The importance of unobserved attributes in hedonic house price models

Purpose – Hedonic models are commonly used in housing markets studies to obtain quantitative measures of various implicit prices. The use of panel data in other fields of research has proved to be valuable when accounting for unobserved heterogeneity. Given that houses are extremely heterogeneous, and given that it is impossible to include all relevant attributes in hedonic models, removing unobserved heterogeneity by basic panel data models sounds appealing. This paper seeks to compare results between models that use pooled cross section data and panel data. The main research question is whether the pooled model gives unbiased estimates on some basic implicit prices. Design/methodology/approach – The paper applies the hedonic methodology. It uses regression analysis and estimate basic and parsimonious models that use either pooled time series and cross section data or panel data. The empirical results when using the two different approaches are compared. Findings – The paper illustrates that the results from the pooled timeseries and cross section model could be biased for some basic implicit prices. With some nuances, it is illustrated that in specific situations the use of a basic panel data estimator could be a simple solution to the problem of misspecification due to omitted, time‐invariant explanatory variables. Research limitations/implications – Most of the included variables do not change over time, however. In these cases potential bias using a basic fixed effects approach could not be checked for. It is also problematic that the variation in some of the time‐varying variables is not reliable and small. Finally, there could be a problem with sample selection bias. This may limit the usefulness of using panel data in disaggregated hedonic house price studies. Originality/value – Hedonic house price models are frequently used in housing market research. It is therefore important to study in various ways whether the traditional approaches provide unbiased results. In this paper models that use panel data are compared to models that use more traditional time series and cross section data. To the author's knowledge, this approach has not been followed before.