Dear all,
I'm running a regression using the European Working Conditions Survey (EWCS).
I'm currently trying to apply the survey weights in the regression but don't know how to (I can't work out what type of weights they are).
The weights are described as this:
"Variable label = w1 - design weight (based on household size)
Variable label = w2 - post stratification weight (based on distributions of weighting variables)
Variable label = w3 - working population in 1000s
Variable label = w4 - final weight (country level); combining w1 and w2; to be applied when running country level analyses"
Since I'm doing a country level analysis I'm guessing w1/w2 or w4 is of interest. w4 takes a value between 0.3 and 3.3)
The document describing the weighting process says this:
"The EWCS weighting scheme had two pillars:
a) selection probability weighting
− on household level
b) post stratification weighting within the national datasets"
I'm not sure how to do this in Stata.
I'm very grateful if anyone has any suggestions on what types of weights they are, and/or what commands I ought to use in Stata. (using Stata 12)
Best,
Kalle
Dear Kalle,
I assume the survey uses a two-stage probability sampling design, i.e. on the first stage a number of households have been selected, wherein (stage two) a sample of employees has been drawn and interviewed. Answers of these employees need to be weighted to represent the population with regard to e.g., age, gender etc. Therefore, you should use the final weight w4, adjusting for both the clustered design (households) and the probability of being sampled.
In Stata, pweights can be used like this:
svyset [pw=w4]
svy: regress y x1 x2 x3
or this (without svyset):
regress y x1 x2 x3 [pw=w4]
Hope this helps,
Regards,
Alex Pabst
Dear Kalle,
I assume the survey uses a two-stage probability sampling design, i.e. on the first stage a number of households have been selected, wherein (stage two) a sample of employees has been drawn and interviewed. Answers of these employees need to be weighted to represent the population with regard to e.g., age, gender etc. Therefore, you should use the final weight w4, adjusting for both the clustered design (households) and the probability of being sampled.
In Stata, pweights can be used like this:
svyset [pw=w4]
svy: regress y x1 x2 x3
or this (without svyset):
regress y x1 x2 x3 [pw=w4]
Hope this helps,
Regards,
Alex Pabst
Hi Kalle
If you are running normal (OLS) or probit/logit types of models I think Alexander gave you a perfect answer. I tend to use w4 when I use the EF data. If you are really interested in (or think) country effects are important, then your modelling should reflect the fact that people (your unit of observation in EFWCS) are nested within countries and use a nested or hierarchical modelling strategy.
Good luck.
Marc
Hi Kalle,
Alexander and Marc already provided great answers.
Just one addition: the dataset also contains a number of cross-national weights 'w5_*'. These correct for the differences between countries in the ratio between the sample size and the population size. They are standardised to average on 1, using various bases (e.g. EU15, EU27, ALL EWCS). If the group of countries you include in your analysis does not match the base for which the weights were standardised this could result in minor distortions. However it is fairly straightforward to create alternative cross-national weights yourself, as the ratio between the sample size and the population size can be obtained using w3.
Hope this is of help. Let me know if you have any further questions/suggestions.
All the best,
Gijs
Hi Alexander, Gijs and Marc,
Thank you so much for the answers!
Does the [pw=w4] also work with zero truncated binomial regressions and fractional logit models?
Such as:
tnbreg y x1 x2 [pw=w4]
or
glm y x1 x2 [pw=w4], family(binomial) link(logit) vce(robust)
Best regards,
Kalle
Hi Kalle,
yes, such models work both with the svy prefix and the [pw=w4] specification.
One more tip: If you have a count variable with a large number of zero values you could also consider a negative binomial-logit hurdle regression (in Stata findit hnblogit), instead of modelling zeros and counts separately. This command also works with pweights.
Also see Cameron and Trivedi (2009). Microeconometrics using Stata, Stata Press.
Best,
Alex
- Badges
- Science method
- Similar topics
- Mathematical Sciences
- Statistics
- Applied Statistics
- Survey Design