Attrition is compensated for by longitudinal
weighting, to re-assert representativity to the initial
population
Item non-response is, in many cases, compensated for by
imputing a value, especially where a household summary (e.g.,
household income) would be affected
The default behaviour in the case of missing values is to
drop the whole case - this is acceptable only if values are
missing completely at random
If not missing completely at random, imputation makes for
better statistical estimates
Two main methods of imputation:
`Hot-decking': take a value at random from those of cases
with identical characteristics
Regression based: fit a model using non-missing covariates
and cases where the variable is not missing, then predict a
value for cases where it is missing
Hot decking introduces some randomness (good!) and ensures
the imputed value is a possible real-world value
Regression-based imputation is in some ways more precise, but the
imputed values have too little variance (i.e.,
)
Special attention is paid to the longitudinal logic:
account is taken of the previous wave's value
but not to the extent to under-represent wave-on-wave
transition rates
If a variable contains imputed values there is a parallel
imputation-flag variable
If the case is imputed, the flag contains the missing value
the original variable used to hold, and is otherwise 0
)
