| Pooling DHS from same country [message #8302] |
Mon, 05 October 2015 13:26  |
Saima
Messages: 10
Registered: October 2015
Location: Bowling Green, Ohio
|
Member |
|
|
Hi,
I am pooling Pakistan Demographic Health Survey of 1990-91 and 2012-13. I am looking at the trends in unintended fertility over time. Do i need to re-normalize the weights before pooling, even when I am taking survey year as a covariate?
Second, my analytical sample consist of those women who had birth in last five years and among them I am restricting my analysis to most recent birth. In that case when I am renormalizing weight, number of women aged 15-49 interviewed should be the total that were interviewed or that subsample.
Saima
|
|
|
|
| Re: Pooling DHS from same country [message #8310 is a reply to message #8302] |
Mon, 05 October 2015 16:58  |
Reduced-For(u)m
Messages: 292
Registered: March 2013
|
Senior Member |
|
|
One easy way to avoid de-normalizing weights would be to just calculate the averages separately by survey round and then look at the "trend" - meaning draw a line between the two points. Then you can just use the original survey weights, and you don't really lose much power (you are just implicitly allowing any covariates to have different effects at different points in time - if you are using covariates, which I don't see a real reason for if you just want to analyze simple trends over time).
|
|
|
|
| Re: Pooling DHS from same country [message #8311 is a reply to message #8310] |
Mon, 05 October 2015 17:56 |
Saima
Messages: 10
Registered: October 2015
Location: Bowling Green, Ohio
|
Member |
|
|
Thanks for your reply. I am looking at how perceived spousal fertility desires affect birth intendedness and I want to see how this relationship has changed over time. For that I am pooling the two datasets and wondering whether in this case I need to de-normalize the weights or not.
Saima
|
|
|
|
| Re: Pooling DHS from same country [message #8366 is a reply to message #8311] |
Thu, 15 October 2015 19:21 |
Reduced-For(u)m
Messages: 292
Registered: March 2013
|
Senior Member |
|
|
By "over time" I would guess you mean from survey round to survey round? Or from cohort to cohort?
In the first case, you can just compute the two statistics separately in each dataset. In the second case, you would have to append the two datasets, and in theory should re-normalize using the guidelines described in other threads. However, some DHS people have suggested that since sample sizes are very similar from round to round within a country (at least usually), adjusting the weights shouldn't matter. Also, if you can assume that the population demographics aren't change much, then another option would be to force each survey's weights to sum up to 1 (so, get your sample, sum up the weights within each survey, and gen new_weight = old_weight/sum_of_weights where sum_of_weights is defined separately for each survey). That would preserve the probability of selection correction within each survey, and make the two surveys each add equal total weight to the final regression.
That make sense?
|
|
|
|
The DHS Program User Forum
Members
Search
Help
Register
Login
Home
