3. Modelling Income Non-Response
Table 4 presents estimates for models which predict income non- response for the
main respondent at sweeps one and two. The dependent variable is 1 if the main
respondent refused to respond or didn’t know their income and 0 if they provided
income data. Those who were not eligible for the income question have been
excluded from this analysis. As the dependent variable is a binary variable these
models have been estimated using a logistic regression (allowing for the survey
design).
The sex of the interviewer is only known for sweep one. The first row of Table 4
shows that being a male interviewer (18% of all interviewers) is not a statistically
significant predictor of income non-response. The strongest and most consistent
predictor is self-employment status. This matches with much of the literature
discussed above. Social class and country have an important effect on non-
response for sweep one only. Northern Ireland has higher odds of non-response
than the reference category, England in sweep one.
For sweep two only, an important predictor is if the main respondent has a partner.
Lone parents who are employed are less likely to respond to the income questions.
In addition, at sweep two, an Indian ethnic background increases the chances of a
non-response at sweep two relative to the reference category, white respondents.
Also the Northern Ireland effect found in column 4 is removed in column 5 once we
condition on response at sweep one. If the main respondent was a non-responder to
the income question at sweep one they are more likely to be a non-responder to the
income question at sweep two. Finally, if the main respondent is the same main
respondent as at sweep one this increases the chance of a non-response at sweep
two.
Table 5 presents the corresponding estimates to Table 4 for the partner respondent
at sweeps one and two. Once again the sex of the interviewer is not a significant
predictor of income non response of the partner. Also, as with the main respondent,
the strongest and most consistent predictor is self employment status across the two
sweeps. Northern Ireland has higher odds of non-response than the reference
category, England consistently across the two sweeps for the partner respondent.
Ethnicity has a positive and significant effect on income non-response across the
sweeps.
For sweep one the older the partner is at interview predicts income non-response.
Also being in the social classes „lower supervisors and technical’ and „semi routine
and routine’ improve the chances of response to the income questions. The more
educated partners are more likely to respond to income questions, as measured by
the NVQ levels. At sweep one Wald tests on social class, NVQ levels, ethnicity and
country find all four sets of variables have an important effect on model fit.
For sweep two those living in owner occupied housing are predicted to be more likely
to respond to the partner income questions. Those who did not respond at sweep
one are more likely not to respond at sweep two. As with the main respondents, this
knocks out the Northern Ireland effect. Finally, if the partner respondent is the same