Conferences

Household panel surveys by definition – and for good reason – collect information from all members of a survey household and trace these individuals over time. Typically, all adult household members are interviewed personally (with the only exception of the US Panel Study of Income Dynamics (PSID), where the household head is asked for proxy information). This design, however, creates the risk of a certain type of non-response if one member (or more) of an otherwise interviewed household rejects to give a personal interview (“partial unit non-response”, PUNR).

In most household panel surveys around the world, an increasing share of households is affected by PUNR. This leads to severe problems for any kind of analysis that is concerned with the aggregation of resources across household members or with the interaction of individuals and their choices and behavior within the household context (e.g. labour supply decisions). The relevance of these phenomena is intensified by the fact that partial unit-non-response is most likely not at random. In this paper, we focus on the problems of income analysis and poverty.

The standard approach in inequality research assumes that all household members pool and share all available resources (i.e., income). Applying an equivalence scale to control for differences in household size and composition yields a welfare measure which is the same for all persons living in that household. However, in cases of partial non-responding households, the aggregation of income falls just short of at least one individual’s income, thus understating the full household income.

Using data on 24 waves of the German Socio-Economic Panel Study (SOEP) we evaluate four different strategies to deal with this phenomenon. (a) Ignoring the fact that a household member and its income information is missing, thus assuming the missing individual’s income to be Zero. (b) Adjusting the equivalence scale by ignoring the person’s contribution to household income as well as household needs, thus assuming the person does not exist at all. (c) Eliminating all households observed with PUNR from the analysis population, thus assuming that these households are missing at random. (d) Imputing the missing income components taking advantage of the longitudinal information, if at all available, and cross-sectional regressions otherwise.

Results on income inequality and even more so on relative poverty may crucially depend on the choice of the aforementioned option, because this decision will affect the income of all individuals living in households which happen to be concerned by PUNR, as well as the relative poverty line to be derived from the national mean or median. In our paper we (a) describe incidence and time trends of PUNR, (b) analyse the selectivity of PUNR, (c) carry out a three-stage longitudinal imputation of missing income components due to PUNR at the individual level, and (d) provide sensitivity analyses showing the variation in the results for income inequality and poverty using alternative imputation strategies.