Question about weights for a cross-sectional dataset using retrospective information

[ayisha (ayisha2)]

Dataversion

SOEP_CORE.v41eu_Stata

Data Set

SOEP

Please describe your problem

Dear SOEP team,

I am currently constructing a cross-sectional dataset of adults using SOEP data, and I have a question regarding the appropriate use of weights.

The final dataset contains one observation per individual. However, to construct the variables, I combine information from different SOEP datasets and different survey waves. In particular, I use retrospective information to obtain characteristics such as individuals’ place of birth, childhood residence, and school location. . I also use individual-level information from the pl dataset.

To construct the final dataset, I merge the different sources and then collapse the information at the individual level. For each variable, I keep the last available non-missing information across the waves. Therefore, the final observation for an individual does not necessarily correspond to a single SOEP survey year: different variables may come from different waves, including retrospective information.

Given this construction, I would like to ask whether there is any possibility of using SOEP survey weights for this type of dataset.

In particular, would it be possible to construct an appropriate weight by similarly collapsing the weights across waves, for example by keeping the weight corresponding to the last available information? Or would there be another recommended approach for weighting a cross-sectional dataset constructed in this way?

I would be very grateful for any guidance on how best to handle the weights in this setting.

Many thanks in advance!

[HWS]

Dear Ayisha,

The short answer

Collapsing the weights in the way you collapse variables is not valid. You should instead choose the cross-section (syear) from which you select your population and outcome variable for the analysis and use the corresponding weight phrf. Information from previous waves / years are unproblematic to use as long as they are fixed person characteristics and do not vary over time.

In more detail

The weights (phrf and hhrf) provided by the SOEP allow for estimating population parameters for the given survey year (syear). For this reason weights are wave-specific for the target population. Over time weights are adjusted for panel attrition from one survey year to the next as well as for the inclusion of new households, new panel members or even new samples. When new samples enter the SOEP weights of old panel members get reduced, so that the old plus the new panel members represent the population. For this reason the weights are not on the same „scale“ over time. Especially with an increase of the size of the population through immigration the total size changes as well as proportions within the population. Finally, the compensation for panel attrition would not be considered properly when collapsing the weights.

For these reasons collapsing weights will lead to biased estimates.

How to proceed

• Choose the population (say: the entire adult population of the SOEP) and your outcome variable for a specific survey year (say: syear == 2023) and use the corresponding weights.

• Restrict the sample to the participants in this survey year (netto %in% 10:19 for the adult population).

• Check that the sum of weights (phrf) is roughly what you would expect and compare it to official statistics (https://genesis.destatis.de/datenbank/online/statistic/12411/details).

  PPATHL |>
  
    filter(netto %in% 10:19 & syear == 2023) |>
  
    summarise(Population = sum(phrf))
  
  

• Retrospective information as well as time-invariant information can be taken from any data set or survey year for which they are available.

• If you are using time-varying information (say: income or life-satisfaction), this should be taken from the survey year you choose. Otherwise your information would relate to different points in time and bias estimates. This bias would be regardless of using weights. To avoid deleting cases from your analysis because of missing values, you can use multiple imputation.

[ayisha (ayisha2)]

@ayisha

Dear SOEP team,

I am currently constructing a cross-sectional dataset of adults using SOEP data, and I have a question regarding the appropriate use of weights.

The final dataset contains one observation per individual. However, to construct the variables, I combine information from different SOEP datasets and different survey waves. In particular, I use retrospective information to obtain characteristics such as individuals’ place of birth, childhood residence, and school location. . I also use individual-level information from the pl dataset.

To construct the final dataset, I merge the different sources and then collapse the information at the individual level. For each variable, I keep the last available non-missing information across the waves. Therefore, the final observation for an individual does not necessarily correspond to a single SOEP survey year: different variables may come from different waves, including retrospective information.

Given this construction, I would like to ask whether there is any possibility of using SOEP survey weights for this type of dataset.

In particular, would it be possible to construct an appropriate weight by similarly collapsing the weights across waves, for example by keeping the weight corresponding to the last available information? Or would there be another recommended approach for weighting a cross-sectional dataset constructed in this way?

I would be very grateful for any guidance on how best to handle the weights in this setting.

Many thanks in advance

Thank you very much for your detailed response. This is very helpful.