1
$\begingroup$

I have a panel dataset of household characteristics (income, expenditure, asset ownership etc.) collected in different waves (30 waves in total). The survey data is stratified, with primary sampling units and post stratification weights provided.

The documentation of the survey package does not mention methods for specifying panel data. From other examples, I gather using the interaction function to specify the different waves would be a possible way.

survey_obj_1 <- svydesign(id=~PSU_ID, weights=~weight, 
                          strata=~interaction(Stratum, Wave_No), 
                          nest=TRUE, data=dataframe)

to specify that the weights are assigned for individual waves, rather than using

survey_obj_2 <- svydesign(id=~PSU_ID, weights=~weight, strata=~Stratum,
                          nest=TRUE, data=dataframe) 

which does not specify the panel / temporal nature of data.

Would this be correct? Any advice or examples on using panel data survey objects would be welcome.

Both survey_obj_1 and survey_obj_2 successfully created objects. The estimates of the average parameters (like average income) was also same for both objects, but the standard errors in survey_obj_1 which considered the temporal nature of data was much lower than survey_obj_2

$\endgroup$
0

1 Answer 1

0
$\begingroup$

You should use strata=~Stratum rather than strata=~interaction(Stratum, Wave_No). Observations on the same individual in different waves are not independent and so do not belong in different strata. This is different from multiple waves of a non-panel study such as NHANES, where different waves should be treated as having separate strata.

The design might be easier to for a human to understand if you wrote

survey_obj_2 <- svydesign(id=~PSU_ID+subject_id, weights=~weight, strata=~Stratum,
                          nest=TRUE, data=dataframe)

showing that individual participants are effectively a finer level of clustering than PSUs. Since you don't have population sizes (fpc= argument), specifying the additional stage of clustering won't affect the results at all, but I still think it's worthwhile.

Incidentally, whoever migrated this was wrong: it's not really an R question, it's a question about the meaning of "strata" in survey analysis.

$\endgroup$
1
  • $\begingroup$ Thanks a lot for the response sir., pardon my late response. "Observations on the same individual in different waves are not independent and so do not belong in different strata". My concern was if this would hold true even if the sample size changes in each wave as in the dataset I have. Probably, as you suggested, using id = PSU_ID+subject_id would take care of this. Please correct me if I am wrong. $\endgroup$ May 5 at 23:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.