Modelling household decisions using longitudinal data from household panel surveys, with applications to residential mobility

Steele, F., Clarke, P.S. and Washbrook, E. (2013) Modelling household decisions using longitudinal data from household panel surveys, with applications to residential mobility. Sociological Methodology, 43(1): 225-276. DOI 10.1177/0081175013479352

Abstract

Many events occurring to married or cohabiting individuals are the result of decisions made jointly by both partners. However, studies of life-course events usually take an individual or head-of-household perspective, and so do not explicitly reflect the joint nature of these decisions. Household panel studies and population registers are a rich resource for the study of household events, but analysing such data presents major analytical challenges. Models should ideally allow for the influence of both partners in a couple’s decision-making, and be flexible enough to handle that individuals can change their partners and have periods when they are not in co-residential unions. In this paper, we propose two types of multilevel random effects model to address some of these issues: a ‘multiple-membership’ model in which the outcome depends on a weighted combination of the random effects for each decision-maker, and a random coefficients model which allows different random effects for individuals when they are single and partnered. All methods are discussed in terms of a binary household outcome before describing more general discrete-choice models for nominal outcomes. The proposed methods are compared with previously used approaches via a simulation study, and illustrated in analyses of residential mobility using data from the British Household Panel Survey.

Further details

The fact that many outcomes measured on married or cohabiting individuals are the result of decisions made jointly by both partners is one that has received little attention in life-course research. Event-history analyses of outcomes such as moving house, childbearing and union dissolution generally ignore the influence of other household members on an individual’s decisions. This individual focus is perhaps a consequence of the heavy reliance on data from birth cohort studies, which are important because of the in-depth information collected on individuals’ childhoods and early adult circumstances. Nevertheless, a limitation of these studies is that very little information is collected on the cohort member’s partner, or on any other household members. Although less informative about early life outcomes than cohort studies, household panel studies typically provide information on every adult household member, and so are particularly suited to studying processes involving choices made by couples rather than individuals. Furthermore, panel studies follow individuals as they move between households of different types and change partners, so that individual and couple effects can potentially be disentangled. However, the appropriate analysis of household-panel data still presents major challenges because individuals can change partners and have periods when they are not in co-residential unions.

In this paper, we have several aims: first, we discuss the implications of choosing a particular unit of analysis when using longitudinal household data; a second aim is to develop a general statistical framework for the analysis of individual and couple decision-making using household panel data, which can be implemented in existing software; and finally, to present analyses of residential mobility, and offer some practical guidance for applied researchers.

The general framework we propose for modeling the timing of events that occur to all decision-makers in a household is based on two flexible families of multilevel models. First, we propose non-hierarchical ‘multiple-membership’ models in which the joint household outcome is modelled as a function of individual-level and household-level covariates, and a weighted sum of the individual random effects associated with each household member. Second, we extend previous models for the head of household to allow separate random effects for single-person and multiple-person households, and autocorrelation between two households where an individual left one household to form another. We also show how previous approaches can be viewed as special cases within this general framework, which allows methods to be compared with respect to their underlying assumptions about the residual structure, and in particular the relative size of the between-couple and between-individual variances. These differences in residual structure have important implications for assumptions about household decision making. For example, including records for both partners in the analysis file treats couples as two independent singletons rather than joint decision-makers, and implies that the variance between couples (in the log-odds of an event) is twice the variance between individuals. Furthermore, this approach leads to substantial underestimation of standard errors. Another common strategy is to analyse singletons and couples together, representing a couple by the head-of-household; this implies that the head-of-household is the primary decision maker and that the between-couple variance and the between-individual variance are equal.

Using a simulation study we demonstrate the potential impact of failing to allow correctly for the unmeasured preferences of both partners when modelling couple-level outcomes. This showed that different assumptions about the residual structure of models for household outcomes can lead to substantially different estimates of the between-individual and between-couple variances. We also showed that although there was little difference in estimates of subject-specific covariate effects across models, incorrect specification of the residual structure has a greater impact on population-averaged effects derived from random effects models.

The different types of model are applied and compared in analyses of residential mobility in Britain. We consider models for a binary indicator of any change in residence between two waves, and extensions to nominal decision indicators which incorporate information on the destination of the move (distinguishing local and migratory moves).