David, To be honest, what I am really after is to compute predicted values of y after dropping the specific observation for which I want to make a prediction.

I used the code you originally sent data outb; do replicate = 1 to num; do rec = 1 to num; set aaa nobs=num point=rec; if replicate ^= rec then output; end; end; stop; run;

and then tried the following

proc sort data=outb; by gvkey; run;

proc reg data=outb; model price_adjusted = data18PPCTOTALatadjustedshares BV_ADJ_adjustedshares; by replicate; output out=b p=yhat; run; quit;

to get the out of sample predicted values i am interested in. I am not sure though whether the specific observation is dropped when computing the out of sample predicted value.

Thanks for the additional points you made. I have already trimmed the outliers. What exactly do you mean by validation dataset?

Thanks,

Marina

On Fri, 15 Dec 2006 15:41:00 -0800, David L Cassell <davidlcassell@MSN.COM> wrote:

>mkekrou@YAHOO.CO.UK wrote: >>On Thu, 14 Dec 2006 14:43:43 -0800, David L Cassell <davidlcassell@MSN.COM> >>wrote: >> >> >mkekrou@YAHOO.CO.UK wrote: >> >> >> >>Hello everybody, >> >> >> >>I want to estimate out-of-sample predicted values using jacknife >> >>regressions. I am using the following code which ideally want to adapt >>to >> >>estimate predicted values in a similar vein to proc reg >> >>(proc reg data=test; >> >>model y=x; >> >>output p=pred; >> >>run; >> >>quit;) >> >> >> >>Any ideas about how to estimate predicted values will be much >>appreciated. >> >> >> >>DATA one; >> >> >> >>set a END=lastcase; >> >> >> >>IF lastcase THEN CALL SYMPUT ('ncases', _N_) ; >> >> >> >>RUN; >> >> >> >>*Macro portion of program begins here; >> >> >> >>%MACRO JackReg ; >> >> >> >>%DO I = 1 %TO &ncases ; >> >> >> >>DATA temp&I ; >> >> >> >>SET one ; >> >> >> >>IF _N_ NE &I ; >> >> >> >>RUN; >> >> >> >>PROC REG OUTEST = loopIest adjrsq; >> >> >> >>Omits&I: MODEL y = x w >> >>*Specify your model in the line above; >> >> >> >>RUN ; >> >> >> >>PROC APPEND BASE = RegEsts NEW = loopIest ; >> >> >> >>RUN ; >> >> >> >>%END ; >> >> >> >>%MEND JackReg; >> >> >> >>*Macro portion of program ends here; >> >> >> >>%JackReg; >> >> >> >>*this statement actually runs the macro JackReg; >> >> >> >>* End of jackknife regression program ; >> >> >> >>*Calculate mean coefficient and R-square estimates; >> >> >> >>proc means data=Regests; >> >> >> >>run; >> >> >> >> >> >>Thanks >> > >> >Let me suggest that you scrap the macro approach *entirely*. >> >Things will be simpler - and faster - if you write this as: >> > >> >[1] a process to build all your jackknife data sets in one long data >> > set; >> >[2] a single PROC REG with a BY statement; and then >> >[3] your PROC MEANS at the end. >> > >> >Although I might use a PROC UNIVARIATE at the end, so I >> >could get nonparametric confidence intervals. >> > >> >In addition, let me suggest that jackknifing and bootstrapping >> >will not solve all your problems here anyway. It depends on what >> >you are trying to do (the big picture, not "I want a jackknife") >> >and what your data are like. In fact, you might have better >> >success running the data through PROC ROBUSTREG instead of >> >doing all this. It depends on the data, the data sources, the >> >data features, the study purpose, the . . . >> > >> > >> >So let me show you how I would do the jackknife here, >> >and you can decide where to go after that. >> > >> > >> > /* build all N jackknife data sets in OUTB */ >> >data outb; >> > do replicate = 1 to num; >> > do rec = 1 to num; >> > set YourData nobs=num point=rec; >> > if replicate ^= rec then output; >> > end; >> > end; >> > stop; >> > run; >> > >> > /* use by-processing */ >> >proc reg data=outb outest=Regests adjrsq; >> > model y = x w; >> > run; >> > >> > /* then aggregate however you want */ >> >proc univariate data=Regests . . . . >> > . . . . . >> > >> > >> >Now whether using a jackknife will achieve your goals >> >is another matter entirely. >> > >> >HTH, >> >David >> >-- >> >David L. Cassell >> >mathematical statistician >> >Design Pathways >> >3115 NW Norwood Pl. >> >Corvallis OR 97330 >> > >> >_________________________________________________________________ >> >All-in-one security and maintenance for your PC. Get a free 90-day >>trial! >> >http://clk.atdmt.com/MSN/go/msnnkwlo0050000002msn/direct/01/? >>href=http://clk.atdmt.com/MSN/go/msnnkwlo0050000001msn/direct/01/? >>href=http://www.windowsonecare.com/?sc_cid=msn_hotmail >> >>Many thanks for your reply. >> >>Let me explain what I want to do-I basically want to run a horse race >>between 2 models and therefore want to calculate out of sample predicted >>values. I use jacknife because i want to get predictions for each firm >>without using that firm's data to get its predicted value. >> >>Now the code you wrote works ok; when you recommend using by processing i >>guess you mean to run the regression by replicate.BUT the issue is that I >>don't get the predicted values I am looking for. Is there any option i can >>use to generate the predicted value of the dependent variable y? I am >>looking for an option similar to that in prog reg that you can get >>predicted values proc reg data=a; >> model y =x1 x2; >> output out=b >> p=yhat; >> run; >> >>Thanks > >Oh, I see that you are getting the right results with my code. > >Okay, first off, what you are doing is not really a jackknife resampling. >Technically, if you want to be all jargon-y, this is more of a cross- >validation effort. You don't really want to compute N statistics, >each based on N-1 records. You want to get a single point estimate >on each record, and that point estimate can be computed *without* >doing a jackknife because it only requires a tweaking of the original >X'X matrix to do the computations. > >As Ian has sagely pointed out, you do not have to do any resampling >to get what you are after, since the PRESS statistic gives you the >ith residual divided by (1-h), where h is the leverage, and where the >model has been refit without the ith observation. So I think that >Ian's code needs a *tiny* tweak. > >Change his OUTPUT statement to: > > output out=regests press=y_hp h=lever; > >and change the transformation in the following data step to: > > y_pred = y - y_hp*(1-lever) ; > > >Then you don't need any looping process or by-processing at all. > >But I'm not sure that you should be trying this. Now that I see >what you are really after, I think I should point out that you need >to clean up your data before trying any sort of 'horse race'. A >single leverage point or outlier, or any other sort of deviation from >the Ordianary Least Squares regression assumptions can mess this >up. Furthermore, a horse race may end up telling you more about >how well one of your models fits extraneous random noise than >how the model really fits your data. Or how well the model will >work on new data. I would recommend using some 'validation' >data sets for this sort of examination instead. > >HTH, >David >-- >David L. Cassell >mathematical statistician >Design Pathways >3115 NW Norwood Pl. >Corvallis OR 97330 > >_________________________________________________________________ >Get the latest Windows Live Messenger 8.1 Beta version. Join now. >http://ideas.live.com