Upload
others
View
70
Download
0
Embed Size (px)
Citation preview
Teachers Teaching Teachers: The Role of Networks on
Financial Decisions∗
Gonzalo Maturana† Jordan Nickerson‡
January 2016
This paper studies the role of peers in the transmission of information pertinent to household
financial decisions. Specifically, we examine the effect of peers on the mortgage refinancing decisions
of school teachers in Texas by exploiting commonalities in teacher schedules to identify peer groups.
Using this source of within-campus variation to separate campus-level unobservable shocks, we find
that refinancing activity among a teacher’s peers increases her likelihood of refinancing by 9.5%.
Peers also affect a teacher’s choice of lender. Overall, our findings suggest that a household’s cost
of acquiring and processing financial information is greatly reduced by interacting with peers.
JEL classification: D71, G21, H31, R20
keywords : Household Finance, Peer Effects, Mortgages, Refinancing
∗We are grateful to Sumit Agarwal, Felipe Aldunate, Ian Appel, Francisco Barillas, Tarun Chordia,Slava Fos, Andreas Fuster (discussant), Todd Gormley, Clifton Green, John Griffin, Edith Hotchkiss, DalidaKadyrzhanova, Oguzhan Karakas, Benjamin Keys, Darren Kisgen, Sam Kruger, Mauricio Larrain, KyleMangum, Dmitriy Muravyev, Christopher Parsons, Tomasz Piskorski, Oliver Randall, Jonathan Reuter,Breno Schmidt, Jay Shanken, Martin Schmalz, Phil Strahan, Daniel Weagley, Vincent Yao, as well as theseminar participants at Boston College, Emory University, Georgia State University, Pontificia UniversidadCatolica, the 2014 FRA Early Ideas session, the 2015 All-Georgia Finance Conference, and the 2015 AtlantaFED Real Estate Conference for their helpful comments. We also thank Integra FEC for providing uswith refinancing activity data. Additional results are available in an Online Appendix at http://www.
GonzaloMaturana.com or at http://www.JordanNickerson.com.†Emory University, Goizueta Business School, 1300 Clifton Road, Atlanta, GA 30322; telephone:
(404)727-7497. E-mail: [email protected]‡Boston College, Carroll School of Management, 140 Commonwealth Avenue, Chestnut Hill, MA 02467;
telephone: (617)552-2847. E-mail: [email protected]
In recent years, economists have become increasingly interested in how households use finan-
cial instruments to achieve their objectives (Campbell (2006)). Several studies have noted
that households often appear to commit financial mistakes that are inconsistent with rational
models and even seem to contradict practical intuition.1 Thus, understanding the factors
that influence a household’s financial decision-making process is important from a welfare
perspective. One potential explanation for these mistakes is the presence of informational
frictions such as the costly acquisition and processing of pertinent information. Therefore,
this paper studies the exogenous arrival of otherwise costly information and its effect on
household financial decision making.
The seemingly suboptimal mortgage refinancing decisions made by a significant portion
of households is of particular interest in household finance, given the relative importance of
home value in the average household’s wealth.2 Many borrowers refinance at a detrimental
rate, while others wait too long or do not refinance at all (e.g., Agarwal, Driscoll, and Laib-
son (2013); Andersen et al. (2015)). The need to acquire and process information related to
current lending rates and the estimated present value of future savings makes the mortgage
refinancing decision an ideal setting in which to evaluate how information acquisition affects
household financial choices. More specifically, the transmission of information from an indi-
vidual’s peer network provides an interesting laboratory in which we can identify a reduction
in the cost of information and its effect on an individual’s decision-making process.3 Thus,
we study the impact of an individual’s employment network on her mortgage refinancing
decision.
However, studying the information generated by peer networks and their effect on house-
1. For example, many individuals take out excessively expensive loans (Agarwal, Skiba, and Tobacman(2009)) and do not maximize employer-matched 401(k) contributions (Choi, Laibson, and Madrian (2011)).
2. In 2010, home value comprised 43.1% of gross assets for individuals in the U.S., representing the largestportion of household wealth, compared to 28.5% in equities and pension accounts combined (Wolff (2012)).
3. Notably, while peer effects have been explored at length as they relate to the financial decisions ofother agents such as the firm (Leary and Roberts (2014)), financially affluent individuals such as CEOs(Shue (2013)), mutual fund managers (Cohen, Frazzini, and Malloy (2008); Pool, Stoffman, and Yonker(2015)), and sell-side analysts (Cohen, Frazzini, and Malloy (2010)), little is known about the role thatnetworks play in a household’s mortgage refinancing decision.
1
hold financial decisions is challenging. First, the identification of peer effects requires that
the selection of peers does not correlate with their subsequent actions. Second, individuals
in the same peer group may be subject to common, unobservable shocks. Any influence
of these shocks on the outcome variable would lead to biased estimates of the peer group’s
effect.
To address these difficulties, we use a novel micro dataset of employment records for all
public school teachers in the state of Texas, which we merge with county deed records to
obtain their refinancing transactions. While teachers have a choice over the school district to
which they apply, they have little or no influence on the campus to which they are assigned
within the school district. In addition, we observe a second layer of quasi-randomization
that exists at the campus level, and stems from class schedule assignments.
This richness of the data permits us to construct a measure of a teacher’s peer group
based on teaching assignments. More specifically, we identify each teacher’s off-periods (i.e.,
periods when the teacher is not in her classroom) and consider two teachers to be in each
other’s peer group if they have at least 40 minutes of overlap in off-periods per day. This
classification allows for the identification of teachers who are more likely to share information.
Furthermore, this formation of peer groups is unlikely to be correlated with refinancing
decisions. The resulting variation in peer groups both across and within campuses allows us
to control for common unobservable shocks to peer groups, such as district-level employment
effects and variations in loan supply exposure at the campus level.4 To further validate our
empirical strategy, we provide evidence consistent with teachers not choosing their schedule
and controlling their off-periods.
We find a strong effect of peers on the refinancing decisions of teachers. A teacher is 16.2%
more likely to perform a mortgage refinance following a one standard deviation increase in
the refinancing activity among teachers who share the same off-period. In addition, the
effect is not driven by variation in loan supply exposure or by another omitted variable at
4. This identification strategy is similar in spirit to that of Mas and Moretti (2009), which exploits theoverlap in work shifts of cashiers in a national supermarket chain.
2
the campus-level; it remains equal to 9.5% even after the inclusion of campus-month fixed
effects.5 Furthermore, we find the effect to be stronger when peers share common ethnicity.
The previous identification strategy accounts for variation in loan supply exposure at
the campus level, such as the construction of a new billboard near a campus that advertises
a lender’s superior mortgage terms. Our tests essentially compare the effect of a teacher’s
refinancing activity on other teachers who share a common off-period relative to the effect
on the remaining teachers at the same school who do not share an off-period. However,
it does not account for the existence of an unobservable factor at the off-period level that
affects both the outcome variable and the refinancing decisions of peer groups. To address
this unlikely possibility, we also undertake an instrumental variable (IV) approach. More
specifically, we use the average potential savings (realized upon refinancing) of a teacher’s
peers as an instrument for the peer group’s refinancing activity. Intuitively, after explicitly
controlling for a teacher’s savings, her likelihood of refinancing should not be affected by her
peers’ average savings; therefore, the exclusion restriction should hold. Using this approach,
we confirm the previous findings.
To further validate the effect of a teacher’s peers, we turn to another dimension of the
refinancing decision: the teacher’s choice of lender. It is unlikely that information transferred
from a teacher’s peers is restricted to the prevailing rate while other pertinent information,
such as the lender offering such terms, is not disclosed. To investigate whether individuals
who seek to refinance their loan are more likely to choose the same lender as their peers, we
undertake a nonparametric approach. Specifically, we generate a benchmark set of pseudo-
peer groups that are geographically similar to the true peer groups but fall under the null
hypothesis of no information sharing within the peer group. We find the probability that
a teacher and her peer share a common lender is 8.16%, or 12.9 standard deviations larger
than the expected overlap based on the pseudo-peer groups. This strongly suggests that a
5. This result is robust to alternative constructions of peer refinancing activity and estimated savings, andbecomes stronger when excluding months when teachers are less likely to interact. We also confirm the mainresults using a hazard model to account for the dynamic nature and infrequent occurrences of refinances.
3
teacher’s choice of lender is also influenced by her peer group.
After establishing the relevance of peer effects, we test for the heterogeneous effects of
peers across specific individual characteristics. We find evidence that suggests that younger,
less tenured teachers, as well as teachers who live in more expensive homes, incorporate
more information from peers into their refinancing decisions. Finally, we find little evidence
of differences in characteristics between teachers who refinance independently and teachers
who follow their peers.
In addition to the literature on mortgage refinancing, this paper relates to a broader lit-
erature on household finance choices. For example, Grinblatt and Keloharju (2001); Hong,
Kubik, and Stein (2004); and Guiso, Sapienza, and Zingales (2008) study household stock
ownership. Choi, Laibson, and Madrian (2011) study the 401(k) contribution choice of house-
holds. Moreover, Gourinchas and Parker (2002) estimate a structural model of consumption
with labor uncertainty. Related to the effect of peers on household finance choices, Duflo
and Saez (2002) find that peers sharing common characteristics influence retirement plan-
ning choices of university workers. Duflo and Saez (2003) provide further evidence consistent
with peer effects in the same setting using a field experiment. More recently, Bursztyn et al.
(2014), also using a field experiment, separately identify the effect of social learning and
social utility from asset purchases by investors. Beshears et al. (2015) find an oppositional
effect in which individuals decrease their savings when given information about the savings
of their peers. Additionally, Grinblatt, Keloharju, and Ikaheimo (2008) examine the effect
of a neighbor’s actions on the consumption decisions of Finnish households. We add to this
literature by contributing to the understanding of how peers reduce informational frictions
in the context of arguably the most important financial decision faced by a household.6
Finally, this paper also relates to a broader literature that studies peer effects beyond
their effects on financial decisions. Eppel and Romano (1998) study the increase in posi-
6. More broadly, this paper also relates to the classic literature that studies the role that informationacquisition plays in financial decision making (e.g., Grossman and Stiglitz (1980); Ippolito (1989); andEllison and Fudenberg (1995)).
4
tive externalities that a high-ability student receives from her peers as the result of tuition
vouchers. Sacerdote (2001) finds a continuation of this peer effect in a post-secondary setting,
where a college student’s class performance and participation in social activities is influenced
by her randomly assigned freshman roommate. More recent work has shed light on the role
of peer effects in fraud among financial advisors (Dimmock, Graham, and Gerken (2015))
and CEOs (Khanna, Kim, and Lu (2015)). Brogaard, Engelberg, and Parsons (2014) find
an increased ability of journal editors in screening their colleagues’ papers. Additionally,
Mas and Moretti (2009) find the introduction of a highly productive worker increases the
productivity of other workers.
This paper is organized as follows: Section I provides some background on mortgage
refinancing. Section II describes the data sources, the procedure we use to match employment
and county deed records, the sample selection process, and the empirical challenges that this
paper addresses. Section III describes the identification strategy. Section IV presents our
main findings. Section V confirms the main findings using two alternative methodologies.
Section VI concludes.
I. Background on Refinancing
I.A. Institutional Details of Refinancing
Mortgage loans can be used to purchase a new property or to refinance an existing loan
(on a previously purchased property). This paper focuses on mortgage loans that are used
for refinancing. Loans for refinancing are classified into two main groups depending on the
borrower’s objective. In a rate or term refinance, the interest rate and/or the term of the
loan is modified, usually with the goal of lowering the mortgage payment. In contrast, in a
cash-out refinance, the new mortgage amount exceeds the existing mortgage amount, since
the objective of this type of refinance loan is to extract equity from the house. Typically, rate
refinances are motivated by interest rate decreases, while cash-out refinances are motivated
by house price increases.
5
Households purchase their homes at different points in the business cycle. Therefore,
refinancing the interest rate of an existing mortgage can result in significant savings for those
borrowers who financed their purchases in a period of high interest rates. A key determinant
for the amount of savings that follows a rate refinance is the spread between the interest
rate of the existing loan and the current rate. Typically, industry practitioners recommend
refinancing if this spread is sufficiently large such that the present value of savings exceeds
the costs of refinancing.7
However, the present value of savings that results from an interest rate differential is
not the only relevant factor for a rate refinance. Rate refinances also depend heavily on the
expectations of future interest rates. Even if interest rate differentials would result in positive
net savings today, it may be more advantageous for the borrower to wait and refinance later
if interest rates are expected to continue to decline in the near future. Additionally, even if
a borrower wants to refinance, she may not be able to obtain the loan if she is “underwater”
(i.e., if the house price has declined enough since the purchase that the borrower currently
has negative equity in the house).
I.B. Optimal Refinancing
There is a large body of literature that more rigorously studies the mortgage refinancing
decision described above using option valuation models.8 One puzzling result discussed in
this literature is the finding that borrowers seem to refinance their mortgages suboptimally.
Some borrowers appear to leave substantial amounts of money on the table by either waiting
too long to refinance or by simply not refinancing at all, even when the associated benefits
justify doing so (referred to in the literature as an error of omission). Other borrowers seem
7. The costs of refinancing, or closing costs, include application fees, appraisal fees, credit report fees,legal fees, home inspection fees, and origination fees, among others. The most relevant fee is the originationfee, which typically equals around 1% of the loan amount. This is the reason for the widely known rule ofthumb for refinancing recommended by industry experts, which recommends refinancing after interest rateshave declined by at least 200 bp.
8. See, for example, Dunn and McConnell (1981a,b); Stanton (1995); Dunn and Spatt (2005); and Agar-wal, Driscoll, and Laibson (2013).
6
to choose an unfavorable interest rate (referred to in the literature as an error of commission).
For example, Agarwal, Rosen, and Yao (2015b) find that about 59% of the borrowers in their
sample refinanced suboptimally. Likewise, Andersen et al. (2015) document substantial costs
for households due to errors of omission in a large dataset of Danish households.9
II. Data, Sample Selection, & Empirical Challenge
II.A. Data Description
We obtain data from multiple sources regarding teacher-level employment records, home
mortgage actions, and housing characteristics. We briefly describe each data source.
II.A.1. Teacher Employment Records
To begin, we obtain payroll records for all public school employees in Texas from the Texas
Education Agency (TEA). These data cover the period from 1999 to 2011 and represents
1,305 unique school districts. Each district is made up of an average of 7.98 campuses. While
all employees are included in the payroll records, we restrict the sample to include only
teachers. The result is a sample of over 3.6 million teacher-year observations across 585,000
teachers. Each employment record observation consists of the teacher’s name, demographics,
date of birth, education level, tenure, salary, and teaching assignments.
II.A.2. County Deed Records
We obtain property characteristics and residential transaction information from DataQuick’s
Assessor and History files. DataQuick compiles and cleans county deed records, and it is
one of the main real estate data providers in the U.S. The Assessor file contains property
addresses and characteristics. The History file contains detailed information on slightly more
than 11 million property transactions in the state of Texas between 2002 and 2011, which
9. For other recent studies of this refinancing puzzle see Agarwal et al. (2015a); Keys, Pope, and Pope(2014); and Johnson, Meier, and Toubia (2015).
7
involve around 4.6 million properties from the Assessor file. Examples of variables available
in the History file include the transaction date, the transaction purpose (i.e., purchase or
refinance), the loan amounts, the interest rate type (i.e., fixed or variable), and the names
of the buyer and the lender.
II.A.3. Registered Voter Rolls
Central to our study is the matching of teachers with their corresponding refinancing
transactions. To aid in the matching process, we obtain the full electoral roll of registered
voters from the Texas Secretary of State. Each record contains the voter’s name, date of
birth, mailing address, and physical address.
II.A.4. Additional Data Sources
Since DataQuick does not provide interest rates, and because house transaction price
coverage is low in Texas, we use two additional datasets. For interest rates, we use the
monthly 30-year fixed rates from Freddie Mac’s Primary Mortgage Market Survey (PMMS).10
Additionally, as a proxy for individual house prices, we use ZIP code-level indices from
Zillow.11
II.B. Matching Description
As mentioned above, in order to evaluate the role of peer effects on the refinancing
decisions of teachers in our sample, we must first map our teacher employment records to
county deed records. We describe each step of the matching procedure below.
10. Freddie Mac surveys lenders weekly and records interest rates on first-lien prime conventional purchasemortgages. Currently, about 125 lenders are included in the survey. Additional information about thissurvey can be found at http://www.freddiemac.com/pmms/abtpmms.htm.
11. Zillow’s indices are based on median home values. A more detailed description of their methodologycan be found at http://www.zillow.com/research/zhvi-methodology-6032.
8
II.B.1. Teacher to Voter
While both the TEA payroll records and county deed records provide names, match-
ing teacher records to mortgage transactions using legal names is problematic for multiple
reasons. One particular concern is the matching of teachers who possess common names.
While the distance between the teacher’s campus and each deed holder’s physical address
can help alleviate this concern, it does not fully address the issue. For instance, our sam-
ple contains four occurrences of the name Michael Cunningham listed on active mortgage
records in Denton County alone. In addition, the use of names alone would not now allow us
to link the teachers in our sample to mortgage records in cases in which only the teacher’s
spouse was listed on the deed. For these reasons, we turn to the electoral rolls of registered
voters, which serves as the linchpin in our matching process. The individual’s date of birth
and legal name are included in our teacher employment records. Thus, we can use both
name and date of birth to identify the voting record of every teacher in our sample who is
a registered voter. We are able to match 76.2% of the payroll records to the voter rolls.12
Therefore, the representative nature of our analysis is limited to homeowners who are also
registered voters.
II.B.2. Voter to County Deeds
We then use the resulting voting records, which contain both a physical and mailing
address for the individual, to establish a link between our teacher records and county deed
records. To do this, we require an exact match on the individual’s last name. We also require
an exact match between the address number prefix and the ZIP code for either the physical
address or the mailing address. Finally, we require that the Levenshtein ratio between the
address names exceeds 0.9. Figure 1 provides an illustration of the matching process.
[INSERT FIGURE 1 HERE]
12. Note, the matching rate is likely larger when conditioning on home ownership.
9
II.C. Final Sample
After matching registered teacher voters with the county deed records, we obtain slightly
more than 1.2 million teacher-year observations over the period from 1999 to 2011. The
transactions associated with these teachers include purchases, refinances, subdivisions, and
other less common types of transactions. We consider only purchases and refinances. We use
purchase transaction records to identify the time and the terms of the loan when a teacher
buys a house. The purchase date is needed to estimate the existing mortgage’s rate and the
potential savings from refinancing. Thus, for a teacher to remain in the sample, we need to
observe when the teacher purchased the house; that is, the teacher must show a purchase
transaction in the county deed records between 2002 and 2011.13 In addition, because
the refinancing of adjustable-rate mortgages (ARM) could occur due to borrower-specific
reasons, we omit teacher-property observations in which the teacher bought (or refinanced)
the property using an ARM.
Since our identification strategy exploits cross-sectional variation in teachers’ schedules,
we exclude campuses with an off-period shared by 90% or more of all teachers. Finally, to
increase the likelihood of teachers sharing information, we restrict the sample to campuses
with an average peer group of fewer than 50 teachers.14 The final sample consists of 32,465
teacher-year observations. Table 1 describes the final sample, along with the samples cor-
responding to each of the main steps of the matching procedure. Overall, all samples of
teacher-years show similar salaries, tenure, demographics, and education level.
[INSERT TABLE 1 HERE]
13. Consequently, we omit from the sample all teachers who do not show any purchases or teacher-yearobservations before the purchase of the property.
14. Additionally, for the accuracy of our measures, we also require that campuses have at least 70% oftheir teachers registered as voters, on average.
10
II.D. Estimation of Savings
To quantify the savings realized upon refinancing, we take the difference between the
present value of a straight annuity discounted at the prevailing interest rate and the estimated
remaining principle balance of the current mortgage. We assume the original mortgage is a
30-year term, and we set the length of the annuity to the minimum of the remainder of the
mortgage’s life and 13 years, which is the average time that buyers remain in their homes
(Emrath (2013)). To calculate the estimated net savings, closing costs of 1.5% of the total
loan amount are considered.15
II.E. Empirical Challenge
We face a twofold empirical challenge in this study. First, studying household financial
decisions is challenging because it is difficult to obtain the necessary data. Ideally, one would
like accurate data that cover a representative sample of the population and follow households
over time (Campbell (2006)). The second challenge is that peer effects studies are difficult
from an identification standpoint. We describe below why our data are advantageous for
addressing these challenges.
II.E.1. Data Accuracy and Representativeness
Mortgage-related decisions are arguably the most important financial decisions for the av-
erage household, since their house is their most valuable asset. As opposed to other datasets
of household investments in which the information is self-reported, county deed datasets are
more accurate and complete because, by law, all house transactions are required to be reg-
15. Estimated lending rates used in the calculation are based on the PMMS 30-year fixed rates at thetime of the last and the current transactions. Note that the closing costs of some mortgage refinances areincorporated into the lending rate. While we cannot see for which observations this is the case, we willsubtract the estimated closing costs from our estimates. If both products are fairly priced, the resulting netsavings should be equivalent. Earlier versions of this paper considered a fixed closing cost of $5,000, and theresults remained unchanged.
11
istered with the county.16 In addition, county records provide a history of transactions for
each property, allowing us to construct a data panel of refinances for each matched teacher,
rather than just a cross-section. Finally, Texas teachers represent a large group of house-
holds, providing thorough coverage across different characteristics such as age, ethnicity, and
wealth.17
II.E.2. Identifying Peer Effects
Studying peer effects is challenging for multiple reasons. First, it is difficult to identify
links between individuals merely because they share a common environment. Second, indi-
viduals tend to self-select themselves into peer groups. Third, individuals within the same
peer group are subject to common shocks. Our empirical setting is useful to address these
challenges in identifying peer effects. Teacher employment records allow us to identify sub-
sets of teachers who are likely to interact regularly and thus are likely to share information
about their financial decisions. Additionally, the Texas public school setting is advantageous
because teachers apply for jobs at the district level, but they have little or no choice regard-
ing which campus they are assigned to. Furthermore, our identification strategy involves
a second layer of quasi-randomization that occurs at the campus level, as discussed in the
following section. This mitigates concerns about individuals self-selecting into peer groups
(i.e., teacher networks are less likely to be formed endogenously). Furthermore, in contrast
to other settings where participation is voluntary, employment records include all the teach-
ers in the system, thus eliminating concerns about sample bias. Finally, the richness of
our micro-level data allows for the exploitation of variation across and within campuses to
account for common shocks to the peer group. Next, we describe our identification strategy.
16. Although the information requirements vary across states, it is a crime to purposely report falseinformation to a county clerk’s office.
17. While teachers’ salaries do not vary considerably and are correlated with tenure, there is variation inhouse value across our sample, which can proxy for family income.
12
III. Identification Strategy
We want to study whether an individual’s refinancing decision is influenced by the refi-
nancing activity of his or her peer group. While the use of within-district variation helps to
address endogeneity concerns related to common unobservable employment shocks that may
influence both a teacher and her peer group’s decision to refinance, some concerns remain. Of
particular concern is the potential variation in campus-level loan supply exposure, resulting
in the clustering of mortgage refinancing at the campus level. For example, the construc-
tion of a new billboard that advertises a lender’s superior mortgage terms near a campus
would provide an alternative information channel which could be mistakenly attributed to
a teacher’s peers. Therefore, to disentangle the possibility of campus-level variation in loan
supply exposure from the role of an individual’s peers, we exploit a source of within-campus
variation when constructing the peer groups: common off-periods.
While the teachers within a school campus share a common set of buildings, this does not
ensure that all teacher-pairs within the campus interact daily. Rather, it is much more likely
for an individual to acquire and incorporate information from a peer with whom they have
frequent contact. While two teachers may be employed at the same campus, their amount
of interaction is largely dictated by the teacher’s class schedule. More specifically, a given
school day is typically divided into multiple class periods. While the teachers in our sample
are assigned a class during the majority of the class periods, each teacher is also allotted one
or more periods which are designated as off-periods. These non-teaching periods are used to
prepare lectures, grade assignments, use office resources such as the school’s copier, and visit
the teacher’s lounge. More importantly, off-periods do not typically occur at a uniform time
for all teachers at a given campus. Therefore, the extent to which two teachers interact daily,
and thus share information, is significantly influenced by the amount of overlap between the
teacher’s off-periods.18
18. We find anecdotal support for this claim based on multiple conversations with public school teacherswho acknowledge that they interact primarily with other teachers while in the teacher’s lounge and teacherworkroom during their off-period.
13
Fortunately, our employment records include detailed information on each teacher’s class
assignments, which includes the beginning and ending times for each class. From this infor-
mation, we are able to recover the off-periods for each teacher in our sample. Using this data
on teacher off-periods, we then construct a more refined measure of a teacher’s peer group.
For each pair-wise combination of teachers within a campus, we first identify all overlapping
off-period segments and then compute the number of minutes of overlap for each segment.
We exclude any segments which overlap by less than 10 minutes, as it is unlikely that teach-
ers are able to interact in a meaningful way in such a brief amount of time before returning
to their class assignments. We define a teacher’s peer group as any teacher with an average
of 40 minutes or more per school day of commonality in off-periods.19
Figure 2 illustrates the class assignment schedule for three teachers and our classification
of each teacher’s peer group. The colored blocks denote a teacher’s off-periods, while the
hatched regions indicate an overlap in off-periods with another teacher. The figure indicates
that both the Bob–Kim and Kim–Ava pairs share 45 minutes of common off-periods per
day, while there are only 15 minutes of common off-periods for the Bob–Ava pair. This
figure illustrates the asymmetric nature of each teacher’s peer group. For instance, while
Bob’s peer group consists only of Kim, Kim’s peer group is made up of both Bob and Ava.
We now turn to the importance of peer effects on a teacher’s refinancing decision using this
measure of one’s peers.
[INSERT FIGURE 2 HERE]
19. In Texas, school districts have discretion over the format of the class schedule. While some districtshave a standardized set of class periods across all days, other districts utilize an alternating, or block, schedule.It is possible for two teachers to share an off-period block during some days of the week but have no overlapon other days. Thus, for each teacher-pair, we compute the average number of overlapping minutes perschool day. Finally, some campuses have a uniform campus-wide off-period. So, as mentioned in SectionII.C, we exclude any campus-year in which more than 90% of teachers have at least 40 minutes of off-periodoverlap per day.
14
IV. Peer Effects on the Refinancing Decision
First, we exploit novel microdata of teacher class schedules to better understand the
transmission of information across individuals. Second, we validate the empirical strategy,
which relies on the assumption of quasi-random schedules. Third, we study the possibility of
commonality in lenders across peers. Fourth, we study the cross-sectional intensity of peer
effects. Finally, we study additional features of peer refinancing.
IV.A. Peer Effects Matter
If informational frictions do play a role in the suboptimal refinancing behavior of house-
holds, and if we assume that teachers share information about their refinancing decisions
with their peers, then there should be a higher probability that a teacher will refinance
her house after a refinance occurs in her peer group. Conversely, there should be a lower
probability of a teacher refinancing when no one in her peer group has recently refinanced.
Figure 3 plots the relation between the probability of a teacher refinancing her existing
mortgage and her estimated net savings. Consistent with the intuition above, we separate
teachers into two groups. The first group (indicated by the black line with hollow markers)
consists of teacher-month observations in which no individual in the teacher’s peer group
performed a rate refinance in the trailing 3-month period.20 The second group (indicated by
the red and blue line with solid markers) consists of the remaining teacher-month observa-
tions in which at least one individual in the teacher’s peer group refinanced within the last
three months. As we are interested in individuals for whom refinancing is a feasible option,
we do not consider teachers whom we estimate to be underwater (i.e., having negative equity
in their home).21
20. We do not take into account refinances that appear to be cash-outs. More specifically, based on thePMMS 30-year fixed rate at the time of the last transaction, and assuming a 30-year loan, we estimate themonthly unpaid principal balance of each mortgage in our sample. We ignore all transactions associatedwith the teacher’s property if the loan amount of a teacher’s refinance is more than 1.05 times the remainingprincipal balance of the existing loan, and if the estimated net savings are negative.
21. For this, we estimate house price fluctuations based on Zillow’s ZIP code-level indices.
15
The figure shows a positive relation between the probability of refinancing and net savings
for both groups. However, the monthly probability of a teacher choosing an interest rate that
will result in negative savings is non-negligible (about 0.3%). The probability of refinancing
conditional on a peer refinancing is greater than the probability for the peer non-refinancing
group for all net saving bins. For example, for an estimated net savings of $5,000, the
likelihood of an individual refinancing increases from 0.383% to 0.965% (about 2.5 times)
when the number of peers who recently refinanced a mortgage increases from zero to one
or more. This difference tends to increase with net savings. Interestingly, although peer
refinancing has a more important effect on individuals who benefit from refinancing (i.e.,
those who have positive net savings), peer effects also appear to increase the probability
of refinancing for some teachers with negative net savings. We revisit this issue in Section
IV.E. Overall, these facts strongly support the hypothesis that peer effects are a significant
determinant of the probability of refinancing among Texas teachers.
[INSERT FIGURE 3 HERE]
We now turn to a more formal framework to better evaluate the economic and statistical
significance of peer effects while controlling for other determinants of refinancing activity in
addition to savings. We estimate OLS regressions in which the dependent variable is a 0/1
indicator for refinancing. Recall that the peer group for each observation contains all other
teachers who have at least 40 minutes of overlap in off-periods per school day. Thus, the
primary variable of interest, Peer Refinances, equals the number of teachers with overlapping
off-periods who have undertaken a mortgage refinance in the previous 3-month period, scaled
by the size of the peer group.22 We standardize Peer Refinances for ease of interpretation,
and we report effects in basis points.
The first regression in Table 2, which controls for other determinants of refinancing such
as (a) the estimated savings, (b) a 0/1 indicator for whether the teacher is underwater, and
22. We obtain quantitatively similar results when scaling by the average peer group size across all teachersin a given campus-year.
16
(c) the underwater amount (in percent); confirms the results in Figure 3. To address the
possible concern that refinancing activity is being driven by variation in local loan supply
exposure or local economic conditions, the first specification includes Metropolitan Statis-
tical Area (MSA)-Year fixed effects. The coefficient of 11.088 (t-stat=4.82) indicates that
a one standard deviation increase in the percentage of an individual’s peer group with re-
cent refinancing activity is associated with an 11.1 bp increase in the individual’s monthly
probability of refinancing. To add additional economic perspective, the unconditional prob-
ability of a teacher refinancing her mortgage in a given month is 68.5 bp in the same sample.
Therefore, a one standard deviation increase in Peer Refinances increases the probability of
refinancing by 16.2% of its unconditional value.
However, other factors beyond local economic conditions may influence an individual’s
decision to refinance her mortgage. For instance, the positive correlation between the re-
financing activity of a teacher and her peer group could be the result of the exposure of
both sets of individuals to common employment risk shocks. Fortunately, the risk of unem-
ployment is the result of economic hardship at the school district level. Thus, the second
specification in Table 2 includes school district-year fixed effects. The coefficient on Peer
Refinances is 10.171 (t-stat=4.41). However, the previous specifications suffer from the
shortcoming that the measured effect of Peer Refinances could be driven by campus-level
loan supply exposure (e.g., the construction of a new billboard that advertises a lender’s
superior mortgage terms near a campus). Fortunately, the within-campus variation in our
definition of Peer Refinances based on off-periods allows us to separate campus-level effects
from the effects of peers. For this reason, the third specification includes campus-year fixed
effects. The results remain largely unchanged, with a coefficient of 9.204.
While these results are consistent with a teacher’s refinancing decision being influenced
by her peers, even with campus-year fixed effects, we cannot rule out the possibility that
the teachers within a campus are exposed to a common shock that dissipates within the
school year. To address this concern, we include campus-month fixed effects in the fourth
17
specification. Thus, we are relying only on the variation in Peer Refinances generated from
teachers within the same campus at the same point in time, and only differ in the set of
coworkers with common off-periods. Following the inclusion of these effects, the coefficient
of Peer Refinances decreases to 6.42 bp. Finally, we include controls for employment and
individual characteristics. The effect remains virtually unchanged (i.e., a coefficient of 6.495),
which indicates that a one standard deviation change in the refinancing activity of a teacher’s
peers increases the probability that the teacher refinances by 9.48% of the unconditional
probability of refinancing.23 To provide additional economic content, an increase of $10,000
in net savings to an individual from refinancing is associated with a 56 bp increase in the
likelihood of refinancing. Therefore, a one standard deviation increase in Peer Refinances
has an effect equivalent to approximately a $1,150 increase in the net savings realized upon
refinancing.24 Overall, the regression analysis shows that peer effects have an economically
significant effect on the probability of refinancing among Texas teachers.
[INSERT TABLE 2 HERE]
IV.B. Self-Selection of Off-Periods
By exploiting overlap in teacher off-periods, we are able to disentangle the role of peers
from other confounding factors, such as campus-level variation in supply effects. Ultimately,
the identification of peer effects in this empirical setting relies on the exogenous formation of a
teacher’s peer group. Specifically, this requires that the assignment of a teacher’s schedule is
uncorrelated with other factors that affect the probability of refinancing. A plausible scenario
that violates this assumption would be the self-selection of teachers into peer groups based on
teacher discretion in the formation process of class schedules. However, conversations with
school officials provide anecdotal evidence that is inconsistent with this view; specifically, the
23. Additionally, Table IA.2 shows that the previous results are similar when Peer Refinances takes intoaccount rate refinances in the previous 2-month period, or when we use an alternative measure of savings,assuming a 30-year life for the loan.
24. More specifically, 6.49556.30 · $10, 000 = $1, 153.
18
fact that teacher assignments are determined by a scheduling algorithm designed to optimize
the scheduling needs of students and the menu of classes available. However, the possibility
remains that optimizing class assignments results in peer groups that share a set of common
characteristics.
To address this concern, we undertake a nonparametric approach to contrast the com-
monality of characteristics among the observed peer groups with that of the counterfactual,
in which peer groups are randomly formed. Specifically, we first compute the probability
that an individual and another teacher from her peer group share a common characteristic,
such as being a registered voter. We then generate a pseudo-peer group for each teacher by
randomly drawing peers from the pool of all teachers at the campus in that year. Each peer
group is constructed to be of equal size with the teacher’s true peer group. Finally, using
the pseudo-peer groups, we compute the probability of sharing a common characteristic. We
repeat this process 10,000 times to generate the distribution of commonality under the of
randomly selected peer groups.
[INSERT FIGURE 4 HERE]
Figure 4 reports the simulation results across multiple individual characteristics. Panel
A reports the likelihood of a teacher and her peer either having a common voter registration
status, having the same self-reported ethnicity, both possessing the same degree level, or
being of the same gender. Across the four estimates, the observed probability of sharing a
common characteristic falls within the bounds of the distribution under the null hypothesis.
The likelihood of two teachers in the same peer group sharing a common gender appears to
differ most substantially from the null, with slightly more than 3% of the pseudo-peer groups
having a greater likelihood of sharing a common gender. However, the difference between
the observed probability and the mean under the null is 0.23%, which suggests that the
difference is not economically important. Panel B repeats the previous analysis on contin-
uous individual characteristics such as annual salary. Here, we examine the mean absolute
difference between a teacher’s characteristic (e.g., her salary) and that of her peer group.
19
If teachers self-select into peer groups of similar pay, we would expect the mean absolute
difference in pay to be less than that of a randomly generated set of peer groups. The figure
indicates that the differences in tenure at the current school, years of teaching experience,
and ages tend to be less than that of a randomly generated peer group. However, the dif-
ferences are economically insignificant with the largest divergence (Tenure) representing a
difference of 0.6 months. Additionally, the absolute difference in annual salary appears to
be larger than that of the randomly generated peer groups. One possible explanation for
this result is the choice of some teachers to supplant an off-period with additional duties in
exchange for increased pay. This would result in a decreased likelihood of two such teachers
sharing an off-period relative to sharing an off-period with another teacher who did not take
on additional duties, thus earning less pay.
Additionally, we test the predictive power of a teacher’s peer group in forecasting observ-
able characteristics using OLS regressions. Internet Appendix Table IA.3 presents the results
when regressing a teacher’s characteristic on the average of her peer group, Peer Group Av-
erage. Alternatively, the possibility exists that clustering occurs in teacher characteristics
due to geographic or institutional effects. Most notably, teacher pay is likely correlated with
a school district’s tax base, resulting in district-level variation in teacher salaries. Further-
more, self-selection of teachers into student age groups could result in variation in teacher
gender across campuses within a district, such as differences between elementary school and
high school campuses. Thus, to control for this possibility, we also include the average of all
teachers who are within the campus, but outside the teacher’s peer group, Campus Aver-
age. Standard errors are clustered at the campus level. These results reinforce the previous
analysis, as Peer Group Average is not economically important determinant for any of the
characteristics tested. Overall, these findings yield no evidence of significant clustering along
teacher characteristics within off-period groups, which supports the use of off-periods in the
main identification strategy of this paper.
20
IV.C. Do Peers Share the Same Lender?
The previous findings are consistent with an individual’s peer group influencing his or
her actions. However, if this effect is the result of information transmission within the peer
network, then the information passed between agents need not be confined to the prevailing
lending rate in the market and to the potential savings associated with refinancing.
The primary determinant in an individual’s refinancing decision is the available lending
rate at which he or she can secure a new loan. While the time-series of mortgage rates
are certainly correlated across lenders, it is likely that there is a large amount of cross-
sectional variation in lending rates.25 Thus, the effort required to solicit rates from multiple
lenders may provide an additional friction that decreases the amount of refinancing activity.
However, this friction may be reduced if the information gained from a teacher’s peers
includes the local lender that currently offers the most beneficial terms. To examine this
possibility, we now turn to the commonality of lender choice within peer groups.
Specifically, if upon refinancing, the information passed from a teacher to a peer who
also refinanced included the lender that currently offers a superior rate, we would expect
the likelihood of both teachers using the same lender to be greater than random chance.
Therefore, for each observation included in Table 2 in which we observe a refinance, we
first select all teachers who share a common off-period and who also performed a mortgage
refinance over the trailing 3-month period. Using this peer group, we then compute the lender
overlap, which is the percentage of refinancing peers who use the same lender as our original
teacher. We find that the average of lender overlap (which we refer to as common lender % )
across our sample is 8.16%.26 However, it is possible that the refinancing activity of a given
lender is either temporally or geographically concentrated. Such concentration would result
in a value of common lender % that is greater than the probability of two randomly selected
25. Scharfstein and Sunderam (2013) show that the sensitivity of local lending rates to MBS yields isrelated to the concentration of mortgage lenders in the area.
26. Note that this is conditional on having at least one peer who refinances over the trailing 3-monthperiod.
21
refinances sharing a common lender. Thus, we need an appropriate counterfactual to serve
as a benchmark for our results in order to evaluate its economic significance. To do this, we
generate pseudo-peer groups under the null hypothesis that there is no information sharing
between individuals, and we then compute common lender % under this null. To address
the concern of lender concentration across time or geographic region, we define the candidate
pool of potential peers for each observation as all teacher refinances over the trailing 3-month
period within the same MSA but outside the true peer group. From this candidate pool, we
randomly select a peer group that is the same size as the peer group in the nonrandomized
sample. We repeat this randomized sampling procedure for 10,000 iterations to approximate
the distribution of common lender % under the null of no information sharing within the
peer group.
Panel A of Figure 5 reports the histogram of the results from the pseudo-peer group
simulations. The figure indicates that, under the null of no information dissemination within
peer groups, the expected value of common lender % is 3.05%. In addition, the maximum
value across the 10,000 iterations performed is 4.76%, or 8.6 standard deviations less than
the value of 8.16%, which we find when computing common lender % using the true peer
groups.
[INSERT FIGURE 5 HERE]
The previous panel suggests that in addition to the refinancing decision, a teacher’s
choice of lender is also influenced by his or her peer group. However, the possibility re-
mains that the abnormally large percentage of common lenders within peer groups is driven
by geographically concentrated lender activity within an MSA. Alternatively, such a result
could be explained by strategic partnerships between lenders and individual school districts.
Ideally, these concerns would be addressed by further restricting the candidate pool from
which pseudo-peer groups are drawn, so that it includes only teachers from the same school
district or even from the same campus. Unfortunately, such constraints are too restrictive;
they result in an insufficient number of candidates for some observations when a candidate
22
is required to have performed a mortgage refinance in the trailing 3-month period. Thus, we
repeat our analysis after relaxing the latter constraint.
Specifically, for each observation in our sample in which we observe a refinance, we
define lender overlap as the percentage of teachers who share a common off-period and
whose existing lender matches that of the teacher performing the refinance. Additionally,
when constructing the pseudo-peer group, we consider only teachers who work at the same
campus but do not share a common off-period. Thus, our focus is now on the percentage of
common lenders among the stock of all mortgages rather than in the flow of new refinances.
Panel B of Figure 5 reports the histogram of common lender % under the pseudo-peer
groups. The figure indicates that the expected value of lender overlap when drawing peers
from all teachers who do not share a common off-period is 1.71% compared to 2.57% when
using the true peer groups. Additionally, in only 0.24% of the iterations does the value of
common lender % in the randomized groups exceed the value when considering teachers
with a common off-period.
IV.D. Sample Splits
We now turn to cross-sectional variation in the effect of peers on a teacher’s refinancing
decision. We segment the sample examined in Table 2 into mutually exclusive subsets based
on tenure, age, pay, education level, and the house price index level of the teacher’s ZIP code.
We then run the previous analysis again. The results are presented in Table 3. To control
for unobservable factors, we include campus-month fixed effects in each model specification
considered. The regressions begin to suggest that less tenured, younger teachers, and teachers
who live in more expensive homes seem to incorporate more information from peers into their
refinancing decisions. However, these results should be interpreted with caution, as F -tests
indicate that none of the coefficient pairs differ statistically from one another.
[INSERT TABLE 3 HERE]
23
IV.E. Negative Savings
Up to this point, we have taken the broad view that the actions of the teacher’s peer group
contribute information (e.g., the potential benefits of refinancing, prevailing market rates for
refinanced mortgages). While the incorporation of additional information is traditionally
associated with more efficient outcomes, Figure 3 suggests that this is not always the case
for the teachers in our sample. Although the effects of peer refinancing are, on average, higher
for teachers whom we estimate will benefit from refinancing, the figure also shows that peer
refinancing increases the average refinancing probability of teachers with estimated negative
net savings. This suggests that the effect of a teacher’s peers could result in an information
cascade in which the knowledge that a peer recently refinanced could lead a teacher to falsely
conclude that she should also refinance her mortgage.
We now examine the effect of a teacher’s peer group within this subsample. Table 4 shows
the effect of peers for the subsample of observations in which the estimated net savings of
refinancing are negative. Without the inclusion of campus-month fixed effects, the effect
of a one standard deviation increase in the refinancing activity of a teacher’s peers on her
probability of refinancing ranges from 6.19 bp (Specification 1) to 5.34 bp (Specification 2)
across the specifications considered, while being statistically significant in each. While the
point estimates are substantially lower than those from Table 2, the results in this panel are
consistent with a teacher simply mimicking the behavior of peers and refinancing regardless of
the benefits and costs. However, the coefficient of Peer Refinances is statistically insignificant
with the inclusion of campus-month fixed effects in the final two specifications.
The results in Table 4 need to be interpreted with caution. While this paper focuses
on mortgage refinances that are initiated to capture savings on interest payments, mortgage
refinances also serve as a means for a homeowner to extract home equity, thereby increasing
their household leverage. Thus, this result is also consistent with an individual’s choice to
increase their household leverage being affected by the actions of her peers. Given the lack of
some loan details in the data, we are forced to make simplifying assumptions when estimating
24
the potential savings of a mortgage refinance, which could result in the misassignment of
some refinances into the wrong subsample based on net savings.
[INSERT TABLE 4 HERE]
IV.F. Conditional Effects of Peers
The results presented thus far have demonstrated the impact of a teacher’s peer group
on her refinancing behavior. Table 5 extends this analysis by examining how the magnitude
of this effect changes when exploiting institutional features of the public school system and
the homogeneity of peer networks. The underlying mechanism within our setting is the
communication of information between teachers during their off-periods. However, public
schools in Texas do not operate for the entire year. Instead, there is a 2-month break during
the summer. During this time, it is less likely that teachers interact, possibly attenuating
the estimated effect of a teacher’s peer group. Thus, in Panel A of Table 5, we exclude the
first two months of the school year, since a teacher is less likely to interact with peers in the
preceeding 3-month period. Following this refinement, we find an increase in the coefficient
of Peer Refinances. For instance, in the final specification, the point estimate increases by
20% to 7.828 bp, relative to the coefficient of 6.495 bp reported in Table 2.
Panel B of Table 5, in turn, examines how the effect of a teacher’s peers varies with
the degree of commonality in characteristics. Thus, we segment Peer Refinances into two
mutually exclusive groups. The first group contains peers with the same ethnicity as the
teacher, while the second group contains peers who differ in ethnicity. The results indicate
that refinancing among a teacher’s peers is associated with an increase in the teacher’s
probability of refinancing, regardless of shared or differing ethnicities when controlling for
local economic and employment effects (Specifications 1–3). However, following the inclusion
of campus-month fixed effects in the final two specifications, only the refinancing activity of
peers who share a common ethnicity remains as a statistically significant factor in a teacher’s
25
refinancing behavior.27
IV.G. First Movers and Followers
Another question of interest is whether the characteristics of teachers who refinance
independently of their peers differ from those of teachers who follow their peers. We explore
this question in Table IA.6, and find that the first movers, or “triggers,” are quite similar
to followers in terms of tenure, age, gender, education level, whether they teach a technical
class, the size of their loans, and the amount of savings when refinancing. The only dimension
in which the two types differ statistically is pay. First movers earn $853 more per year than
followers, on average, a difference equivalent to 1.8% of the mean salary in our sample.28
V. Evidence from Alternative Methods
One assumption that underlies the previous tests is that the actions of the teacher’s peer
group are exogenous with respect to her refinancing decision. To address this concern, we
begin this section by undertaking an instrumental variable approach. Next, we conduct
a survival analysis to account for the dynamic nature and the infrequent occurrences of
refinances.
V.A. Instrumented Refinancing among Peers
To address concerns that an unobservable factor is influencing both the outcome variable
and the refinancing decisions of peer groups, we exploit within-campus variation in peer
groups. However, the use of off-periods only alleviates this endogeneity concern if the unob-
servable variable exists at the campus level, such as a billboard placed near a school campus
which advertises a lender’s superior mortgage rates. In contrast, we would be unable to iden-
27. However, F -tests fail to reject the null of no difference between the two coefficients. Internet AppendixTable IA.4 reports additional tests that examine alternative constructions of commonality based on age andgender.
28. Finally, the lack of sufficient events prevents a reliable examination of possible learning by followers.
26
tify the true effect if the variation in the unobservable variable exists at the off-period level.
One example of such an unobservable variable, which we find to be particularly entertaining,
is the possibility that bank loan officers visit campuses to discuss refinancing benefits with
teachers. However, due to some unforeseen circumstance, the loan officers are only able to
speak to teachers during one off-period before leaving the campus.
Also, by definition, the first teacher within each peer group to refinance does so without
being influenced by any individual in her peer group. The decision of this first mover
or “trigger” to refinance could be influenced by the peer group of the teacher’s spouse,
the teacher’s social network formed outside of her work environment, or another source of
information such as media advertisements. Regardless, we do not observe the reason for this
first action. Thus, up to this point, we have taken the view that this initial refinance is
exogenous in nature, a view which some may find unsatisfactory.
Thus, we next seek an instrumental variable that is able to generate exogenous variation
in the refinancing decision of a teacher’s peers to justify this assumption of exogeneity. While
this paper focuses on the transmission of information within peer groups, the role that peer
effects play on an individual’s refinancing decision is likely secondary when compared to
the effect of potential savings as a result of refinancing. Intuitively, while an individual
may gather information regarding the importance of refinancing and the prevailing rates
in the market, such information is of no use to a teacher who does not currently benefit
from refinancing. Thus, a natural choice for an instrumental variable is the variation in
the potential savings among a teacher’s peers to predict the refinancing activity of this peer
group. In doing so, we are explicitly assuming (when the teacher’s savings are also included
in the regression) that the average savings of a teacher’s peer group only affects the teacher’s
refinancing decision through the refinancing activity of her peers.
Specifically, we compute net savings for each individual in a teacher’s peer group and
take the average across all the teacher’s peers to instrument for Peer Refinances, yielding
Peer Refinances. It is possible that an unobservable characteristic exists that is correlated
27
among the teachers of a peer group and results in correlated refinancing choices in the group.
However, the estimated quantity Peer Refinances utilizes only the information contained
in the savings of a teacher’s peers and not this unobservable factor.
We now turn to the empirical evidence when instrumenting for the refinancing activity
of a teacher’s peer group using the peer groups’ average savings. Table 6 reports the results
from the second stage of 2SLS regressions for the full sample.29 As before, the variable Peer
Refinances has been scaled by its standard deviation. The results are quite striking. When
we include MSA-year fixed effects in the first specification to control for geographic effects,
the coefficient indicates that a one standard deviation increase in the refinancing activity of
a teacher’s peers increases her probability of refinancing by 36.31 bp (t-stat=6.20). Thus,
the effect of a one standard-deviation increase in Peer Refinances is equivalent to slightly
more than a $7,590 increase in a teacher’s savings. The effect remains relatively stable
with the inclusion of district-year fixed effects (Specification 2), campus-year fixed effects
(Specification 3), and teacher demographics (Specification 4).30
[INSERT TABLE 6 HERE]
At first glance, these results appear to be quite substantial, especially when benchmarked
against the coefficient estimates of Table 2. For instance, when viewed relative to the effect
of Savings, the point estimate for the effect of a one standard deviation increase in Peer
Refinances in the first specification is roughly 3.35 times the estimated effect from the pre-
vious table, which did not include an instrument variable.31 Alternatively, the coefficient of
36.31 in the first specification represents an increase of 54.2% relative to the mean monthly
probability of refinancing, which is 67.0 bp in this sample. However, recall that the in-
strumental variable predicts the refinancing activity of a teacher’s peer group based on the
29. The first stage of each regression is reported in Internet Appendix Table IA.5. The partial F -statisticfor the instrument variable ranges from 23.04 to 24.01, indicating that there should not be concern about aweak instrument problem.
30. Note that Table 6 differs from previous tables and does not include campus-month fixed effects, astheir inclusion results in a weak instrument.
31. The relative effect for the IV estimate is $7,590, while the effect in Table 2 is $2,265.
28
group’s potential savings. Thus, the estimated value of Peer Refinances is at its largest when
a teacher’s peers would benefit the most from refinancing, which is when the teacher would
plausibly be the most eager to discuss her savings with her peers. In contrast, the results
from the previous subsection include all peer refinancing activity that we cannot explicitly
classify as a cash-out refinance. It is unlikely that a teacher who initiates a cash-out refi-
nance would discuss the benefits of an interest-rate refinance with peers. Thus, the inclusion
of peer refinances initiated for a reason other than the savings benefits would result in a
downward attenuation of the estimated effect.
V.B. Hazard Model Results
While the previous tables present our main findings, they do so in the context of a
linear probability model. Doing so provides the advantage of being able to control for
unobservable variables using fixed effects, sometimes at a highly granular level, such as the
campus-month level, while still presenting the results within a standard framework. However,
for complementarity, given the relatively rare occurrence of mortgage refinances and the
dynamics of this decision faced by households, we seek to augment our current findings
with results from a survival analysis framework. Thus, Table 7 presents the results of Cox
Proportional Hazard model estimations, analogous to the results of Table 2.32 Reported are
hazard ratios with standard errors clustered at the MSA-year level.
[INSERT TABLE 7 HERE]
In the first specification, the coefficient of 1.041 (z -stat=4.03) indicates that a one stan-
dard deviation increase in the percentage of an individual’s peer group with recent refinancing
activity is associated with a 4.1% increase in the hazard rate of an individual refinancing.
In contrast to the previous analysis, the hazard model does not easily accommodate high
32. Recall that the peer group for each observation contains all other teachers who have at least 40 minutesof overlap in off-periods per school day. Thus, the primary variable of interest, Peer Refinances, equals thenumber of teachers with overlapping off-periods who have undertaken a mortgage refinance in the previous3-month period, scaled by the size of the peer group. Similar to Table 2, we standardize Peer Refinances forease of interpretation.
29
density fixed effects. Therefore, as a control for the effect of local loan supply exposure or
economic conditions, we include as an additional variable the percentage of all mortgages
that were refinanced in the teacher’s MSA-year. The results remain relatively unchanged
when controlling for the possibility of employment effects or differences in local supply expo-
sure with the inclusion of variables measuring the percentage of refinancing activity in the
district-year and campus-year, respectively.
In the fourth specification, the coefficient increases to 1.102 (z -stat=5.29) with the in-
clusion of a measure of refinancing activity constructed at the campus-month level. This
specification indicates that a one standard deviation change in the refinancing activity of a
teacher’s peers increases the hazard rate of the teacher’s refinancing behavior by 10.2%. To
provide some economic content, note that a $10,000 increase in the effective net savings upon
refinancing increases the hazard rate by 63.3%. Thus, a one standard deviation increase in
the refinancing activity of a teacher’s peers is roughly equivalent to a $1,600 increase in the
savings realized upon refinancing. The effect remains unchanged in the final specification
with the inclusion of controls for employment and individual characteristics, while increasing
the statistical significance slightly to a z -stat of 5.52. Overall, the results presented in Table
7 confirm the previous findings that a teacher’s peer group is an important determinant of
her refinancing behavior.
VI. Conclusion
Using a comprehensive data set of employment records of Texas teachers, we exploit the
quasi-random assignment of teacher class schedules to identify the effect of peers on an im-
portant household financial decision. We find that a one standard deviation increase in the
amount of recent mortgage refinances by a teacher’s peers increases her likelihood of refinanc-
ing by 9.5%. Our setting allows us to control for other determinants of refinancing, teacher
characteristics, and even unobservable campus-level shocks. The results are confirmed using
an IV estimation approach and robust to the consideration of a hazard model. We find that
30
a teacher’s choice of lender upon refinancing is also influenced by her peer group, a finding
that is consistent with information being transmitted within her peer network. Additionally,
we find evidence that suggests that younger, less tenured teachers, as well as teachers who
live in more expensive homes, incorporate more information from peers into their refinancing
decisions. Finally, we find little evidence of differences in characteristics between teachers
who refinance independently and teachers who follow their peers.
Overall, our results support the hypothesis that informational frictions can be detrimental
to household financial decisions, and that the effects of peers reduce these frictions. Sub-
stantial savings for households could be achieved by educating homeowners on the benefits
of refinancing and by making interest rate information widely available.
31
References
Agarwal, Sumit, John C. Driscoll, and David I. Laibson, 2013, Optimal mortgage refinancing:
A closed-form solution, Journal of Money, Credit and Banking 45, 591–622.
Agarwal, Sumit, Crocker Liu, Walter Torous, and Vincent Yao, 2015a, The mistakes people
make: Financial decision making when buying and owning a home, Working paper.
Agarwal, Sumit, Richard J. Rosen, and Vincent Yao, 2015b, Why do borrowers make mort-
gage refinancing mistakes?, Management Science, forthcoming .
Agarwal, Sumit, Paige M. Skiba, and Jeremy Tobacman, 2009, Payday loans and credit
cards: New liquidity and credit scoring puzzles?, American Economic Review 99, 412–
417.
Andersen, Steffen, John Y. Campbell, Kasper Meisner Nielsen, and Tarun Ramadorai, 2015,
Inattention and inertia in household finance: Evidence from the danish mortgage market,
Working paper.
Beshears, John, James J. Choi, David Laibson, Brigitte C. Madrian, and Katherine L.
Milkman, 2015, The effect of providing peer information on retirement savings decisions,
Journal of Finance 70, 1161–1202.
Brogaard, Jonathan, Joseph Engelberg, and Christopher Parsons, 2014, Network position
and productivity: Causal evidence from editor rotations, Journal of Financial Economics
111, 251–270.
Bursztyn, Leonardo, Florian Ederer, Bruno Ferman, and Noam Yuchtman, 2014, Under-
standing mechanisms underlying peer effects: Evidence from a field experiment on finan-
cial decisions, Econometrica 82, 1273–1301.
Campbell, John Y., 2006, Household finance, Journal of Finance 61, 1553–1604.
Choi, James J., David Laibson, and Brigitte C. Madrian, 2011, $100 bills on the sidewalk:
Suboptimal investment in 401 (k) plans, Review of Economics and Statistics 93, 748–763.
Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, 2008, The small world of investing:
Board connections and mutual fund returns, Journal of Political Economy 116, pp. 951–
979.
Cohen, Lauren, Andrea Frazzini, and Christopher Malloy, 2010, Sell-side school ties, Journal
of Finance 65, 1409–1437.
32
Dimmock, Stephen, Nathaniel Graham, and William C. Gerken, 2015, Is fraud contagious?
Career networks and fraud by financial advisors, Working paper.
Duflo, Esther, and Emmanuel Saez, 2002, Participation and investment decisions in a retire-
ment plan: The influence of colleagues choices, Journal of Public Economics 85, 121–148.
Duflo, Esther, and Emmanuel Saez, 2003, The role of information and social interactions in
retirement plan decisions: Evidence from a randomized experiment, Quarterly Journal of
Economics 118, 815–842.
Dunn, Kenneth B., and John J. McConnell, 1981a, A comparison of alternative models for
pricing GNMA mortgage-backed securities, Journal of Finance 36, 471–484.
Dunn, Kenneth B., and John J. McConnell, 1981b, Valuation of GNMA mortgage-backed
securities, Journal of Finance 599–616.
Dunn, Kenneth B., and Chester S. Spatt, 2005, The effect of refinancing costs and market
imperfections on the optimal call strategy and the pricing of debt contracts, Real Estate
Economics 33, 595.
Ellison, Glenn, and Drew Fudenberg, 1995, Word-of-mouth communication and social learn-
ing, Quarterly Journal of Economics 93–125.
Emrath, Paul, 2013, Latest calculations show average buyer expected to stay in a home 13
years, Special studies, January 3.
Eppel, Dennis, and Richard E. Romano, 1998, Competition between private and public
schools, vouchers, and peer-group effects, American Economic Review 88, 33–62.
Gourinchas, Pierre-Olivier, and Jonathan A. Parker, 2002, Consumption over the life cycle,
Econometrica 70, 47–89.
Grinblatt, Mark, and Matti Keloharju, 2001, How distance, language, and culture influence
stockholdings and trades, Journal of Finance 56, 1053–1073.
Grinblatt, Mark, Matti Keloharju, and Seppo Ikaheimo, 2008, Social influence and con-
sumption: Evidence from the automobile purchases of neighbors, Review of Economics
and Statistics 90, 735–753.
Grossman, Sanford J., and Joseph E. Stiglitz, 1980, On the impossibility of informationally
efficient markets, American Economic Review 393–408.
33
Guiso, Luigi, Paola Sapienza, and Luigi Zingales, 2008, Trusting the stock market, Journal
of Finance 63, 2557–2600.
Hong, Harrison, Jeffrey D. Kubik, and Jeremy C. Stein, 2004, Social interaction and stock-
market participation, Journal of Finance 59, 137–163.
Ippolito, Richard A., 1989, Efficiency with costly information: A study of mutual fund
performance, 1965-1984, Quarterly Journal of Economics 1–23.
Johnson, Eric, Stephan Meier, and Olivier Toubia, 2015, Money left on the kitchen table:
Exploring sluggish mortgage refinancing using administrative data, surveys, and field ex-
periments, Working paper.
Keys, Benjamin J., Devin G. Pope, and Jaren C. Pope, 2014, Failure to refinance, Working
paper.
Khanna, Vikramaditya, Han E. Kim, and Yao Lu, 2015, CEO connectedness and corporate
fraud, Journal of Finance 70, 1203–1252.
Leary, Mark T., and Michael R. Roberts, 2014, Do peer firms affect corporate financial
policy?, Journal of Finance 69, 139–178.
Mas, Alexandre, and Enrico Moretti, 2009, Peers at work, American Economic Review 99,
112–145.
Pool, Veronika, Noah Stoffman, and Scott Yonker, 2015, The people in your neighborhood:
Social interactions and mutual fund portfolios, Journal of Finance, forthcoming.
Sacerdote, Bruce, 2001, Peer effects with random assignment: Results for Dartmouth room-
mates, Quarterly Journal of Economics 116, 681–784.
Scharfstein, David S., and Adi Sunderam, 2013, Concentration in mortgage lending, refi-
nancing activity and mortgage rates, Working paper.
Shue, Kelly, 2013, Executive networks and firm policies: Evidence from the random assign-
ment of MBA peers, Review of Financial Studies 26, 1401–1442.
Stanton, Richard, 1995, Rational prepayment and the valuation of mortgage-backed securi-
ties, Review of Financial Studies 8, 677–708.
Wolff, Edward N., 2012, The asset price meltdown and the wealth of the middle class,
Working paper.
34
TeacherEmployment
Records
Voting Records
NameDOB
Gender
Registered TeacherVoters
County DeedRecords
AddressLast Name
Teacher-Transactions
+ Employment information+ Demographic information
+ Residential transactions+ Property characteristics
Phase 1
Phase 2
FIGURE 1. Data Matching Procedure
This figure illustrates the process used to map teacher employment records to county deed records. First,the TEA payroll records are matched with the electoral records of registered voters based on name, date ofbirth, and gender, in order to obtain teachers’ addresses (Phase 1). Then, we use these addresses and lastnames to match the resulting teacher voter records with county deed records (Phase 2). We require an exactmatch between the address number prefix, ZIP code, and individual’s last name; we also require that theLevenshtein ratio between the address name listed on the county and voting records exceeds 0.9.
35
9:00am 10:00am 11:00am 12:00pm 1:00pm 2:00pm 3:00pm
Ava
Kim
Bob
Ava Kim Bob
4515
4545
15 45
-
45
15
45
-
45
15
45
-
Teacher SchedulesMinutes Overlap
Time
FIGURE 2. Off-Periods Classification
This figure illustrates an example of the class assignment schedule for three teachers and our classificationof each teacher’s peer group using overlapping off-periods. The colored blocks denote a teacher’s off-periods,while the hatched regions indicate an overlap in off-periods with another teacher. Both the Bob-Kim andKim-Ava pairs share 45 minutes of common off-periods per day, while there are only 15 minutes of commonoff-periods for the Bob-Ava pair. This figure illustrates the asymmetric nature of each teacher’s peer group:While Bob’s peer group consists of only Kim, Kim’s peer group is made up of both Bob and Ava.
36
0.5
11.
52
2.5
mon
thly
pro
babi
lity
of re
finan
cing
(%)
−10 −5 0 5 10 15 20net savings ($1,000)
FIGURE 3. Net Savings and Refinancing Probability
This figure shows the relation between the probability of an individual refinancing her mortgage and herestimated net savings upon refinancing when using overlapping off-periods to identify peer groups. Teachersare separated into two groups. The first group (indicated by the black line with hollow markers) consists ofteacher-month observations in which no individual in the teacher’s peer group performed a mortgage refinancein the trailing 3-month period. The second group (indicated by the red and blue line with solid markers)consists of the remaining teacher-month observations, in which at least one individual in the teacher’s peergroup refinanced within the last three months. Teachers whom we estimate to be underwater (i.e., havenegative equity in the house) are not included.
37
Panel A: % Peers with Common Characteristic
65.25 65.50 65.75 66.00 66.25Voter
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5Density
74.75 75.00 75.25 75.50 75.75Ethnicity
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
69.25 69.50 69.75 70.00 70.25Degree
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Density
64.50 64.75 65.00 65.25 65.50 65.75Gender
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
FIGURE 4. Off-Periods Selection SimulationThis figure illustrates the degree of commonality for individual characteristics among peer groups relativeto pseudo-peer groups. The observed degree of commonality in our peer groups are indicated by the dashedvertical blue line, while the histogram and kernel density estimate report the distribution of commonality fromrandomly generated peer groups. Panel A reports the probability of having the same voter registration status(Voter), self-reported ethnicity (Ethnicity), degree level (Degree), and gender (Gender). Panel B reportsthe mean absolute difference between a teacher and her peer group with respect to annual compensation(Pay), years of tenure at the current school (Tenure), total years of teaching experience (Experience), andcurrent age as of September 1 of the current school year (Age). Pseudo-peer groups are formed by randomlydrawing peers from the pool of all other individuals within a teacher’s campus but outside her peer group.
38
Panel B: Mean Absolute Difference with Peers
7575 7600 7625 7650 7675 7700 7725Pay
0.000
0.005
0.010
0.015
0.020
0.025
0.030Density
7.60 7.65 7.70 7.75Tenure
0
5
10
15
20
25
9.95 10.00 10.05 10.10 10.15Experience
0
5
10
15
20
Density
12.05 12.10 12.15 12.20 12.25Age
0
5
10
15
20
39
Panel A
1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5Probability of Common Lender (%)
0.0
0.2
0.4
0.6
0.8
1.0
1.2Density
FIGURE 5. Common Lenders SimulationThis figure illustrates the probability of a teacher and her peer sharing a common lender relative to pseudo-peer groups. Panel A reports the probability of a common lender between a refinancing teacher and herpeer, conditional on the peer also refinancing in the prior 3-month period (indicated by the dashed blueline). The figure also reports the distribution of probabilities for pseudo-peer groups. Pseudo-peer groupsare formed by constructing a pool of candidate peers by selecting all individuals who refinance within theprior 3-month period and who are within a teacher’s MSA but not in her peer group. Then, candidates arerandomly selected from the pool to match the true size of the teacher’s peer group. Panel B reports theprobability of a refinancing teacher having the same lender as her peer. Pseudo-peer groups are formed bydrawing candidates from the pool of all other individuals within a teacher’s campus but outside of her peergroup.
40
Panel B
0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 2.75Probability of Common Lender (%)
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6Density
41
TABLE 1–Summary Statistics
All Teachers Registered Voters Teacher-Transactions Final SampleMean Std. Dev. Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.
Employment:Pay 43998.67 (9558.94) 44339.12 (9678.51) 45330.83 (9730.22) 47813.23 (8245.75)Tenure 7.65 (8.05) 8.23 (8.20) 7.70 (7.79) 6.19 (6.86)
Ethnicity:Caucasian 0.694 (0.461) 0.705 (0.456) 0.700 (0.458) 0.742 (0.438)African American 0.090 (0.287) 0.089 (0.285) 0.100 (0.300) 0.115 (0.320)Asian 0.010 (0.100) 0.007 (0.084) 0.009 (0.094) 0.013 (0.111)Hispanic 0.199 (0.400) 0.193 (0.395) 0.185 (0.388) 0.127 (0.333)Other 0.006 (0.079) 0.006 (0.075) 0.006 (0.079) 0.003 (0.053)
Other Characteristics:Bachelor’s Degree 0.764 (0.425) 0.757 (0.429) 0.742 (0.438) 0.728 (0.445)Advanced Degree 0.227 (0.419) 0.234 (0.423) 0.250 (0.433) 0.264 (0.441)Age 42.49 (11.19) 43.31 (10.97) 42.72 (10.88) 41.40 (10.51)Female 0.772 (0.420) 0.771 (0.420) 0.771 (0.420) 0.656 (0.475)
N 3,672,937 2,798,802 1,258,810 32,465
This table describes the different samples that correspond to each of the main steps of the matching procedure and the sample selection process. Morespecifically, the table provides the mean and standard deviations of multiple employment and demographic variables of 1) all teachers in the TEApayroll records, 2) teachers who are registered voters, 3) teachers who were matched to county deed transaction records, and 4) teachers in the finalsample. The units of observation are teacher-years.
42
TABLE 2–Peer Effects on the Decision to Refinance
(1) (2) (3) (4) (5)
Peer Refinances 11.088*** 10.171*** 9.204*** 6.422* 6.495**(4.82) (4.41) (3.64) (1.96) (1.99)
Savings ($, ×10,000) 48.961*** 49.299*** 53.049*** 56.306*** 56.299***(6.47) (6.48) (6.75) (5.46) (5.45)
1(Underwater) -7.661 -9.099 -7.893 -5.428 -4.744(-1.10) (-1.16) (-0.93) (-0.57) (-0.50)
Percent Underwater -9.846 5.737 15.127 -33.014 -58.476(-0.12) (0.05) (0.12) (-0.24) (-0.42)
Teacher Characteristics N N N N YMSA-Year FE Y N N N NDistrict-Year FE N Y N N NCampus-Year FE N N Y N NCampus-Month FE N N N Y Y
N 303,467 303,467 303,467 303,467 303,467R2 0.024 0.032 0.055 0.242 0.242
This table shows OLS regressions in which the dependent variable is a 0/1 indicator of refinances, and themain variable of interest is Peer Refinances, a variable that captures the number of peers having undertakena mortgage refinance in the previous 3-month period, scaled by the size of the peer group. Reported arethe effects of a one standard deviation change in Peer Refinances. The regressions also include controlsfor other determinants of refinancing such as the estimated savings, a 0/1 indicator of whether the teacheris underwater, and the underwater amount (in percent). Depending on the specification, MetropolitanStatistical Area (MSA)-year fixed effects, district-year fixed effects, campus-year fixed effects, campus-monthfixed effects, or teacher characteristics such as gender, ethnicity, highest academic degree obtained, and salaryare also included. Reported t-statistics in parentheses are heteroskedasticity-robust and clustered by MSA-year. ***p<0.01, **p<0.05, *p<0.1.
43
TABLE 3–Cross-Sectional Splits
Tenure Age Pay Grad Degree House PriceAbove Below Above Below Above Below Yes No Above Below
Peer Refinances 2.393 9.806** 6.930 11.895* 5.215 7.770* 4.778 7.048 7.770* 11.365*(0.30) (2.49) (1.29) (1.86) (0.82) (1.70) (0.54) (1.61) (1.70) (1.75)
Savings ($, ×10,000) 51.743*** 57.376*** 49.276*** 61.566*** 55.728*** 50.529*** 54.713*** 53.659*** 50.529*** 52.548***(4.45) (4.96) (4.00) (5.36) (4.83) (4.81) (6.63) (4.62) (4.81) (5.45)
1(Underwater) -12.960 -11.054 -8.449 -1.435 -10.439 11.759 -1.156 -4.906 11.759 -8.380(-1.00) (-0.73) (-0.62) (-0.12) (-0.95) (0.49) (-0.08) (-0.44) (0.49) (-0.79)
Percent Underwater 63.303 -90.848 -100.055 -135.588 2.840 -182.053 -275.092 79.780 -182.053 35.420(0.41) (-0.35) (-0.53) (-0.60) (0.03) (-0.56) (-1.03) (0.42) (-0.56) (0.14)
Teacher Characteristics Y Y Y Y Y Y Y Y Y YCampus-Month FE Y Y Y Y Y Y Y Y Y Y
N 131,964 171,503 139,921 163,546 143,597 159,870 79,536 223,931 143,268 160,199R2 0.409 0.347 0.394 0.368 0.370 0.392 0.507 0.289 0.376 0.387
This table shows OLS regressions in which the dependent variable is a 0/1 indicator of refinances, while the main variable of interest is Peer Refinances,a variable that captures the number of peers having undertaken a mortgage refinance in the previous 3-month period, scaled by the size of the peergroup. The regressions also include controls for other determinants of refinancing such as the estimated savings, a 0/1 indicator of whether theteacher is underwater, and the underwater amount (in percent). Campus-month fixed effects and teacher characteristics such as gender, ethnicity,highest academic degree obtained, and salary are also included. The sample has been segmented into two mutually exclusive groups based on whetherthe teacher is above or below the median in terms of tenure, age, pay, or house price value or based on whether the teacher holds a graduatedegree (Yes) or does not (No). Reported are the effects of a one standard deviation change in Peer Refinances, and t-statistics in parentheses areheteroskedasticity-robust and clustered by MSA-year. ***p<0.01, **p<0.05, *p<0.1.
44
TABLE 4–Negative Net Savings
(1) (2) (3) (4) (5)
Peer Refinances 6.189*** 5.338*** 5.827*** 4.525 4.670(3.21) (2.70) (2.73) (1.27) (1.31)
Savings ($, ×10,000) 3.100 1.159 -1.544 6.565 6.080(0.62) (0.24) (-0.26) (0.66) (0.61)
1(Underwater) -4.370 -5.495 -5.717 -0.767 -0.656(-0.58) (-0.71) (-0.63) (-0.08) (-0.06)
Percent Underwater 109.134 149.026 87.799 -33.722 -51.563(0.71) (1.03) (0.58) (-0.19) (-0.29)
Teacher Characteristics N N N N YMSA-Year FE Y N N N NDistrict-Year FE N Y N N NCampus-Year FE N N Y N NCampus-Month FE N N N Y Y
N 172,756 172,756 172,756 172,756 172,756R2 0.013 0.023 0.055 0.299 0.299
This table shows OLS regressions in which the dependent variable is a 0/1 indicator of refinances, and themain variable of interest is Peer Refinances, a variable that captures the number of peers having undertaken amortgage refinance in the previous 3-month period, scaled by the size of the peer group. The main differencewith Table 2 is the inclusion of only observations in which estimated net savings are negative. Reportedare the effects of a one standard deviation change in Peer Refinances, and t-statistics in parentheses areheteroskedasticity-robust and clustered by MSA-year. ***p<0.01, **p<0.05, *p<0.1.
45
TABLE 5–Peer Effects on the Decision to Refinance: Additional Tests
Panel A: Excluding September and October(1) (2) (3) (4) (5)
Peer Refinances 11.985*** 11.122*** 9.597*** 7.740** 7.828**(4.91) (4.25) (3.28) (2.17) (2.19)
Savings ($, ×10,000) 48.407*** 48.597*** 52.215*** 56.491*** 56.448***(6.63) (6.64) (6.73) (6.00) (6.00)
1(Underwater) -8.973 -9.931 -9.722 -6.959 -6.166(-1.19) (-1.17) (-1.07) (-0.68) (-0.61)
Percent Underwater -29.107 -29.750 4.188 -56.815 -87.348(-0.31) (-0.23) (0.03) (-0.35) (-0.53)
Teacher Characteristics N N N N YMSA-Year FE Y N N N NDistrict-Year FE N Y N N NCampus-Year FE N N Y N NCampus-Month FE N N N Y Y
N 255,808 255,808 255,808 255,808 255,808R2 0.025 0.034 0.061 0.237 0.237
Panel B: Refinement of Peer Groups Based on Common Ethnicity(1) (2) (3) (4) (5)
Peer Refinances 10.218*** 9.484*** 8.483*** 7.058** 7.288**common ethnicity (4.42) (4.13) (3.18) (2.00) (2.08)
Peer Refinances 6.851*** 6.166*** 5.715*** 1.399 1.054different ethnicity (3.12) (3.00) (2.85) (0.40) (0.30)
Savings ($, ×10,000) 49.044*** 49.464*** 53.021*** 55.956*** 55.906***(6.48) (6.48) (6.67) (5.42) (5.41)
1(Underwater) -7.803 -9.165 -8.588 -5.403 -4.754(-1.09) (-1.15) (-1.01) (-0.54) (-0.47)
Percent Underwater -34.850 -24.934 7.458 -54.166 -80.997(-0.41) (-0.22) (0.06) (-0.35) (-0.53)
Teacher Characteristics N N N N YMSA-Year FE Y N N N NDistrict-Year FE N Y N N NCampus-Year FE N N Y N NCampus-Month FE N N N Y Y
N 298,397 298,397 298,397 298,397 298,397R2 0.024 0.033 0.056 0.242 0.242
This table shows OLS regressions in which the dependent variable is a 0/1 indicator of refinances, and themain variable of interest is Peer Refinances, a variable that captures the number of peers having undertakena mortgage refinance in the previous 3-month period, scaled by the size of the peer group. In Panel A, themain difference with Table 2 is the exclusion of September and October from each year of the sample. PanelB, in turn, segments Peer Refinances into two mutually exclusive groups. The first group contains peers withthe same ethnicity as the individual, while the second group contains peers who differ in ethnicity. Reportedare the effects of a one standard deviation change in Peer Refinances, and t-statistics in parentheses areheteroskedasticity-robust and clustered by MSA-year. ***p<0.01, **p<0.05, *p<0.1.
TABLE 6–Instrumental Variable Estimation
(1) (2) (3) (4)
Peer Refinances 36.309*** 37.323*** 42.474*** 42.518***(6.20) (5.73) (6.56) (6.59)
Savings ($, ×10,000) 47.817*** 48.513*** 51.894*** 51.477***(7.77) (7.69) (7.45) (7.38)
1(Underwater) -4.676 -4.367 -5.184 -5.489(-0.69) (-0.59) (-0.69) (-0.75)
Percent Underwater -25.072 -8.416 37.119 17.932(-0.39) (-0.09) (0.27) (0.13)
Teacher Characteristics N N N YMSA-Year FE Y N N NDistrict-Year FE N Y N NCampus-Year FE N N Y Y
N 245,200 245,167 244,937 244,937F -statistic 24.01 23.62 23.04 23.23
This table shows the second stage of 2-stage least squares regressions in which the dependent variable is a0/1 indicator of refinances, and the main variable of interest is the estimated quantity Peer Refinances, avariable that captures the number of peers having undertaken a mortgage refinance in the previous 3-monthperiod, scaled by the size of the peer group. In the first stage, we use the average net savings conditional onrefinancing of a teacher’s peer group as an exogenous instrument to estimate Peer Refinances. The first-stagepartial F -statistics of the IV is reported, while the full first stage of each specification is reported in InternetAppendix Table IA.5. Reported are the effects of a one standard deviation change in Peer Refinances, andz -statistics in parentheses are heteroskedasticity-robust and clustered by MSA-year. ***p<0.01, **p<0.05,*p<0.1.
47
TABLE 7–Cox Proportional Hazard Model
(1) (2) (3) (4) (5)
Peer Refinances 1.041*** 1.045*** 1.054*** 1.102*** 1.102***(4.03) (4.19) (5.69) (5.29) (5.52)
Savings ($, ×10,000) 1.550*** 1.586*** 1.624*** 1.623*** 1.633***(16.71) (12.12) (12.47) (12.77) (12.46)
1(Underwater) 0.961 0.899 0.890 0.916 0.927(-0.53) (-1.00) (-1.02) (-0.79) (-0.67)
Percent Underwater 0.699 1.241 1.318 1.086 0.668(-0.41) (0.24) (0.30) (0.09) (-0.42)
Teacher Characteristics N N N N YMSA-Year Refi Control Y N N N NDistrict-Year Refi Control N Y N N NCampus-Year Refi Control N N Y N NCampus-Month Refi Control N N N Y Y
N 252,706 252,703 252,660 241,131 241,131
This table shows hazard ratios of Cox Proportional Hazard model regressions in which the dependent variableis a 0/1 indicator of refinances, and the main variable of interest is Peer Refinances, a variable that capturesthe number of peers having undertaken a mortgage refinance in the previous 3-month period, scaled by thesize of the peer group. The regressions also include controls for other determinants of refinancing such asthe estimated savings, a 0/1 indicator of whether the teacher is underwater, and the underwater amount (inpercent). Depending on the specification, a Metropolitan Statistical Area (MSA)-year refinancing activitycontrol, a district-year refinancing activity control, a campus-year refinancing activity control, a campus-month refinancing activity control, or teacher characteristics such as gender, ethnicity, highest academicdegree obtained, and salary are also included. Reported are the effects of a one standard deviation changein Peer Refinances, and z -statistics in parentheses are heteroskedasticity-robust and clustered by MSA-year.***p<0.01, **p<0.05, *p<0.1.
48
Teachers Teaching Teachers: The Role of Networks on Financial Decisions
Gonzalo Maturana and Jordan Nickerson
Internet Appendix
Additional Tables and Figures
Table IA.1–Correlation Matrix of Teacher Characteristics
Pay Tenure Grad Degree Age Female
Pay 1Tenure 0.52 1Grad Degree 0.29 0.17 1Age 0.47 0.48 0.24 1Female -0.16 0.04 -0.02 -0.0001 1
This table shows the correlation coefficients among teacher characteristics.
1
Table IA.2–Robustness for Table 2
Panel A: Alternative measure of Peer Refinances(1) (2) (3) (4) (5)
Peer Refinances 11.854*** 10.851*** 9.492*** 6.554** 6.618**(5.55) (5.08) (3.97) (2.05) (2.08)
Savings ($, ×10,000) 48.903*** 49.243*** 53.010*** 56.310*** 56.307***(6.46) (6.46) (6.74) (5.45) (5.45)
1(Underwater) -7.638 -9.089 -7.887 -5.400 -4.712(-1.10) (-1.16) (-0.93) (-0.57) (-0.50)
Percent Underwater -10.605 5.204 14.607 -33.544 -58.994(-0.13) (0.05) (0.12) (-0.24) (-0.42)
Teacher Characteristics N N N N YMSA-Year FE Y N N N NDistrict-Year FE N Y N N NCampus-Year FE N N Y N NCampus-Month FE N N N Y Y
N 303,467 303,467 303,467 303,467 303,467R2 0.024 0.032 0.055 0.242 0.242
Panel B: Alternative measure of Savings(1) (2) (3) (4) (5)
Peer Refinances 11.075*** 10.153*** 9.164*** 6.281* 6.343*(4.81) (4.39) (3.62) (1.93) (1.96)
Savings ($, ×10,000) 44.621*** 45.032*** 48.668*** 54.139*** 54.126***(6.66) (6.70) (7.00) (5.41) (5.40)
1(Underwater) -9.967 -11.383 -10.167 -8.621 -7.888(-1.45) (-1.46) (-1.21) (-0.93) (-0.85)
Percent Underwater -18.839 -7.812 -3.728 -47.085 -72.377(-0.23) (-0.07) (-0.03) (-0.33) (-0.51)
Teacher Characteristics N N N N YMSA-Year FE Y N N N NDistrict-Year FE N Y N N NCampus-Year FE N N Y N NCampus-Month FE N N N Y Y
N 303,467 303,467 303,467 303,467 303,467R2 0.024 0.032 0.055 0.242 0.242
This table shows OLS regressions in which the dependent variable is a 0/1 indicator of refinances, and themain variable of interest is Peer Refinances, a variable that captures the number of peers having undertakena mortgage refinance in the previous 3-month period, scaled by the size of the peer group. In Panel A, themain difference with Table 2 is the definition of Peer Refinances, which only considers refinances done bypeers in the last two months (instead of three months). Panel B, in turn, uses an alternative measure ofsavings which assumes a term of 30 years. Reported are the effects of a one standard deviation change inPeer Refinances, and t-statistics in parentheses are heteroskedasticity-robust and clustered by MSA-year.***p<0.01, **p<0.05, *p<0.1.
2
Table IA.3–Teacher Characteristics and Group Characteristics
Employment Characteristics DemographicsAge Pay Grad Degree Female White Black Asian Hispanic(1) (2) (3) (4) (5) (6) (7) (8)
Peer Group Average 0.030*** 0.042*** 0.006 0.075*** 0.026** 0.009 -0.001 0.011(3.31) (3.95) (0.66) (8.10) (2.36) (0.63) (-0.11) (0.87)
Campus Average 0.424*** 0.830*** 0.407*** 0.469*** 0.863*** 0.891*** 0.139*** 0.878***(26.35) (86.22) (20.64) (30.65) (71.43) (56.50) (4.19) (57.23)
N 187,369 187,369 187,369 187,369 187,369 187,369 187,369 187,369R2 0.024 0.254 0.017 0.038 0.287 0.286 0.001 0.303
This table shows OLS regressions of a teacher’s characteristic on the average of her peer group (Peer Group Average) and the average of all teacherswithin the campus but outside the teacher’s peer group (Campus Average), using all teachers in the TEA records. Reported t-statistics in parenthesesare heteroskedasticity-robust and clustered by campus. ***p<0.01, **p<0.05, *p<0.1.
3
Table IA.4–Alternative Refinements of Peer Groups
Panel A: Common Age(1) (2) (3) (4) (5)
Peer Refinances 7.245*** 6.336*** 4.987** 3.779 3.842common age group (3.75) (3.32) (2.39) (1.41) (1.42)
Peer Refinances 9.251*** 8.940*** 8.983*** 6.132 6.202different age group (4.93) (4.58) (4.05) (1.57) (1.59)
Savings ($, ×10,000) 49.060*** 49.470*** 53.023*** 55.955*** 55.912***(6.47) (6.48) (6.67) (5.41) (5.41)
1(Underwater) -7.799 -9.174 -8.613 -5.382 -4.734(-1.09) (-1.15) (-1.01) (-0.54) (-0.47)
Percent Underwater -36.040 -25.252 7.299 -54.718 -81.225(-0.42) (-0.22) (0.06) (-0.36) (-0.53)
Teacher Characteristics N N N N YMSA-Year FE Y N N N NDistrict-Year FE N Y N N NCampus-Year FE N N Y N NCampus-Month FE N N N Y Y
N 298,397 298,397 298,397 298,397 298,397R2 0.024 0.033 0.056 0.242 0.242
Panel B: Common Gender(1) (2) (3) (4) (5)
Peer Refinances 8.508*** 7.939*** 6.515*** 4.487 4.562common gender (4.33) (3.87) (2.77) (1.37) (1.40)
Peer Refinances 9.737*** 8.743*** 9.141*** 6.964 7.041different gender (3.44) (3.06) (3.01) (1.60) (1.60)
Savings ($, ×10,000) 49.065*** 49.475*** 53.030*** 55.958*** 55.910***(6.48) (6.48) (6.67) (5.42) (5.41)
1(Underwater) -7.788 -9.162 -8.577 -5.358 -4.714(-1.09) (-1.15) (-1.01) (-0.53) (-0.47)
Percent Underwater -36.140 -25.328 6.800 -55.000 -81.479(-0.42) (-0.22) (0.05) (-0.36) (-0.53)
Teacher Characteristics N N N N YMSA-Year FE Y N N N NDistrict-Year FE N Y N N NCampus-Year FE N N Y N NCampus-Month FE N N N Y Y
N 298,397 298,397 298,397 298,397 298,397R2 0.024 0.033 0.056 0.242 0.242
This table shows OLS regressions in which the dependent variable is a 0/1 indicator of refinances, and themain variable of interest is Peer Refinances, a variable that captures the number of peers having undertaken amortgage refinance in the previous 3-month period, scaled by the size of the peer group. The main differencewith Table 5 is the segmentation of Peer Refinances into two mutually exclusive groups. The first groupcontains peers with the same characteristic as the individual, while the second group contains peers whodiffer in the characteristic. In Panel A, peers are defined as belonging in the same age group if the absolutedifference in ages is less than 10 years. Panel B, in turn, segments the peer group based on having the samegender as the individual. Reported are the effects of a one standard deviation change in Peer Refinances, andt-statistics in parentheses are heteroskedasticity-robust and clustered by MSA-year. ***p<0.01, **p<0.05,*p<0.1.
Table IA.5–First Stage Regressions
(1) (2) (3) (4)
Avg. Peer Savings 0.385*** 0.381*** 0.400*** 0.401***(4.90) (4.86) (4.80) (4.82)
Savings ($, ×10,000) -0.074*** -0.070** -0.034 -0.027(-2.82) (-2.60) (-0.90) (-0.71)
1(Underwater) -0.044 -0.057 -0.071 -0.054(-0.64) (-0.83) (-1.16) (-0.87)
Percent Underwater -0.490 -0.016 -0.114 -0.071(-0.50) (-0.02) (-0.14) (-0.09)
Teacher Characteristics N N N YMSA-Year FE Y N N NDistrict-Year FE N Y N NCampus-Year FE N N Y Y
N 245,201 245,201 245,201 245,201R2 0.032 0.080 0.211 0.211
This table reports the first stage of the 2SLS IV regressions of Table 6. The dependent variable, PeerRefinances, instrumented for in the second stage, is the number of a teacher’s peers who have undertakena mortgage refinance in the previous 3-month period, scaled by the size of the teacher’s peer group. Weuse the average net savings conditional on refinancing of a teacher’s peer group, Avg. Peer Savings, as anexogenous instrument to estimate Peer Refinances. The variables Avg. Peer Savings and Peer Refinanceshave been standardized. Reported t-statistics in parentheses are heteroskedasticity-robust and clustered byMSA-year. ***p<0.01, **p<0.05, *p<0.1.
5
Table IA.6–First Mover and Follower Characteristics
Trigger Follower Difference t-statistic
Unpaid Balance ($) 138,700.9 138,190.5 510.4 0.13Savings ($) 6,087.8 6,600.1 -512.3 -1.07Pay ($) 47,807.3 46,954.8 852.5 2.08Tenure 6.17 6.33 -0.167 -0.48Technical Class 0.26 0.27 -0.001 -0.06Grad Degree 0.27 0.26 0.008 0.37Age 41.64 41.33 0.307 0.61Female 0.67 0.63 0.041 1.72
This table compares the characteristics of teachers that refinance independently of their peers with those ofteachers who follow their peers.
6