RDP 2014-04: Home Price Beliefs in Australia Appendix B: Estimating Homeowner Recollection Bias

The advantage of our data is that we can estimate any valuation bias, abstracting from recollection bias. In this appendix, we estimate the degree of home price recollection bias. Home price recollection bias occurs when surveyed homeowners incorrectly recall the purchase price of their homes. This could happen if, for example, the owner simply has imperfect recall because they bought the home a long time ago. In the HILDA Survey, homeowners are asked to recall the purchase price of their home every four years (specifically in the 2002, 2006 and 2010 surveys). About 13 per cent of homeowners report a different purchase price from one interview to the next despite apparently not moving home between interviews. This suggests that recollection bias might be a problem.

A nice feature of our dataset is that we can directly estimate such a bias. We proxy recollection bias as the difference between what the homeowner recalls for their purchase price and the purchase price inferred from the APM dataset using a hedonic regression model. To the best of our knowledge, we are the first to directly estimate recollection bias in home prices.

We estimate the following hedonic regression model:

where S_i^*pt represents the sale price of home i^* in postcode p in year t (where i^* includes sales prices from the APM transactions dataset and recalled sale prices from the HILDA Survey).^[15] The key explanatory variable is a dummy variable (PURCHASE_i^*pt) which is equal to 1 if the sale price is a reported purchase price from the HILDA Survey and is equal to 0 if the sale price is from the APM dataset.

The specification includes the same set of hedonic controls as before (for example, bedrooms and the type of housing) as well as postcode-time fixed effects. The specification effectively ‘stacks’ all the sales prices from APM with the reported purchase prices from the HILDA Survey. Our main interest is in the coefficient on the intercept term PURCHASE_i^*pt , which captures the average difference between sale prices and reported purchase prices after controlling for the location and characteristics of the property at the time of purchase. If the coefficient on this term is positive it indicates that surveyed homeowners overstate the purchase price of their homes, on average. If the coefficient is negative, surveyed homeowners understate their home purchase prices, on average. The intercept therefore provides an estimate of the average recollection bias.

The results of estimating Equation (B1) suggest that homeowners understate the purchase price of their homes by 3.4 per cent on average (Table B1). This negative bias is statistically significant. Melser (2013) finds an average positive valuation difference of about 4.5 per cent by comparing current estimates of home prices to initial purchase prices in the HILDA Survey. But if homeowners undervalue the initial purchase price by 3.4 per cent, then this suggests that Mesler's estimated valuation difference would mainly reflect recollection bias. Taking the Melser estimates at face value, and adjusting for this recollection bias, we would find an average positive valuation difference of about 1 per cent. This is very similar to our own estimate of the average valuation difference across all postcodes.

Footnote

This model specification uses annual data rather than quarterly data because the HILDA respondents only report the year of purchase and not the exact date. [15]