Adjusting data for missing observations I have an unbalanced panel data set of 40 cities and 20 years.  It is unbalanced because the data are not collected for certain cities for every year.  The data are then balanced after these 20 years.
Suppose the time periods are 1960 to 1979 and then 1980 to 1999.  I want to compare the data for 1979 with the data from 1980, but this is somewhat difficult to do because the 1979 data have some missing observations.  
How do I do this?  Should I predict what the 1960 to 1979 data would have been had the data been balanced using the unbalanced data?  Then forecasting the data would make the 1979 data balanced so that I can compare the 1979 data with 1980.  
My concern is that predicting the data for 1979 using the 1960-1979 data may cause the series to inherently not be continuous with the 1980 data.  
 A: Multiple imputation is an important and common practice. Assuming your data are normally distributed and missing more or less at random, I would recommend you impute the missing values.
A good package that I use quite frequently is Amelia for R. http://gking.harvard.edu/amelia/ The package uses various probabilistic methods to impute the missing values, and provides pretty good diagnostic tools as well.
A: As @gmacfarlane suggests, you should probably use multiple imputation.  
However, it's not necessary that anything be normally distributed.  Most good software, including the Amelia package will impute other kinds of data as necessary.  Indeed one of the nice things about Amelia is that it can also take into account the fact that your data is a time series.
For multiple imputation to work your data should be 'missing completely at random' (MCAR) or 'missing at random' (MAR) which are (very regrettably) technical terms in this field.  See other questions on this site for a definition of MCAR and MAR.  Alternatively slides 14-19 of @guest's link provide some more details.
Specific to your question: you should use anything that might be informative to impute your missing data.  Specifically that includes the complete data you have from 1980 and later.  (Counterintuitively, it also includes any dependent variable values you might be working with.)  Using all the information you have to impute will ensure you are not actively introducing any discontinuities.  The only constraint is that you should ideally not have a final analysis model with variables not available to the imputation model; that really might cause artifactual discontinuities.
After you've imputed your half dozen completed data sets, you run the analysis you originally wanted to do on each imputed dataset and combine the results.  The multiple imputation FAQ discusses how to do this, and also the issues above.
