By Henrik Zaunbrecher and Jurre Thiel – Netherlands Bureau for Economic Policy Analysis (CPB)
There is broad agreement that there is a shortage of affordable housing in the Netherlands. However, there is less agreement on how serious the problem is, what kind of houses are missing, and what the right steps are to solve the problem. How big the shortage is and for what kind of housing can only be answered, if we know what the demand for housing actually is. Likewise, we can only assess if and how much prices decrease if more housing is built if we know how sensitive demand is to price changes. Thus we first need to know the housing demand function of households to properly assess the housing problem as well as evaluating solutions for it.
In this project the researchers estimate the housing demand in the Netherlands with a discrete choice model which allows them to estimate what influence different house and household characteristics have on the housing choices that people in the Netherlands make. Once the preferences of households are estimated, it is also possible investigate the effect of different policies, such as changing of the lending norms, on the housing market.
One obstacle in the estimation of the model is that rental data is only available for a small sample of the private/liberalized rental market. The researchers thus first use gradient boosted trees to impute the rent and liberalization status of houses for which rent and liberalization status are not known. To simplify the problem of millions of households making choices over millions of houses to something computationally more manageable, they use k-means clustering to classify houses based on property value, size, and location into housing types. The discrete choice model is then estimated with households making a choice over the housing type that they wish to live in.
The dataset that the researchers work with is quite large, as they use linked registry data about individuals, households, and houses over several years. This means that the working memory requirements for both the machine learning techniques we use and the estimation of the discrete choice model make it difficult, if not impossible, to perform them in CBS’s normal remote access environment. They thus use the ODISSEI supercomputer for the estimation of the housing demand model as well as for the preparation of the dataset that is used to estimate the model.
Relevant links
Photo by Tom Podmore on Unsplash