>imputation introduces bias as well. The main assumption is that, for the
>missing cases, we can specify at least the distribution ( be it the marginal
>one or a conditional one). In the case of missing data in agronomic research
>and the like, those assumptions MAY be tenable. In the case of human
>respondents not providing answers, the bias is already there by virtue opf
>the fact that they did not repond. Imputation is NOT going to correct for
>that. Our theories of human behaviour with regards to responding to surveys
>are just not good enough.
>John de Vries
Whether, or how well, imputation works depends on how well the imputation
assumptions match the reasons for "missingness". With human data the data
will be missing for different reasons in different contexts, so there is no
particular reason to feel that imputation will always fail, but it isn't a
panacea. With survey responding I'd argue that reasons for missing data are
probably less problematic than drop-out in a longitudinal study (given that
most people just forget to send questionnaires back in time!).
Thom
Thom Baguley
Human Experimental Psychology
Human Sciences, Loughborough University
Loughborough, Leicestershire LE11 3TU, UK.
Tel: +44(0)1509 223049 Fax: 223940
http://www-staff.lboro.ac.uk/~hutsb/
|