Downweighting observations that are on

[Anna Persson]

Hello

I have a dataset, with observations that I want to analyze on (elderly care) unit level. However some of the observations are recorded on section level (for some reason I do not know), which is sub-unit level. I am now wondering if I can downweight these observations in a regression using aweight, where I create variable aw as my weighting variable. In the example below I have weighted the observations in unit 3 with a third since a unit that should be represented by one observation is instead represented by three. Does this make sense? Or should I simply aggregate up to unit level instead using means?

unit | section | aw=aweight
1 | 1 | 1
2 | 1 | 1
3 | 1 | 1/3
3 | 2 | 1/3
3 | 3 | 1/3

Thank you for your help!ing

[Charlie Joyez]

Anna, if you know that for the units that only appear once, the value of your variable is already a mean of sub-units levels, then the best is to keep only one observation of repeated units, with the value being the mean of sub-units levels.

However, if the units that appear only once are only made of one sub-unit observation (they are not means), then this is a different issue. Then weighting solution is not the best solution, since all observation should be treated equally. Then You should consider changing your reference level (section and no longer units). Depending on your data and precise question, you could then either add unit fixed effects, or cluster the observations around units.

In both cases aggregating around the mean value will result in information loss, but in the first case, you assume this information to be lost for most/some of observation, and then you adjust the precision of the information to the same level for all units. In the latter case, generating mean will generate information loss that could have been saved.

Best,
Charlie

[Oded Mcdossi]

if you want to analyze your data only by unit level, then you can do two things:
1. you might need to collapse your data by that level. type help collapse for more information.
2. without seeing your data, if there is no different values within unit by section, than you can tag your data by unit and use regression with the if qualifier (if tag==1), type help egen (and search for tag).

*Added: Charlie was faster, but I think the 2 posts point to different alternatives.