Multilevel models have a similar motivation to
fixed-/random-effects models
To deal with `nested' data where observations may not be
independent
It's a more general approach
In panels, nesting may feature wave-observations within
individuals within households within regions
MLM can cope with such grouping, apportioning the variance to
each level
It can also use variables at each level: e.g., household-level
variables enter at the household level to explain variance
between households, and not at the individual level
If we have regional-level variables (e.g., regional
unemployment rate, industry mix) these can enter at the regional
level and not at the household or individual level where the
observations within region are not independent.
The simpler form of multi-level models simply apportions the
variance across the levels
A more complex form allows `random coefficients': the
estimated parameters can vary across units within a given level
Thus individuals in a one household may be estimated to, for
instance, spend a higher proportion of their earnings, than
individuals in a household with different characteristics.
Multi-level models can be fitted in MLn or
MLwin, by proc mixed in SAS, or using the
free add-on gllamm in Stata.
A good introduction is Ch 6 in Dale and Davies (by Ian
Plewis).
