Working Paper No 158

An Analysis of Addictive Expenditure: A Panel Data Model with Fixed Effects and Time Analysis

Laurence Lester

The panel survey of an activity that can both recreational and addictive (henceforth Activity)—for example, alcohol consumption or exercise—provides a unique opportunity to examine the behaviour of both recreational and addict participants (henceforth “Participants”)—e.g. social drinking versus binge drinking—in ways that are not available in cross-sectional data which are prone to confuse time (cohort) and individual effects. The key distinguishing feature of panel data is that individuals are followed over time, thus allowing the study the dynamics of Participants’ spending. Thus, the panel data provides a rare opportunity to correct for the correlation of independent variables with unobservable and unchanging factors influencing Participants’ spending—or individual heterogeneity.

Longitudinal (panel) survey data are recorded observations of variables obtained by sampling the same subjects (e.g. Participants) at different times. For panel data with unacceptable between-wave attrition, sample refreshment attempts to maintain the validity of the sample.

Panel data helps explain the causes of change in Participants’ behaviour (in this case weekly spending on the Activity). Panel analysis assesses both the level and flows between various amounts spent, and so establishes links of causal relationships among different Participant’s attributes and their sequences of spending.

Panel data can overcome the problems associated with bias in econometric model estimates due to unobserved heterogeneity of Participants that may occur when analysing cross-sectional data. Analysis based on cross-sectional data may assume that individuals within the sample, and across different samples, are similar, and so are equally likely to spending the same average amount per period when undertaking the Activity. Observed differences in spending may then be attributed to observed differences (for example changes in income). If however, Participants differ in their propensity to undertake the Activity (that is, there is unobserved heterogeneity) the impact of, for example, income increases may be under or overemphasised. Similarly, changes in exogenous factors, such as government policy, may be confounded by changes in unobserved Participants’ attributes. Panel data can account for the inter-temporal aspect (differences in observed attributes across time) and spatial aspect (that is information pertaining to individuals or the cross-sectional at a point in time) of Participants behaviour, and so allows causality to be attributed to changes (or differences) in individuals’ characteristics, or to exogenous changes (such as policy). Dealing with omitted variable bias has been one of the prime motivators for developments in panel data analysis.

As well as dealing with both within and between sample heterogeneity, panel data reduce the possibility that a particular cross-section is atypical and hence that estimated model information is misleading when used to assess population traits. When used in econometric modelling, panel data analysis also alleviates aggregation bias, improves model efficiency, examines more complicated behaviour, allows a more correctly specified econometric model, and so allows more reliable evaluation of how changes to individuals’ measured attributes, and exogenous effects, influence Activity expenditure.

Finally, we note that little economic theory formally considers addictive behaviour, and expenditure on such Activities has not been rigorously investigated. Consequently, the following analysis is based on an econometric approach—the models are tested and if successful are accepted, if they are not successful adjustments based on model test statistics and other model information to modify the econometric model.

Click here to download this paper (Requires Acrobat Reader)