| Re: Creating a panel dataset [message #4086 is a reply to message #4083] |
Fri, 27 March 2015 15:50   |
Reduced-For(u)m
Messages: 292
Registered: March 2013
|
Senior Member |
|
|
Karlo,
There is no guide (I am almost positive) because this is not precisely the kind of analysis the DHS were designed for, which is essentially within-survey analysis (as opposed to analysis between surveys). That said, there are a number of important points that have come up in the past on the forum, and I can summarize most of that here:
1 - you have to think about your survey weights in one form or another, either using the standard de-normalization procedure (see the Weighting threads), by making them sum up to 1 within a survey, or at least checking that the sum of weights between the two rounds is similar (so as not to weight one survey against another improperly).
2 - you have to adjust your PSU variables to get good inference. Standard practice is to create new PSU numbers that are unique across survey rounds so that you cluster at the right levels. That said, if you are saying "difference-in-difference across two surveys" then I'm guessing you'll be using some sort of geographic regions to define treated/not-treated groups, in which case it is more common to cluster at the region level across both surveys.
3 - just a quick note that confuses some people and has come up a lot: you don't "merge" the two datasets you are using, you "append" them, and it is very easy.
4 - you have to make sure your regions and region numbers are the same across survey rounds, otherwise you are matching regions across time improperly.
I think those are the key bits for doing DnD using DHS that aren't just key bits you should know about setting up any DnD problem, so that should get you started.
|
|
|
|
The DHS Program User Forum
Members
Search
Help
Register
Login
Home
