Effect of homeownership on Unemployment Status and Unemployment Duration Takuya Hasebe∗ Ph.D Program in Economics, City University of New York Graduate Center

Catherine Lau† Economics and Finance Departments, Carthage College

THIS IS A PRELIMINARY DRAFT. PLEASE DO NOT CITE.

June 30, 2013

Abstract Using data from the Current Population Survey (CPS), we examine the relationship between homeownership and unemployment in order to better understand the current U.S. labor market. Our research adds to the previous body of work by controlling for the endogeneity of being a homeowner and the selection into unemployment for the analysis of unemployment duration: we estimate the probabilities of being a homeowner and of being unemployed and the duration of unemployment simultaneously. Our results show that the duration of unemployment is longer for homeowners than renters while homeowners are as likely to be unemployed as renters. We use the data from the year of 2012 for our benchmark analysis, but also extend our work to include the data from the year prior to the financial crisis of 2008. We find that homeowners and renters are adversely affected equally.

Keywords: Homeownership, Unemployment, Mobility JEL classification: J61, J64, R2 ∗

Email:[email protected] Email:[email protected] The authors thank Leigh Ann Leung and participants at the 39th Eastern Economic Association Annual Conference for their valuable comments. †

1

1

Introduction

In general, homeownership has been viewed in a positive light, increasing stability of families and neighborhoods, and causing investment in properties. Additionally, there is a dearth of good quality rental stock in many areas of the United States, increasing the importance of homeownership. One of the ways of promoting homeownership in the United States is through permitting mortgage interest payments to be deducted from federal income tax liabilities. However, this tax treatment of mortgage interest has come under increasing scrutiny in recent months due to its regressive nature, the need to reduce the U.S. government deficit, and the high amount of mortgage debt it encourages. It is therefore of particular interest to look at other possible effects of homeownership. Homeownership may reduce mobility, which in turn may impair the efficient functioning of the labor market. There are two main causes for this lower mobility. The first concerns personal characteristics that cause a person to self select home ownership: people who choose home ownership tend to want to stay in one location. The second has to do with the higher moving costs associated with both buying and selling a home which lowers the probability that homeowners will move for a job opportunity. Since costs of selling a house also depend on homeowners’ equity, home equity can affect mobility and labor market outcomes. The current stubbornly high unemployment rate in the United States (7.9%, October 2012, Bureau of Labor Statistics), may also be tied to the housing market: under the current environment of low home prices and tight mortgage market conditions, the economic cost of selling a home may make unemployed homeowners less mobile than they would be in stronger housing markets. This would strengthen any positive relationship between homeownership and both probability of unemployment and duration of unemployment. Previous research examining the correlation of homeownership and unemployment has produced mixed results on the sign of the relationship.

1

Using a micro-level data from the Current Population Survey (CPS), we examine the effects of being a homeowner on both the probability of being unemployed and unemployment duration. Our research adds to the previous body of work by controlling for the endogeneity of homeownership. We also consider the sample selection issue of unemployment duration by examining the probability of being unemployed and the duration of unemployment simultaneously. When we employ joint estimation to control for the endogeneity and the selectivity, the effect of homeownership on the probability of being unemployed is insignificant while homeownership statistically and economically significantly increases unemployment duration. While we use 2012 as our benchmark analysis, we extend our work to include the year before the financial crisis of 2008 to see how the crisis affects our analysis. We find that the financial crisis affects homeowners and renters equally. This paper is organized as follows. Section 2 provides the brief review of the literature. Section 3 outlines the empirical methodology of this paper. Section 4 explains the dataset followed by the estimation results in Section 5. Section 6 concludes the paper.

2

Literature Review

The existing body of literature on the relationship between homeownership and unemployment is inconclusive on the sign of the relationship. The idea that homeownership was positively correlated with unemployment was first espoused by Oswald (1996), who theorized that the high homeownership percentage in industrialized Europe impaired mobility, leading to high unemployment. He found that countries with higher rates of homeownership had higher rates of unemployment: he credited this to the higher costs of moving for homeowners restraining their ability to accept jobs which necessitated a household move. Oswald and Blanchard’s current research (2013), reinforces and expands Oswald’s original hypothesis by using AR lags to demonstrate a link between high levels of home ownership in an area and later high levels of joblessness in that same geographical area. Their research 2

also shows that states with higher levels of home ownership have lower mobility, longer commute times, and lower rates of business formation, all of which could be tied to unemployment, though no attempt is made to link these three variables to unemployment. To delve further into this theory, we first look at research which focuses solely on homeowners’ mobility. Genosove and Mayer (2001) , Chan (2001), and Ferreira et al. (2010) focus on a lock-in effect: homeowners mobility is constrained by changes in home values. The first two articles are limited in geographic scope, while Ferreira et al. (2010) uses a national dataset. Genosove and Mayer (2001) find that Boston condominium owners list homes at prices higher than market values after a fall in prices, thereby extending the duration of the sales period, reducing mobility. Loss aversion as the mechanism for this lower mobility is identified through the relationship among asking prices, market prices and original purchase prices: those homeowners who will suffer a nominal loss, i.e., the market price at the time they are selling their condominium is lower than their purchase price, list their homes at higher prices relative to the market than owners who will not suffer losses. Loss aversion is an extension of prospect theory for Genosove and Mayer (2001). Chan (2001) employs Chemical Bank mortgage data from November 1989 and January 1994 and finds that homeowners whose mortgage loans are greater than 50% of their home values are less likely to move, with the effect increasing as loan to value (LTV), rises. Chan (2001) attributes the lower probability of a move, the lock-in effect, to the inability of these homeowners to pay off their current mortgages and have a down payment for a new home from the proceeds if they were to sell their homes. Ferreira et al. (2010) also examine the effect of house values on mobility; they employ the American Housing Survey data from 1985 to 2007, and develop a latent index to capture the various variables affecting households decisions to move. Through a probit model, they find negative equity having the largest negative effect on mobility, but falling house prices and increasing interest rates also reduce mobility. Their argument is one of financial frictions decreasing mobility. To control for possible measure3

ment error and the resulting attenuation bias from the self-reported nature of house values in the AHS, Genosove and Mayer (2001) use IV for house values using purchase price and appreciation/depreciation from the Freddie Mac repeat sales price index. Engelhart (2003) also IVs for self-reported home values with variables based on Freddie/Fannie indices of average MSA appreciation rate, and finds that the sign on the coefficient for loss in home value changes when the index is employed. That is, when he uses self-reported home values from the NLSY79, declines in home value increase the probability of a household moving, but when he uses a home price index as an instrumental variable, declines in the index cause the likelihood of moving to go down. Henley (1998)’s results add to this literature by showing a nonlinear relationship between equity and mobility. He employs a hazard model on the probability that a residence spell ends using owner occupiers in 1991 from the British Household Panel Survey (BHPS), and finds the hazard rate converges on zero at over 20,000 pounds of negative equity and rises rapidly to 10% at zero equity before starting to fall slowly as equity increases. The coefficient on regional unemployment is negative, signifying less mobility for homeowners in areas with high unemployment. Coulson and Grieco (2013) add to the literature on the mobility of homeowners by looking at both mobility at different equity levels and comparing renters to homeowners. Though their results also find homeowners less mobile than renters, in direct contradiction to the earlier research (Genosove and Mayer, 2001; Chan, 2001; Ferreira et al., 2010), Coulson and Grieco (2013) find that homeowners with high LTVs are more likely to move. This relationship is most pronounced for those with negative equity located in the boom states of California, Nevada, Florida, and Arizona. These contradictory results could stem from a number of sources. Coulson and Grieco (2013) use self-reported home values for their equity calculation while the other research IV’s for self reported home values; Engelhart (2003) actually finds negative changes in house values increase the probability of a move when he uses self-reported home values. (These perplexing results when IV is employed is 4

one of the reasons we do not IV in our research.) Secondly, the results of Coulson and Grieco (2013) may point to an endogeneity issue that none of these papers explicitly addresses: there could be an unobserved variable that simultaneously causes a homeowner to have a larger mortgage relative to his home price and be more likely to move, e. g., a characteristic of being less risk averse. Another possible reason lies in the recent nature of the mortgage market. Since the market rose and fell rapidly, homeowners who are finding themselves underwater may only have been in their homes a short time and not had the time to build neighborhood relations which are thought to increase the stability of homeowners. While time in home is a control variable Coulson and Grieco (2013), it has a large standard deviation and therefore the effect could be driven by those with the shortest time in their homes. Additionally, some of the moves could be the direct result of foreclosures. In summary, this strand of literature finds that homeowners are overall less mobile than renters; if lower mobility harms labor market outcomes, this research prima facie lends support to the idea that homeownership and unemployment or the duration of unemployment should be positively correlated. To see whether this lower mobility of homeowners causes the expected frictions in the labor market, we next review research on unemployment and homeownership, which also ties more closely to our work. The work of Green and Hendershott (2001) finds a positive correlation between unemployment and homeownership rates, at least for the middle-aged, supporting the earlier work by Oswald. Green and Hendershott (2001) look at the relationship between unemployment and homeownership both at the state and household level in the United States; their research mainly supports Oswald. Coulson and Fisher (2002), on the other hand, reject Oswald’s hypothesis that homeownership negatively affects the labor market though models that use both macro and micro level data. They find the unconditional and conditional probability of unemployment is lower for homeowners than for renters, and unemployment spells are longer for renters than owners. Neither Oswald (1996) nor Green and Hendershott (2001) nor 5

Coulson and Fisher (2002) control for selectivity into homeownership. However, in a later paper, Coulson and Fisher (2009), a two-stage model is employed which first predicts homeownership, then uses the predicted value to test against unemployment and wages. To get a predicted value for homeownership, they use the state marginal tax rate on the mortgage interest deduction, reasoning that the higher the tax rate, the more valuable the deduction, the more likely that residents become homeowners. However, other research by Hilber and Turner (2010), has shown that the mortgage interest tax deduction is a poor predictor of home ownership, in some cases having a negative effect on ownership rates. When van Leuvensteijn and Koning (2004) model the probability of job mobility and homeownership simultaneously, they find homeowners less likely to become unemployed which they attribute to the higher income reduction of unemployment for homeowners versus renters. They do not find an effect of homeownership on job to job mobility or probability of leaving the labor force. Munch et al. (2006) similarly control for selectivity through a bivariate distribution, but also employ a competing risks model with two different employment destinations from unemployment. That is: transitions from unemployment to employment can occur in either the local market or in a market which would require a move. This enables them to test the theory that it is the higher relative cost of moving that keeps homeowners from moving, not being a homeowner per se. They find that homeowners are less likely to transition into jobs which would require a move, but more likely than renters to transition to jobs in the local market. The latter dominates so that, overall, homeowners have shorter spells of unemployment than renters. The key therefore seems to be in distinguishing between employment that does not necessitate a move and employment that does. There are two forces working on homeowners: one lowers their probability of moving; the other increases their probability of being employed. Whichever dominates determines the sign on the coefficient of homeownership in a model 6

predicting employment status. van Leuvensteijn and Koning (2004) limit their work to the probability of becoming unemployed; Munch et al. (2006) focus on the duration of unemployment once unemployed. We contribute to this research by looking at the relationship between homeownership and both unemployment status and unemployment duration simultaneously.

3

Empirical Methodology

The main question of this paper is whether homeownership has an impact on the unemployment status and unemployment duration. There are two important empirical issues which need to be addressed. First, the status of homeownership is potentially endogenous to the outcome we are interested in: unemployment status. Many of the traits associated with homeownership, such as education, marital status also affect employment. While we can control for these known variables, there is also the possibility of an unquantifiable trait that makes a person both more likely to be employed and a homeowner, perhaps an innate stability. Prior research has mainly attempted to control for this endogeneity through the use of instrumental variables. van Leuvensteijn and Koning (2004) find that the potential for bias from the homeownership coefficient is small when estimating movements into different jobs or out of the labor force, but not into unemployment. That is, they find that unobserved characteristics which make employees less likely to fall into unemployment make them more likely to be homeowners and vice versa. They use the regional homeownership rate as an instrumental variable for homeownership, since there is a strong link between regional homeownership rate and probability of owning a home. However, the regional level of the instrument means the fitted values from the first stage regression will be highly correlated with the aggregate homeownership rate in the second stage of the model. Additionally, the relationship Blanchflower and Oswald (2013) found between the level of homeownership and 7

unemployment makes regional homeownership an inappropriate instrument, at least for the United States. Munch et al. (2006) and Battu et al. (2008) have panel data that allow them to identify the effect of homeownership on transition out of unemployment by observing multiple spells of unemployment for individuals who during some spells are homeowners (treated) and during some, renters (untreated). They argue that this technique identifies the effect of homeownership without the need for an instrumental variable. Second, there is a possibility of a sample selection issue. Clearly, unemployment is not a random event. Unobservable characteristics may affect the probability of being unemployed and unemployment duration. For example, more capable persons may be less likely to lose a job than less capable persons, but once they are unemployed, they may set higher reservation wages, which lead to longer unemployment durations. Ignoring such factors result in inconsistent estimates. Even though the previous studies consider the endogeneity problem of homeownership’s effect on employment status, they have paid little attention to the selectivity issue. Although Flatau et al. (2003) attempt to fix the issue by including the inverse Mill’s ratio, it is not a technically appropriate approach. Moreover, some studies replace actual homeownership status with the predicted probability of being a homeowner, which is obtained from a probit estimation. van Leuvensteijn and Koning (2004) employ a probit equation with homeownership as the dependent variable. This procedure is not appropriate to measure the effect of homeownership, either. In this paper, we adopt a more appropriate econometric model, which controls for both endogeneity and selectivity simultaneously. Our econometric model comprises three equations. One equation is for homeownership status and another equation is for employment status. The last equation is for unemployment duration. The first two equations are formalized as a usual binary choice model. The dummy variable hi indicates homeownership

8

status: for an observation i, i = 1, ..., N ,

hi = 1(Xhi ′ βh + εhi ≥ 0),

where 1(·) is an indicator function: it is 1 if the argument is true, and it is 0 otherwise. Xhi is a set of observable characteristics and εhi is an unobservable error. Likewise, the dummy ui indicates whether an observation i is unemployed:

ui = 1(αh hi + Xui ′ βu + εui ≥ 0).

Homeownership appears as an explanatory variable along with other observable characteristics Xui . It may be endogenous when an unobservable error in this equation, εui , is not independent of the error in the homeownership equation, εhi . An unemployment spell is observable only when an individual is unemployed. Let yi be an observed unemployment duration and yi∗ be latent unemployment duration. Then, yi = ui × yi∗ . In this study, we assume that the latent variable yi∗ follows the Weibull distribution conditional on observable characteristics, Xyi , and homeownership status, hi . The distribution function can be written as ∗ 1/γ

Fy (yi∗ ) = 1 − e−(λyi )

,

′

where γ is a shape parameter and λ = e−αy hi −Xyi βy . The expected value of yi∗ is obtained as ′

E(yi∗ ) = γΓ(γ)eαy hi +Xyi βy ,

where Γ(·) is the gamma function. As is clear from this expression, the partial effects of explanatory variables have the same signs as the coefficients αy and βy . We also estimate

9

the model under the assumption that yi∗ follows exponential, gamma, log-logistic, and lognormal distributions. The results from these distributional assumptions are similar to the result under the Weibull distribution assumption.1 The main empirical issue here is that even after controlling for observables, the latent variable yi∗ may not be independent of the error in the homeownership equation, εhi , so that homeownership status hi is endogenous. In addition, when yi∗ is not independent of the error in the unemployment status equation, εui , a sample selection problem arises. As mentioned above, the endogeneity problem may also arise if εhi and εui are not independent. To estimate the model consistently, it is essential to consider a joint distribution of the unobservables. The joint normality of εhi and εui leads to the well-known recursive bivariate probit model (Greene, 2008). In order to allow dependence across yi∗ and these errors, we follow the approach proposed by Lee (1983) and transform a non-normal distribution Fy (yi∗ ) into a normal variate. The transformed variate yei = Φ−1 (Fy (yi∗ )) is assumed to be jointly normally

distributed with εhi and εui : 

 εhi   ε  ui  yei





 

  0   1 ρhu ρhy      ∼ N  0  ,  ρ     hu 1 ρuy     0 ρhy ρuy 1



   ,  

where the variances are set to 1 for identification, and thus, the ρ’s are correlation coefficients. Note that the correlation coefficients ρhy and ρuy do not measure the correlation between the error terms εhi and εui and the latent variable yi∗ , but the correlation between the error terms and the transformed variate yei . However, when these correlation coefficients equal zero, the

errors are independent of yi∗ . In other words, we can check the independence of a pair of

the variables by testing the null hypothesis that the corresponding correlation coefficient is Conceptually, our method described below can be extended to the case where yi∗ is a more flexible Cox-type or discrete distribution, but we leave this as a scope of future research. 1

10

Table 1: Descriptive Statistics of the Year of 2012 Whole Sample Variable HOMEOWNER Unemp. Status Unemp. Durationa AGE FEMALE MARRIED WORKING SPOUSE MSA CHILD18 HS COLLEGE GRAD WHITE BLACK ASIAN HISPANIC Number of Obs.

Mean 0.647 0.067 35.989 40.595 0.448 0.566 0.434 0.816 1.058 0.249 0.536 0.132 0.793 0.118 0.055 0.163

Homeowner

( S.D. ) ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( (

Mean

0.478 ) 0.251 ) 32.965 ) 8.095 ) 0.497 ) 0.496 ) 0.496 ) 0.388 ) 1.167 ) 0.433 ) 0.499 ) 0.338 ) 0.405 ) 0.323 ) 0.228 ) 0.370 )

( S.D. )

0.046 35.976 42.132 0.423 0.689 0.555 0.794 1.127 0.231 0.562 0.154 0.845 0.078 0.050 0.119

34,916

( ( ( ( ( ( ( ( ( ( ( ( ( ( (

0.209 ) 32.862 ) 7.663 ) 0.494 ) 0.463 ) 0.497 ) 0.405 ) 1.160 ) 0.421 ) 0.496 ) 0.361 ) 0.362 ) 0.269 ) 0.218 ) 0.324 )

22,577

Renter Mean 0.107 35.999 37.782 0.494 0.340 0.211 0.856 0.933 0.283 0.490 0.092 0.699 0.192 0.063 0.243

( S.D. ) ( ( ( ( ( ( ( ( ( ( ( ( ( ( (

0.309 ) 33.058 ) 8.107 ) 0.500 ) 0.474 ) 0.408 ) 0.351 ) 1.168 ) 0.451 ) 0.500 ) 0.289 ) 0.459 ) 0.394 ) 0.243 ) 0.429 )

12,339

Unemployment duration in weeks among unemployed workers.

zero.2 By assuming the transformed variate is normally distributed, we estimate the entire model by maximum likelihood estimation. Appendix A describes the maximum likelihood estimation in more detail.

4

Data

Our data source is the Current Population Survey (CPS), a monthly survey of households conducted by the Bureau of Census for the Bureau of Labor Statistics. The CPS provides data on the labor force, employment, unemployment, persons not in the labor force, and other demographic information and is therefore a good source for a study on employment status. We use the March supplements since homeowner status and unemployment duration in weeks are available only in these supplements. However, the spells of unemployment in To measure the dependence across εhi , εui and yi∗ , we can calculate Kendall’s τ , which is straightforward to compute from the correlation coefficients. It is given as τ = 2 sin−1 (ρ)/π. 2

11

the CPS data are right-censored for all observations in the CPS survey since only currently unemployed persons are asked the relevant question. This fact makes identification of the parameters impossible. In order to address this problem, we merge the March supplements with the April data and identify those who are unemployed in March but employed in April. The unemployment spells of those observations are interval-censored within a fourweek period. Otherwise, we treat the spells as right-censored. The unemployment spells of the outgoing rotation group are also right-censored. We restrict the samples to heads of household whose ages are between 25 and 55 years . Table 1 shows the descriptive statistics for the samples from the 2012 survey. We see that approximately 65% of the observations own their homes. The unemployment rate is more than double among renters than homeowners in the sample. This may be partly because homeowners are older and attain higher levels of education than renters. Heads of households with more children under 18 years old tend to be homeowners. Those who live in metropolitan areas are more likely to be renters than those who do not live in MSAs. There are also differences in prevalence of homeownership across racial and ethnic groups. The model estimation using this 2012 sample yields our benchmark result. In addition to this benchmark estimation, we also add the sample from the year 2007, which is prior to the financial crisis of 2008, in order to investigate the impact of the financial crisis. Figure 1 shows the trends of homeownership rate, unemployment rate, and unemployment duration (mean and median) from 2000 to 2012. The share of homeowners increased from 2000, peaked in 2005, and then declined. As of 2012, the share was below the level in 2000. The deteriorating conditions in the labor market caused by the financial crisis can be seen in the unemployment rate, which increased steeply from 2008 to 2010. Then, the unemployment rate started a gradual decline. Unemployment duration, which is measured in weeks, was prolonged after the financial crisis. Interestingly, the unemployment rate for homeowners has been consistently below that 12

0

65

5

%

%

10

70

15

75

Figure 1: Time trend of homeownership rate, unemployment rate, and unemployment duration

60

2000 2000

2002

2004

2006 year

2008

2010

2002

2012

2004

2006 year

Renter

2010

2012

Homeowner

(b) Unemployment Rate

15

10

20

15

week

week 25

20

30

35

25

(a) Home Ownership Rate

2008

2000

2002

2004 Renter

2006 year

2008

2010

2012

2000

Homeowner

2002

2004 Renter

(c) Unemployment Duration (Mean)

2006 year

2008

2010

2012

Homeowner

(d) Unemployment Duration (Median)

for renters. The gap seems to be widening toward the end of 2009. On the other hand, the means and medians of unemployment duration for homeowners and renters have been overlapping.

5

Result

This section shows the estimation results. First, we present the results from the estimation of each equation separately in Table 2. Table 2 shows the result from separate estimations of each equation. The variable of interest, HOMEOWNER, has different impacts on unemployment status and unemployment 13

Table 2: Result from Separate Estimation Unemployment Duration a Variable HOMEOWNER

Coef.

Unemployment Statusb

( S.E. )

0.0092

Coef.

( 0.1642 )

( S.E. )

-0.2995 ***

Home Ownership Statusb Coef.

( S.E. )

( 0.0271 )

AGE

0.0331 ***

( 0.0100 )

0.0021

( 0.0015 )

0.0434 ***

( 0.0010 )

FEMALE

0.5529 ***

( 0.1471 )

0.0347

( 0.0247 )

-0.0810 ***

( 0.0162 )

MARRIED

0.0418

( 0.2916 )

( 0.0357 )

0.4081 ***

( 0.0256 )

WORKING SPOUSE

-0.0817 **

-0.2829

( 0.2647 )

-0.0751 ***

( 0.0287 )

0.4850 ***

( 0.0352 )

MSA

0.0169

( 0.2005 )

-0.0131

( 0.0337 )

-0.1134 ***

( 0.0311 )

CHILD18

0.1061

( 0.0769 )

0.0061

( 0.0108 )

0.0542 ***

( 0.0088 )

HS

0.1571

( 0.2658 )

-0.2409 ***

( 0.0334 )

0.3128 ***

( 0.0264 )

COLLEGE

0.4342

( 0.2995 )

-0.4486 ***

( 0.0351 )

0.5630 ***

( 0.0290 )

GRAD

0.1833

( 0.3670 )

-0.7328 ***

( 0.0481 )

0.7439 ***

( 0.0351 )

WHITE

-0.7075

( 0.4623 )

-0.2160 ***

( 0.0513 )

0.2917 ***

( 0.0556 )

BLACK

-0.1916

( 0.5181 )

( 0.0561 )

-0.3289 ***

( 0.0509 )

ASIAN

-0.3200

( 0.6752 )

-0.2794 ***

( 0.0802 )

-0.0158

( 0.0711 )

-0.0602

( 0.0420 )

-0.3577 ***

( 0.0681 )

HISPANIC

0.3313

( 0.2685 )

UI REPLACE

0.2710

( 2.0755 )

γ

1.1762

( 0.0404 )

ln L Number of Obs.

0.0589

-1,238,17

-8,065.93

-18,014.55

2,353

34,916

34,916

a

The asterisks ***, **, * indicate significance at the 1% level, 5% level, and 10% level, respectively. The unemployment duration equation includes dummy variables for division of residence and a constant term.

b

The equations of homeowner status and unemployment status includes dummy variables for state of residence and a constant term.

duration. The results say that while homeownership has a significantly negative effect on the probability of being unemployed, its effect on unemployment duration is insignificant. All the demographic and economic variables included in the homeownership equation are significant, except for ASIAN. Similarly, most of the socioeconomic variables we include have a significant impact on whether or not an observation is employed, but most of the variables are insignificant in the unemployment duration model. Only age and gender have a significant effect on the duration of unemployment. For example, higher educational attainment makes it less likely for workers to be unemployed, but does not affect the duration of unemployment once a worker is no longer employed. Since we suspect endogeneity of homeownership and selectivity of unemployment, both of 14

Table 3: Result from Joint Estimation Unemployment Durationa Variable HOMEOWNER

Coef.

( S.E. )

0.4702 ***

Coef.

( 0.1739 )

0.0189 -0.0025

AGE

0.0316 ***

( 0.0104 )

FEMALE

0.5186 ***

( 0.1513 )

MARRIED

0.1845

( 0.2911 )

WORKING SPOUSE

Unemployment Statusb

0.0431 * -0.1276 ***

( S.E. )

Home Ownership Statusb Coef.

( S.E. )

( 0.0876 ) ( 0.0018 )

0.0434 ***

( 0.0010 )

( 0.0237 )

-0.0814 ***

( 0.0162 )

( 0.0342 )

0.4080 ***

( 0.0256 )

-0.1923

( 0.2660 )

-0.1225 ***

( 0.0324 )

0.4860 ***

( 0.0353 )

MSA

0.0408

( 0.2186 )

-0.0066

( 0.0319 )

-0.1145 ***

( 0.0312 )

CHILD18

0.0890

( 0.0765 )

0.0007

( 0.0106 )

0.0539 ***

( 0.0088 )

HS

0.5259 **

( 0.2610 )

-0.2712 ***

( 0.0341 )

0.3128 ***

( 0.0263 )

COLLEGE

1.0950 ***

( 0.3224 )

-0.5028 ***

( 0.0375 )

0.5631 ***

( 0.0289 )

GRAD

( 0.4324 )

-0.8027 ***

( 0.0502 )

0.7433 ***

( 0.0349 )

WHITE

-0.4687

1.3269 ***

( 0.4275 )

-0.2399 ***

( 0.0519 )

0.2912 ***

( 0.0553 )

BLACK

-0.3657

( 0.4932 )

( 0.0554 )

-0.3297 ***

( 0.0507 )

ASIAN

0.0256

( 0.6854 )

-0.2726 ***

( 0.0789 )

-0.0156

( 0.0709 )

-0.0262

( 0.0418 )

-0.3575 ***

( 0.0682 )

HISPANIC

0.3979

( 0.2731 )

UI REPLACE

0.0958

( 1.9982 )

1.0556

( 0.0470 )

γ ρuy

-0.7314 ***

( 0.0557 )

ρuh

-0.1878 ***

( 0.0498 )

ln L

0.0940 *

-27310.45

Number of Obs.

34,916

See the footnotes of Table 2.

which could bias our results, we estimate all three equations together as outlined in Section 3. Then, we test the independence of the error terms. Although we can reject the nulls of ρhu = 0 and ρuy = 0, we fail to reject the null ρhy = 0.3 That is to say, when paired, the variables unemployment duration and homeownership are independent, but neither homeownership and unemployment status nor unemployment status and duration of unemployment are independent. The estimation result presented in Table 3 is the model with ρhy = 0. First of all, both of the estimated ρhu and ρuy are negative. The value of ρbuy indeed is

large and indicates a strong dependence between the probability of being unemployed and 3

The test statistics of the Wald and Likelihood Ratio tests are 0.31 and 0.44, respectively. Their p-values are 0.58 and 0.51, respectively.

15

the duration of unemployment. Possible explanations are that more capable persons are less likely to become unemployed, but stay unemployed longer due to setting a higher reservation wage, or due to the scarcity of jobs at higher skill levels. Second, correcting potential endogeneity and selectivity yields an interestingly different result from previous research which ignored such econometric issues. On the one hand, in contradiction to previous results, being a homeowner has a statistically significantly positive effect on unemployment duration. The size of the effect is also economically significant. Compared with a renter, a homeowner stays unemployed longer by about 73% (= e0.5464 −1). This result is in direct contradiction to the overall results of Munch et al. (2006) who find homeowners have a 34% higher transition rate out of unemployment than do renters. However, when Munch et al. (2006) employ a competing risks duration model, which distinguishes between finding a job locally and finding a job in a different market, they do find that homeowners are more likely to find a job locally than renters, but less likely to transition to a job in a different market. The extension of our research to allow competing risks is one of our future tasks. On the other hand, the significance of homeownership in the unemployment status equation disappears after considering endogeneity. Coulson and Fisher (2002) find a significantly negative relation between homeownership and unemployment, which do not consider the potential endogeneity of being a homeowner. Once endogeneity is corrected by an IV probit approach, Coulson and Fisher (2009) find a marginally significantly negative effect on employment probability. Our finding contradicts these findings. Next, we extend our analysis to investigate the effect of the financial crisis of 2008: whether the recent financial crisis affects homeowners and renters differently. Since the financial crisis has its roots in the housing market collapse, it might affect renters and homeowners differently, reducing the mobility of homeowners whose home values declined. Our extension is simply adding the sample of the year before the crisis and comparing the 16

Table 4: Estimation Results from the 2007 and 2012 samples Unemployment Duration Variable

Coef.

Unemployment Status

( S.E. )

Coef.

HOMEOWNER

0.5464**

( 0.2186 )

Year of 2012

0.4623**

( 0.1813 )

HOMEOWNER × Year of 2012

-0.0598

( 0.0249 )

( 0.2317 )

0.0046

( 0.0333 )

( 0.0073 )

-0.0006 0.0260

0.0244***

FEMALE

0.4815***

( 0.1225 )

MARRIED

0.1884

( 0.1993 )

-0.1297***

Coef.

( S.E. )

( 0.1368 )

0.3337***

AGE

WORKING SPOUSE

-0.1470

( S.E. )

Home Ownership Status

-0.1527***

( 0.0107 )

( 0.0022 )

0.0429***

( 0.0007 )

( 0.0174 )

-0.0767***

( 0.0112 )

( 0.0354 )

0.4729***

( 0.0179 )

-0.1285

( 0.2017 )

-0.0946***

( 0.0328 )

0.4434***

( 0.0173 )

MSA

0.0893

( 0.1463 )

-0.0650***

( 0.0234 )

-0.0904***

( 0.0154 )

CHILD18

0.1055*

( 0.0540 )

( 0.0081 )

0.0668***

( 0.0051 )

HS

0.3164*

( 0.1782 )

-0.2892***

( 0.0309 )

0.3479***

( 0.0206 )

COLLEGE

0.7721***

( 0.2167 )

-0.4888***

( 0.0360 )

0.5956***

( 0.0199 )

GRAD

( 0.3819 )

-0.7716***

( 0.0489 )

0.7446***

( 0.0250 )

WHITE

-0.8161

( 0.3386 )

-0.2359***

( 0.0432 )

0.2952***

( 0.0291 )

BLACK

-0.5063

( 0.3616 )

( 0.0480 )

-0.2802***

( 0.0327 )

ASIAN

-0.2339

( 0.4688 )

-0.2491***

( 0.0573 )

-0.0270

( 0.0366 )

0.3631**

( 0.1663 )

-0.0252

( 0.0304 )

-0.3671***

( 0.0163 )

0.9530

( 0.0658 )

HISPANIC γ

1.0492***

0.0127

ρuy

-0.7333**

( 0.1022 )

ρuh

-0.0971

( 0.0793 )

ln L

0.0460

-51,791.14

Number of Obs.

72,819

See the footnotes of Table 2.

year effect and its interaction with HOMEOWNER. Here, we add the year of 2007. We also try the year of 2005 and the year of 2006, and the results are similar to the result presented here.4 Table 4 presents the result of this estimation. The results show that workers are less likely to be homeowners but more likely to be unemployed in 2012 than in 2007, and that unemployment duration tends to be longer in 2012 than in 2007. However, the interaction 4

Furthermore, we previously pooled the samples from 2000 to 2012. We created a dummy variable for the crisis, which is defined as the year 2008 and after, and we interact it with homeowner status. This also yields similar results, but we are unable to consider the potential endogeneity of HOMEOWNER in the unemployment equation due to a computational reason. The evaluation of the trivariate normal distribution with a large data set takes an incredibly long time.

17

term between the year 2012 dummy and HOMEOWNER is insignificant, which indicates that the financial crisis adversely affected the labor market outcomes of renters and homeowners equally. These results are supported by Farber (2012) who claims that the hypothesis that the financial crisis made it more difficult for unemployed homeowners to move for a new job is not consistent with the mobility patterns over years. Neither our study nor the study by Farber (2012) examines possible cross-sectional variations of the impact of the financial crisis. Although overall housing values declined since the financial crisis, the rates of decline are not geographically uniform. Exploiting such variations may reveal some interesting aspects of the impact of the financial crisis on the labor markets. This is currently our primary follow-up task.

6

Conclusion

It is clear from past research that homeownership in general reduces mobility, and that this reduced mobility causes frictions in the labor market. What has not been previously made clear is whether this results in higher probability of unemployment and longer duration of unemployment for homeowners versus renters. In our separate estimations of homeownership status, unemployment status, and unemployment duration, homeownership does not have a significant effect on the duration of unemployment, but homeownership is significantly and negatively correlated with the probability of being unemployed. However, when we control for endogeneity and selectivity through joint estimation of homeownership status, unemployment status, and unemployment duration, homeownership no longer has a statistically significant effect on the likelihood of being unemployed. Yet a homeowner remains unemployed on average approximately 73% more weeks than a renter. The conflicting results among our research and previous research may support the economic theory of rational behavior. When homeowners are significantly underwater on their mortgages and residing in states experiencing severe economic downturns as in Coulson and Grieco (2013), 18

the probability that they move increases. This could lead to a negative correlation between homeownership and unemployment duration. When homeowners reside in diversified economic areas like Boston Genosove and Mayer (2001) or New York Chan (2001), they are more likely to wait out a housing market downturn. This could lead to a positive correlation between homeownership and unemployment duration. An area for further study is therefore to look at homeowners’ mobility cross-sectionally, controlling for changes in home values and other geographic specific economic variables.

Appendix A

Maximum Likelihood Estimation

This appendix describes the maximum likelihood estimation. As mention ed in the text, we assume that εhi , εui , and yei are jointly normally distributed. Under this distributional assumption, we implement maximum likelihood estimation.

If an observation is employed, the contribution to the likelihood by this observation, Li , is

Li =

"Z ×

−Xhi ′ βh −∞

"Z

Z

∞ −Xhi ′ βh

−Xui ′ βu

φ2 (εh , εu , ρhu )dεu dεh −∞

Z

#1−hi

−αu −Xui ′ βu

φ2 (εh , εu , ρhu )dεu dεh −∞

# hi

where φ2 (·) is a pdf of the standard bivariate normal distribution with the correlation coefficient ρhu . As described in Section 4, an unemployment spell is either right-censored and interval

19

censored for an unemployed observation. The contribution to the likelihood by an observation with a right-censored unemployment spell yi is;

Li =

"Z ×

−Xhi ′ βh Z ∞

∞

−Xui ′ βu yei

−∞

Z

Z

∞

Z

∞

φ3 (εh , εu , ye; Σ)de y dεu dεh Z

#1−hi

∞

−Xhi ′ βh −αu −Xui ′ βu yei

φ3 (εh , εh , ye; Σ)de y dεu dεh

 hi

where φ3 (·) is a pdf of the trivariate standard normal distribution with the covariance (correlation) matrix Σ. The elements of this matrix are in Section 3. The contribution to the likelihood by an observation with an interval-censored unemployment spell is;

Li =

"Z ×

−Xhi ′ βh Z ∞

yei+

−Xui ′ βu yei

−∞

"Z

Z

∞

Z

∞

φ3 (εh , εu , ye; Σ)de y dεu dεh

Z

#1−hi

yei+

−Xhi ′ βh −αu −Xui ′ βu yei

φ3 (εh , εh , ye; Σ)de y dεu dεh

# hi

,

where yei+ = Φ1 (Fy (yi+ )) and yi+ is the upper limit of the interval.

Prieger (2002) discusses an alternative parametric approach to the sample selection model

of a duration outcome. Its limitation is that it allows only a moderate degree of dependence between the errors. Our results show the degree of dependence beyond it. Actually, our approach and the approach by Prieger (2002) are special cases of the copula approach (Smith, 2003). See also Trivedi and Zimmer (2007) for the copula method. We will consider a further sophisticated econometric methodology such as allowing multiple destinations (competing risks) in future research.

20

References Battu, H., Ma, A., Phimister, E., 2008. Housing tenure, job mobility and unemployment in the UK. Economic Journal 118 (422), 311–328. Blanchflower, D. G., Oswald, A. J., 2013. Odes high home-ownership impair the labor market? Chan, S., 2001. Spatial lock-in: Do falling house prices constrain residential mobility? Journal of Urban Economics 49, 567–586. Coulson, N. E., Fisher, L. M., 2002. Tenure choice and labuor market outcomes. Housing Studies 17, 35–49. Coulson, N. E., Fisher, L. M., 2009. Housing tenure and labor market impacts: The search goes on. Journal of Urban Economics 65, 252–264. Coulson, N. E., Grieco, P. L., 2013. Mobility and mortgages: Evidence from the PSID. Regional Science and Urban Economics 43, 1–7. Engelhart, G., 2003. Nominal loss aversion, housing equity constraints, and household mobility:evidence from the United States. Journal of Urban Economics 49, 171–195. Farber, H. S., 2012. Unemployment in the great recession: Did the housing market crisis prevent the unemployed from moving to take jobs? American Economic Review 102, 520–525. Ferreira, F., Gyourko, J., Tracy, J., 2010. Housing busts and household mobility. Journal of Urban Economics 68, 34–45. Flatau, P., Forbes, M., Hendershott, P. H., Wood, G., 2003. Homeownership and unemployment: The roles of leverage and public housing. Working Paper 10021, NBER. Genosove, D., Mayer, C., 2001. Loss aversion and seller behavior: Evidence from the housing market. The Quarterly Journal of Economics 116, 1233–1260. Green, R. K., Hendershott, P. H., 2001. Home-ownership and unemployment in the US. Urban Studies 38, 1509–1520. Greene, W. H., 2008. Econometric Analysis, 6th Edition. Pearson Prentice Hall, Upper Saddle River,N.J. Henley, A., 1998. Residential mobility, housing equity, and the labour market. Economic Journal 108, 414–427. Hilber, C. A. L., Turner, T. M., 2010. The mortgage interest deduction and its impact on homeownership decisions. Discussion Paper 55, SERC. 21

Lee, L.-F., 1983. Generalized econometric models with selectivity. Econometrica 51 (2), 507– 512. Munch, J. R., Rosholm, M., Svarer, M., 2006. Are homeowners really more unemployed? Economic Journal 116, 991–1013. Oswald, A. J., 1996. A conjecture on the explanation for high unemployment in the industrialized nations: Part 1. Working Paper 475, University of Warwick. Prieger, J. E., 2002. A flexible parametric selection model for non-normal data with application to health care usage. Journal of Applied Econometrics 17 (4), 367–392. Smith, M. D., 2003. Modelling sample selection using Archimedean copulas. Econometrics Journal 6 (1), 99–123. Trivedi, P. K., Zimmer, D. M., 2007. Copula modeling: An introduction for practioners. Foundations and Trends in Econometrics 1 (1), 1–111. van Leuvensteijn, M., Koning, P., 2004. The effect of home-ownership on labor mobility in the Netherlands. Journal of Urban Economics 55, 580–596.

22