diradj

 

*! program diradj version 2.0.0 – 03jan2000 – GvM

*!

*! usage: diradj varlist

*!

*! first var in varlist is <main> regressor (sought effect)

*! followed by other regressors (that are to be adjusted for)

*!

*! data must be -stset- first; all regressors must be binary

*!

*! performs “direct adjustment” as described in Maruch (1981) in order to

*! obtain better adjustment than via classical “mean subject” (see Ghali etal)

*!

*! computed as total probability= sum( Pr(T>t|pattern i) * Pr(pattern i) )

*!

*! where the survival probabilities are obtained via Cox regression

*! and Pr(pattern i) is estimated as (N in i / total N)

*! thus producing weighted averages suW0 (for main reg=no) and suW1 (yes)

*!

*! program also computes Crude survival: Cru0 and Cru1

*! as well as “mean patient” adjustment: suM0 and suM1

*!

*! –> final dataset contains: _t (time) Cru0 Cru1 suM0 suM1 suW0 suW1

*!

prog def diradj

*

version 6.0

 

st_is 2 analysis

syntax varlist

 

unab vL: `varlist’

gettoken z vL:vL /* main regressor : e.g. Diab */

loc regs `vL’ /* other regressors: e.g. creat CHF male */

 

loc nreg: word count `regs’

loc xreg=`nreg’+1

 

tempname m A hrCru Bz Bm hr

qui {

* —- patterns

 

sca npat= 2^`nreg’ /* possible number of patterns */

g np=0

loc j 0

sca `m’=2^`nreg’

while `j'<`nreg’ {

sca `m’= int(`m’/2+.01)

loc j=`j’+1

loc x: word `j’ of `regs’

qui replace np= np+`m’ if `x’

}

mat `A’=J(npat,1,0) /* will contain the Ni’s */

loc NN 0

loc p 0

while `p'<npat {

count if np==`p’ & _st

loc NN=`NN’+r(N)

loc p =`p’+1

mat `A'[`p’,1]=r(N)

}

* —- cox reg

 

stcox `z’ , basesurv(Cru0) /* Cru0= crude baseline surv */

sca `hrCru’=exp(_b[`z’])

 

stcox `z’ `regs’ , basesurv(s0) /* s0= full reg baseline surv */

sca `Bz’=_b[`z’]

sca `Bm’= 0

 

* — adjust: mean patient –> haz h(t)= h0(t) * exp(Xbar’Beta)

loc j 0

while `j'<`nreg’ {

loc j=`j’+1

loc x: word `j’ of `regs’

su `x’ if _st, meanonly

sca `Bm’=`Bm’ + _b[`x’]*r(mean)

}

sort _t

qui {

by _t: keep if _n==_N

keep _t Cru0 s0

g double Cru1= Cru0^`hrCru’

g double suM0= s0^exp(`Bm’)

g double suM1= s0^exp(`Bm’+`Bz’)

order _t C* su*

g double suW0=0

g double suW1=0

}

* — adjust: = direct adjustment

loc p 1

while `p'<=npat {

if `A'[`p’,1] { /* some patterns may be empty */

loc j =`nreg’

sca `hr’=0

loc i=`p’-1

while `i’ {

loc x: word `j’ of `regs’

sca `hr’=`hr’+ mod(`i’,2)*_b[`x’]

loc j=`j’-1

loc i= int(`i’/2+.01)

}

replace suW0= suW0+`A'[`p’,1] * s0^exp(`hr’)

replace suW1= suW1+`A'[`p’,1] * s0^exp(`hr’+`Bz’)

}

loc p=`p’+1

}

drop s0

replace suW0= suW0/`NN’

replace suW1= suW1/`NN’

#delimit ;

drop if Cru0==Cru0[_n-1] & suM0==suM0[_n-1] & suW0==suW0[_n-1]

& Cru1==Cru1[_n-1] & suM1==suM1[_n-1] & suW1==suW1[_n-1]

; #delimit cr

}

show

 

end

* ———–

prog def show

/*

shows name of surv curve next to curve

stata won’t do that alone so it must be programmed low-level

instead of calling this <show> could just say

 

gr su* C* _t, xla yla s(..oo..) c(llllll)

 

to produce the 6 curves

instead of that 1 line need 46 lines! to obtain same plus the name and axes

*/

loc su suM0 suM1 suW0 suW1 Cru0 Cru1

loc wc: word count `su’

tempvar x y

qui g `x’=.

qui g str8 `y’= “”

loc grmin= 1

loc j 0

qui while `j'<`wc’ {

loc j=`j’+1

loc v: word `j’ of `su’

loc va=`v'[_N]

loc grmin= min(`grmin’,`va’)

replace `x’= `va’ in `j’

replace `y’=”`v'” in `j’

}

loc grmin=int(50*`grmin’)/50

loc a `su’ _t, s(..dd..) c(llllll) xla yla(`grmin'(.02)1) t1(” “)

 

sort `x’

tempname X

mkmat `x’ in 1/`wc’, mat(`X’)

loc tx

loc ht

loc k=`wc’+1

while `k’>1 {

loc k= `k’-1

loc u= `x'[`k’]

loc ht “`ht’ `u'”

loc u= `y'[`k’]

loc tx “`tx’ `u'”

}

sort _t

loc bb bbox(0,0,23000,28900,700,300,0)

gr `a’ `bb’ border

gph open

gr

gph pen 1

loc k 0

while `k'<`wc’ {

loc k=`k’+1

loc w: word `k’ of `tx’

loc h: word `k’ of `ht’

loc h=18000*(1-`h’)/(1-`grmin’)+2200

gph text `h’ 29000 0 -1 `w’

}

gph close

graph, saving(adj.gph)

end

*

exit