Tenure Choice and the Future of Homeownership

Transcription

1 WORKING PAPER Tenure Choice and the Future of Homeownership By Kevin A. Park, Chris Herbert and Roberto G. Quercia November 2014 Funding provided by the Ford Foundation

2 The UNC Center for Community Capital at the University of North Carolina at Chapel Hill is the leading center for research and policy analysis on the transformative power of capital on households and communities in the United States. The center's in-depth analyses help policymakers, advocates and the private sector find sustainable ways to expand economic opportunity to more people, more effectively. Roberto G. Quercia, Director Lucy S. Gorham, Executive Director UNC Center for Community Capital The University of North Carolina at Chapel Hill 1700 Martin Luther King Blvd. Campus Box 3452, Suite 129 Chapel Hill NC (877) (919)

3 Tenure Choice and the Future of Homeownership Abstract Using the Survey of Income and Program Participation, this paper estimates the likelihood of homeownership. A shift share analysis reveals the dramatic swings in homeownership over the past fifteen years were not driven by demographics but rather by the general housing environment, including housing affordability. A sub sample of recent movers reveals that housing affordability was not a statistically significant determinant of homeownership at the height of the housing bubble. Finally, the different general housing environments studied are used as scenarios to project the future number of homeownership and homeownership rates based on a variety of demographic forecasts. By 2035, the homeownership rate could be as low as 57 percent if restrictive housing conditions persist. Keywords: homeownership; tenure; demographics; projections; credit Kevin A. Park Corresponding Author UNC Center for Community Capital 1700 Martin Luther King Blvd., Suite 129 Chapel Hill NC kapark@ .unc.edu (828) Chris Herbert Joint Center for Housing Studies Harvard University Roberto G. Quercia UNC Center for Community Capital

4 Tenure Choice and the Future of Homeownership Introduction The United States is experiencing significant demographic changes that will affect its housing markets and the future of homeownership for its people the American Dream. As the Baby Boom generation ages, the Census Bureau projects that the population over the age of sixty five will increase from 15 percent of the population to 21 percent over the next twenty years. Older households typically have higher rates of homeownership, and have weathered the recent declines better than younger households (Figure 1A). In fact, the homeownership rate for households over sixty five years old was higher in 2012 than in Meanwhile, households between twenty five and fifty four had the lowest level of homeownership since records began in 1976 (Joint Center for Housing Studies 2013). At the same time, the Census Bureau projects that the country will become majority minority by Minority households typically have lower rates of homeownership due in part to less wealth and legacies of racial discrimination. For example, the homeownership gap between white and black households reached over 30 percentage points by the end of 2013 (Figure 1B). The arrival of new immigrants, who are typically younger and minority, will affect the rate of these demographic changes. Myers and Pitkin (2013) estimate that immigration will account for nearly onethird of the growth in all households between 2010 and 2020, including nearly 36 percent of the growth in homeowners. Nevertheless, demographics are not destiny. Although population trends provide a strong baseline for estimating housing demand and the likelihood of homeownership, other factors are also important. For example, not every household that would like to own a home has sufficient income and assets to purchase a house outright or qualify for a mortgage under prevailing underwriting standards, which can be affected by public policy decisions and overall conditions in the housing market. The housing cycle, which is not a single phenomenon but the result of multiple, mutually reinforcing dynamics, causes fluctuations around long term demographic trends. The homeownership rate reached a record high in 2004 at 69.2 percent, but has since fallen over the past decade to 64.8 percent (as of 2014Q1) the lowest rate since 1995 (Figure 2). Nationally, house prices continued to rise for another two years after the homeownership rate peaked. The ratio of house prices to owners equivalent rent rose to 45 percent above its long term average before collapsing. The declines in both homeownership and house prices were exacerbated by a tightening of underwriting standards for residential mortgage credit during the financial turmoil of 2008 and Both demographics and the dynamics of the mortgage market are important when considering the future of homeownership. This paper estimates the likelihood of homeownership at different points in the housing cycle between 1997 and Possible future trajectories of homeownership in the United 1

5 States are then projected using the results of these tenure choice models and the expected changes in the country s demographics, particularly in the age and racial composition of households. Literature Review Homeownership has proven benefits, including stronger families (Grinstein Weiss et al. 2010), increased social cohesion (DiPasquale and Glaeser 1999), and greater community engagement (Manturuk et al. 2010). Further, there is even evidence of the financial benefits of homeownership, when properly managed, during a dramatic decline in house prices (Riley et al. 2009). Nevertheless, the foreclosure crisis has made abundantly clear that the financial risks of homeownership may outweigh the benefits for some households. And even households that may like to own their own home may be prevented by income and wealth barriers. This paper presumes that the mortgage underwriting environment is a major determinant of the likelihood of homeownership. Mortgage underwriting scenarios are explicitly modeled after different years in the housing cycle as well as an explicit affordability indicator. There is a twenty five year history in housing econometrics of identifying income and wealth constraints as determinants of tenure choice, while controlling for relevant demographic and household characteristics that might affect demand for homeownership. Linneman and Wachter (1989) model homeownership in 1977 and 1983 as a function of income and wealth constraints using data on recent movers from the Survey of Consumer Finances. To define wealth and income constraints, they use two common underwriting criteria. The first is that the loan tovalue ratio the mortgage amount as a share of the home s purchase price should not exceed 80 percent of the home s purchase price. Consequently, a household s net wealth must be sufficient to cover a 20 percent downpayment. In other words, the maximum purchase price can be no more than five times household net wealth. The second criterion is that the front end debt to income ratio the annual payment of principal and interest on a mortgage as a share of income should equal no more than 28 percent. The maximum purchase price of a home, therefore, is equivalent to 35 percent (0.28/0.80) of household income divided by the prevailing mortgage interest rate. Linneman and Wachter model optimal desired house values using a sample of unconstrained households, defined as households who have purchased a home valued at less than 85 percent of either constraint. If a household s optimal desired house value, as derived from this model, is close to or exceeds the maximum value established by the income or wealth constraint, then the household is considered constrained. Categorical definitions of moderately and highly constrained, by both income and wealth, as well as the gap for highly constrained households between optimal and maximum purchase prices, are then used in a tenure choice model. Both constraints are found to be statistically significant, with the wealth constraint proving more binding. Haurin, Hendershott, and Wachter (1997) examine entry into homeownership among young adults (twenty to thirty three years old) between 1985 and 1990 using the National Longitudinal Survey of Youth. The authors build on Linneman and Wachter by, in particular, allowing LTV to be endogenous to 2

6 the tenure choice decision, which lowers the number of constrained households, and by using more sophisticated econometric modeling techniques. Quercia, McCarthy, and Wachter (2003) use the American Housing Survey to analyze the effect of underwriting standards on a variety of populations of interest. Income and wealth constraints are first defined and their effects estimated. But then, the constraints are redefined under less binding standards. The coefficients from the baseline estimation are applied using these new constraints to determine how rates of homeownership would change under the alternative underwriting regimes. However, the authors had to impute wealth by capitalizing non wage income. Like Linneman and Wachter (1989), Gabriel and Rosenthal (2005) use Survey of Consumer Finances data (from the years 1983 to 2001), but use an alternative method of identifying credit constraints. They note that the likelihood of being credit constrained and the likelihood a family prefers to own in the absence of credit constraints are both latent variables 1. Families who are not credit constrained are identified as having had no difficulty obtaining credit in the previous three to five years, but remaining families are only potentially constrained. And the preference for homeownership is only fully observed among those households which are not credit constrained. As noted, there is a persistent gap in homeownership rates between white, black, and Hispanic households; however, some share of these differences can be explained by correlated differences in age, family type and other demographics. For example, Wachter and Megbolugbe (1992) find 81 percent of the white black homeownership gap, and 78 percent of the gap with Hispanics, can be attributed to differences in income and household demographic. Gabriel and Rosenthal (2005) find that credit constraints account for no more than five percentage points of the racial gap in homeownership. Most of the gap is attributable to other household characteristics. In contrast, Quercia, McCarthy and Watcher (2003) find that after controlling for underwriting constraints, blacks and Hispanics actually exhibit higher rates of homeownership than whites. Gabriel and Rosenthal (2005) also examine the extent to which changes in the homeownership rate in the 1980s and 1990s were driven by demographics, as opposed to by the economic environment and underwriting regimes. The authors employ a shift share analysis: they predict the homeownership rate while holding constant either the cohort of households or the tenure choice model coefficients from a single year, but not both. For example, the demographic composition of households in 1998 is used to simulate the homeownership rate in 1992, 1995, and 1998 based on changes in the coefficients of the tenure choice model. This effectively demonstrates the cumulative effect of changes in the general homeownership environment, including economic factors, mortgage underwriting standards, and household tastes, while controlling for demographic changes. Assuming that tastes for homeownership remain unchanged, coefficients from different years capture the influence of yearspecific macroeconomic and lending market conditions that affect housing tenure decisions, including interest rates, business cycle risk and uncertainty, and innovations in housing policy and mortgage 1 To further complicate matters, Stein (1995) raises the connection between credit constraint, mobility and reservation prices. Droes and Hassink (2014) find that credit constrained households expect to sell their house for a higher price. This may confound the ability to estimate optimal house values. 3

7 finance (Gabriel and Rosenthal 2005, p. 105). By contrast, holding estimated coefficients constant across cohorts captures changes in the socio demographic attributes of the population over time. Gabriel and Rosenthal (forthcoming) undertake a similar shift share analysis using census microdata from 2000, 2005, and 2009, but find that model coefficients capturing household attitudes and market conditions have greater impact than socioeconomic demographics the reverse of their earlier study. There are also several improvements in the analysis, including model stratification by age, which is possible given the size of the census database, model re estimation using only a sub sample of recent movers, and the creation of household attitude variables. Specifically, the authors include median house values, expected price change, and house price volatility. On the other hand, census data does not provide information on wealth and assets. As part of an ongoing series, Wilson and Callis (2013) use the Survey of Income and Program Participation to determine how many families could afford to buy a modestly priced home. Affordability was based on sufficient income and assets to cover a 5 percent downpayment as well as 28 percent front end and 36 percent back end debt to income ratios. A modestly priced home is defined as the lowest quartile value of owner occupied homes in the American Community Survey. The authors find 50.3 percent of families could afford to buy a modestly priced home in 2009, the lowest estimate since the Census Bureau s Who Could Afford to Buy a Home? series began in On the other hand, the Census Bureau s affordability constraint is not necessarily determinative; some households that could afford to buy will instead choose to rent, while some household that may be considered constrained by the affordability indicator will nevertheless purchase a home. Psychology and household preferences play a role. Case and Shiller (1988) find homebuyers focus more on expectations of future house price increases in boom markets than market fundamentals. At the other extreme, several articles in the aftermath of the housing crisis proclaimed that the American Dream no longer includes homeownership, and that the United States is becoming a renter nation (e.g., Epstein 2010; Olick 2012; Haynes, Craighill and Clement 2014; Woellert 2014). Nevertheless, Fannie Mae s National Housing Survey found in 2013 that the large majority of renters thought homeownership was preferable for people wanting control, privacy, security, seeking to raise a family or investing wisely. Belsky (2013) notes responses in the University of Michigan s Survey of Consumers to whether it is a good time to buy a home is highly cyclical, but that the recent drop is less severe than in previous recessions. Meanwhile, the ability of affordability constrained households to purchase their homes has been clearly affected by changes in underwriting standards that have restricted mortgage credit. Goodman, Zhu and George (2014) quantify the decline in mortgage originations due to tighter underwriting standards using HMDA data and CoreLogic information on FICO credit scores. While originations fell across the range of credit scores, purchase originations to borrowers with credit scores under 660 fell 70 percent between 2001 and 2012 compared to just 18 percent among credit scores over 750. Using this information, Goodman, Zhu and George estimate that between 273,000 and 1.2 million purchase mortgages (or between 12% and 55%, respectively, of the decline between 2001 and 2012) were missing in 2012 due 4

8 to tighter underwriting. Similarly, Bhutta (2012) finds the trillion dollar decline in mortgage debt between 2009 and 2011 was driven more by a reduction in new loans, particularly among first time homebuyers with poor credit history, rather than an increase in outflows through mortgage amortization and default. Household formation is also related to the decline in homeownership and mortgage originations. Paciorek (2013) finds that rising housing costs contributed to declining headship rates between 1980 and 2000, and that high unemployment led to a sharp decline between 2006 and He also finds that poor credit scores and a history of foreclosure negatively impact household formation. Himmelberg et al. (2014) note the decline in housing turnover using American Community Survey data. Only 3 percent of the population moved into their current owned home in 2012, compared to 4.5 percent in 2001, corresponding to 1.3 million fewer home purchases. Further, of those that did purchase new homes, fewer reported a mortgage in 2012 than in 2001 (74% to 79%, respectively). The decline in mortgages is more pronounced among low income households, households with less education, and Hispanics. Himmelberg et al. estimate that approximately half of the decline in housing turnover between 2001 and 2012 was due to credit standards, with another 35% driven by cyclical factors and the remainder due to demographics. If changes in household formation disproportionately occur among owner occupied households relative to renter occupied households, then the observed homeownership rate will be affected. Haurin and Rosenthal (2007) find lower headship rates are associated with lower homeownership rates and that declining household formation between 1970 and 2000 contributed to a lower homeownership rate among younger households. In contrast, however, Yu and Myers (2010) find that declining household formation between 1990 and 2006 elevated homeownership rates. We know lack of affordability reduces the level of homeownership. This paper builds on the existing literature by updating the research in light of the dramatic swings in underwriting standards and homeownership levels over the past fifteen years. In addition, this paper uses existing demographic projections to forecast possible ranges of future American homeownership. Methodology The analysis consists of two parts. First, a tenure choice model is used to estimate the likelihood of homeownership in a given year after accounting for household demographic factors and affordability. Following Gabriel and Rosenthal (forthcoming), the tenure choice model is stratified by age and survey year: Stratifying the sample in this manner greatly enriches the analysis as it provides a clear indication of how the drivers of homeownership vary over the lifecycle. The probability of homeownership can be estimated using a logistic regression model, which can be represented as where 1 1 5

9 for age stratum in survey year. Household characteristics ( ) include total household income, 2 race/ethnicity, age (continuous variable within each stratum), family type, educational attainment, employment status, veteran status, and whether the household is inside a metropolitan area. The effect of income and wealth constraints is captured by the coefficient on, a binary variable indicating whether the household could afford a modest home given their available income and assets and common underwriting standards. Fixed effects capture regional (census division) factors, which may include unmeasured economic conditions, state regulations (e.g., homestead exemptions, deficiency judgments, etc.), and household preferences. In addition to focusing on the effects of affordability and race/ethnicity, the analysis includes a shiftshare analysis that decomposes variation in the overall homeownership rate into changes from demographics versus changes in model coefficients. This is accomplished by simulating the homeownership rate of each cohort of households using the coefficients from each tenure choice model. By comparing the homeownership in same year cohort coefficient combinations to estimated homeownership rates from combinations in which either cohort or model coefficients (but not both) are different, the effect of demographics or the general homeownership environment, respectively, can be observed in isolation. The second part of the analysis uses the range of homeownership rates estimated by the tenure choice model to project future homeownership rates given changing demographics. Specifically, the range of homeownership rates was estimated for 240 demographic buckets created by the combination of four race/ethnicity categories, five family types, and twelve age groups. These combinations correspond to demographic categories used in household projections created by Harvard University s Joint Center for Housing Studies (McCue 2014). A scenario of future homeownership is obtained by applying the homeownership rate from a tenure choice model for a given year to each demographic bucket and allowing the number of households in each bucket to change in accordance with the demographic projections. Data The Survey of Income and Program Participation (SIPP) is particularly well suited for tenure choice modeling. A topical module on assets and liabilities provides detailed information on wealth that is not available in many other datasets commonly used to model the likelihood of homeownership, such as the American Housing Survey (Quercia, McCarthy, and Wachter 2003) and American Community Survey (Gabriel and Rosenthal forthcoming; Himmelberg et al. 2014). In addition, SIPP has a larger sample size than the Survey of Consumer Finance, used in other analyses (e.g., Linneman and Wachter 1989; Gabriel 2 Household income is related to the affordability status of the household, potentially creating a collinearity problem; however, analyses with and without household income did not substantively alter the primary findings. This paper includes household income to be consistent with previous literature that uses household income to help capture socioeconomic differences in the demand for homeownership. 6

10 and Rosenthal 2005). In particular, SIPP oversamples lower income households, leading to more accurate inferences on the subpopulation most likely to be affected by problems with affordability. SIPP has been used in the Census Bureau s Who Could Afford to Buy a Home? series (Wilson and Callis 2013). This analysis follows the Census Bureau s methodology in defining the affordability status of observed households based on qualified income and assets, and the value of a modestly priced home; however, some differences exist. Most importantly, the Census Bureau s study identifies constrained households but does not determine the degree to which lack of affordability negatively affects the likelihood of homeownership. The fact that the homeownership rate consistently exceeds the estimated share of households that can afford a modestly priced home indicates that there are some households who cannot afford to buy a home but are nevertheless homeowners. At the same time, not everyone that can afford to buy a home chooses to do so. The affordability indicator is not always binding and household preferences affect tenure choice. These factors can be measured in a formal econometric model. Other differences exist in the definition of qualified income and assets. For example, the Census Bureau does not include payments from unemployment insurance as a component of available income; however, following Goodman (1988), we do include unemployment compensation given its relation to human capital and previous earnings as well as the assumption that unemployment is temporary. This is particularly important because of the structure of the SIPP data collection, which only asks interviewees for income over the past four months, leading to greater recall accuracy but potentially greater variability in employment. For a list of all SIPP variables included in the available income and assets calculations, see Appendix A and Appendix B, respectively. In addition, the public SIPP data is top coded to protect privacy, while the Census Bureau has access to the unedited data. Data from five survey years are used to capture changes in the general homeownership environment over the past fifteen years. Based on the availability of the Assets & Liabilities topical module, the following years (SIPP panel waves) are included in the analysis: 1997 (1996 6), 2001 (2001 3), 2005 (2004 6), 2009 (2008 4), and 2011 ( ). As noted, each survey year is stratified by age in order to allow coefficients to vary over the lifecycle of a household. Three age strata are used in the analysis based on the age of the household head (reference person): 15 to 39 years, 40 to 64 years, and 65 years or older. The age cutoffs were chosen to roughly separate households into typical first time homebuying years, peak earning years, and retirement. With five survey years stratified by three age groups, fifteen separate tenure choice models are estimated with sample sizes ranging from 5,564 to 17,656, with an average of 10,523. Table 1 presents summary demographic characteristics for each cohort. Home equity presents a challenge when using available assets to predict homeownership. House values constitute a major component of household wealth, at least among homeowners. Consequently, declines in house values can greatly diminish net worth. According to the 2011 SIPP data, nearly 8.6 million households (11.0 percent of all homeowners or 7.4 percent of all households) were underwater, meaning debt on their home exceeded its market value. Yet by definition, these households were still homeowners. Homeowners typically have greater wealth than renters, but the decline in house prices reduces their home equity and overall net worth, making that wealth less predictive of homeownership. 7

11 Home equity could be excluded from the available assets calculation, but that ignores a major component of household wealth and important decisions regarding portfolio allocation. One method to address this endogeneity is to use a sub sample of households that recently moved. Most homeowners sell their existing home when relocating, liquidating home equity to make the downpayment on the purchase of a new home. Whereas a sample of all households may include homeowners who purchased their home many years earlier, recent movers are going to be most directly affected by prevailing underwriting standards. Because SIPP tracks households over time, we can identify those which changed addresses between survey waves. On the other hand, the sub sample of recent movers is much smaller; consequently, we choose not to stratify the model by age group and to use Census division in place of state. More problematically, the decision to move is not random but may itself be a function of income, wealth, or other factors. In fact, the share of households that moved in the previous year fell from nearly 15 percent in 1997 to 10 percent in Consequently, the recent mover analysis may suffer from selection bias. Attrition within SIPP may also be disproportionate among recent movers, creating another possible source of bias. The longitudinal nature of the SIPP data allows us to use household characteristics before the move to model homeownership a year later. This can include prior tenure choice, which may help account for unobservable homeownership preferences. We can also examine whether renter households change tenure status after one year. Like recent movers, renters are not affected by the endogeneity of wealth and homeownership. Again, the selection of survey years depends on the availability of the Assets & Liabilities topical module. The following years (SIPP panel waves) are included in the recent movers and renter analyses: 1996 (1996 panel wave 3), 2004 ( ), and 2010 (2008 7). The dependent variable is the homeownership status one year (three survey waves) later, or 1997, 2005 and 2011, respectively. Sample sizes range from 2,527 to 9,385. Table 2 presents summary demographic characteristics for each recent mover and renter cohort. Definition of Affordability Status By equating available assets ( ) with downpayment and rearranging the loan to value ratio, it is possible to see the maximum value ( ) a household can afford with a given loan to value ratio. 1 1 For example, a 5 percent minimum downpayment (95 percent maximum loan to value ratio) allows a household to purchase a home valued at up to twenty times available assets. Similarly, the maximum value consistent with a debt to income ratio ( ) can be obtained by rearranging the formula for a fully amortizing debt payment. 8

12 is assumed to be 30 years and is the effective mortgage interest rate in a given year and state taken from the Federal Housing Finance Agency s Monthly Interest Rate Survey. We define the value of a modestly priced home ( ) as the 25th percentile of home values reported by homeowners in the SIPP data in a given year and state (the Census Bureau used the 25th percentile of owner occupied homes in the American Community Survey). The affordability indicator ( ) is determined by whether is less than the maximum value a household could afford given available assets, income, and the traditional underwriting standards of a 5 percent downpayment (the lowest typically allowed with private mortgage insurance) and a 28 percent debt to income ratio. 1, if and/or 0, otherwise If this affordability indicator was binding, it would perfectly predict homeownership; some households which could afford to buy will choose instead to rent, but none that could not afford to buy would be homeowners. Although this affordability indicator captures two of the three C s of underwriting collateral and capacity to repay it also misses one important factor: creditworthiness, typically measured by a credit score. Although Park (2013) finds credit history was cited less often over time as a reason for denying a purchase mortgage application, it is undeniable that the credit profile of loan originations has shifted. The average credit score on purchase mortgages insured by FHA increased from 633 in late 2007 to over 700 in early In the conventional market, the average FICO score on a purchase mortgage loan was 763 in 2012 according to Ellie Mae; the average on a denied application was 733. The lack of information on household credit scores is a serious problem for evaluating how underwriting standards changed over the course of the housing cycle. Nevertheless, by including this affordability variable in an econometric tenure choice model, we can determine not only how many households were constrained, but also the effect of affordability on the likelihood of homeownership, controlling for household preferences and other factors. Household Projections Harvard University s Joint Center for Housing Studies regularly produces projections of household growth using data from the Census Bureau. The latest projections are based on the Census Bureau s updated 2012 population projections, which incorporate the results from the 2010 decennial census and extend through 2035 (McCue 2014). These projections come with assumptions of low, middle, and high rates of immigration. The Joint Center for Housing Studies converts these population projections into household projections by applying a constant headship rate, calculated by age and race/ethnicity using 9

13 the Census Bureau s population estimates and the household count from the Current Population Survey s March Supplement. 3 Based on these projections and headship rates, the Joint Center for Housing Studies projects that the United States will add 1.0 to 1.3 million households per year for the next twenty years, increasing from a total of roughly million households today to between million and million in Minorities will account for roughly 84 percent of that increase while seniors will account for 72 to 86 percent, depending on immigration assumptions. The assumption of a constant headship rate is useful for long term demographic projections, but may not accurately reflect the effect of the economic environment on the number of households and on homeownership rates. Recent years have shown that household formation and homeownership, particularly first time homeownership, are related (Haurin and Rosenth 2007; Yu and Myers 2010). Consequently, it is reasonable to assume that a scenario with a low likelihood of homeownership would also be characterized by lower levels of household formation, meaning the absolute number of homeowners would be even lower than projected. Tenure Choice Model Table 3 shows the number and share of households with sufficient available income and assets to afford a modestly priced house in their state given a 5 percent downpayment requirement and maximum 28 percent debt to income ratio. The share of households that met this standard fell from 52.0 percent in 1997 to 46.1 percent in 2005, but then increased somewhat to 48.5 percent by The results of the logistic regression for each cohort and age group are presented in Table 4. The F statistic for each model s coefficients and standard errors are exponentiated into odds ratios. Affordability Status The odds of homeownership are consistently and substantially higher among households with sufficient income and assets to afford a modestly priced home under common underwriting standards. These effects are statistically significant across all cohorts and age groups. In 1997, the odds of homeownership were 8.7 times greater among younger (15 to 39 years old) households that met the affordability standard, 12.7 times higher among middle aged households (40 to 64 years old), and 4.7 times higher among senior households (65 years or older). As noted, fewer households could afford to buy a home at the height of the housing bubble. The effect of the affordability indicator on the likelihood of homeownership was also greater in the early 2000s. The odds ratio among younger and middle aged households increased between 1997 and 2005, indicating a more binding affordability constraint at the height of the housing bubble. For example, the odds of homeownership were nearly 12 times higher among younger households that met the affordability standard in However, this trend reversed as house prices fell. By 2011, the odds of homeownership were only 5.5 times higher among this unconstrained group, and 9.9 times higher among unconstrained middle aged households still a substantial effect, but smaller than observed in 3 The population count from the Current Population Survey is not used because it does not count people living in institutional group quarters. 10

14 earlier years. By contrast, the odds ratio increased among senior households, rising from roughly 5.0 in 2005 to 8.9 in 2011, meaning affordability is an increasing concern for older households To further aid in their interpretation, these results can be converted into average marginal effects: the predicted homeownership rate is calculated first when every observation is simulated as unconstrained by the affordability standard, and then when every observation is simulated as constrained. In Table 5, the average marginal effects are equivalent to the estimated difference between these two rates. Among younger households, the difference in affordability status was equivalent to almost 41 percentage points in the homeownership rate in The difference grew to over 46 percentage points in 2005, but fell to 32 percentage points by Among middle aged households the difference has been continuously declining, from 39 percentages points in 1997 to 34 percentage points in In contrast, the difference among older households has been generally increasing from 17 percentage points to 26 percentage points. Weighting these marginal effects by the size of each age group, the homeownership rate among households that were not constrained by the affordability measure was almost 32 percentage points higher on average in 2011 than for households unable to afford a modestly priced home under common underwriting standards. While substantial, the effect is actually 3.5 percentage points smaller than in Racial Homeownership Gap The odds of homeownership among black households were 45 to 61 percent lower than for non Hispanic white households and generally falling, after accounting for differences in income, wealth, and other households characteristics. Again using a weighted average of the marginal effects to summarize differences across age groups, the homeownership gap between white and black households (presented in Table 6) was nine percentage points in 1997 and increased to twelve percentage points in 2009 before declining somewhat in The gap between white and Hispanic households has bounced around between seven and ten percentage points. However, there was greater evidence of generational differences. Older Hispanic households had odds of homeownership 55 to 65 percent lower than white households in their same age group, while younger Hispanic households had odds only 32 to 43 percent lower. Shift Share Analysis We can simulate the experience of each survey cohort using the estimated coefficients of each survey year. This effectively isolates the effect of the coefficients from the effect of the demographics and socioeconomic characteristics of the population. Figure 3 presents the results of this shift share analysis. Following a given column across cohorts shows the effect of changes in the composition of the cohort, controlling for the effect of the coefficients. For example, the first column (1997 coefficients) shows a decline between 1997 and 2009 before increasing in On the other hand, examining the columns within a cohort shows the effect of the coefficients while controlling for cohort characteristics. For example, within the 1997 cohort, the estimated homeownership rises from 68.5 percent when using the 1997 coefficients to 71.5 percent when using the 2005 coefficients, before falling to 68.3 percent by The darker columns correspond to the estimated homeownership rate for same year cohort and coefficient combinations (i.e., the estimated actual homeownership rate). The shift share analysis 11

15 reveals that the observed changes in the homeownership rate between 1997 and 2011 first rising and then falling were entirely due to changes in the importance of certain socioeconomic characteristics, exemplified by changes in underwriting standards, rather than changes in the characteristics themselves. In fact, the effect of cohort composition has tended to run counter to the observed change in the homeownership rate. Recent Movers The results of the tenure choice model using recent movers are presented in Table 7. Because recent movers are directly subject to prevailing economic conditions, the sub sample provides a more accurate depiction of the effect of changes in underwriting standards and other policies. Among recent movers, the odds of homeownership among households with sufficient income and assets to afford a modestly priced home were roughly twice those of constrained households in 1997 (Table 7). The effect was substantially smaller and not statistically significant in 2005, before increasing slightly in These results support the conventional wisdom that affordability constraints were less binding during the housing bubble. Since then, the strength and statistical significance of the affordability measure has increased, although the effect has not returned to the level seen in the late 1990s. Converting the results to average marginal effects, the difference in affordability status was equivalent to a difference in the homeownership rate of 12 percentage points among recent movers in 1997 (Table 9A). This difference narrowed to 5 percentage points in 2005 before increasing again to 6.5 percentage points. The percentage point difference among recent movers is smaller than in the general tenure choice model because the homeownership rate in general is lower among recent movers. Only the white black homeownership gap is consistently significant among recent movers. The odds of homeownership were 52 percent lower among non Hispanic black movers relative to non Hispanic white movers in 1997, 41 percent lower in 2005, and then 54 percent lower In terms of percentage points, the gap was 11.6 percentage points in 1997, 8.9 percentage points in 2005, and 11.2 percentage points in The shift share analysis can also be applied to recent movers. Figure 4A confirms that the effect of the model coefficients, controlling for the composition of the cohort, contributed to a higher homeownership rate in 2005 than in either 1997 or Unlike in the general model, the cohort effect reinforced this pattern. Renters The results of the tenure choice model using renters are presented in Table 8. Among renters, the odds of homeownership a year later were 77 percent higher in 1997 when the household had sufficient income and assets to afford a modestly priced home under common underwriting requirements. The effect of the affordability measure diminished in 2005 to 52 percent before increasing slightly again in 2011 to 64 percent, but is statistically significant each year. Average marginal effects are presented in Table 9B, but the comparability of these effects is somewhat compromised by the overall decline in the 12

16 homeownership rate (i.e., a large relative decline indicated by odds ratios may correspond with a small absolute decline indicated by average marginal effects if the base rate has declined substantially). Hence, the odds ratio suggests affordability was more binding in 2011 than in 2005, but the general decline in homeownership means this relative measure corresponds to a smaller absolute difference in homeownership. Again, only the white black homeownership gap is consistently significant, with the odds of homeownership 43 percent to 51 percent lower among black households. The shift share analysis (Figure 4B) suggests that renters were most likely to become homeowners between 1996 and Since then, the effects of both the coefficients and cohort characteristics have worked in concert to lower the rate of transition into homeownership, particularly between 2010 and Homeownership Projections The Joint Center for Housing Studies at Harvard University estimates the future number of households to the year 2035 based on demographics and assumptions about household formation and immigration. These household projections are provided by race/ethnicity, age, and family type categories. Using these estimates, we can also project the future number of homeowners and the homeownership rate by assuming a homeownership rate for each demographic combination. Three survey years from the general tenure choice model were selected to serve as the basis for projecting the future number of homeowners and the homeownership rate. The 1997 survey year serves as a normal homeownership environment, whereas 2005 and 2011 represent the extremes of the housing bubble and bust. These scenarios are phased in between 2013 and 2025 using a weighted 4 average of a base rate derived from the latest SIPP data and the scenario rate for each demographic combination. Changes in homeownership between 2025 and 2035 are driven entirely by changes in household composition projected by the Joint Center. One issue with this projection is that no scenario worse than 2011 is available, when we know that the actual homeownership rate continued to fall. We take two approaches to simulate alternative worst case scenarios. First, following Quercia, McCarthy, and Wachter (2003), we redefine the affordability indicator using a maximum 79 percent loan to value ratio and a 21 percent front end debt to income ratio, which correspond to the average characteristics of new conventional purchase mortgage originations in 2012 according to Ellie Mae s Origination Insight Report. We then apply the coefficients from the 1997 tenure choice model to estimate what the homeownership rate would be if fewer households could afford a modestly priced home under more stringent underwriting standards. The second alternative is based on the proportional decline in homeownership in the recent mover tenure choice models. There was approximately a 20 percent decline in the homeownership rate among 4 0,,12 13

17 recent movers based on changes in model coefficients (see Figure 4A); however, there was substantial variation by population subgroups defined by age, race/ethnicity, and family type. The greatest average decline was among middle aged (45 to 49 years old) black non married households with children. The homeownership rate among recent movers in this group declined from 26.1 percent to 16.1 percent, a decline of over 40 percent. On the other hand, some groups actually experienced an increase in their likelihood of homeownership. These proportional declines in homeownership rates are then applied to the estimated homeownership rate for the entire population under the 1997 scenario. This extrapolation is meant to account for the fact that renters are more likely to move than homeowners, meaning recent movers typically have lower homeownership rates than the general population. However, it does not address the fact that the likelihood of moving may have changed disproportionately for some groups. Table 9 presents the projected number of homeowners and the projected homeownership rate by year and immigration scenario. The difference between a homeownership environment like 2005 and one like 2011 corresponds to a difference of 5.3 million homeowners by 2035 under the baseline (medium) immigration scenario. The differences between the 2011 environment and the worst case alternatives corresponds to differences of 3.3 and 11.2 million homeowners in By comparison, the difference between the high and low immigration scenario is roughly only 2.5 million homeowners. While the overall number of homeowners is projected to continue increasing with population growth, the homeownership rate is actually projected to fall in the long term as minorities become a greater share of the population (see Figure 5). Nevertheless, the homeownership environment matters, as the homeownership rate under the 2005 underwriting scenario is projected to be 68.1 percent in 2035, compared to 64.6 percent under the 2011 scenario. The worst case scenarios lead to rates as low as 62.3 percent and 57.0 percent. While higher immigration increases the overall number of homeowners, it actually decreases the homeownership rate by increasing the minority, particularly Hispanic, share of the population, given their lower rates of homeownership even after adjusting for other socioeconomic factors. Table 11 breaks down projected levels of homeownership in 2035 by race and ethnic group. The number of non Hispanic white homeowners is projected to range from a high of 68.6 million, based on the 2005 homeownership rate scenario and high immigration rates, to 65.7 million, assuming low immigration and the 2011 scenario. The worst case alternative further lowers the estimate to just 55.8 million. The homeownership rate for white households could be as high as 77.4 percent or as low as 63.5 percent. The number of non Hispanic black homeowners ranges from 10.2 to 9.4 million using the 2005 or 2011 scenarios, respectively, or as low as 8.6 million under the lowest projection. The homeownership rate for black households could be as high as 50.9 percent or as low as 43.6 percent. Hispanics show the greatest variation in the projected number of homeowners. Assuming high immigration and the 2005 scenario, the number of Hispanic homeowners could be as high as 15.1 million, while using low immigration and the 2011 scenario, the number of Hispanic homeowners would reach just 12.2 million, a difference of 2.9 million. Under the worst case scenario the number of owners would further fall to 11.6 million. Across these three scenarios the homeownership rate for Hispanic households could be as high as 54.2 percent or as low as 44.2 percent. 14

18 Conclusion The future of American homeownership is being pushed in contradictory directions by demographic changes. The aging of the Baby Boomers would support higher levels of homeownership, but the increasing minority share of the population and the persistent racial gap in homeownership acts to reduce homeownership. Layered on top of these demographics is an uncertain mortgage underwriting environment affected by policy decisions and general economic conditions. This paper found that housing affordability, defined as sufficient income and assets to purchase a modestly priced home under common underwriting standards, had a statistically significant effect on the likelihood of homeownership. Estimates from the general population are confounded by the endogeneity of wealth, but a sub sample of recent movers confirmed that the effect of constraint was greatest in the late 1990s and weakest at the height of the housing bubble. The tenure choice model found a strong and possibly widening homeownership gap between white and black households, even after controlling for other socioeconomic factors. However, the gap between white and Hispanic households, particularly among younger households, was less evident. Depending on assumptions concerning immigration and the underwriting environment, this paper projects that the number of homeowners will increase from roughly 82 million in 2015 to between 83 and 102 million by The homeownership rate is not expected to be substantially different from today s rate of almost 65 percent under normal homeownership environments, but could reach as high as 68 percent if the loose underwriting environment of 2005 recurs, or as low as 57 percent if the current restrictive environment is institutionalized. Future research should attempt to incorporate credit score and address issues of endogenous wealth and household formation. The former can be addressed by examining only recent movers, but must consider the selection bias in such a sub sample. Households that are more likely to move are likely to be systematically different, and the likelihood of moving is likely to vary over time based on economic conditions. Similarly, household formation is likely to vary over time. These changes are likely to be correlated with tenure choice. One approach might be to model the owner headship rate used by Yu and Myers (2010) instead of the traditional homeownership rate. Acknowledgements The authors would like to acknowledge the following for their assistance in undertaking this research: George McCarthy and the Ford Foundation for providing financial support, Daniel McCue and Eric Belsky for their contributions to the analysis, and Janneke Ratcliffe for editorial comments and suggestions. Works Cited Belsky, Eric S. The Dream Lives On: The Future of Homeownership in America. Cambridge, MA: Joint Center for Housing Studies, Harvard University,