Estimating an amenity rent premium from listing data: compare within the same submarket and watch for absence-of-evidence bias
Context: estimating whether apartment communities with an amenity (pool, gym) ask higher rent, using listing data plus a flag for whether the amenity appears in some other source, such as a set of scans or photos.
Pitfalls I hit: - The naive comparison (all communities with vs without) is confounded by age and size. Communities with the amenity were 8-28 years newer and larger, so most of the gap was the building, not the pool. - "Not in the scan or photo set" does not mean "does not have it". Treating missing evidence as absence pushes the estimate toward zero and mislabels the control group. Confirm absence from an independent source (e.g. aerial detection) before using it as a control.
What helped: - Normalise rent per square foot ((minRent+maxRent)/2 divided by average unit size), and drop rows with 0 rent, since 0 often means no rent was given. - Compare within the same submarket: take the median ratio per submarket where both groups have at least 2 members, then the median of those ratios. The premium fell from a larger raw gap to about 4%. - Report the number of submarkets used (here 8-22). Present the result as an association, not a causal effect. - The next step would be matching on year-built band and unit count, or a regression with submarket fixed effects.