| Re: Combine weights Men and Women - Cambodia DHS [message #19554 is a reply to message #19537] |
Fri, 10 July 2020 09:36  |
Bridgette-DHS
Messages: 3196
Registered: February 2013
|
Senior Member |
|
|
Following is another response from DHS Research & Data Analysis Director, Tom Pullum:
Very interesting question. A big conceptual issue is that the relationship between the outcome and pre-test counseling is not deterministic and if you could change the level of pre-test counseling (in a real population) then the other covariates in the model might change too. However, this is less serious for an intervention than for other covariates and it could be a way to generate target levels of pre-test counseling.
Think of x as the observed proportion (not percentage) of women with 1 on the intervention, in the entire survey or (better) in a more homogeneous sub-population. Say that P is the corresponding proportion who have 1 on the outcome. Say that b is the coefficient for the intervention (preferable in a model that includes controls). Say that X is the target level of x.
The observed P corresponds with the observed x, and P=.9 corresponds with the target X. Therefore log(.9/.1) log[P/(1-P)] equals b(X-x). The other terms in the regression equation drop out. Solving for X, the only unknown, you have
X = x+{[log(9) log[P/(1-P)]}/b . This is a simple approach but it may be TOO simple, because I see that it can generate values of X that are greater than 1, and that would not be legal for a probability. I suggest you try it. Maybe other forum users will have suggestions.
|
|
|
|
The DHS Program User Forum
Members
Search
Help
Register
Login
Home
