Home » Data » Weighting data » Weighting in IV approach (Egypt - panel data - 2005-2008-2014)
Weighting in IV approach [message #26507] |
Mon, 27 March 2023 16:56 |
Alanood
Messages: 18 Registered: November 2022
|
Member |
|
|
Hello
I am doing my first paper as PhD student and I am using panel data for Egypt - KR file for 2005, 2008, and 2014.
I append them in one file and I extract all the childern above two years old.
I am testing the realtionship between maternal employment and child health (stunting, wasting, underweight, and overweight).
The Analysis Approach: 1- OLS 2- IV-Instrument variable.
I usually start my analysis by add svy: in the begining, which is conducted by:
***********************svy*************
gen strata=v023
gen psu=v021
gen sampwt=v005/1000000
svyset psu, strata(strata) weight(sampwt) vce(linearized) singleunit(centered)
.
.
.****************************STATA command********************
.****************Analysis****************
****OLS*****
svy: regress stunted motheremployed child characteristics mother characteristics partner characteristics Household characteristics (fixed effect (year) (governorates) )
svy: regress wasted motheremployed child characteristics mother characteristics partner characteristics Household characteristics (fixed effect (year) (governorates) )
svy: regress overweight motheremployed child characteristics mother characteristics partner characteristics Household characteristics (fixed effect (year) (governorates) )
svy: regress underweight motheremployed child characteristics mother characteristics partner characteristics Household characteristics (fixed effect (year) (governorates) )
*****IV******
******IV-unweighting****
ivregress 2sls stunted child characteristics mother characteristics partner characteristics Household characteristics (fixed effect (year) (governorates) ) (motheremployed= instrument ), robust first
.
.
*******IV- wieghting*******
.
ivregress 2sls stunted child characteristics mother characteristics partner characteristics Household characteristics (fixed effect (year) (governorates) ) (motheremployed= instrument ) [pweight = sampwt], vce(cluster v021) first
.........****************************************
My question here is relted to the result that I got from in IV approach, the result are differ between weighting and unweighting IV appreoach
IV- weighting
the variable mother employment 0/1 - is significat at 5% in(wasted- in IV weighted regresion)
IV- not weighting
the mother employment 0/1 - is significat at 5% in(stunted -wasted-overweight- in IV unweighted regresion )
So I am wondering that
1- why the result is differ between weighting and unweighting.
and 2- In my paper can I include the unweighting regression or it yields unseginificant result.
Thank you and I appreciate your guidance.
|
|
|
|
Re: Weighting in IV approach [message #26517 is a reply to message #26513] |
Tue, 28 March 2023 17:35 |
Alanood
Messages: 18 Registered: November 2022
|
Member |
|
|
Thank you for your reply.
could you please clarify more " I'd be very cautious in the interpretation of borderline significance. "Cautious" means interpreting a borderline value as "not significant", even if it's disappointing." ?
Regarding the reason of using panel, it is because i want to test the imapct of maternal employment and child health in multiple years, and include years fixed effect in the regression, Even though there are not the same group of people interviewd, its still interesting to test for it.
I have no issue with including "svy" with OLS, but when i used ivreg "svy" not allowed.
I tried to include:
gen strata=v023
gen psu=v021
gen sampwt=v005/1000000
[pweight = sampwt], strata vce(cluster psu)
.
however STATA does not allow "ivreg" with strata (v023)!!
I only can include weight and psu.
[pweight = sampwt], vce(cluster psu)
Any way regarding the weighting, I found two views: the svy (weight+psu+strata),
1- Some mentioned that If we have the svy information in the dataset, we should always use it.
"To avoid any bias in the coefficients from the over-sampled population and adjust the standard errors for the survey clustering effect (Deaton, 1997), all the regression analyses have corrected for the survey design, i.e. the sampling weight, the cluster, and the strata were all taken into account."
2-The other view as found in this reply " https://userforum.dhsprogram.com/index.php?t=tree&goto=2 6509&#page_top[/url] " and in this paper " https://www.nber.org/system/files/working_papers/w18859/w188 59.pdf " by Gary Solon,Steven J. Haider, and Jeffrey Wooldridge, 2013,WHAT ARE WE WEIGHTING FOR?.
they stated that
" If you want a population average, you have to use the weights. That is a general truth about representative sampling and the sampling structure of the DHS>
But, if you want a causal estimate, it gets a little murkier. If you believe (read: assume) that every person, regardless of their characteristics, will have the same response to some causal input, then you do not need to weight your regressions, because it doesn't matter who was in the sample."
.
.
In conclusion:
1- I am thinking now, Is the option of using "svy" in the analysis up to the author ? or this will yield biased results.!!
2- Also If I have to use the weight+psu+strata in the analysis, Do you know how to include strata in the IV regression command.?
*******************************command********************** *******
here is the command for IV weighted and Unweighted.
weighted IV
ivregress 2sls ....(outcome variable) (controls) (fixed effect) ......... (endogenous variable = instrument)[pweight = sampwt], vce(cluster v021) first
unweighted IV
ivregress 2sls ....(outcome variable) (controls) (fixed effect) ......... (endogenous variable = instrument), robust first
********************
Thank you
[Updated on: Tue, 28 March 2023 18:57] Report message to a moderator
|
|
|
|
Re: Weighting in IV approach [message #26527 is a reply to message #26518] |
Wed, 29 March 2023 08:47 |
Bridgette-DHS
Messages: 3208 Registered: February 2013
|
Senior Member |
|
|
Following is a response from Senior DHS staff member, Tom Pullum:
Some estimation commands do not allow for a full svyset adjustment. If ivreg does not allow strata, then I'd recommend proceeding without that part of svyset.
There have been other posts on whether or not to use weights. If you do not use weights, then all estimates, not just for the means, but all coefficients, will be biased toward the subpopulations that have been over-sampled and away from the subpopulations that have been under-sampled. Comparisons between surveys in different countries or at different times in the same country will be affected.
Some people, mainly economists, do not like weights because they tend to increase standard errors. If you want to align with that view, you are certainly free to do so.
There is actually a hybrid option, to use the weights for the estimates of coefficients, but not use the weights for the estimates of standard errors. However, this gets messy and is not practical.
I recommend the full svyset adjustment, including weights, but I recognize that some people do not like weights. It's up to you and whoever will review your analysis.
|
|
|
|
Goto Forum:
Current Time: Thu Dec 12 23:31:27 Coordinated Universal Time 2024
|