A house is not one good. It is a bundle — floor area, age, the walk to the nearest station, the school it is zoned for, the flood risk of its plot — sold at a single price. The hedonic idea (Rosen, 1974) is that the market price of the bundle reveals an implicit price for each attribute: what buyers, in aggregate, are willing to pay at the margin for one more unit of it.
Formally, regress log price on the attributes:
ln P_i = α + β·A_i + γ·S_i + δ·L_i + ε_iwhere A is accessibility (walking time to a school or station), S is structural (area, age), and L is locational (distance to a centre, environmental risk). Each coefficient is a marginal implicit price. A negative coefficient on “walking minutes to school” means buyers pay more as that walk gets shorter — proximity is capitalised into the price.
Why log price? Three reasons: (1) prices are right-skewed and the log pulls them toward normality; (2) a coefficient then reads as a semi-elasticity — roughly the percent change in price per unit of the attribute; (3) the fit is usually better.
Two classic ideas explain why location should be priced at all:
- Bid-rent (Alonso, 1964). Households trade the cost of a location against the convenience of access. Land closer to a valued facility commands a premium because it saves travel; prices fall with distance.
- Tiebout (1956). Households “vote with their feet,” sorting into places whose public services match their preferences. When schooling is tied to residential zone, school quality and access are bundled into the house — and the price capitalises that revealed preference.
A few conventions keep the estimates honest:
- Don’t collapse distinct kinds of access into one index. Access to a primary school and to a secondary school can capitalise very differently (younger children are more spatially constrained); merging them hides that.
- Mean-centre logged regressors, so the intercept is the predicted price at the average bundle and collinearity with quadratic (non-linear distance) terms eases.
- Stratify genuinely different submarkets (apartments vs detached houses) rather than pooling — their price-formation mechanisms, and their spatial structure, differ.
The essential caveat: a hedonic coefficient is a market correlation, not a causal effect. Unobserved neighbourhood quality, and the sorting of households across space, contaminate it — the same sorting that makes a location valuable also makes it endogenous. The hedonic surface is where the analysis starts, not where the causal claim ends. And because the whole thing lives on a map, its errors and its coefficients are rarely well-behaved in space — which is exactly why the rest of these notes exist.