Home » Topics » Domestic Violence » Pakistan 2014 IPV analysis  multilevel weights
Re: Pakistan 2014 IPV analysis  multilevel weights [message #8521 is a reply to message #8507] 
Sun, 08 November 2015 16:40 
ReducedFor(u)m
Messages: 292 Registered: March 2013

Senior Member 


I'm not sure there is a "right" answer here, although surely some wrong ones. I'll give a couple of thoughts:
1  a weight is just a person's probability of selection (rescaled in a DHSmanner, but still basically a probability weight). This does not vary by community, because there is random sampling within a cluster. So everyone in each PSU will have the same weight, and so you can just use the individual weights.
2  Even adding in information from the Men, if the unit of analysis is still the individual woman you would want the women's weights.
3  if you are trying to estimate "causal effects" of the community level variables, and you think those relationships are independent of the criteria used for sampling villages (so that, say, all women respond the same to a 1sd change in their "decision making" then you don't need to weight. Weighting buys you populationlevel mean estimates, but if there are "constant treatment effects" regardless of covariates, you may not need to weight*.
4  clusterlevel measurements are based on too few observations to be meaningful in and of themselves  as you say, there are wildly underpowered. A couple of things you could do: a) by averaging over many clusters, you can still get good estimates of community level variables, but each individual clusterlevel pointestimate would be very, very noisy. But they may still mostly "agree" in some sense; b) so if in your hierarchical model you allow each cluster an unconstrained clusterspecific effect (like treating each cluster as a miniexperiment), you could look at those individual pointestimates on a scatter plot (say Beta across some variable you think would affect Beta); c) and then you could start restricting those Betas to have some particular distribution (a random slope model) and see how that changes your overall point estimate as you make your priors on the distribution of Beta more/less informative. I think this makes sense as a kind of modelchecking or informal/additional inference procedure. A leaveoneout crossvalidation approach might make sense too, depending on how you end up thinking about each of these withincluster estimates.
I hope something there helped.



Goto Forum:
Current Time: Thu Oct 28 00:05:04 Coordinated Universal Time 2021
