In the UKHLS course in Essex, we were told that in order to account for the complex survey design, we needed to use this formula:
svyset psuvar [pweight = weightvar], strata (tratavar)
However, the variables "psu" and "strata" are not in the currently released version of the data. Will these be available in the new release? And for both waves?
How should we use the above formula when the Waves are used together?
Thank you for your help!
#1 Updated by Redmine Admin over 7 years ago
- Status changed from New to In Progress
- % Done changed from 0 to 50
There is some advice on this in the new user guide, "If your analysis uses data from both Wave 1 and Wave 2, select the "lw" (longitudinal) version of the weight ...." (to appear in the section, Weighting adjustments for the Wave 2 release).
#4 Updated by Carolina Zuccotti over 7 years ago
Having said this, I have a further question. Currently I am only working with Wave 1 (and with individuals/no proxy), for which I am applying this formula:
svyset a_psu [pweight = a_indinus_xw], strata (a_strata)
However, when running some regressions I get no standard errors, and this Note: "missing standard errors because of stratum with single sampling unit".
And I assume that the more variables I add to the regression, the more likely I will have this problem of missing standard errors, is this correct?
In the ESSEX course we were explained a way to merge strata and then avoid this error. But this example simple calculates a mean age by gender. What about more complex analysis such as a regression with 4 or 5 independent variables? Should I use the "pweight" at the end of the regression commands, instead? Or should I merge the strata every time? In which other way could I deal with this?
Thank you very much in advance!
#5 Updated by Olena Kaminska over 7 years ago
Thank you for your question. This is a common issue which is related to missingness in the variables included in your regression. The issue occurs when there is at least one strata value with only one cluster (psu). If you tabulate psu and strata variables excluding all cases with missing values for all the variables in your model, you will find which strata values cause the issue. Theoretically, because strata values are ordinal you should combine the adjacent values of the strata. The course example should have given you a code to do this. In your situation, instead of female variable you should use var1 described below.
There are two simpler alternatives:
First, (not recommended) is to drop strata variable. The values will be unbiased, but conservative, i.e. there may be situation that you would have enough power to detect significant difference but it would appear non-significant because you omit strata. Ommitting psu or weight is wrong and will introduce bias to your estimates. The point is that if you omit strata (the code will run then) and you find significance, then you are safe. If you have marginal significance or nonsignificance, then you may still find significant difference with strata.
Second (recommended) is to use subpop option within svy command. I suggest that first, you create a variable indicating whether there is no missing value on any of the variables in your model (var1=1 and 0 otherwise). Now, keeping all the cases in the dataset (make sure you don't delete the ones that are not used in the model), run the usual svyset command. When running regression use the following syntax:
svy, subpop(var1): reg y1 x1 x2
This also will work if you are interested in a subgroup, e.g. only female. In this situation, make sure to keep all people in the dataset, and var1 will indicate females with no missing values on any of the variables.
Hope this helps,