Home » Data » Weighting data » Query on Cluster-Level Modeling with DHS Data and Sampling Weights (Building models whereby the response variable and covariates, are aggregated at the cluster level.)
Query on Cluster-Level Modeling with DHS Data and Sampling Weights [message #30047] |
Mon, 16 September 2024 04:21 |
sayianka
Messages: 2 Registered: September 2024
|
Member |
|
|
Greetings DHS Forum,
I'm working on modeling the prevalence of health outcomes (e.g., diarrhea) as a function of covariates such as literacy and wealth index using DHS data. My approach focuses on cluster-level analysis (v001) with the goal of producing modeled map surfaces similar to the geospatial map surfaces DHS creates for certain indicators.
My current process is as follows:
1. I compute cluster-level proportions for variables (e.g., proportion of women uneducated, proportion from "poor" wealth index) using the weighted mean approach as per DHS guidelines, and sum the total number of children in the cluster eligible (total_children) and the number of "successes" (num_sick_children):
ddply(dhs_dataset, ~v001, summarise, mean = weighted.mean(x = my_variable, w = v005))
2. For model building, I'm considering a GLM structure like this:
glm(cbind(num_sick_children, total_children - num_sick_children) ~ prop_poor + prop_illiterate,
data = my_dhs_data, family = "binomial")
My questions are:
1. In building cluster-level models, how should I utilize the sampling weights (v005)?
- Should I use the unique v005 per cluster?
- Should I use the total v005 in the cluster?
- Or should cluster-level models not use the v005 weighting variable at all?
2. I've noted a quote from a DHS expert in post #9779 in response to #9772, and a related #6672:
"If you calculate a cluster-level mean, proportion, standard deviation, etc., it will be the same whether or not you use weights. However, for analyses that include the clusters as units, you do need to save the total weight for the cluster."
How does this apply to my GLM approach? Should I be incorporating cluster weights, and if so, how?
Thank you in advance for your guidance.
|
|
|
Re: Query on Cluster-Level Modeling with DHS Data and Sampling Weights [message #30076 is a reply to message #30047] |
Fri, 20 September 2024 06:41 |
Bridgette-DHS
Messages: 3172 Registered: February 2013
|
Senior Member |
|
|
Following is a response from Senior DHS staff member, Tom Pullum:
You are using R. Below I will paste a simple example in Stata, showing what I would do. The example shows how the weights and number of cases come into play with a binary outcome and a glm model. Substantively, this would not be a good analysis of the data, but it is just intended as an example of the setup. Hope you can convert to R and hope this is helpful.
* Example of individual-level and cluster-level analysis with the same variables
* Kenya 2014 DHS survey
use "...KEIR81FL.DTA" , clear
* construct a binary outcome variable for 4+ children
gen nch4plus=0
replace nch4plus=1 if v201>=4
* construct dummies for wealth quintiles
xi i.v190
rename _I* *
* Individual-level analysis
svyset v001 [pweight=v005], strata(v023) singleunit(centered)
glm nch4plus v190_* , family(binomial) link(logit) eform
* Cluster-level analysis; first switch to clusters as units
gen cases=1
collapse (first) v005 v023 (sum) nch4plus cases (mean) v190_*, by(v001)
svyset [pweight=v005], strata(v023) singleunit(centered)
glm nch4plus v190_* , family(binomial cases) link(logit) eform
|
|
|
|
Re: Query on Cluster-Level Modeling with DHS Data and Sampling Weights [message #30101 is a reply to message #30090] |
Mon, 23 September 2024 14:55 |
Bridgette-DHS
Messages: 3172 Registered: February 2013
|
Senior Member |
|
|
Following is a response from Senior DHS staff member, Tom Pullum:
When I said "this" would not be a good analysis of the data, I was referring to my own example. I was not judging what you are doing!
In the glm command that I suggest, the outcome is the numerator of a proportion (which is the mean of a 0/1 variable) and the option "family(binomial cases)" specifies the denominator (as "cases"). This is equivalent to a model in which the cluster-level outcome is a proportion and it is weighted by the number of cases in the denominator. A fitted proportion for a cluster would be the fitted frequency divided by "cases".
This part of the collapse command: "(mean) v190_*" will construct five proportions that add to one. These will be the proportions of "cases" that are in wealth quintiles 1, 2, 3, 4, 5. On the right hand side of the estimation command you could include all five of those proportions (as I did); one will be aliased because of the linear constraint (they add to 1). However, I would recommend a recode to a single proportion, such as the proportion in the bottom two quintiles, which will give just one coefficient and be easier to interpret.
You could perhaps include other covariates after the "mean" portion of the collapse command, but you could also just use your geospatial variables.
Hope this helps but let us know if you have questions specifically for the geospatial team.
|
|
|
Re: Query on Cluster-Level Modeling with DHS Data and Sampling Weights [message #30109 is a reply to message #30101] |
Tue, 24 September 2024 10:08 |
Bridgette-DHS
Messages: 3172 Registered: February 2013
|
Senior Member |
|
|
See the following from Senior DHS staff member, Tom Pullum:
A slightly modified version of yesterday's Stata program is attached. Dummies for all 5 of the wealth quintiles (whether you need them or not) are constructed by adding "noomit" to the "xi" command. Also, I show how to construct the observed and fitted proportions with 1 on the outcome variable.
|
|
|
Goto Forum:
Current Time: Thu Oct 10 20:14:21 Coordinated Universal Time 2024
|