RDP 2016-04: Housing Prices, Mortgage Interest Rates and the Rising Share of Capital Income in the United States 3. Data

The data used in this paper come from four main sources: the Bureau of Economic Analysis (BEA), the Federal Housing Finance Agency (FHFA), the US Census Bureau, and Carrillo, Early and Olsen (2014). The data are all collected on a calendar year basis and typically cover the period from the early 1980s to at least 2012. More detailed information is provided in Appendix A.

I examine several measures of housing capital income that are disaggregated by either state or metropolitan statistical area (MSA). The state-level estimates of gross housing capital income are based on the gross operating surplus of the real estate sector. The real estate sector consists of housing and ‘other real estate’, which essentially measures the output of the commercial real estate sector. This implies that trends in the state-level estimates of real estate output will not perfectly capture the national estimates of housing output. However, at the national level, the output of the housing sector makes up more than 80 per cent of the output of the real estate sector, and this share has been fairly constant over time. This suggests that the state-level estimates will mainly capture housing output rather than other types of real estate. Moreover, if the state-level estimates of real estate gross value added are aggregated to the national level and we compare growth rates to the national estimates of gross value added for the housing sector we find a correlation coefficient of 0.85 for the period from 1963 to 2013.

The BEA does not produce state-level estimates of housing depreciation, so it is not possible to construct state-level estimates of net operating surplus for the housing sector – the measure preferred by Piketty (2014) and Rognlie (2015). This could be problematic if depreciation rates vary over time and differ across states in a systematic way. The American Housing Survey suggests that the age of the housing stock can vary a lot across states – the ‘Sand States’ (e.g. Nevada and Arizona) tend to have much newer housing stock than states in the north-east (e.g. New York). However, this variation in the age of the housing stock has been fairly constant over time. As will be shown, any confounding time-invariant variation in depreciation rates are dealt with in the statistical analysis. Moreover, at the national level, the trend in the share of housing gross value added closely follows that of housing net operating surplus, suggesting that similar relationships hold at more disaggregated levels of geography.

I also collect state-level estimates of net housing profits (or the ‘rental income of persons’) for both owner-occupied and tenant-occupied property. These data are provided by the BEA as part of the State Personal Income accounts.

For the econometric analysis, I augment these data with information on housing prices by state which are available as part of the FHFA Housing Price Index (HPI).[4] The HPI is based on transactions for single-family properties that involve conforming conventional mortgages purchased or securitised by the Government-sponsored enterprises (Fannie Mae or Freddie Mac).[5] The HPI is a weighted index with the weights based on the shares of one-unit detached properties in each state. The HPI is a repeat-sales index so it measures average price changes based on repeat sales or refinancing of the same properties. In effect, the price index abstracts from changes in the composition of housing sold.[6]

To examine relative trends in rents and housing prices, I also utilise a source of state-level information on rents and non-housing prices. I obtain state-level price indices for rents and for all other goods and services from Carrillo et al (2014). The ability to track the prices of all goods and services in each state allows me to construct state-specific relative price estimates for both housing prices and rents. The state-level price indices are produced by first creating cross-sectional price indices for the year 2000 for around 400 metro areas and then using time-series price indices provided by the Bureau of Labor Statistics (BLS) to create a panel of prices. More detailed information on the construction of the price indices can be found at Edgar Olsen's website (<http://eoolsen.weebly.com/price-indices.html>).

I supplement these data with disaggregated information on mortgage lending rates, housing supply elasticities and real GDP growth.

The state-level estimates of mortgage interest rates are obtained from the FHFA. Based on a sample of mortgage lenders, the FHFA obtains information on the terms and conditions of all single-family (non-farm) mortgages that lenders close during the last five business days of each month. The survey includes conventional mortgages. It excludes multi-family loans and refinanced mortgages.

To assess how the elasticity of housing supply varies across states I use the index developed by Saiz (2010). This index is based on the coefficients estimated from a regression of housing price growth on measures of regulatory and physical constraints, as well as pre-determined population levels. The regression is estimated at the MSA level. The measure of regulatory constraints is based on the Wharton Residential Urban Land Use Regulatory Index (Gyourko, Saiz and Summers 2008). The measure of physical constraints compiles information on local geographic characteristics to capture the amount of developable land in a given area. More specifically, Saiz uses satellite-generated data on water bodies, land elevation, and slope steepness at the MSA level to compile an index of land constructability for each metropolitan area that has at least 500,000 inhabitants. Where necessary, I aggregate the ‘elasticity index’ to the state level by weighting the MSA-level indices within a given state (where the weights are given by the amount of land in each state accounted for by each MSA).

Table 2 summaries the state-level estimates of some of the key variables used in the paper. The correlations between some of the key variables are summarised in Appendix B.

Table 2: State-level Summary Statistics
  Obs Mean Median Std dev Min Max
Housing gross value added (% of GDP) 1,784 10.2 9.9 2.4 4.2 18.6
Net owner-occupied housing income (% of GDP) 1,784 0.7 0.7 0.7 −0.5 35.6
Net tenant-occupied housing income (% of GDP) 1,784 0.7 0.6 0.3 0.0 28.7
Population growth (%) 1,784 1.0 0.8 1.1 −6.2 8.3
Personal income growth (%) 1,784 6.1 5.9 3.4 −9.5 27.0
Real GDP growth (%) 1,784 2.6 2.6 3.1 −17.9 17.8
Nominal mortgage interest rate (%) 1,784 8.2 7.7 2.7 3.5 17.1
Real mortgage interest rate (%) 1,784 4.3 4.4 2.5 −3.6 12.8
CPI inflation (%) 1,784 3.8 3.1 2.9 −1.5 13.3
Housing supply elasticity (index) 1,714 2.4 2.1 1.0 0.9 4.5
Rent to housing prices (index) 1,478 100.7 100.2 12.5 79.3 137.0
Relative housing prices (index) 1,784 10.8 10.6 2.7 4.6 19.3

Notes: All estimates are at the state level; the sample period is that underpinning the regression estimates, which covers the years from 1978 to 2012; the elasticity index is not available for two states – Alaska and Hawaii

Sources: Bureau of Economic Analysis; Carrillo et al (2014); Federal Housing Finance Agency; Saiz (2010)


The FHFA provides housing price information for between 350 and 400 MSAs in the period since 1990 and between 150 and 350 MSAs for the period prior to 1990. [4]

Conventional mortgages are those that are neither insured nor guaranteed by the FHFA, the US Department of Veterans Affairs, or other federal government entities. Mortgages on properties financed by government-insured loans are excluded, as are properties with mortgages that exceed the conforming loan limit. Mortgage transactions on condominiums, cooperatives, multi-unit properties, and planned unit developments are also excluded. [5]

Given that the housing price estimates are based on transactions involving conforming mortgages there is likely to be sampling bias in using this housing price index (Garner and Verbrugge 2009). However, previous research has suggested that the direction of the bias is unclear, with both the lower and upper end of the housing market potentially being under-represented. Moreover, repeat-sale indices can suffer from renovation or ‘flip’ bias; renovations typically improve the quality of a home and lead to higher measured prices, so a repeat-sales index does not fully abstract from quality changes. [6]