Home » Data » Weighting data » Pooled datasets - weighting data problem
Pooled datasets - weighting data problem [message #2028] |
Wed, 16 April 2014 08:10 ![Go to next message Go to next message](/theme/default/images/down.png) |
lukassg
Messages: 12 Registered: January 2014
|
Member |
|
|
Hey all,
I have already extensively searched the forum but couldn't find a definite answer that's why I would like to raise my question:
I am pooling DHS survey, always two rounds per country, but all countries still separate. However since I pool two years together, I still need to de-normalize m weights before I do the actual pooling as far as I understood from various sources.
I am using birth recode files and thus I should proceed like this:
V005*=V005×(total females age 15-49 in the country at the time of the survey)/(number of women age 15-49
interviewed in the survey)
Now I read the one-pager by Ruilin Ren but I still don't know where to find "total number of females age 15-49 in the whole country at the time of the survey".
I guess the only way is to go and check all the censuses that have been used for my DHS surveys at hand and look for that specific number there? Is there a database at DHS maybe where all that information is stored already?
Subsequently I have a question for the procedure. I open both datafiles, let's say Ethopia 2000 and Ethopia 2011. For both datasets I de-normalize the weight according to the formula above (before that I divide v005 by 1,000,000 of course). Then I assume I can pool the two datasets by appending one to the other.
But now that I have them pooled and want to specify my svyset, do I just use the weight variable as it is, e.g. "svyset psu [pweight=v005], strata(strat_id)" ?
Or do I have to make any other adjustments?
Thanks a lot for your help, this forum is really helpful always!
Lukas
|
|
|
Re: Pooled datasets - weighting data problem [message #2047 is a reply to message #2028] |
Thu, 17 April 2014 14:14 ![Go to previous message Go to previous message](/theme/default/images/up.png) ![Go to next message Go to next message](/theme/default/images/down.png) |
Bridgette-DHS
Messages: 3151 Registered: February 2013
|
Senior Member |
|
|
Following are comments by our Senior Sampling Specialist, Ruilin Ren:
Quote:Now I read the one-pager by Ruilin Ren but I still don't know where to find "total number of females age 15-49 in the whole country at the time of the survey".
Comment: You need to find this piece of information from other sources, such as the country's statistical office or census office, or from the UN's World Population Prospects: http://esa.un.org/wpp/unpp/panel_indicators.htm
Quote:I guess the only way is to go and check all the censuses that have been used for my DHS surveys at hand and look for that specific number there? Is there a database at DHS maybe where all that information is stored already?
Comment: DHS does not store the population information for any country.
Quote:...now that I have them pooled and want to specify my svyset, do I just use the weight variable as it is, e.g. "svyset psu [pweight=v005], strata(strat_id)" ?
Or do I have to make any other adjustments?
Comment: The way you pool the data together is fine. You do not need to rename the weight variable, but you do need some treatment for the PSU/cluster (HV001 or V001 or MV001) and the stratification variables (HV022, V022, MV022, or HV024, V024, MV024 if you use region cross urban/rural for stratification). The idea is that each PSU/cluster and each stratum from each survey must stand alone. If you do not treat them properly, for example, the software cannot distinguish cluster 001 (HV001=001) from survey one and survey two, the system will simply merge them as one cluster. This is not true. The easiest way to do is to add 10000 to HV001 from survey one, add 20000 to HV001 from survey two. So cluster 10001 and 20001 will be treated as two different PSUs/ clusters. Similarly you need to treat HV022, or HV024 if you use HV024 cross HV025 for stratification.
|
|
|
|
|
Re: Pooled datasets - weighting data problem [message #2242 is a reply to message #2225] |
Sun, 01 June 2014 15:36 ![Go to previous message Go to previous message](/theme/default/images/up.png) ![Go to next message Go to next message](/theme/default/images/down.png) |
Reduced-For(u)m
Messages: 292 Registered: March 2013
|
Senior Member |
|
|
I think one issue we haven't dealt with (yours) is re-normalizing weights when using a sub-population. For a simple solution, I'd suggest going with the recommendation from the DHS staff (buried somewhere in this thread) for when you have multiple survey rounds from the same country - just use the regular weights*.
The idea is that, if the sample sizes are similar across survey rounds, the re-normalizing shouldn't matter much (you are implicitly weighting each survey by sample size when using the regular weights). And with the sub-pop command, I'm not sure how you would want to re-normalize anyway (that bit of survey design inference I'm not real teched up on and the Stata documentation isn't super helpful to me - I think the right re-normalizing might somehow relate to the ratio of the prevalence of the sub-population to the full population, probably across strata or something really difficult and nuanced).
*Note - you want to use the "subpop" command and you want to create new identifiers for "cluster" and "strata" that are survey-round specific (say, replacing cluster "10" with cluster "svy2011_10" or something like that.
**Additional option: just take out all the sub-pop observations from both rounds, sum (within round) all the weights, and divide the old weights by the sum of the new weights. Then you have sub-populations only in each survey round, and each round has the same total weight, but nothing is "population representative", it's just corrected for selection probability across the sub-population.
Any thoughts from the DHS staff on either of these options?
|
|
|
Re: Pooled datasets - weighting data problem [message #2280 is a reply to message #2242] |
Mon, 02 June 2014 12:50 ![Go to previous message Go to previous message](/theme/default/images/up.png) ![Go to next message Go to next message](/theme/default/images/down.png) |
geoK
Messages: 56 Registered: May 2014
|
Senior Member |
|
|
O.K. thank you! I'll try what you suggest...Just a last quick question please: is there a SUBPOP command in SPSS? What If I have only selected the subpop cases and create a new dataset (then pooled the two together?). How can I apply subpop? Still creating a Country-specific variable (like 0;1) ? Thanks!!!
[Updated on: Mon, 02 June 2014 13:01] Report message to a moderator
|
|
|
Re: Pooled datasets - weighting data problem [message #2281 is a reply to message #2280] |
Mon, 02 June 2014 21:16 ![Go to previous message Go to previous message](/theme/default/images/up.png) ![Go to next message Go to next message](/theme/default/images/down.png) |
Reduced-For(u)m
Messages: 292 Registered: March 2013
|
Senior Member |
|
|
If you only have a dataset with the sub-population observations, you don't need to use any special "sub-pop" command, in SPSS or in anything else. I would just do it unweighted, and then weighted with regular (probability) weights, and then by the "survey round weights sum to 1" method, and compare the three. If they are similar, you should be fine (and they will probably all be similar).
Also - "country specific variable?" I thought you only had one country.
|
|
|
Re: Pooled datasets - weighting data problem [message #2340 is a reply to message #2225] |
Wed, 04 June 2014 16:45 ![Go to previous message Go to previous message](/theme/default/images/up.png) ![Go to next message Go to next message](/theme/default/images/down.png) |
Bridgette-DHS
Messages: 3151 Registered: February 2013
|
Senior Member |
|
|
Comment from Ruilin Ren, Senior DHS Sampling Expert:
Yes, it is true that even when pooling two different surveys from the same country, you need to de-normalize the sampling weight because the sampling fractions of the two surveys might be very different. But it does not matter if your study units are all women 15-49 or a sub-set of them, or even if your study units are children under five from the women's data file. You need to de-normalize V005 using the population of women 15-49 at the time of the survey, or the best estimates you can get.
[Updated on: Wed, 04 June 2014 16:47] Report message to a moderator
|
|
|
Re: Pooled datasets - weighting data problem [message #2345 is a reply to message #2340] |
Fri, 06 June 2014 12:19 ![Go to previous message Go to previous message](/theme/default/images/up.png) |
geoK
Messages: 56 Registered: May 2014
|
Senior Member |
|
|
Thank you for your kind replies.
Also, even in the case of pooled dataset, is it always advisable the use of svyset, when I perform any regression analysis? Should I use as STRATA 'country specific strata' or, given that I am working with the same country, can I merge the strata of the two?
thanks
regards.
|
|
|
Goto Forum:
Current Time: Sun Sep 1 01:31:27 Coordinated Universal Time 2024
|