The DHS Program User Forum
Discussions regarding The DHS Program data and results
Home » Data » Dataset use in Stata » Data analysis using multiple countries different survey years
Data analysis using multiple countries different survey years [message #3111] Thu, 16 October 2014 07:15 Go to next message
busi41 is currently offline  busi41
Messages: 1
Registered: July 2013
Member
Hello

I am using a DHS dataset that merges the latest women's individual datasets from 31 different countries. It is therefore a cross-sectional dataset but with survey years ranging from 2004 to 2013 and DHS phases 4 to 6. I would like to use matching to find the best balance of measured covariates before running a regression, and will match using a mixture of individual variables such as the respondent's religion or educational attainment and country level variables such as GDP or Maternal Mortality ratio. I remember reading that cross-sectional datasets should use variables from one point in time, but this is not really the case here. So what's the best approach for choosing country level variables.
1. For example should I choose values from one year e.g. 2011 for all countries
2. Or values from one year before each interview year. Some countries have interviews in two years e.g. some women in the country were interviewed in 2010 and others in 2011 so GDP values would be from 2009 and 2010 depending on the year of the interview.
3. Alternatively, how would one handle lagging in this instance. For example, if I think the Maternal Mortality rate when they were 18 affects the outcome at the time of the survey, then would I use MM values for when each respondent was 18 years old? My age range is 15 to 36 so that would be GDP values going back 18 years for each of the 31 countries?

Also can you think of any issues I need to watch out for from the different phases?

Any ideas would be very welcome

thank you


BM
Re: Data analysis using multiple countries different survey years [message #3114 is a reply to message #3111] Thu, 16 October 2014 22:17 Go to previous message
Reduced-For(u)m
Messages: 292
Registered: March 2013
Senior Member


I think the answer here will depend a lot on what is the ultimate aim of your analysis. What is your outcome? Are you looking to identify the effect of some particular covariate across countries, or a whole bunch of determinants of some covariate? Matching as you are describing it is usually used to identify some particular causal effect, but apparently you want to match across countries instead of within them (note: GDP will match perfectly within countries) and match on country-year level covariates? Or are you collapsing these surveys down into one observation?

As for your list:

If you think that MM at age 18 is what matters, women should have a value of the appropriate MM at age 18 for their covariate. If you think the MM the year prior to the survey is what matters, you should include that.

Same with GDP - is it the GDP they faced at 18 that matters (even for a 15 year old?) or is it the GDP they faced last year.

All these will answers will depend on your outcome and your theoretical framework. But regardless of what you are trying to estimate, matching a woman in Ghana in 2007 with a woman in Zambia in 2007 (or 2011, or whenever) sounds like an odd approach.

Some things to definitely look out for:

1 - You will have to re-normalize your survey weights in some way or another. See the discussions in the thread.

2 - Many variables are not comparable across countries, an obvious example being household wealth index which is only country and survey round specific.

3 - You will have to adjust your standard errors to deal with the aggregate variables in your regressions, perhaps clustering on something like age-by-country. Clustering on PSUs will not be sufficiently flexible here.

4 - Anything that varies with age will, in this context, vary with time as well. GDP keeps going up for the most part over time. If there is a general secular trend in your outcome variable over age, this will be artifactually correlated with a secular trend over "year" or "time", and thus you can easily end up with a spurious correlation between your GDP measure and your outcome which is actually a function of age-at-measurement.

Those are a few things to definitely look out for, but I don't think anyone here can provide you with better help on your matching/year problem without understanding what you are trying to estimate and why you think matching is appropriate.


Previous Topic: Migration data for NDHS2011
Next Topic: how to reproduce malawi 2004 published sampling errors and deft values
Goto Forum:
  


Current Time: Sun Nov 24 22:44:29 Coordinated Universal Time 2024