|
Missing data can bias
results and can also reduce sample size. Analysis of variables
having missing values will affect statistical tabulations and
tests of significance.
Multivariate procedures,
such as linear regression, require complete information on variables
that are in the model. Observations having missing values in model
variables are deleted; this can seriously reduce the degrees of
freedom. In tabular analysis, different combinations of variables
with missing values will yield tables based on different sample
sizes. The number of cases in the study varies from table to table.
Sometimes it is possible to infer that a "missing" value
really means "none" of the property being measured,
in which case missing can become the number zero.
In categorical variables
it is possible to code "missing" as a response category
for the variable. Another alternative is to replace the missing
value with the mean response of subjects that had similar demographic
or other relevant characteristics. Preserving sample size is very
important. A variety of statistical techniques such as imputation
are also used to handle missing data (Faris
et al., 2002; Groves et al.,
2002; Rubin, 1987).
|