Stata The Stata listserver
[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: Newbie: Case selection problem

From   Taavi Lai <>
Subject   Re: st: Newbie: Case selection problem
Date   Wed, 02 Mar 2005 15:55:21 +0200

Thanks again,

The code in the original letter worked OK, but as David points out there was a problem naming and identifying all the combinations and it didn't quite solve my problems. What I finaly came up with follows
As I don't feel myself comfortable with foreach and forvalues I did it the long way using a sample from full file.

there was ~9000 persons identified by donorid and the total number of different diagnoses was 17

gen whatever=1
collapse (sum) whatever, by(donorid)
drop whatever
save temp1.dta
save temp2.dta
append using temp1.dta /* 17 times */
sort donorid
egen count=seq(), by(donorid)
save temp2.dta
use original.dta
egen count=seq(), by(donorid)
joinby donorid count using temp2.dta, unm(b)
drop _merge
reshape wide dgn, i(donorid) j(count)
egen combination=concat( dgn_1 ... dgn_17), punct(", ")
egen combination2=ends(combination), punct(", ,") head
drop combination
tab combination2

The resulting frequency tabel is a starting point for a health state valuation workgroup which has to select most common diseases and combinations for their work. The full dataset has ~1800 different diag values, so its hard to imagine the final frequency tabel.

Best regards,

David Kantor wrote:

This is a different problem. How many possible combinations of k potential diagnoses are there?

The answer is 2^k, and there is a natural (one-to-one) mapping from these combinations to the integers from 0 to 2^k -1. But if k is large, then how do you name all the combinations? You may be stuck with just using the resulting integer.

First, you need to have your diag values to be in a small range of non-negative integers, such as 0-k (with minimal gaps in this range). If they already are in such a form, okay. Else, you need to map them (one-to-one) into such a set of integers. (If your diag values are string, you can -encode- them and use the encoded values.)

Next, suppose that diag is that variable (or a one derived by an appropriate mapping).

summ diag
local diagmax = r(max) // get the maximal value (corresponds to k in the above)
assert r(min) >=0 // we really don't want any negatives

sort personid
forvalues n = 0 / `diagmax' {
egen byte hasdiag`n' = max(diag==`n'), by(personid)
/* That is like what I wrote in the previous reply -- but compacted under a -forvalues-. */

/* At this point you can condense to one observation per person; this is optional. */
bysort personid: keep if _n==1

/* Generate the identifier of all combinations. */
gen long combination = 0
forvalues n = 0 / `diagmax' {
replace combination = combination + 2^`n' if hasdiag`n'

There may be other (better) ways to express that computation.
Also, be warned, I have not tested this.
And, if `diagmax' is large, you may need double rather than long -- for the type of combination.

If this has been done correctly, then each value of combination should uniquely correspond to a distinct combination of diag values. The correspondence is that for each diag value of n, that diag value is present if and only if there is a 1 in the nth bit of the binary representation of combination (counting from the right, starting with 0) -- but only when represented as an integer (not float or double).

Again, I hope this helps.

*   For searches and help try:

© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index