I've followed the instructions laid out in this thread, and created 'group' and 'time' variables. Below is a small subsample of my data. The set is longitudinal and in long format.

FIPS    eighteenplus    year    group   time
1001    24158           1990    1001    1
1003    72825           1990    1003    1
1005    17979           1990    1005    1
1007    11747           1990    1007    1
1009    29140           1990    1009    1
1011    7620            1990    1011    1
1013    15306           1990    1013    1
1015    86988           1990    1015    1
1017    27317           1990    1017    1
1019    14920           1990    1019    1

In total, there are 3,143 id variables (FIPS) and 27 time periods (annual from 1990-2016). I'm missing data for the periods 1991-1999, 2001-2004, and want to impute 'eighteenplus' values for these time periods.

Is this data structure anissue? Should the 'eighteenplus' variable (which represents population) be transformed to log form (or some other scaling)?

Any help would be appreciated! Also, 'Amelia' never finishes executing for me, and has ran for more than 75 minutes before I quit. Is this timing typical for a set with <80,000 observations which needs 30,000+ imputed values?


  • $\begingroup$ As suggested in comments to your earlier question about this, this is not an imputation problem: it's one of interpolation. Unless you have datasets of this variable or a related variable--perhaps for geographic units other than counties--that have actual data for the intervening years, then imputation is going to be fruitless (or wrong). $\endgroup$ – whuber Jan 4 '18 at 19:04
  • $\begingroup$ Thanks for the feedback. Per those comments, would you also recommend using 'Amelia' or 'akima'? Or is there another solution you'd suggest? $\endgroup$ – SteveKoller Jan 5 '18 at 14:32
  • $\begingroup$ It all depends on what additional data you might have (or could obtain), as well as what "eighteenplus" represents and the reasons why you are trying to do all this. I doubt the solution will be as simple as downloading an R package and running a command. It could be as simple as downloading intermediate Census estimates for these years, if you can find them. $\endgroup$ – whuber Jan 5 '18 at 14:36
  • $\begingroup$ We're hoping to use the interpolated 'eighteenplus' data as a control variable in regression analysis, since it represents the 18+ population in each FIPS county (and captures variation in labor market size). As for additional data, I haven't been able to find any county-level 18+ population estimates produced by the Census or other sources. Given the unavailability of these data, would you recommend tracking down other data types, or taking another approach to complete the interpolation? $\endgroup$ – SteveKoller Jan 5 '18 at 14:52
  • $\begingroup$ Right--the Census typically balks at making estimates or projections at detailed geographic levels. Occasionally you can find projections made by states, but they will be too spotty. One approach is to find state or even national level projections for each year for various population variables and use those to impute the county-level values. For that to work, you need some kind of annually measured county-level variable that could be related to population change. $\endgroup$ – whuber Jan 5 '18 at 14:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.