**cd g:\philippe

use drs_mi, clear
list in 1
compress

***CHECK DATA
tab cluster, m
tab sex, m
tab age, m
tab treat_history, m
tab h, m
tab r, m
tab e, m
tab s, m

***MDR IS h and r 
count if h==1 & r==1
gen mdr = h+r
tab mdr, m
gen mdr_gp = mdr
recode mdr_gp 2=1 1=0
tab mdr_gp

***PERCENTAGE WITH MISSING MDR DATA, BY CLUSTER, AGE GROUP, SEX, AND TREATMENT HISTORY
gen mdr_missing = 0
replace mdr_missing = 1 if mdr_gp==.
***GROUP AGE
histogram age, start(10) width(5)
tab age
gen agegrp = age
recode agegrp min/29=1 30/39=2 40/49=3 50/max=4 
tab agegrp
***% missing by age, sex, cluster, treatment history
tab agegrp mdr_missing, row chi2
tab sex mdr_missing, row chi2
tab cluster mdr_missing, row chi2
tab treat_history mdr_missing, row chi2

***PATTERNS BY AGE, WITH TREATMENT HISTORY AND MDR
tab agegrp treat_history, row chi2
tab agegrp mdr_gp, row chi2
***COMPARISONS AMONG AGE GROUPS, OVERALL AND STRATIFIED BY TREATMENT HISTORY
mhodds mdr_gp agegrp, c(2,1)
mhodds mdr_gp agegrp, c(2,1) by(treat_history)
mhodds mdr_gp agegrp, c(3,1)
mhodds mdr_gp agegrp, c(3,1) by(treat_history)
mhodds mdr_gp agegrp, c(4,1)
mhodds mdr_gp agegrp, c(4,1) by(treat_history)

***PATTERNS BY SEX, WITH TREATMENT HISTORY AND MDR
tab sex treat_history, row chi2
tab sex mdr_gp, row chi2
mhodds mdr_gp sex, c(1,0)
mhodds mdr_gp sex, c(1,0) by(treat_history)

***SEEMS THERE IS MALE/FEMALE DIFFERENCE IN % MDR, AND NOT ALL EXPLAINED BY TREATMENT HISTORY
***LITTLE DIFFERENCE BY AGE GROUP FOR MDR, BUT SOME VARIATION FOR TREATMENT HISTORY

****PATTERNS BY TREATMENT HISTORY
tab treat_history mdr_gp, row chi2
****DOES ODDS RATIO FOR TREATMENT HISTORY 1 VS 0 VARY BY CLUSTER?
mhodds mdr_gp treat_history, c(1,0)
mhodds mdr_gp treat_history, c(1,0) by(cluster)
***LITTLE EVIDENCE THAT ODDS RATIO FOR MDR TB, COMPARING TREATMENT HISTORY ///
1 VS 0, VARIES BY CLUSTER
****MDR TB BY CLUSTER, AND TREATMENT HISTORY BY CLUSTER
tab cluster mdr_gp, row chi2
tab cluster treat_history, row chi2

*****CLUSTER MEANS, FOR INDIVIDUALS WITH DATA ON MDR, BY TREATMENT HISTORY
preserve
drop if mdr_gp==.
collapse (sum) mdr_gp (count) id, by(treat_history cluster)
list
rename id total
gen mdr_prev = mdr_gp / total
hist mdr_prev, by(treat_history) start(0) width(0.05) fraction
sort treat_history
***COEFFICIENT OF VARIATION OF CLUSTER MEANS IS LOWER FOR RETREATMENT THAN NEW CASES
by treat_history: summ mdr_prev
restore

***COUNT NUMBER OF INDIVIDUALS PER CLUSTER, SEPARATELY FOR NEW AND RETREATMENT CASES
sort treat_history cluster
by treat_history cluster: gen totrec=_N
by treat_history cluster: gen recnum=_n
tab totrec if recnum==1 & treat_history==0
tab totrec if recnum==1 & treat_history==1

***COUNT NUMBER OF INDIVIDUALS PER CLUSTER, OVERALL
sort cluster
by cluster: gen totsize=_N
by cluster: gen recnum1=_n
tab totsize if recnum1==1

gen pweight = 27/totsize

****1) ANALYSIS RESTRICTED TO INDIVIDUALS WITH DATA ON MDR TB

****FIRST DO ANALYSIS SEPARATELY BY TREATMENT HISTORY

****NB FOR RETREATMENT CASES, RANDOM-EFFECTS MODEL NOT APPROPRIATE? /// 
****FOR CLUSTERS WITH RELATIVELY FEW RETREATMENT CASES, THIS REFLECTS THAT A LOW ///
****PROPORTION OF TOTAL CASES IN THE CLUSTER ARE RETREATMENT CASES. ///
****BUT IN THE RANDOM-EFFECTS MODEL RESTRICTED TO RETREATMENT CASES, ///
****A RETREATMENT CASE IN A CLUSTER WITH FEW OTHER ///
****RETREATMENT CASES WILL GET MORE WEIGHT IN THE ANALYSIS THAN A RETREATMENT CASE IN ///
****A CLUSTER WITH MANY RETREATMENT CASES. ///
****WHEREAS, GIVEN THAT FOR CLUSTERS WITH FEW RETREATMENT CASES AT THE TIME THE ///
****TARGET SAMPLE SIZE FOR THE CLUSTER WAS REACHED, THERE WAS EXTRA RECRUITMENT OF ///
****RETREATMENT CASES, IF ANYTHING THE RETREATMENT CASES ///
****FROM CLUSTERS WITH FEW RETREATMENT CASES SHOULD BE DOWN-WEIGHTED RATHER THAN UP-WEIGHTED

preserve
****CASES WITH NO PREVIOUS TREATMENT
keep if treat_history==0
***1) NO ALLOWANCE FOR CLUSTERING
xi: logit mdr_gp
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
***2) ALLOWANCE FOR CLUSTERING
xi: logit mdr_gp, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
***3) RANDOM-EFFECTS MODEL TO ALLOW FOR CLUSTERING
xi: xtlogit mdr_gp, re i(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
quadchk, nooutput
restore

preserve
****PATIENTS WITH PREVIOUS TREATMENT
keep if treat_history==1
***1) NO ALLOWANCE FOR CLUSTERING
xi: logit mdr_gp
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
***2) ALLOWANCE FOR CLUSTERING
xi: logit mdr_gp, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
***3) RANDOM-EFFECTS MODEL TO ALLOW FOR CLUSTERING
xi: xtlogit mdr_gp, re i(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
***SEEMS THIS MODEL DOES NOT FIT WELL
quadchk, nooutput
restore

****NOW DO COMBINED ANALYSIS OF ALL PATIENTS, BUT INCLUDE ///
TREATMENT HISTORY AS EXPLANATORY VARIABLE AND THEN PREDICT PREVALENCE ///
SEPARATELY FOR PREVIOUS TREATMENT NO VS YES
***NO ALLOWANCE FOR CLUSTERING
xi: logit mdr_gp i.treat_history
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itreat_his_1])/(1+exp(_b[_cons]+_b[_Itreat_his_1]))
***ALLOWANCE FOR CLUSTERING
xi: logit mdr_gp i.treat_history, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itreat_his_1])/(1+exp(_b[_cons]+_b[_Itreat_his_1]))
***RANDOM-EFFECTS MODEL
****THIS MODEL DOES NOT FIT WELL
xi: xtlogit mdr_gp i.treat_history, re i(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itreat_his_1])/(1+exp(_b[_cons]+_b[_Itreat_his_1]))
quadchk, nooutput
***4) SURVEY METHOD, ALLOWING FOR CLUSTERING AND WEIGHTING FOR TOTAL CLUSTER SIZE
***OVERALL, CONSIDER THAT FOR DRS DESIGN WEIGHTING IS NOT NECESSARY/APPROPRIATE
svyset cluster [pw=pweight]
xi: svy: logit mdr_gp i.treat_history
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itreat_his_1])/(1+exp(_b[_cons]+_b[_Itreat_his_1]))

***TRY TO UNDERSTAND DIFFERENCES BETWEEN THE MODELS
gen clustersize_gp = totsize
tab totsize
tab totsize if recnum1==1
recode clustersize_gp min/39=1 40/45=2 46/max=3
tab totsize clustersize_gp
tab clustersize_gp mdr_gp, row
tab clustersize_gp mdr_gp if treat_history==0, row
tab clustersize_gp mdr_gp if treat_history==1, row
tab clustersize_gp treat_history, row

****OVERALL CONCLUSION, IGNORING MISSING DATA ISSUE:
****1) BEST NOT TO USE RANDOM-EFFECTS AS DOES NOT WORK WELL FOR RETREATMENT CASES
****AND SINCE ONLY 30 CLUSTERS IN DRS DESIGN, NOT RELIABLE ENOUGH TO RECOMMEND
****SIMPLER METHOD OF ADJUSTING STANDARD ERROR FOR CLUSTERING IS TRANSPARENT
****AND APPROPRIATE HERE
****2) NO NEED FOR WEIGHTING, GIVEN THE DRS DESIGN
****3) MUST ADJUST THE CI FOR CLUSTERING, EVEN IF THE DESIGN EFFECT IS SMALL, GIVEN THE ///
****DRS DESIGN IS A CLUSTER SAMPLE SURVEY

****SO: STICK TO SIMPLE MODEL THAT ALLOWS FOR CLUSTERING IN THE STANDARD ERROR/CI
****BUT GIVES EQUAL WEIGHT TO EACH CASE IN THE ESTIMATE OF MDR TB

**********************************************
**********************************************
****NOW ADDRESS MISSING VALUES WITH IMPUTATION
**********************************************
**********************************************

****MV PATTERN IS VIRTUALLY "ALL OR NOTHING" FOR DRS RESULTS
****ON THIS BASIS DO NOT USE E AND S TO IMPUTE
mvpatterns id cluster age sex
mvpatterns h r e s
mvpatterns h r

***NO EVIDENCE OF VARIATION IN MDR PERCENTAGE BY CLUSTER, BUT NB SMALL CLUSTER SIZE
tab cluster mdr_gp, row chi2

compress

***DUMMY VARIABLES FOR USE IN IMPUTATION MODEL
gen cluster_2 = 0
replace cluster_2 = 1 if cluster==2
gen cluster_3 = 0
replace cluster_3 = 1 if cluster==3
gen cluster_4 = 0
replace cluster_4 = 1 if cluster==4
gen cluster_5 = 0
replace cluster_5 = 1 if cluster==5
gen cluster_6 = 0
replace cluster_6 = 1 if cluster==6
gen cluster_7 = 0
replace cluster_7 = 1 if cluster==7
gen cluster_8 = 0
replace cluster_8 = 1 if cluster==8
gen cluster_9 = 0
replace cluster_9 = 1 if cluster==9
gen cluster_10 = 0
replace cluster_10 = 1 if cluster==10
gen cluster_11 = 0
replace cluster_11 = 1 if cluster==11
gen cluster_12 = 0
replace cluster_12 = 1 if cluster==12
gen cluster_13 = 0
replace cluster_13 = 1 if cluster==13
gen cluster_14 = 0
replace cluster_14 = 1 if cluster==14
gen cluster_15 = 0
replace cluster_15 = 1 if cluster==15
gen cluster_16 = 0
replace cluster_16 = 1 if cluster==16
gen cluster_17 = 0
replace cluster_17 = 1 if cluster==17
gen cluster_18 = 0
replace cluster_18 = 1 if cluster==18
gen cluster_19 = 0
replace cluster_19 = 1 if cluster==19
gen cluster_20 = 0
replace cluster_20 = 1 if cluster==20
gen cluster_21 = 0
replace cluster_21 = 1 if cluster==21
gen cluster_22 = 0
replace cluster_22 = 1 if cluster==22
gen cluster_23 = 0
replace cluster_23 = 1 if cluster==23
gen cluster_24 = 0
replace cluster_24 = 1 if cluster==24
gen cluster_25 = 0
replace cluster_25 = 1 if cluster==25
gen cluster_26 = 0
replace cluster_26 = 1 if cluster==26
gen cluster_27 = 0
replace cluster_27 = 1 if cluster==27
gen cluster_28 = 0
replace cluster_28 = 1 if cluster==28
gen cluster_29 = 0
replace cluster_29 = 1 if cluster==29
gen cluster_30 = 0
replace cluster_30 = 1 if cluster==30
gen cluster_31 = 0
replace cluster_31 = 1 if cluster==31
gen cluster_32 = 0
replace cluster_32 = 1 if cluster==32
gen cluster_33 = 0
replace cluster_33 = 1 if cluster==33
gen cluster_34 = 0
replace cluster_34 = 1 if cluster==34
gen cluster_35 = 0
replace cluster_35 = 1 if cluster==35
gen cluster_36 = 0
replace cluster_36 = 1 if cluster==36
gen cluster_37 = 0
replace cluster_37 = 1 if cluster==37
gen cluster_38 = 0
replace cluster_38 = 1 if cluster==38
gen cluster_39 = 0
replace cluster_39 = 1 if cluster==39
gen cluster_40 = 0
replace cluster_40 = 1 if cluster==40

***DUMMY VARIABLES FOR AGE TO USE IN IMPUTATION MODEL
gen agegrp2=0 if agegrp!=.
replace agegrp2=1 if agegrp==2
gen agegrp3=0 if agegrp!=.
replace agegrp3=1 if agegrp==3
gen agegrp4=0 if agegrp!=.
replace agegrp4=1 if agegrp==4

****NEED TO INCLUDE CLUSTER, TREATMENT HISTORY, SEX IN MODEL
****NO STRONG EVIDENCE NEED TO INCLUDE AGE, BUT INCLUDE TO BE "SAFE"
****ALSO AGE MAY BE AN IMPORTANT PREDICTOR IN OTHER SETTINGS
****WANT TO KNOW THAT THE IMPUTATION HAS ALLOWED FOR AGE GROUP AND SEX, TREATMENT HISTORY, AND CLUSTER
****PREDICT AGE AS A CONTINUOUS VARIABLE WHEN IMPUTING MISSING VALUES FOR AGE,
****BUT USE AGE AS A CATEGORICAL VARIABLE (DUMMY VARIABLES) FOR PREDICTING OTHER VARIABLES
****TRY H,R AND MDR; AND ALSO JUST MDR IN THE IMPUTATION MODEL

***SORT ON CLUSTER AND ID TO ENSURE MISSING VALUE IMPUTATION IS REPRODUCIBLE
sort cluster id
***H AND R AND MDR_GP, DRY RUN
/*
ice age agegrp2 agegrp3 agegrp4 sex treat_history cluster_* h r mdr_gp, passive(mdr_gp: h*r \ agegrp2: age>=30 & age<=39 \ agegrp3: age>=40 & age<=49 \ agegrp4: age>=50 & age!=. ) ///
eq(h:agegrp2 agegrp3 agegrp4 sex treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40 r, ///
r: agegrp2 agegrp3 agegrp4 sex treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40 h, ///
sex: agegrp2 agegrp3 agegrp4 treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40 h r mdr_gp) dryrun
*/
****THIS IS MORE EFFICIENT
ice age agegrp2 agegrp3 agegrp4 sex treat_history cluster_* h r mdr_gp, passive(mdr_gp: h*r) ///
substitute(age: agegrp2 agegrp3 agegrp4) dryrun

***H AND R AND MDR_GP, IMPUTE
sort cluster id
/*
ice age agegrp2 agegrp3 agegrp4 sex treat_history cluster_* h r mdr_gp, passive(mdr_gp: h*r \ agegrp2: age>=30 & age<=39 \ agegrp3: age>=40 & age<=49 \ agegrp4: age>=50 & age!=.) ///
eq(h:agegrp2 agegrp3 agegrp4 sex treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40 r, ///
r: agegrp2 agegrp3 agegrp4 sex treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40 h, ///
sex: agegrp2 agegrp3 agegrp4 treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40 h r mdr_gp) ///
saving(drs_clusterimpute_new, replace) m(10) cycles(10) seed(101)
*/
****THIS IS MORE EFFICIENT
ice age agegrp2 agegrp3 agegrp4 sex treat_history cluster_* h r mdr_gp, passive(mdr_gp: h*r) ///
substitute(age: agegrp2 agegrp3 agegrp4) saving(drs_clusterimpute_new2, replace) m(10) cycles(10) seed(101)


****MDR_GP ONLY, DRY RUN
sort cluster id
/*
ice age agegrp2 agegrp3 agegrp4 sex treat_history cluster_* mdr_gp, passive(agegrp2: age>=30 & age<=39 \ agegrp3: age>=40 & age<=49 \ agegrp4: age>=50 & age!=.) ///
eq(mdr_gp: agegrp2 agegrp3 agegrp4 sex treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40, ///
sex: agegrp2 agegrp3 agegrp4 treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40 mdr_gp) dryrun
*/
***THIS IS MORE EFFICIENT
ice age agegrp2 agegrp3 agegrp4 sex treat_history cluster_* mdr_gp, ///
substitute(age: agegrp2 agegrp3 agegrp4) dryrun
****MDR_GP ONLY, IMPUTE
sort cluster id
/*
ice age agegrp2 agegrp3 agegrp4 sex treat_history cluster_* mdr_gp, passive(agegrp2: age>=30 & age<=39 \ agegrp3: age>=40 & age<=49 \ agegrp4: age>=50 & age!=.) ///
eq(mdr_gp: agegrp2 agegrp3 agegrp4 sex treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40, ///
sex: agegrp2 agegrp3 agegrp4 treat_history cluster_2 cluster_3 cluster_4 cluster_5 cluster_6 cluster_7 cluster_8 cluster_9 cluster_10 cluster_11 cluster_12 cluster_13 cluster_14 cluster_15 cluster_16 cluster_17 cluster_18 cluster_19 cluster_20 cluster_21 cluster_22 cluster_23 cluster_24 cluster_25 cluster_26 cluster_27 cluster_28 cluster_29 cluster_30 cluster_31 cluster_32 cluster_33 cluster_34 cluster_35 cluster_36 cluster_37 cluster_38 cluster_39 cluster_40 mdr_gp) ///
saving(drs_clusterimpute_mdrgp, replace) m(10) cycles(10) seed(101)
*/
***THIS IS MORE EFFICIENT
ice age agegrp2 agegrp3 agegrp4 sex treat_history cluster_* mdr_gp, ///
substitute(age: agegrp2 agegrp3 agegrp4) saving(drs_clusterimpute_mdrgp2, replace) m(10) cycles(10) seed(101)



****CAN ALSO TRY IMPUTING SEPARATELY FOR TREATMENT HISTORY 0 AND 1
****THIS MIGHT MAKE A DIFFERENCE, AS COEFFICIENT OF VARIATION OF CLUSTER MEANS IS ///
****LOWER FOR RETREATMENT THAN FOR NEW CASES
****BUT: NO EVIDENCE THAT ASSOCIATION BETWEEN TREATMENT HISTORY AND MDR VARIES BY ///
****CLUSTER - IMPLYING THAT NO EVIDENCE THAT THE EFFECT OF CLUSTER VARIES BY NEW VS ///
****RETREATMENT CASE.
****ALSO NO EVIDENCE THAT MALE/FEMALE DIFFERENCE VARIES BY TREATMENT HISTORY
****OR THAT EFFECT OF AGE VARIES BY TREATMENT HISTORY
****SO OVERALL, LITTLE SUGGESTION THAT NEED TO MODEL NEW AND RETREATMENT CASES SEPARATELY

*****CAVEAT: THIS MIGHT BE DIFFERENT IN OTHER DATASETS
*****SAFEST IS TO DO BOTH COMBINED AND SEPARATE ANALYSIS OF NEW AND RETREATMENT CASES ///
***** - AND CHOOSE BETWEEN THEM BASED ON OBSERVED PATTERNS IN THE NON-MISSING DATA

*****IF NO EVIDENCE THAT ASSOCIATION BETWEEN TREATMENT HISTORY AND MDR VARIES ///
BY CLUSTER, SEX, OR AGE GROUP,THEN IT IS MOST EFFICIENT TO FIT AN OVERALL MODEL ///
COMBINING NEW AND RETREATMENT CASES - THIS IS PARTICULARLY THE CASE FOR THE IMPUTATION

****MDR PREVALENCE AFTER IMPUTING MISSING DATA, IMPUTING H AND R AND THEN MDR PASSIVELY
/*
clear
use drs_clusterimpute_new
rename treat_history trt_hist
xi: micombine logit mdr_gp i.trt_hist, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itrt_hist_1])/(1+exp(_b[_cons]+_b[_Itrt_hist_1]))
***compare to original dataset with missing data
xi: logit mdr_gp i.trt_hist if _mj==0, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itrt_hist_1])/(1+exp(_b[_cons]+_b[_Itrt_hist_1]))
*/

clear
use drs_clusterimpute_new2
rename treat_history trt_hist
xi: micombine logit mdr_gp i.trt_hist, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itrt_hist_1])/(1+exp(_b[_cons]+_b[_Itrt_hist_1]))
***compare to original dataset with missing data
xi: logit mdr_gp i.trt_hist if _mj==0, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itrt_hist_1])/(1+exp(_b[_cons]+_b[_Itrt_hist_1]))

/*
****MDR PREVALENCE AFTER IMPUTING MISSING DATA, IMPUTING MDR AND IGNORING SEPARATE H AND R VALUES
clear
use drs_clusterimpute_mdrgp
rename treat_history trt_hist
xi: micombine logit mdr_gp i.trt_hist, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itrt_hist_1])/(1+exp(_b[_cons]+_b[_Itrt_hist_1]))
***compare to original dataset with missing data
xi: logit mdr_gp i.trt_hist if _mj==0, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itrt_hist_1])/(1+exp(_b[_cons]+_b[_Itrt_hist_1]))
*/

****MDR PREVALENCE AFTER IMPUTING MISSING DATA, IMPUTING MDR AND IGNORING SEPARATE H AND R VALUES
clear
use drs_clusterimpute_mdrgp2
rename treat_history trt_hist
xi: micombine logit mdr_gp i.trt_hist, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itrt_hist_1])/(1+exp(_b[_cons]+_b[_Itrt_hist_1]))
***compare to original dataset with missing data
xi: logit mdr_gp i.trt_hist if _mj==0, robust cluster(cluster)
nlcom exp(_b[_cons])/(1+exp(_b[_cons]))
nlcom exp(_b[_cons]+_b[_Itrt_hist_1])/(1+exp(_b[_cons]+_b[_Itrt_hist_1]))