/* Again, the preliminaries */

clear
set memory 100m
set more off


/* As the outcome of the lottery is known for each indivdidual,
   we can use a binary "choice" model to predict an individual-
   specific probability of being drawn.

   The estimates of this model may then be applied to the associated
   variables in the random draws dataset to generate a prediction of 
   lottery-odds for each alternative in each individual's choice set. */ 



/* The following estimates a probit model of lottery outcome (drawn/not drawn) 
   We do not use the random draws dataset; instead we use the original dataset. */

use c:\lotteries_b\1997\data\hunts_97.dta
keep huntcode licenses apps95 apps96
rename huntcode choice_1
sort choice_1
save c:\lotteries_b\1997\intermediate\hunts_97_B.dta, replace
clear

use c:\lotteries\1997\intermediate\elk_res_1997_2.dta
drop name_l name_f name_m street city state orig_zip resident dob phone resobsno
generate obsno=group(_N)
drop if obsno>18708
generate outcome=1 if choice_1==drawn
replace outcome=0 if outcome==.
generate outcome2=1 if (purchase~=. & purchase~=0) & choice_1~=drawn
replace outcome2=0 if outcome2==.
generate male=1 if gender=="M"
replace male=0 if male==.

sort choice_1
merge choice_1 using c:\lotteries_b\1997\intermediate\hunts_97_B.dta

generate age_sqd=age*age
drop _merge gender
save c:\lotteries_b\1997\intermediate\elk_res_1997_6.dta, replace



/* This is the actual "probit" statement with outcome (1/0 variable
   indicating whether or not the individual was drawn in the lottery.
   We also hold on to the estimates as they will be used with the random
   draws dataset to generate predicted lottery-odds across each alternative
   in each individual's choice set */

probit outcome age age_sqd male licenses apps96
estimates hold b_hat



/* Now, we take the probit estimates to the random draws dataset and use
   them to calculate a new variable called "prob_prd".

   Important: this prob_prd variable might be simply included as a regressor
              in the destination choice model, implying the model is simply
              a traditional "random utility model".  Alternatively, prob_prd
              could be interacted with regressors, with the resulting model
              being a form of "random expected utility model".  */ 

use c:\lotteries_b\1997\intermediate\10_sorted_draws_B.dta, clear

drop prob_prd
estimates unhold b_hat
predict prob_prd



/* The following uses the pred_prob variable to generate interaction variables */
/*
generate pi_tcost=tcost*prob_prd
generate pi_prob=prob_tru*prob_prd
generate pi_lics=licenses*prob_prd
generate pi_harv=harvest*prob_prd
generate pi_ne=ne*prob_prd
generate pi_nw=nw*prob_prd
generate pi_sw=sw*prob_prd
generate pi_se=se*prob_prd
generate pi_qual=quality*prob_prd
generate pi_bull=bull*prob_prd
generate pi_open=opening*prob_prd
generate pi_late=late*prob_prd
generate pi_holid=holiday*prob_prd
generate pi_age=age*prob_prd
generate pi_male=male*prob_prd

save c:\lotteries_b\1997\intermediate\10_sorted_draws_B.dta, replace
*/
#delimit ;




/* And finally, two destination choice models are estimated using the
   clogit (conditional logistic regression) command.  The first is the
   random utility model and the second is the random expected utility model */

clogit choice tcost prob_prd licenses harvest ne sw se quality bull opening late holiday,
group(id);

clogit choice pi_tcost pi_lics pi_harv pi_ne pi_sw pi_se pi_qual pi_bull pi_open 
pi_late pi_holiday, group(id) ;


/* Enjoy! */