you on the right track here.  If this is time series data, getting a whole bunch of t values, say one for each time point, and then averaging them falls into the category of a very bad idea.  First, the values are not independent and have some sort of serial correlation. Second, how could you possibly calculate the correct degrees of freedom to come up with a p value or standardized effect size?  I think you need to look at the ETS suite, especially PROC PANEL, if this describes your dataset. Try this for a start--look at PROC MIXED for a repeated measures analysis, with country and time as CLASS variables.  Your model should include the interaction between these to get separate estimates by country and time.  I don't know what the repeated nature of the time variable is like, but an AR(1) structure probably fits if your data are equally spaced in time.  It may be heterogeneous by country, so that is something else you will have to investigate.  In any case, the Type III tests are "averaged t-tests" in this context.  Specific t values can be obtained with an LSMEANS statement. Good luck. Steve Denham ... View more

Could post the answer You found ? Thanks. Ksharp ... View more

PGstats and Ksharp, thanks! ... View more

You can establish a 60 row (60 month) array.  The first 60 months of your data go into this array, and then are written out to make data for the first 5-year window.  Then the next 12 months can be read in, overwriting rows 1 through 12 of the array (i.e. replacing year 1 data with year 6 data).  At that point write out the full array again, yielding data for the 2nd windows.  Read another 12 months into rows 13-24 (replacing year 2 data with year 7 data), and the write out the 3rd 60-month window.  Et Cetera.  No subsequent sorting is required, which (in combination with making a data VIEW instead of a data FILE) should make things faster. data want (keep=id endyear endmo y x1 x2)/ view=want;   array vars {60,3}  /*60 rows for Y, x1 and x2 */;   do n=1 by 1 until (last.id);     set have (keep=id date y x1 x2);     by id;     row=mod(n-1,60)+1;   /* for N=1 .. 60 , row=1..60, then N=61==>row=1, N=62==>row=2 etc. */     vars{row,1}=y;  vars{row,2}=x1;  vars{row}=x{3};     if n>=60 and mod(n,12)=0 then do;  ** If we are at least at obs 60 and it is divisible by 12 ... **;       endyear=year(date);  endmo=month(date);       do row=1 to 60;         y=vars{I,1}; x1=vars{I,2} x2=vars{I,3};         output;       end;     end; run; proc reg data=want noproint outest=regout;   by id endyear endmo;   model y=x1 x2; quit; Now you might be concerned that while the data for the first window (year-year5) will have 60 rows in chronological order, not all windows will (for instance for the second window, the data will be ordered year6, year2 ... year5).  But PROC REG gives the same result no matter the order. But if you do care, then, instead of the "do row=1 to 60" block, you can do this:    do J=N-59 to N;      row=mod(J-1,60)+1;      y=vars(row,1};  x1=vars(row,2};  x2=vars{row,3);     output;   end; Now even though WANT is a data VIEW and not a data FILE, this can take a long time.  It would likely to be a good deal faster if you could use the DATA step to make rolling sum-of-squares-and-cross-products, which can be read by PROC REG.  Not only is less data passed from one sep to the next, but making a rolling SSCP in the data step ought to be faster.  I'm working on a paper to show this.  Hoping to present it at NESUG 2012 ... View more