6. Implications for Analysis
Longitudinal data are collected in order to measure and to model change and so the
results from this paper need to be considered in this light. We should also bear in
mind that both unit and partner non-response in sweep one was associated with
income in that poorer families and, in addition, poorer partners were more likely to be
missed (Plewis, 2004). This process continued as we moved from sweep one to
sweep two in that, as we have seen, poorer families were more likely to drop out
(although we do not, of course, know whether the process of becoming either poorer
or richer between sweeps was related to attrition). The under-representation of
poorer families, although partially compensated for by the strategy of over-sampling
more disadvantaged areas, could have implications for model estimates if, for
example, the relation between an outcome of interest and income was non-linear. In
a similar vein, the lack of information about the income of the self-employed might
have an impact on model estimates. There are, of course, techniques for adjusting
for non-response - weighting, multiple imputation and selection modelling for
example - but discussion of these goes well beyond the purposes of this paper. What
we have shown here is that members of households containing a young child do not
always report their income and their reluctance or inability to do so is related to how
they earn a living, who they live with (if anyone), where they live and what their ethnic
background is. These are all factors to bear in mind when using income either as a
response or, perhaps more commonly, as an explanatory variable in models of
change.
19
19