Aggregate Lm2 model using first-level t-stats

[ejolly]

Should add an option to use t-statistics from first-level OLS models instead of just coefficients. This has the effect if implicitly down-weighting noisy first-level estimates based on their standard-errors. However, it only makes sense if the dataset is balanced with respect to each first-level group (e.g. each individual has the same number of observations at the first level).