The Economic Consequences of Mergers Between Real Estate

The Economic Consequences of Mergers Between

Real Estate Agencies and Mortgage Lenders

Rebecca A. Jorgensen

November 15, 2023

(Click Here for Latest Version)

Abstract

This paper studies the consequences of joint ownership between real estate agencies

and mortgage lenders for consumers, lenders, and mortgage market structure. I con-

struct a novel data set which matches home buyers’ real estate agencies, lenders, and

loan characteristics while tracking ownership of lenders and agencies over time. Using

hand-collected data for over 100 mergers involving real estate agencies or lenders, I

implement a staggered diﬀerences-in-diﬀerences strategy that compares lender-agency

pairs which are jointly owned due to horizontal mergers between real estate agencies to

lender-agency pairs that are never jointly owned. After merging, lenders double their

loan shares within jointly owned real estate agencies with little impact on a lender’s

CBSA market share. Buyers who use a lender jointly owned with their real estate

agency pay interest rates 9 basis points higher, amounting to 225 in additional in-

terest per year on the average loan. However, I ﬁnd no evidence that home buyers’

credit characteristics, delinquency rates, or transaction speed change following these

mergers. Finally, I develop a structural model of the mortgage market to study the

welfare implications of mergers under counterfactual policies. I ﬁnd that completely

banning mergers harms consumers, while allowing mergers that promote competition

can improve consumer welfare.

I would like to thank my committee, Ben Keys, Maisy Wong, and Fernando Ferreira for their advice

on this project. I would also like to thank Sasha Indarte, Lu Liu, Jessie Handbury, James Vickery, Bryan

Stuart, Ryan Goodstein, You Suk Kim, Charles Taragin, David Benson, and seminar participants at the

Wharton Urban Lunch, the Federal Reserve Bank of Philadelphia, the FDIC, and the Board of Governors

of the Federal Reserve System for their insights.

Disclaimer: The opinions and analyses contained herein are solely of the users/authors of any data

analyses or papers, and the FHFA cannot and does not attest to nor vouch for the quality, accuracy,

or timeliness of the data, or analyses derived from these data after the data has been retrieved from

FHFA.gov. This material is based upon work supported by the National Science Foundation Graduate

Research Fellowship Program under Grant No. DGE-1845298. Any opinions, ﬁndings, and conclusions or

recommendations expressed in this material are those of the author and do not necessarily reﬂect the views

of the National Science Foundation. The author received a travel grant from the IAAE to present this paper

in June 2023.

1 Introduction

The increasing sophistication of technology has led to its integration in all aspects of life,

including ﬁnance. Such “FinTech” advances have led to online-only banking, investing, and

lending. With respect to mortgages, Buchak et al. (2018) documents that non-banks, includ-

ing technology-rich online-only lenders, have gained signiﬁcant market share in the post-crisis

period. These non-banks lack the pre-existing customer base of traditional banks and thus

must ﬁnd a diﬀerent way to reach customers. As a solution, some of these non-bank lenders

have merged with real estate agencies. However, very little is known about how these merg-

ers aﬀect mortgage markets.

This paper studies the consequences of joint ownership between real estate agencies and

mortgage lenders. For many households, a mortgage represents the largest debt and debt

ﬁnancing obligation they will face. When a real estate agency and a lender share a common

parent company, understanding how this joint ownership aﬀects how consumers access loan

products and the ensuing competitive eﬀects in the mortgage market will be crucial given

the size and scope of the mortgage market. However, determining the consequences of joint

ownership in this market is challenging due to identiﬁcation and data issues. Beginning with

data, the main challenge is that there is no single data set that contains all the information

necessary to analyze the potential consequences of joint ownership. In terms of identiﬁca-

tion, simply comparing outcomes before and after a lender and agency are jointly owned is

not suﬃcient, because the ﬁrms that decide to merge are likely to be selected.

First, to address the data issue, I construct a novel data set that matches loans to the

originating lenders, the underlying home purchase, and the buyers’ real estate agencies by

merging four CoreLogic data sets. I then hand-match mergers involving real estate agencies

and/or lenders into this data set to obtain the ownership structure over time. By construct-

ing this data set, I am able to identify whether the lender and agency were jointly owned at

any point in time in my data, including at the time of the home purchase and loan origina-

tion. I observe more than 100 mergers over the span of my data, and more than one million

home purchases matched to the loan origination funding the purchase.

Second, to solve the identiﬁcation issue, I exploit the fact that horizontal mergers between

real estate agencies occur frequently, and that these horizontal mergers indirectly change

which lenders are jointly owned with which real estate agencies, what I will call “indirect in-

tegration”. If these mergers between agencies occur to consolidate the involved ﬁrms’ agency

businesses and the lender’s books are not the primary concern, then these mergers create

quasi-random variation in joint ownership from the perspective of the lender. Of the mergers

I identify, 87 mergers ﬁt these criteria. This large number of mergers and observations allow

me to include rich ﬁxed eﬀects (over 50,000 total) to control for lender, agency, time, or

geographic factors which could bias my results.

I use this data set and research design in a staggered diﬀerence-in-diﬀerences model to ana-

lyze the consequences of joint ownership between real estate agencies and mortgage lenders

for market shares, prices, speed of transactions, characteristics of borrowers, and ex-post

loan performance. Lender-agency pairs impacted by indirect integration are treated, while

other lender-agency pairs that are never jointly owned function as a control group. To the

best of my knowledge, this paper is the ﬁrst to study this important class of mergers in a

comprehensive manner.

Beginning with the market structure eﬀects, I ﬁnd that lenders jointly owned with a real

estate agency more than double their loan share at that agency upon integration, consistent

with the theory that jointly owned agencies direct buyers to their sibling lender. Doubling

their loan share is a signiﬁcant change for the lender’s portfolio, and suggests that agencies

direct buyers to the in-house lender when it exists, giving the lender market power over those

consumers. When looking instead at the lender’s market share in the core-based statisti-

cal area (CBSA), lenders jointly owned with a real estate agency gain only 0.54 percentage

points higher market share. While this represents a 16% increase in the average lender

market share, making it signiﬁcant for the lender, the overall market structure is nearly

unaﬀected due to the fragmented nature of the mortgage market.

With respect to interest rates, I ﬁnd that buyers using a lender jointly owned with their

real estate agency pay 9 basis points more on average, which is consistent with the theory

that joint ownership allows lenders to “capture” some borrowers and charge these borrow-

ers higher interest rates that dominate any cost reductions from joint ownership passed on

to borrowers. This result holds even after accounting for the relationship between search

behavior and interest rates, as documented by Bhutta et al. (2020). I argue that 9 basis

points is a large eﬀect in the context of mortgage rates; it is similar to the 7 basis point

premium that minority buyers pay as found by Bartlett et al. (2021), more that 15% of the

total interest rate dispersion documented by Bhutta et al. (2020), and amounts to an addi-

tional 225 in interest payments per year on the average loan. In sum, joint ownership has

important market share consequences for lenders and leads to higher interest rates for buyers.

While buyers pay more when going to a merged lender-agency pair, doing so does not change

how fast they can close on their mortgage or the credit proﬁle they need to get a loan. I

ﬁnd no evidence of the possibility suggested by theory that lenders are cream-skimming the

best credit proﬁles from their jointly owned agency. The lack of change in borrower charac-

teristics also suggests that lenders are not receiving additional information from the agency

which the lender then uses to lend to riskier buyers, which is another possibility suggested

by theory. I also ﬁnd no eﬀect on ex-post loan performance. This result contrasts with the

ﬁndings in Stroebel (2016), where home buyers using a builder’s lender were less likely to

default on their mortgages.

I further enrich these reduced form results with a structural model of the mortgage mar-

ket which allows me to test for marginal cost eﬃciencies from joint ownership and to run

regulatory policy counterfactuals. I construct a logit demand model for loans with consumer-

speciﬁc choice sets. To estimate the demand model, I ﬁrst use maximum likelihood to recover

mean utility for each loan product. I then use two stage least squares to recover the con-

tribution of each product characteristic to a product’s overall mean utility. The parameter

estimates from this structural model are consistent with my reduced form ﬁndings that con-

sumers value jointly owned agency-lender pairs, pre-existing relationships with lenders, and

faster closing times. I pair this demand model with a supply model of proﬁt-maximizing

lenders oﬀering diﬀerentiated mortgage products under Bertrand competition.

Using the estimates from my structural model, I study three diﬀerent counterfactual reg-

ulatory policies the government could enact. All of these policies regulate lender behavior

and consequently inﬂuence borrowers’ access to loan products. In the ﬁrst two, joint own-

ership between real estate agencies and mortgage lenders is banned, but the two diﬀer in

the choices available to buyers previously choosing the jointly owned products. In the third,

joint ownership is permitted, but lenders are forbidden from oﬀering diﬀerent products to

consumers based on what real estate agency they use. I ﬁnd that banning mergers slightly

decreases consumer welfare in both of the ﬁrst two counterfactuals, but that welfare gains

come from permitting mergers but requiring that lenders oﬀer all products to all consumers.

In short, a simple ban on mergers reduces welfare, while more nuanced regulatory policy can

promote competition and thus improve welfare.

This paper contributes to several literatures within household ﬁnance and banking: First, I

contribute to the literature on mortgage price dispersion and the role of steering in ﬁnan-

cial markets. This paper oﬀers a novel channel to explain the dispersion in mortgage rates

documented and partially explained in prior papers such as Bhutta et al. (2020), Bartlett

et al. (2021), and Bhutta and Hizmo (2021). It also complements studies on other forms

of steering in mortgage and home sales [Guiso et al. (2021), Agarwal et al. (2016), Barwick

et al. (2017), Woodward and Hall (2012)]. I connect steering across mortgages and home

purchases, and examine a particular kind of steering: bringing another good in-house. To the

best of my knowledge, only one other paper looks at steering linking lenders and real estate

agencies, that of Lopez et al. (2019). Relative to this earlier paper, I look at the buyer’s

agent (who buyers are more likely to trust) instead of the seller’s agent, while including a

much larger geography, additional outcomes, a diﬀerent identiﬁcation strategy that utilizes

variation at a level above the transaction, and a structural model to estimate policy counter-

factuals. Joint ownership is also common between car dealerships and auto lenders, another

large household ﬁnancing decision. Grunewald et al. (2020) ﬁnd that the discretion given

to dealers in these arrangements incentivizes them to increase the rate over the minimum

speciﬁed by the lender. While the settings are not identical, the results in this paper on the

mortgage market can provide insight into consequences in the auto lending market as well.

Lastly, this paper contributes to the literature on industrial organization and banking and

to the debate on whether, and how, policy makers should regulate mergers involving banks

and non-banks, as well as the consequences of the rise of non-banks and integration. Prior

literature has shown that mergers in ﬁnancial industries can have costs and beneﬁts and

inﬂuence pricing behavior [Berger et al. (1999), Stroebel (2016), Robles-Garcia (2019)]. Re-

lated papers, such as Buchak et al. (2018), have documented the rise of non-banks in spheres

traditionally occupied by banks. Here, I oﬀer a rich study of the consequences of mergers

in one particular ﬁnancial market while studying a broad set of potential consequences with

more mergers and richer outcomes than the current literature.

For a more general discussion of complementary goods mergers among non-ﬁnancial products see Akg¨un

et al. (2020), Choi (2008), and Ershov et al. (2018).

The rest of this paper is organized as follows. Section 2 provides institutional background

for the setting of my paper. Section 3 presents survey evidence exploring the interaction

between real estate agents, mortgage lenders, and borrower search behavior. Section 4 dis-

cusses a merger case study. Section 5 discusses my data; Section 6 presents the reduced form

identiﬁcation strategy, and Section 7 presents the results. Section 8 presents and estimates

my structural model, and Section 9 concludes.

2 Institutional Background

This paper requires institutional background on real estate agencies, mortgage lenders, and

the mortgage market. Beginning with real estate agencies, real estate agencies employ real

estate agents to work with home buyers and sellers. In the United States, 86% of buyers

used a real estate agent in 2020.

In general, buyers are satisﬁed with the quality of the

service they receive from their real estate agent, with 89% saying they would use the same

agent again or recommend the agent to others.

Buyer’s agents are paid out of the seller’s

agent’s commission when the clients ﬁnd a home. Thus, agents are not paid until after the

client closes on a home. Besides helping buyers ﬁnd homes and negotiate the purchase price

of their home, real estate agents will recommend providers of related services to their clients,

such as title companies and lenders.

Real estate agents cannot receive kickbacks from the ﬁrms they recommend to their clients

under the Real Estate Settlements Procedure Act (RESPA), but they are allowed to share

costs if the ﬁrms share things such as advertising or oﬃce space. Companies have violated

this; in August 2023 the Consumer Financial Protection Bureau ﬁned Freedom Mortgage for

violating this and paying more than cost sharing to Revolve Realty. Similarly, ReMax has

See: https://www.nar.realtor/research-and-statistics/quick-real-estate-statistics

been ﬁned in the past for oversharing with mortgage lenders.

Moving to mortgage lenders, mortgage lenders can attract consumers directly through their

retail channel, but they can also use mortgage brokers. Mortgage brokers are third party

agencies which lenders can contract with to help them ﬁnd more borrowers. Mortgage bro-

kers contract with several lenders, and receive a commission from the lender when a buyer

chooses that lender. Thus, borrowers going to a mortgage broker receive loan terms from

several lenders while borrowers going through a lender’s retail channel will see only loans

from that lender.

When buying a home, 87% of buyers ﬁnanced with a mortgage in 2021.

. The vast ma-

jority of mortgages in the United States are ﬁxed rate mortgages, meaning that the rate on

the mortgage will remain constant for the duration of the loan.

Loans in the United States can either be originated through banks or non-bank lenders.

Non-bank lenders originate the loans through lines of credit they hold with traditional banks

(called warehouse lines), which they then securitize and sell. Most loans are backed by the

FHA or government sponsored entities, Fannie Mae and Freddie Mac. These agencies pro-

vide detailed guidance on what the maximum value of the loan can be depending on the

average home price in the area. In addition, the agencies give lenders a matrix of ”Loan-Level

Pricing Adjustments”, which are increases to the interest rate lenders must charge based on

the borrower’s loan-to-value ratio (LTV) and credit (FICO) score.

Prior to the 2008 ﬁnancial crisis, mortgage brokers originated over half of loans in the

See:https://www.consumerfinance.gov/abo ut-us/newsroom/cfpb-penalizes-freedom-mortgag

e-and-rea lty-connect-fo r-illegal-kickbacks/ and https://www.consumerfinance.gov/enforcem

ent/actions/rgc-services-inc-dba-remax-gold-coast-realtors/

See: https://www.nar.realtor/research-and-statistics/research-reports/highlights-f ro

m-the-profile-of-home-buyers-and-sellers

United States. Following federal regulations which made mortgage brokers less proﬁtable,

many mortgage brokers went out of business, and in 2019, mortgage brokers originated only

19% of loans in the United States. Thus, for lenders looking for borrowers, they cannot rely

on mortgage brokers to ﬁnd them. Now, lenders must ﬁnd another channel through which

to ﬁnd borrowers.

At the same time the share of borrowers ﬁnding loans through mortgage brokers is de-

clining, residential real estate agencies are merging with lenders and related services. Home

purchases and mortgages are complementary goods; as the majority of buyers ﬁnance their

home purchase with a mortgage, these two goods go hand in hand. Thus, mergers between

residential real estate agencies and lenders can be proﬁtable for both parties. Residential real

estate agencies provide a set of customers for lenders who can no longer rely on mortgage

brokers, and bringing a lender in house could simplify the loan search process for buyers,

allowing them to close on a home, and the agent to get paid, more quickly.

While this paper focuses on mergers between lenders and residential real estate agencies,

these agencies are also merging with other parts of the home buying process, such as title

insurance. Gino Blefari, CEO of HomeServices of America, which owns multiple lenders, real

estate agencies, and title insurance companies, said that ”[c]reating a seamless all-inclusive

shopping experience for a consumer’s real estate transaction – agency, mortgage, title &

homeowners insurance is critical for both the best consumer experience and the best path to

proﬁtability,” suggesting that merging ﬁrms see the complementarities between these goods

and that merging them improves overall performance.

Figures 1a, 1b, and 1c show how the market has evolved over time. Beginning with Figure

1a, this shows what share of home purchases are made using an agency that has a jointly

owned lender, also referred to as a sibling lender. Since 2014, this share has increased by

nearly 25%, such that in 2019 almost half of home purchases used an agency with a sibling

lender. That is, the agencies which lenders are able to use as a distribution channel are a

large fraction of the market and thus come with a large pool of buyers these lenders could

attract.

The story is similar when looking at lenders. In Figure 1b, we can see that the share of

loans originated by lenders merged with agencies has increased from 3.5% in 2014 to 6% by

2019. Although my data end in 2019, this trend has not stopped in the intervening years.

While the agencies are large players in the market, the lenders have considerably smaller

market share. This is consistent with the state-level licensing to originate mortgages being

a high barrier to entry for lenders. Most lenders are geographically concentrated in at most

a handful of states.

Lastly, in Figure 1c, I show the share of home purchases using a merged agency-lender

Notes: Figure shows share of home purchases which use a real estate agency that has a sibling lender at

time of home purchase, regardless if lender originated the mortgage for that purchase or not.

Figure 1a: Share of Home Purchases with Conglomerate Agency

pair. These are the purchases and loan originations which I am most interested in studying.

The share of home purchases which are using a merged agency-lender pair is still small at

2% of my data, but given the growth of merged agencies and lenders in the previous ﬁg-

ures as well as the continued addition of new joint ﬁrms, this share is likely to grow over

Notes: Figure shows share of purchase loan originations which use a lender that has a sibling real estate

agency at time of home purchase, regardless if agency was involved in the transaction or not.

Figure 1b: Share of Loans Originated by Conglomerate Lender

Notes: Figure shows share of home purchases which use a jointly owned agency-lender pair for home

purchase and mortgage ﬁnancing.

Figure 1c: Share of Purchases using Conglomerate Agency - Lender Pair

time. Furthermore, while 2% of all buyers is a small fraction, that is 33% of the market share

of merged lenders, indicating that this is an important distribution channel for these lenders.

3 National Survey of Mortgage Originations

The National Survey of Mortgage Originations (NSMO) is a quarterly survey conducted by

the FHFA and the CFPB asking recent home buyers a variety of questions about their home

purchase and mortgage acquisition process. Questions include topics such as their under-

standing of mortgage terms, house price expectations, and how the buyer chose a lender. It

is the questions on this last topic which are relevant for this paper and which I discuss in

this section.

When asked to select which features were not at all, somewhat, or very important to them,

34% of buyers said that the recommendation of their real estate agent was somewhat or very

important.

Moreover, 17% of buyers were introduced to their lender by an interested third

party such as their builder or real estate agent. In addition, nearly half (49%) of buyers

report only seriously considering one lender.

When restricting the survey data to ﬁrst-time home buyers, 58% valued their real estate

agent’s recommendation, and 33% were directly introduced to their lender through an in-

terested third party. Similar to the general population, 47% of buyers only seriously look at

one lender.

Similarly, when looking at low credit score borrowers, even more, 64% of borrowers found

their agent’s recommendation somewhat or very important, 18% were introduced by an in-

Other options include: rate (98%), lender reputation (71%), prior relationship with the lender (57%),

local lender (50%), used lender before (39%), recommendation of a friend (36%), and lender is a friend (14%).

terested third party, and 51% only seriously consider one lender.

Taken together, these survey results indicate that, especially for low credit score and ﬁrst-

time home buyers, the recommendation of the real estate agent inﬂuences the lender they

ultimately choose. Thus, lenders and real estate agencies merging has the potential to lead

agents to recommend their sibling lender, which has bearing on the lender chosen by buyers.

Thus, a real estate agency merging with a lender will likely increase the number of buyers

coming to the lender from that agency. However, the eﬀects on overall lender market share

are unclear. If the lender treats these borrowers as perfect substitutes for other borrowers

coming from non-merged real estate agencies, then overall market share will not go up, only

the market share coming from the sibling ﬁrm. However, if the lender treats the sibling bor-

rowers as additional customers, without completely substituting away from buyers coming

from agencies with whom they are not merged, then both the lender’s market share within

their sibling agency and their overall market share will increase.

Similarly, this survey data provides no evidence on price eﬀects as interest rates are not

available in the public use version of the NSMO. However, the lack of search by approx-

imately half of borrowers leaves borrowers open to unknowingly paying supracompetitive

prices as a result of the integration and lack of search as previously discussed.

4 Data

The data to complete this analysis are diﬃcult to compile. No pre-existing data set contains

all the necessary pieces: lender, real estate agency, and loan characteristics. Furthermore,

corporate ownership structure plays a vital role in this project but is not readily matched to

the other pieces. Thus assembling the data explained below is no small feat and represents a

contribution to the ﬁeld as this data set is useful for projects beyond this one. In all ﬁles, I

restrict my sample to 2011-2019 so as to avoid contamination from either the 2008 ﬁnancial

crisis or the COVID-19 pandemic.

CoreLogic Deeds Data The CoreLogic Deeds data include details on residential prop-

erties at the time of sale. This includes characteristics of the property, the deed transfer,

and most notably for this project, the mortgage used to purchase the home. The mortgage

details include such ﬁelds as the amount, the start date, loan term, interest rate, loan pur-

pose and type, borrower name, and lender. While these variables all appear at least once

in the data set, some, most notably interest rate, are missing the vast majority of the time.

Lender name, however, is well populated, allowing me to see the lender who issued the loan.

I will also restrict this analysis to purchase loans, since purchases are when borrowers are

most directly involved with a real estate agent, and when the inﬂuence coming from mergers

between lenders and agencies are likely to be strongest. In comparison, at the point of reﬁ-

nancing, borrowers are not working directly with a real estate agent, and while the decision

of a reﬁnancing lender may be correlated with the the original choice of lender, the actual

link is less clear and I leave to future research.

CoreLogic MLS Data CoreLogic provides a second data product, the MLS data. This

data set covers the MLS feeds for 138 MLSes across the United States. Similar to Zillow,

this contains all ﬁelds from the listing of a home: list price, transaction price, date sold, date

listed, home characteristics such as number of bedrooms and bathrooms, address, square

footage, etc. It also contains additional ﬁelds, including information on the buyer’s real

estate agent, including their name and agency at the time of sale.

I will use the term agent

and agency interchangeably.

For this project, I utilize the agency as opposed to the individual agent. In cases of mergers, directives

to recommend a given lender are likely to come from the parent company and aﬀect all agents. Second, in

cases where a team of agents work with a buyer, it is not clear who the purchase should be attributed to

from the team. It is far more obvious which agency is responsible.

CoreLogic LLMA Originations Data The CoreLogic LLMA Originations data, or Loan-

Level Market Analytics Originations data come from a large loan servicer. This data contain

individual loan characteristics at origination, including the interest rate, amount, origination

month and year, property type, loan type, loan term, location, FICO score of borrower at

origination, and loan-to-value-ratio at origination. From here, the key variables I use are

interest rate, FICO score, and LTV ratio.

CoreLogic LLMA Events Data The CoreLogic LLMA Events data track the major

performance events of a loan including the ﬁrst day of 30, 60, and 90 day delinquency, the

ﬁrst date of a bankruptcy ﬁling, and the ﬁrst date of foreclosure ﬁling. I use these data to

compare the performance of loans by merged lender-agent pairs and those by other lenders

who may have less soft information to base their lending decision on.

CompuStat Transactions Data To identify mergers and back out ownership, I use the

CompuStat Transactions Data. This data set records all mergers and acquisitions back to

2000, the target, the acquirer, the transaction date, and additional details. I have gone

through this data and identiﬁed all relevant transactions by hand.

Due to the large number

of real estate oﬃces in the country, not all mergers are going to be widely publicized. To the

best of my knowledge, my data set is the most comprehensive which tracks solely real estate

ﬁrms.

Home Mortgage Disclosure Act The Home Mortgage Disclosure Act (HMDA) pub-

lic use data are redacted loan-level data originally provided to the federal government in

order to ensure fair lending practices. Across all years of data, this includes borrower char-

acteristics such as race and gender. For the last two years of my sample, this also includes

The similar nature of many real estate company names makes a fuzzy match or other algorithmic

approach impossible.

loan fees and discount points.

Loan fees are fees paid at origination to the lender, and are

a function of the loan amount in most cases. Discount points are an additional fee paid

at origination in order to reduce the interest rate on the loan. While two years of data

is a limited time to view either of these, it does provide some insight into other monetary

considerations for borrowers beyond the interest rate, and the possible points-fees tradeoﬀ

as documented in Bhutta and Hizmo (2021).

Summary statistics can be found in Table 1. Due to outliers, I winsorize loan amount,

time to close, and interest rate at the 5 and 95th percentiles. Just over half of my data come

from purchases made with real estate agencies that ever have a sibling lender. This occurs

because the largest real estate ﬁrms in the United States are the ones that tend to have

lending arms. However, only 5% of my sample uses a lender which ever merges with a real

estate agency. I believe this is due to three things: ﬁrst, lenders obtain licenses to originate

mortgages in each state separately, and so not every lender is licensed in every state even

if their eventual parent company has real estate agents in that state, secondly because the

large bank lenders do not have real estate agent arms; lenders with real estate agent arms

are primarily non-bank lenders, and third because lenders which merge later in the sample

period are going to get less beneﬁt from the merger that I observe. 2% of purhcases use a

lender and agent pair that are ever merged, even if they were not merged at the time of the

purchase and origination.

Looking at the ownership structure at the time of origination, 32% of buyers use a real

estate agency which has a sibling lending arm, while 5% use a lender who has a real estate

agency sibling company. Again, 1.4% of borrowers use a merged pair. Similar concerns about

the fact that these number are not conditional on licensing apply. Nearly 40% of loans in

my sample are classiﬁed as coming from an agent preferred lender, a deﬁnition I will explain

This change is due to a 2015 change in the reporting requirements which required reporting the additional

variable beginning in 2018.

in Section 6.2. The average interest rate is 4.08%, reﬂecting the fact that interest rates were

generally low in the post-crisis period I study. The average FICO score in my sample is 736,

with an average loan-to-value ratio (LTV) of 87%.

Table 1: Summary Statistics

Mean Std. Dev Min Max

Agent Ever Merged 0.5 0.5 0 1

Lender Ever Merged 0.05 0.21 0 1

Agent/Lender Ever Merged 0.02 0.12 0 1

Agent Merged 0.44 0.5 0 1

Agent/Lender Merged 0.01 0.12 0 1

Lender Merged 0.05 0.21 0 1

Agent Preferred Lender 0.37 0.48 0 1

Interest Rate 4.08 0.45 3.25 5.00

FICO Score 735 55 300 900

LTV Ratio 87 14 1 200

Time to Close 41 16 7 88

Notes: Real estate Oﬃce/Lender Ever Merged equals one if a buyer’s real estate agent and mortgage

lender are ever merged, regardless of if they were at the time of the buyer’s purchase and origination.

Real estate oﬃce preferred lender equals one if the lender meets the criteria I deﬁne for likely being a

lender recommended by the buyer’s real estate agent.

4.1 Breakdown of Mergers

Here, I present a breakdown of the mergers I observe in my data set involving real estate

agencies. I observe a total of 137 mergers involving at least one real estate agency. Of those,

87, or the majority occur between two agencies where at least one agency has a lending arm.

These are the mergers which result in what I call “indirect integration.” As a result of the

merger that was consolidating agency business, the agency which previously did not have

a sibling lender now does. Another 48 of the mergers are between two real estate agencies

when neither one has a lending arm at the time of the merger. Thus, these do not cause

indirect integration. Finally, I observe 2 cases where a lender and an agency merge directly.

Comparing the two types of agency-agency merger, the lender-less agency involved in merg-

ers which include a lending are are approximately the same size as the agencies involved in

the mergers without a lending arm. On average the lender-less agencies have a 3-4% market

share.

Table 2: Merger Breakdown Statistics

N Avg. Mkt Share Std. Dev of Mkt. Share

Agency-Agency with Lender 87 0.04 0.06

Agency-Agency No Lender 48 0.03 0.03

Agency-Lender 2 0.001 0.0001

Notes: Agency-Agency with Lender mergers are those where one of the two real estate agencies merging

has a sibling lender. Agency-Lender mergers are those where a lender and agency merge directly (ie:

there is no second agency engaged in a horizontal merger). The average market share represents the

average market share of the acquired ﬁrm.

5 Case Studies

I will now present a case study to examine one merger more closely and see if there are ad-

ditional implications beyond those I can analyze in the reduced form and structural analysis.

In July 20XX, the real estate services subsidiary of a large conglomerate, BigFirm acquired

AgencyA.

This merger gave BigFirm signiﬁcant market share in portions of the United

States. At the time of the merger, LenderA, a mortgage company was a subsidiary of

AgencyA, and was acquired in this merger by BigFirm. Popular press and the company an-

nouncements of the merger suggest that the main reason for the merger was BigFirm gaining

market share in the residential real estate market. Indeed, if the acquisition of LenderA is

mentioned at all, it is later in the article, often seeming like an afterthought.

Following the merger, BigFirm and it’s subsidiary real estate agencies are jointly owned

with LenderA. Prior to the merger, BigFirm owned LenderB, another mortgage lender fol-

Identifying details have been removed.

lowing a merger with AgencyB in August 2013.

For the purposes of this case study, I will

ignore LenderB.

I begin by plotting the share of buyers who use LenderA to originate their mortgage, split

by AgencyA, all other BigFirm buyers, and all other buyers in Figure 2. As can be seen

here, prior to the merger, LenderA originated very few loans for purchases made with any

agencies other than AgencyA, consistent with the idea that AgencyA buyers were already

directed to LenderA as they were already jointly owned. Following the merger, we see no

change in the number of buyers from agencies that are not owned by BigFirm, while the

BigFirm share increases signiﬁcantly. This graph also does not show the geographic expan-

sion of LenderA following the merger. I have restricted this graph to only states in which

LenderA is licensed before the merger, but following it’s acquisition by BigFirm LenderA

obtains licenses in many new states.

In Figure 3a, I map the market shares of BigFirm’s real estate agencies in 20XX, the

year it acquires AgencyA and LenderA. BigFirm had real estate agencies in most states in

this year, with signiﬁcant market shares in the Mid-Atlantic states as well as Minnesota,

Missouri, and Kansas.

For comparison, Figure 3b shows the market share of LenderA in each state in 20XX, the

year of the merger. LenderA has non-zero market share in only ﬁve states: Texas, Pennsyl-

vania, Virginia, Maryland, and Delaware, reﬂecting the fact that lenders must be licensed in

BigFirm has several subsidiary real estate agencies. For the purposes of this case study, I will use

BigFirm to refer to all residential real estate agencies which were a subsidiary of BigFirm real estate services

subsidiary in a given year.

This is not a major issue; LenderB was a very geographically concentrated lender, with 80% of loans

coming from the Philadelphia Metro area between 2015 and 2017 according to the Consumer Financial

Protection Bureau.

. Furthermore, LenderB stopped originating new loans in late 2020, and the patterns

in my data suggest that they were winding down their origination business before this. Lastly, Lender2 has

been ﬁned for discriminatory lending practices, making it a poor choice for a case study.

Notes: Figure shows LenderA loan share at BigFirm agencies, AgencyA agencies, and all other agencies.

Years are relative to merger year, denoted at 0 and with red line.

Figure 2: LenderA Loan Share by Real Estate Agency Owner

each state and LenderA was not licensed in most states in the year of the merger.

However, in Figure 3c, I report the market shares of LenderA in each state two years after

the merger. Now, LenderA is lending in considerably more states, and has gained market

share in some states that is sizable given the relatively small market share of most lenders.

These ﬁgures suggest that LenderA expanded into the states where BigFirm had a substan-

tial presence already. Thus, bringing a lender in-house with a real estate agency inﬂuences

the geographic expansion of the lender.

Notes: Figure shows BigFirm’s market share in each state immediately prior to acquiring AgencyA and

LenderA. Darker states have higher market shares for BigFirm.

Figure 3a: BigFirm Agency Market Shares in Merger-Year by State

Notes: Figure shows LenderA’s market share in each state immediately prior to being acquired by

BigFirm. Darker states have higher market shares for LenderA.

Figure 3b: LenderA Market Shares in Merger-Year by State

Figure 3c: LenderA Market Shares Two Years Post-Merger by State

Notes: Figure shows LenderA’s market share in each state two years after acquisition by BigFirm. Darker

states have higher market shares for LenderA.

6 Reduced Form Speciﬁcation

6.1 Identiﬁcation

In this section I will discuss how I intend to estimate the causal eﬀects of mergers between

mortgage lenders and real estate agencies. I will begin with a discussion of the identiﬁcation

strategy before moving into estimating the equations. I present the results in the next section.

The ideal experiment to estimate the eﬀects of mergers would to randomly assign agen-

cies and lenders to be jointly owned for a period of time and then randomly reallocate them.

This is not a feasible research design. Thus, I will exploit natural variations in the ownership

structure of ﬁrms stemming from mergers and acquisitions.

As mergers occur at diﬀerent points and not all ﬁrms merge, this creates a treatment group

of lenders that are ever merged with real estate agents to compare with lenders who never

merge, as well as the buyers who use a given agent and lender pair. To ﬁx ideas, when

Berkshire Hathaway (a large real estate agency that acquired a real estate ﬁrm with a

lender, Prosperity) merges with another real estate agency, is Prosperity more likely to ﬁnd

borrowers through this link with the new agency? How does this aﬀect the characteristics

of the buyers coming to Prosperity from this agency as well as the loans these buyers obtain?

The identifying assumption is that the merger is motivated by Berkshire’s desire to ex-

pand its real estate agency business, and not because Berkshire was interested in acquiring

Prosperity speciﬁcally. In this case, the merger is exogenous as far as the lender is con-

cerned, simply changing how it acquires customers and where it expands, as is seen in the

case study. To further control for lender or agent-speciﬁc eﬀects, I include ﬁxed eﬀects in all

my speciﬁcations.

The main threat to causality is that buyers who use merged agency-lender pairs are dis-

tinct from others, even after controlling for observables, and that it is not joint ownership

leading to the eﬀects I ﬁnd but rather these unobservable characteristics. To ﬁx this, I pro-

pose the following solution. First, I assume that buyers select a real estate agent and then

a mortgage lender. Given that all real estate agents oﬀer buyers a list of suggested lenders,

and anecdotal evidence indicates that only lenders on that list will return buyer calls, this

is a reasonable assumption. Even if this direction does not hold, it does not invalidate the

eﬀects I ﬁnd, merely changes the interpretation. Now, instead of the agent referring buyers

to a lender, it is the lender referring buyers to an agent. Either way, the channel which

allows this is joint ownership.

Then, assuming that buyers ﬁrst chose an agent, I assume that they pick this agent based

on characteristics of the agency/agent, not due to a merger set up. Agents provide a list

of recommended services to buyers, including recommended mortgage lenders, regardless

of merger status. Furthermore, many merged ﬁrms do not share common names, making

it unlikely that buyers would be aware of the relationship.

Thus, buyers are selecting

agents on criteria other than sibling lenders, and the change in status would not inﬂuence

the type of buyers choosing to use a given real estate agency. I will further validate this

assumption by showing that borrowers do not change on most observables following a merger.

The second threat I face is misattributing the mechanism. Buyers who use a merged lender-

agent pair are implicitly not searching while buyers who go to a merged ﬁrm and do not

use the sibling lender may be shopping around. The prior literature demonstrates that price

dispersion is prevalent in consumer ﬁnancial markets, and that search mitigates this. Thus,

The set-up of my model does require that buyers choose an agent ﬁrst, but the reduced form results do

not.

Merged ﬁrms will often advertise for each other on their websites. However, they refer to each other as

”partner” or ”preferred,” the text stating the nature of the relationship is ﬁne print at the bottom, which is

likely to be ignored by most buyers.

failing to control for search behavior could lead me to mistakenly attribute a price eﬀect to

a merger when in fact it is a result of lack of search, which is not the goal of this paper. To

remedy this, I construct a proxy for lack of search, which I will discuss below.

6.2 Proxy for Search (or Lack Thereof)

No matter which real estate agency buyers go to, they will receive a list of recommended

lenders. I do not have access to this list, but I can proxy for it based on outcomes. Since

agent recommendations are something buyers pay attention to, the lenders agents suggest

will comprise a larger share of the loans matched with purchases from those real estate

agencies than agencies where they are not on the list of recommended lenders. Thus, I

construct the following measure to determine if a lender is likely to be recommended by an

agent:

1. Within a real estate agency-year, calculate what share of loans each lender originates,

conditional on originating at least one purchase loan. That is, lenders who do not

originate any loans at an agency are not considered.

2. Withing a CBSA-year, calculate lender market shares, conditional on a lender having

a non-zero market share in that CBSA-year.

3. Deﬁne a lender as recommended by an agency if the lender’s loan share within the

agency is more than two standard deviations above the lender’s loan share within the

county, and the lender originated at least eight loans in the CBSA-year.

In other words, I deﬁne a lender as recommended by an agency if I can reject the null hy-

pothesis that the lender’s share of loans within the CBSA-year and agency-CBSA-year are

the same at the 95% conﬁdence level. Using this criterion, 39% of loans in my sample are

from agency recommended lenders, and 92% of loans which come from jointly owned pairs

Eight represents the bottom decile of lender-CBSA-year loan counts in my data. I have experimented

with adjusting this threshold and the results are robust.

meet this criteria.

With this set up, I will look at four outcomes: lender market shares at the CBSA level

and within a agency, total number of loans originated in a county, borrower characteristics,

and prices (interest rates). To estimate all of these outcomes, I will use one speciﬁcation for

the CBSA market shares, and another for everything else.

7 Results

7.1 Lender Market Shares: Within Agency

The ﬁrst outcome of interest is lender market shares within a real estate agency. If mergers

lead agencies to direct clients to their sibling company in a way that is not true without

mergers, and this recommendation is inﬂuential, then we expect that the within-agency

share for the merged lender to increase post merger. This leads to the speciﬁcation:

ijkt

= β

+ β

T reat ∗ P ost

ijt

+ β

T reat

+ γX

ijkt

+ ϵ

ijkt

(1)

Here, Share

ijkt

is the share of purchases at agency i made with loans from lender j in CBSA

k at time t. T reat ∗ P ost

ijt

is equal to one if agency i and lender j are merged at time

t. T reat

takes the value one if agency i and lender j are ever merged. X

is a vector of

controls depending on the exact column in the table..

The results for this speciﬁcation can be found in Table 3. Before controlling for lender

and agent ﬁxed eﬀects, the T reat coeﬃcient is negative and highly signiﬁcant, but after

controlling for lender and agent ﬁxed eﬀects in column (2), the magnitude drops and is sta-

tistically insigniﬁcant. This suggests that the ﬁrms which merge have lower within agency

market shares merge, however, following the merger that these relationships are stronger.

This suggests that there is some strategy in which ﬁrms decide to merge; namely, ﬁrms

which have less of a relationship strengthen that relationship by bringing it ”in house.” The

coeﬃcient on T reat ∗ P ost is 0.15 in column (2), which indicates that following a merger,

the share of loans originated by the sibling lender of a real estate agency increases by 15

percentage points above the already existing relationship.

Merged lenders obtaining a larger market share from their sibling real estate agency post-

merger is consistent with the idea that mergers change the referral pattern of real estate

agencies. Now, agencies are referring more clients to their sibling lender, which results in

additional originations from the real estate agency for the lender.

Recent literature has shown that staggered diﬀerence in diﬀerence with two-way ﬁxed ef-

fects are subject to unequal weighting of treatment cohorts and traditional two way ﬁxed

eﬀects models do not recover the average treatment eﬀect. Thus, I employ the method

suggested by Sun and Abraham (2021) as well. This method reports a coeﬃcient for each

relative time dummy included in the regression, which I have aggregated up to an average

treatment eﬀect. In the last row of the table, I report the average coeﬃcient on the event

study coeﬃcients. This result is robust to the method suggested by Sun and Abraham

(2021); in fact the results are stronger. Now the coeﬃcient on T reat ∗ P ost is 0.28, or 28

percentage points. This is consistent with the event study plot found in Figure 4.

While the eﬀect on within-agency shares is interesting in its own right, it also provides

the ﬁrst stage result for later loan-level analysis. It demonstrates that mergers change the

referral pattern of agents in a way that aﬀects the lenders buyers choose. First, as agents

now have an ”inside track” to the lender, they may use this to help borrowers who otherwise

struggle to get a loan; however, they may instead use this to ”cream skim” and send only

the best borrowers to their sibling lender. Second, the lender may price loans to this group

of home buyers higher as they have a form of market power over the home buyer stemming

from the lack of search and the merger.

I will investigate the consequences of mergers on the characteristics of borrowers and the

interest rate in subsequent sections.

Table 3: Within-Agency Lender Shares

(1) (2)

Treat*Post 0.15*** 0.15***

(0.014) (0.012)

Treat -0.10*** -0.0094

(0.011) (0.0095)

Sun & Abraham (2020) 0.29*** 0.28***

(0.04) (0.04)

FE CBSA, Year CBSA, Year,

Lender,Agency

R-squared 0.11 0.61

N 3,176,993 3,176,993

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is a lender’s loan share within a CBSA-year-real estate agency. Unit of

observation is a lender-agency-CBSA-year. Treat*Post represents the coeﬃcient of interest and equals

one if the lender and agency are a merged pair. Treat equals one for all lender-agency pairs which merge

at any point in time. Standard errors are clustered at the agency level.

7.2 Lender Market Shares: CBSA Level

In addition to market share eﬀects at the company level, mergers may have eﬀects on over-

all market shares for lenders. If, for example, mergers make companies more eﬃcient, this

may allow them to process more loans in the same amount of time and thus attract more

customers. Then, their market share overall may increase.

To test this possibility, I use the below speciﬁcation:

jkt

= β

+ β

T reat ∗ P ost

+ β

T reat

+ γX

jkt

+ ϵ

jkt

(2)

Notes: Dependent variable is a lender’s loan share within a CBSA-year-real estate agency. Unit of

observation is a lender-agency-CBSA-year. Treat*Post coeﬃcients are plotted. Treat*Post represents the

coeﬃcient of interest and equals one if the lender and agency are a merged pair. Standard errors are

clustered at the agency level.

Figure 4: Lender Agency Market Shares Event Study Plot

Now, the left-hand side is the lender market share in a given CBSA-year, while T reat ∗

P ost

and T reat

turn on if the lender is merged with any real estate agency at time t or

at any point in time, respectively. This way, I am comparing lender market shares when

a lender is always merged to lenders who merge at some point in my sample. The results

for this can be found in Table 4. Here, after controlling for CBSA trends, there are no

signiﬁcant diﬀerences in lender market shares following a merger, until I add in lender ﬁxed

eﬀects. With the addition of lender ﬁxed eﬀects, merging results in a market share increase

of 0.54 percentage points. While the absolute value of this number is quite small, the aver-

age lender market share in my sample is 3.3% overall, meaning that a 0.54 percentage point

increase in market share represents a 16% increase in market share.

The fact that this is

only signiﬁcant after the introduction of lender ﬁxed eﬀects represents the fact that lenders

expand following mergers, as seen in the case study.

The event study plot version can be found in Figure 5. The results here are considerably

noisier. This is in large part due to the relatively sparse data, especially after ﬁxed eﬀects

The market share for lenders who are ever jointly owned is 3.8% on average, so 0.54 percentage points

represents a 14% increase in market share.

are incorporated. However, the general pattern of the results match the results in the tables,

and there are not obvious pre-trends. My results are even stronger after correcting for the

biases of the two way ﬁxed eﬀects estimator, the eﬀect on lender market share increases to

0.7 percentage points (21% of average lender market share and 18% of the average market

share for lenders who are ever merged with a residential real estate agency.)

This increase in market share is signiﬁcant for the lender, but has relatively little impact

on the overall market structure. While the jointly owned lender gains signiﬁcant market

share with respect to where it started, this is not making them dominant in the market.

The broader market implications of this increase in market share are unclear. On the one

hand, this could represent loans that would otherwise be provided by other lenders, and thus

be business stealing. Alternatively, if merging makes the lender more eﬃcient so that it is

proﬁtable to originate more loans, they could gain market share without taking clients from

other lenders. In this case, the overall access to credit in the market will increase.

Table 4: CBSA Lender Market Shares

(1) (2)

Treat*Post -0.0012 0.0054*

(0.0021) (0.0029)

Treat 0.0051***

(0.00072)

Sun & Abraham (2020) -0.001 0.007*

(0.003) (0.004)

FE CBSA,Year CBSA,Year,Lender

R-squared 0.63 0.64

N 415,764 415,764

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is a lender’s market share within a CBSA-year. Unit of observation is a lender-

CBSA-year. Treat*Post represents the coeﬃcient of interest and equals one if the lender is merged with

any agency at that point in time. Treat equals one for all lenders which are merged with an agency at

any point in time. Standard errors are clustered at the lender level.

Notes: Dependent variable is a lender’s market share within a CBSA-year. Unit of observation is a

lender-CBSA-year. Treat*Post coeﬃcients are plotted. Treat*Post represents the coeﬃcient of interest and

equals one if the lender is merged with any agency at that point in time. Standard errors are clustered at

the lender level.

Figure 5: Lender CBSA Market Shares Event Study Plot

7.3 Interest Rate

For mortgages, interest rates are analogous to prices, and thus could be manipulated by

lenders if joint ownership with a real estate agency confers market power on them that does

not exist otherwise. Studying interest rates is additionally complicated because prior work

by Bhutta et al. (2020) documents that borrowers who apply to more than one lender pay

7 basis points less on average, and that “seriously considering” 3 or more lenders reduces

rates by 9.5 basis points on average. I want to ensure that the eﬀect I ﬁnd is not caused

by this lack of search, but rather the ownership structure of the ﬁrms selected. Thus, I

restrict my sample to only loans I ﬂag as agent recommended as discussed earlier in Section

6.2. Furthermore, in my main analysis, I restrict to loans coming from the lender’s retail

channel, as theory suggests those are the buyers aﬀected by joint ownership. Full sample

results are available in the appendix.

The results for this variable can be found in Table 5. All columns include lender, agency,

CBSA, state, and year ﬁxed eﬀects, and I report the T reat ∗ P ost coeﬃcient from the Sun

and Abraham (2021) methodology at the bottom of each column. The results are robust

to this, and the event study plot can be found in Figure 6. the coeﬃcient on T reat ∗ P ost

is positive and highly signiﬁcant. This means that buyers who use a merged lender-agency

pair and go directly to the lender pay 8 basis points more on average for their loan. This

is consistent with the idea that buyers going to the retail lender who don’t search are most

likely to be aﬀected by the merger. When including FICO score and LTV in column (2),

this increases to 9 basis points.

Table 5: Interest Rate

(1) (2)

Treat*Post 0.079*** 0.092***

(0.019) (0.015)

Treat -0.054*** -0.070***

(0.019) (0.016)

FICO*LTV 0.00042***

(7.5e-07)

FICO Score -0.005***

(0.0001)

LTV -0.030***

(-0.001)

Sun&Abraham(2020) 0.122** 0.138***

(0.036) (0.036)

FE CBSA,Year CBSA,Year

Lender,Agency Lender,Agency

R-squared 0.43 0.45

N 813,284 813,284

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the loan interest rate. Unit of observation is a loan matched to a home

purchase. Treat*Post represents the coeﬃcient of interest and equals one if the lender and agent are a

merged pair. Treat equals one for any lender-agent pairs which ever merge. Standard errors are clustered

at the agency level.

An interest rate eﬀect of 8-9 basis points is quite large in the context of the literature. This

is similar to the eﬀect found by Bartlett et al. (2021) for the premium paid by black borrowers

over white borrowers. It is also approximately 15% of the interest rate dispersion found by

Bhutta et al. (2020) between the 10th and 90th percentile of rates. While their diﬀerence is

signiﬁcantly larger, they further document that search reduces this dispersion substantially.

Notes: Dependent variable is the loan interest rate. Unit of observation is a loan matched to a home

purchase. Treat*Post coeﬃcients are plotted. Treat*Post represents the coeﬃcient of interest and equals

one if the lender and agent are a merged pair. Standard errors are clustered at the agency level.

Figure 6: Interest Rate Event Study Plot

In the case of my ﬁnding, I am analyzing a subset of consumers who I believe did not search.

Thus, the eﬀect I ﬁnd is not due to changes in shopping behavior of consumers, but solely

the legal structure of the company they happen to choose. As most real estate agents will

oﬀer lender recommendations to consumers, this is not a the fee for the convenience of not

having to ﬁnd a lender, any eﬀect from that is already included.

This eﬀect is due purely

to the organizational structure of the ﬁrms in question. Furthermore, for aﬀected buyers,

this is not a small eﬀect. On the average loan in my sample, it represents an extra 225 per

year in interest payments. In high cost of living areas, such as the Washington, DC metro,

where loan amounts are higher, this rises to 289 per year. As home prices continue to rise,

these numbers will grow.

7.4 Time to Close

One feature borrowers might be willing to pay for in the form of higher rates is a loan which

they believe will close more quickly. In particular, in hot housing markets, the ability to

receive ﬁnancing quickly can be the diﬀerence between having an oﬀer accepted and not.

Merged ﬁrms might be able to originate loans more quickly due to the easier communication

The way in which agents recommend lenders to buyers is interesting in its own right, but I leave that

analysis for a future paper.

between lenders and agents.

I cannot directly observe the time it takes for a borrower to receive ﬁnancing; however,

I am able to proxy for it. I observe both the date a property was under contract, that is on

the day that the home had an oﬀer accepted and was no longer up for sale, and the close

date, the date on which the paperwork for the sale was signed and the home formally trans-

fers from one owner to the next. I take the diﬀerence between these two values to obtain the

length of time it takes for the property to move from under contract to closed. The results

of this analysis can be found in Table 6. In column (1), I use time to close directly. In

column (2), I use the probability that a purchase takes more than 45 days to close, 60 days

in (3), and 75 days in column (4). In column, I use only loans ﬂagged as agent referred and

obtained through the lender’s retail channel, which is the sample that accounts for search

intensity and which is most likely to be aﬀected by a merger. I trim time to close at the 5

and 95 percentiles before calculating each variable.

Beginning with column (1), using a merged pair reduces the time to close by 1.4 days or 3%

of the mean. However, buyers may be less concerned with small changes in the level of the

time to close but far more concerned with avoiding a lengthy delay. Thus, I in columns (2)

through (4) I put a dummy variable equal to one if the time to close is more than X days

on the left hand side, making this a linear probability model. There is a 2.3% reduction

in the probability that it takes more than 45 for a purchase to close when using a merged

lender-agent pair, but this is not statistically signiﬁcant. Moving to columns (3) the eﬀect

for buyers going through the retail channel is statistically signiﬁcant at 2.1% reduction in

the probability it takes more than 60 days to close. In column (4), there is no signiﬁcant

change in the probability it takes more than 75 days to close. This is due in part because

less than 10% of transactions take more than 75 days to close.

Two days on a mean of 41 days is not a large eﬀect, nor are any of the probabilistic re-

sults large. Thus, if there is any eﬃciency gain for buyers from choosing a lender merged

with their real estate agency, it is not through faster closing times.

Table 6: Time to Close

(1) (2) (3) (4)

Days 45+ Days 60+ Days 75+ Days

Mean 42.0 0.35 0.16 0.09

Treat*Post -1.4** -0.025 -0.023 -0.021**

(0.59) (0.020) (0.019) (0.0088)

Treat 0.97 0.034 0.020 -0.0037

(0.68) (0.021) (0.022) (0.012)

FE CBSA,Year CBSA,Year CBSA,Year CBSA,Year

Lender,Agency Lender,Agency Lender,Agency Lender,Agency

R-squared 0.31 0.23 0.24 0.22

N 793,231 793,231 793,231 793,231

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the number of days between a home under contract and the home closing

(column (1)), or indicator for if it took more than 45 days (column (2)), 60 days (column (3)), or 75 days

(column (4)) to close. Unit of observation is a loan matched to a home purchase where loan is originated

through the retail channel. Treat*Post represents the coeﬃcient of interest and equals one if the lender

and agency are a merged pair. Treat equals one for any lender-agency pairs which ever merge. Standard

errors are clustered at the agency level. ALl columns include CBSA, Year, Lender, and Agency ﬁxed

eﬀects.

7.5 Borrower Characteristics

Real estate agents and the agencies they work for have a considerable amount of informa-

tion about their clients and their ability to repay a loan. Thus, they can act as conduits

of information between lenders and borrowers. Following the merger, this channel is legally

codiﬁed, making the exchange even easier. There are to two possible uses for this channel:

assisting marginal borrowers or cream skimming. In the ﬁrst case, suppose the agency has

soft information about applicants which lead it to believe that they are a ”better” risk than

their credit information may suggest. They can communicate this to the lender and help

them get a loan. In the second case, the agency could selectively decide to refer only the

best clients to its sibling lender.

While both of the above cases could be true for agencies and lenders who have a rela-

tionship that is not legally codiﬁed, its existence is more plausible following a merger. Now

the agency and lender are working for the same parent company, and as such their incentives

are aligned. Without the merger, the agency is primarily interested in getting a loan for its

clients, not necessarily how well they will be able to repay, and likewise the lender is not

interested in how easy it is for the agency to make a sale, so is unlikely to want to “go the

extra mile” to originate a loan. With the merger, both have reason to care about the others’

goals, making one or both channels possible. I will now discuss results for three ex-ante

characteristics of borrowers: loan amount, FICO score at origination, loan-to-value ratio. In

addition, I have results for loan performance, to check if there are some unobservables which

are inﬂuencing interest rates as found by Stroebel (2016). For these, I test the share of loans

which are ultimately 30, 60, and 90 days delinquent.

Borrower characteristics are the second set of results that may be correlated with search

behavior. That is, borrower characteristics may be correlated with the propensity to shop

around. As such, I report the speciﬁcations keeping only loans which I ﬂag as agent referred

and from the retail channel as discussed in the beginning of the empirical speciﬁcation dis-

cussion. The full sample results are available in the appendix.

These results can be found in Table 7. All six columns are insigniﬁcant, suggesting that

neither the ex-ante borrower characteristics, nor ex-post loan performance are diﬀerent be-

fore and after merger. Thus, I ﬁnd no evidence of cream-skimming or soft information

helping marginal borrowers. If there were cream-skimming, I would expect an improvement

in the borrower characteristics after merger. In reverse, if there were soft information helping

marginal borrowers get loans who would otherwise be denied, I would expect the quality of

borrower to decline following the merger.

Looking at Table 7, neither story appears to hold. The coeﬃcient on loan amount is small

in magnitude and statistically insigniﬁcant. The coeﬃcient on LTV is similarly small and

insigniﬁcant, at 2.2, or just under a 3% change. Similarly, the average FICO score in my

data set is 735, but the coeﬃcient on T reat ∗ P ost is 10 points. While this is marginally

signiﬁcant, it does not represent a large change in the credit score of borrowers.

Moving to the loan performance results in columns (4) through (6), I again ﬁnd no sig-

niﬁcant diﬀerence coming from merger. This suggests that the performance of loans does

not change with merger status. This is further evidence that the merger does not lead to soft

information changing loans. It also helps put the eventual interest rate results in context;

the higher rates are not due to information leading to expected worse loan performance (at

least not that is realized).

Table 7: Borrower Characteristics

(1) (2) (3) (4) (5) (6)

Amount LTV FICO 30 Days 60 Days 90 Days

Mean 254,484 87.00 735 14.2% 7.8% 5.8%

Treat*Post 402 -2.2 10* -0.012 -0.0048 0.0013

(7,118) (1.5) (5.5) (0.011) (0.0075) (0.0051)

Treat -14,694* 2.5* -11* 0.023* 0.0065 -0.0044

(8,318) (1.5) (5.7) (0.013) (0.0095) (0.0070)

FICO*LTV -0.000025*** -0.000021*** -0.000019***

(6.3e-07) (4.9e-07) (4.3e-07)

FICO Score 0.00054*** 0.00081*** 0.00080***

(0.000056) (0.000043) (0.000037)

LTV 0.020*** 0.017*** 0.015***

(0.00049) (0.00038) (0.00033)

FE CBSA,Year CBSA,Year CBSA,Year CBSA,Year CBSA,Year CBSA,Year

Lender,Agency Lender,Agency Lender,Agency Lender,Agency Lender,Agency Lender,Agency

R-squared 0.54 0.21 0.19 0.17 0.15 0.13

N 606,238 830,385 824,923 824,821 824,821 824,821

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is loan amount (column (1)), loan-to-value ratio (column (2)), FICO (credit)

score at origination (column (3)), or indicator for if loan is ever 30 days delinquent (column (4)), 60 days

delinquent (column (5)), or 90 days delinquent (column (6)). Unit of observation is a loan matched to a

home purchase where loan is originated through the retail channel. Treat*Post represents the coeﬃcient

of interest and equals one if the lender and agency are a merged pair. Treat equals one for any lender-

agency pairs which ever merge. Standard errors are clustered at the agency level.

8 Model

While the reduced form results above discuss the implications of mergers between agencies

and lenders for consumers, mortgage market structure, and lender market shares, they are

unable to examine any potential cost savings for the lender, as well as unable to test policy

counterfactual scenarios. Thus, I have constructed a logit demand model for mortgages as

well as Bertrand diﬀerentiated product competition supply so as to recover marginal cost

and estimate counterfactuals.

8.1 Household

Each household i in market t has a choice set C

from which they select one loan, j. For a

given product j, the indirect utility household i receives can be written as:

ijt

= −αr

+ βx

+ ξ

+ ϵ

ijt

(3)

where r

represents the rate paid by the household, and x

is a vector of product charac-

teristics observed by the econometrician. ξ

is the characteristics of the product which are

observable to the household, but unobservable to the econometrician. ϵ

ijt

is the error term.

Then, household i chooses the product in their choice set C

which maximizes their indirect

utility.

8.2 Lender

A given lender l oﬀers products j. I assume that if a lender is jointly owned, it oﬀers two

product in a given market: one to buyers from its sibling agency and another to all other

buyers in the market.

For a given product j which is sold in market t, the lender has proﬁt:

jlt

= (r

jlt

− κ

jlt

(4)

where r

jlt

is the interest rate, and κ

jlt

is the marginal cost of selling that product. Thus,

each product the lender oﬀers is allowed to have a diﬀerent rate and a diﬀerent cost in every

market. s

jlt

represents the choice probability for product j. Thus, the overall proﬁt for ﬁrm

l in market t is:

j∈J

jlt

) (5)

where s

jlt

is the market share for product j. Then, the lender chooses the interest rate r

jlt

for each product to maximize their proﬁt function. The ﬁrst order condition for the optimal

rate on good j from lender l in market t, r

∗

jlt

is:

∗

klt

= κ

klt

−

klt

∂s

klt

∂r

klt

−

j̸=k∈J

jlt

− κ

jlt

∂s

jlt

∂r

klt

∂s

klt

∂r

klt

(6)

9 Model Estimation

9.1 Household Demand

Each household i belongs to a market t. I deﬁne a market as a CBSA-year-above/below me-

dian credit score-above/below median loan-to-value ratio-loan type (Non-conforming, con-

forming, or jumbo). In standard logit demand, households in the same market have the

same choice set. However, in this case, not all households in a given market have access to

the same products. Household-speciﬁc choice sets mean that the standard logit-estimation

strategy using market shares will not work, and instead I will exploit the micro data.

I do not observe the options in a household’s choice set, so I will construct it. First, I

separate all loans into above and below median credit score, and then above and below

median loan-to-value-ratio. Then, I deﬁne a product as a combination of lender, and if the

loan went to buyers from vertically integrated broker-lender pairs. That is, even with the

same credit score, loan-to-value ratio, and lender, I consider a loan a diﬀerent product if the

household i came from the lender’s sibling residential real estate agency or not.

Now, I assume that buyers in market t have access to loans issued in market t that required

the same credit score bin (as deﬁned by above/below median), loan type, and loan-to-value

ratios (above/below median) from all lenders operating in market t. However, if the lender is

jointly owned with the buyer’s real estate agency, they will only have access to that lender’s

vertically integrated product for a given credit score x loan to value ratio, and vice versa.

I construct the outside option, j = 0, by setting it to all lenders who have a 0.05% market

share or less, and normalize it to have mean utility of zero.

This speciﬁcation for the choice set does not take into account heterogeneity in search costs.

For example, large ﬁrms may advertise more aggressively, and gain larger market shares

as a result. To account for this, I will include lender ﬁxed eﬀects. Then, assuming that

conditional on observables borrower composition is not correlated with product attributes,

including the dummy for being a linked lender, not accounting for search costs is unlikely to

drive my results. In the reduced form I ﬁnd little eﬀect of mergers on borrower characteris-

tics, making this assumption plausible.

Then, recall that a household i’s indirect utility from product j in market t can be written

as :

ijt

= −αr

+ βx

+ ξ

+ ϵ

ijt

(7)

where r

is the average rate, and x

are the observable product characteristics: vertical

integration and lender ﬁxed eﬀects, and ξ

are unobservables. I assume that ϵ

ijt

is distributed

Extreme Value Type-I. Then, a consumer i chooses the product j from their choice set C

which maximizes their indirect utility. That is, if household i chooses product j, then:

1. Product j ∈ C

2. V

ijt

> V

ikt

∀k ∈ C

Among all products in its choice set, C

, the household will choose the product which

maximizes its indirect utility. That is, if household i choose product j from lender l in

market t, that implies that:

1. Product j is in the household’s choice set C

2. V

ijt

> V

ikt

∀k ∈ C

Then, the conditional choice probability of household i in market t choosing product j is:

ijt

= P r(j|C

) =

exp(δ

)

k∈C

exp(δ

ikt

)

(8)

Where δ

can be written as:

= αr

+ βX

+ ξ

(9)

In other words, δ

is the indirect utility with the ϵ

ijt

integrated out. δ

is normalized to zero.

Then, then household i’s likelihood function is:

j∈C

(product j chosen)

ijt

(10)

This gives the log-likelihood for household i as:

ln (L

) =

j∈C

ln (s

ijt

) (product j chosen) (11)

I estimate this in two steps. First, I use a non-linear optimization to ﬁnd the δ

which max-

imizes the likelihood for all households. Then, I use two stage least squares to decompose

into each of its pieces. This cannot be done with standard OLS because it is possible

cov(r

, ξ

) ̸= 0. For example, one element of ξ

could be quality. Then, it makes sense

that higher quality products would have higher prices, so cov(r

, ξ

) > 0. Thus, failing to

account for the endogeneity between r

and ξ

will lead to biased estimates.

To deal with the endogeneity of ξ

, the unobservables which may be correlated with price,

I use the 10-year Treasury bond rate interacted with the number of lenders in the market.

The intuition behind this IV is that mortgage rates are often tied to the treasury rate. How-

ever, the degree to which lenders can pass on changes in the federal funds rate to buyers is

dependent on the degree of competition in the market. With rate increases, in more com-

petitive markets lenders cannot pass as much of the treasury rate increase on to borrowers.

I have plotted the 10-year Treasury rate in each year in Figure 7. The rate varies over time,

lending time variation to my IV. Furthermore, the number of lenders in a market varies year

over year; only 11% of my markets have the same number of lenders from one year to the

next, providing additional time variation. There is also signiﬁcant variation in the number

of lenders in a market within a given year, providing geographic variation. A histogram with

the distribution of number of lenders can be found in Figure 8.

1.5 2 2.5 3

10-Year T-Bill

2010 2012 2014 2016 2018 2020

Year

10Yr Treasury Rate Over Time

Notes: Figure shows the annual average of the 10-year Treasury rate over sample period (2011-2019).

Figure 7: 10-Year Treasury Rate over Sample Period

0 .02 .04 .06 .08 .1

Density

0 10 20 30 40 50

Number of Lenders

Distribution of Number of Lenders

Notes: Figure is histogram of number of lenders in market year over sample period (2011-2019).

Figure 8: Histogram of Number of Lenders in Market-Year

The ﬁrst stage results can be found in the appendix in Table 11. The coeﬃcients on

the product characteristics can be seen in Table 10. The coeﬃcient on interest rate is -8.26.

This is consistent with the idea that borrowers prefer to pay less interest, all else equal. The

coeﬃcient on Merged Pair is positive. This is consistent with the reduced form results in

Table 3. When a lender is merged with the borrower’s real estate agent, the borrower is more

likely to choose that lender. This can be thought of as the additional ”brand premium” that

a lender receives when they are jointly owned with a real estate agent. However, the lender

ﬁxed eﬀects are considerably larger, on the order of 40. Thus, the premium for being jointly

owned is not large in comparison. This matches the reduced form results where the eﬀect

on CBSA market share is relatively small.

The coeﬃcient on time to close is negative. This suggests that borrowers prefer to close

quickly, which is again consistent with theory.

Table 8: Mean Utility Parameters

Rate −8.26

∗∗∗

(1.74)

Merged Pair 2.20

∗∗

(0.23)

Time to Close −0.020

∗∗∗

(0.004)

Observations 16,024

* 0.10 ** 0.05 *** 0.01

Notes: Mean utility parameters recovered from nested logit run on a 25% random sample of data. Unit

of observation is a given loan product j in a market t. Merged Pair is a dummy variable equal to 1 if

the residential real estate agency and lender are part of the same conglomerate. Time to close is the

number of days between contract and close. Standard errors bootstrapped on 500 random samples at

market level.

9.2 Supply

Using the ﬁrst order condition for optimal pricing, I can solve for the marginal cost of prod-

uct k in market t, κ

. Then, using the results from the demand estimation, I can compute

conditional choice probabilities s

jnlt

for all products, as well as the partial derivatives. Then,

in every market, I am left with a system of J

equations with J

unknowns, so the system is

just identiﬁed, and a unique set of marginal costs exists.

Summary statistics for marginal cost can be found in Table 9. The average marginal cost

on non-merged loans is 4.00%, while that on merged loans is 3.98%. I cannot reject that

these two are the same, but it is suggestive that merged loans are cheaper to originate than

non-merged loans. The magnitude of the marginal cost on these loans is qualitatively similar

to those found in the UK mortgage market by Robles-Garcia (2019).

Table 9: Marginal Cost

Marginal Cost Markup

Non-Merged 4.00 0.018

(0.03) (0.07)

Merged 3.98 0.02

(0.06) (0.011)

Notes: Average marginal cost and markups over marginal cost calculated from model. Unit of observation

is a product j in market t. Non-Merged products are those available to buyers not coming from the

sibling residential real estate agency of a lender while merged products are only available to buyers

coming from the lender’s sibling agency. Results calculated using a 25% random sample of markets.

Standard errors bootstrapped on 500 random samples at market level.

10 Counterfactual

Now that I have recovered the demand parameters and marginal cost for each type of loan,

I can run counterfactuals. I run a total of three counterfactuals. In the ﬁrst, I estimate a

counterfactual equivalent to breaking up the merged agencies and lenders, but not allowing

borrowers to choose new loan products other than the outside option. All merged products

are removed from the market so only the unmerged products are left; borrowers who previ-

ously chose merged products now choose the outside option. Keeping the utility parameters

from the demand estimation and the recovered marginal cost, I solve for the optimal interest

rate for each product using the lender’s pricing equation:

∗

klt

= κ

klt

−

klt

∂s

klt

∂r

klt

−

j̸=k∈J

jlt

− κ

jlt

∂s

jlt

∂r

klt

∂s

klt

∂r

klt

(12)

In the second counterfactual, I again assume that the government has banned joint own-

ership, but now I allow the aﬀected buyers to choose any product, instead of just the outside

option. Again, I hold the mean utility parameters and all product characteristics other than

rate ﬁxed and re-solve the lender’s problem. These ﬁrst two counterfactuals, the main change

is the composition of the consumer’s choice sets, but the size is fairly constant.

In the third counterfactual, I assume the government has allowed joint ownership, but

now the lender cannot oﬀer diﬀerent products to consumers based on which real estate agency

they chose, they must oﬀer both products to all consumers. Thus, this counterfactual can be

thought of as increasing the size of the choice set but all consumers keep the product they

chose in the status quo in their choice set.

The results of these counterfactuals can be seen in Table 10. In each column, I report the

weighted average of each product characteristics in that counterfactual as well as the utility

implied. These values are weighted by the product market shares in that simulation. I also

report the status quo for comparison.

Beginning with the ﬁrst counterfactual, banning mergers but sending all aﬀected consumers

to the outside option decreases rates by 1 basis point on average. However, because con-

sumers now choose products without the brand premium boost from joint ownership and

with slightly lower unobservable quality, overall utility decreases slightly.

Moving to the last counterfactual, banning mergers but allowing consumers to pick be-

tween all options in the market and the outside option results in higher interests rates on

average, increasing from 4.09% to 4.14%. However, buyers compensate for this by choosing

products which close slightly faster, with higher unobservable characteristics (ξ increases

from an average of 0.99 to 1.32), and from lenders which provide a higher brand premium.

However, these positive characteristics are not enough to oﬀset the disutility from a higher

There are a handful of markets where the choice set does change size because the jointly owned lender

does not oﬀer both a jointly owned and a non-jointly owned product in that market, but correcting for this

does not substantively change the results.

rate, so welfare decreases. The welfare decrease is equivalent to that from an interest rate

increase of 0.3 basis points, or welfare loss 9 at the mean loan amount.

In the second counterfactual, consumers now have access to two products from jointly owned

lenders: the one previously oﬀered to customers coming from the lender’s sibling agency and

the one oﬀered to all other customers. In this instance, rates do not change relative to the

status quo, but buyers substitute to diﬀerent products. Fewer consumers choose the jointly

owned product, 1% instead of 2%, but choose products with higher quality, faster closing,

and more preferred lenders. This means that consumer welfare increases by 18 on average.

Taken together these counterfactuals suggest that regulating the product oﬀerings of lenders

jointly owned with real estate agents would increase consumer welfare if it is done carefully.

Simply banning mergers would harm consumers, but requiring all consumers be oﬀered the

same products rather than banning joint ownership increases consumer welfare on average.

Table 10: Counterfactual Results

Status Quo Merger Ban, Merger Ban Competition

No Re-Sort

Interest Rate 4.09 4.08 4.14 4.11

(0.01) (0.01) (0.01)

Time to Close 39.99 39.95 39.79 39.64

(0.19) (0.23) (0.33)

Jointly Owned 0.02 0.00 0.00 0.01

(0.00) (0.00) (0.001)

ξ 0.99 0.96 1.32 1.21

(0.19) (0.16) (0.18)

Lender FE 44.65 44.68 44.73 44.69

(0.03) (0.05) (0.04)

Utility 11.09 11.05 11.06 11.15

(0.19) (0.21) (0.18)

Equivalent Rate Change(bp) 0.5 0.3 -0.7

(0.34) (0.29) (0.69)

Utility Change (Dollars) -12 -9 18

(23) (16) (35)

Notes: Results of counterfactual simulation on a 25% random sample of markets. Unit of observation is

a product j in a market t. Dollar utility calculated by computing the change in the interest rate which

corresponds to the same change in utility and then multiplying that by the average loan amount. Results

are weighted by the product market share in the counterfactual. Standard errors bootstrapped on 150

random samples at market level

11 Conclusion

In this paper, I construct a novel data set of home purchases matched to the loans which

funded them, and exploit the 100+ mergers I hand-matched to examine the consequences

of jointly owned real estate agencies and mortgage lenders which has been previously un-

studied in the literature. Of these more than 100 mergers, 87 occur between real estate

agencies and indirectly integrate a lender with one of the two agencies. I use this indi-

rect integration in a staggered diﬀerence-in-diﬀerences design to analyze the consequences of

joint ownership between real estate agencies and mortgage lenders for lenders and borrowers.

Beginning with the eﬀects on market share, I ﬁnd that joint ownership has a small im-

pacts on a lender’s market share in a CBSA, and larger eﬀects on their market share within

sibling residential real estate agencies. This suggests that when an agency and a lender

share a parent company, the agent directs clients to her sibling lender, and this increases the

lender’s market share. While this increase in market share represents a signiﬁcant increase

in market share for the lender, the eﬀect on overall market competition is small.

Third, after documenting that mergers change the lender chosen by buyers using a merged

real estate agency, I investigate if interest rates change as a result of using a jointly owned

agency-lender pair, and I ﬁnd that borrowers going directly to a lender pay 9 basis points

more in interest than borrowers with similar characteristics did before the merger, even when

the lender was recommended by the agency in both cases, so search intensity has not changed.

Fourth, I examine if mergers facilitate easier communication between agencies and lenders,

leading buyers close on their mortgage faster, which buyers may be willing to pay for. While

I ﬁnd a two day reduction in the time to close, this is small relative to the average time to

close. Furthermore, there is only slight evidence that joint ownership decreases the proba-

bility of an unusually long time to close. I therefore conclude that joint ownership does not

get home buyers into their homes faster.

Fifth, I examine how mergers change borrower characteristics. By strengthening the ties

between lenders and agencies, mergers could facilitate information exchanges which beneﬁt

marginally qualiﬁed borrowers or result in jointly owned lenders taking only the best bor-

rowers (“cream-skimming”). When looking at FICO score, LTV, and loan amount, I ﬁnd

no large signiﬁcant eﬀects. The fact that lenders do not change behavior with respect to

FICO and LTV scores is consistent with the fact that these are GSE-insured loans where

little of the riskiness of the borrower is borne by the lender. Similarly, I ﬁnd no evidence of

diﬀerential loan performance after origination.

Finally, I supplement my reduced form ﬁndings of higher market shares for lenders and

higher interest rates for buyers with a structural model of supply and demand for loans.

Using a logit demand model for mortgages, I estimate mean utility parameters for prod-

uct characteristics which conﬁrm my reduced form results that buyers prefer jointly owned

agency-lender pairs, faster closing time, and that there is a signiﬁcant ”known brand” pre-

mium. I couple this demand model with a supply model where lenders are proﬁt maximizing,

face exogenous marginal cost, and must choose an interest rate for each loan product. From

this supply model paired with my demand model results, I recover marginal costs. I fail to

reject that marginal cost is identical for merged and non-merged loans. Using this model, I

next estimate three policy counterfactuals for what could happen if joint ownership between

real estate agencies and mortgage lenders were further regulated.

In the ﬁrst counterfactual, I assume that joint ownership is banned and force all of the

buyers choosing jointly owned loan products to choose the outside option. In this case,

interest rates fall, but so does average utility. In the second counterfactual, I again assume

joint ownership is banned, but now consumers can re-sort among all products, not just con-

sumers who previously chose jointly owned products, and not just to the outside option.

In this case, the decrease in competition increases interest rates, but consumers are able to

choose products that have other desirable characteristics, so while utility falls, it falls by

less than in the ﬁrst counterfactual. In the third and ﬁnal counterfactual, I assume that

joint ownership is permitted, but that jointly owned lenders must oﬀer all their products to

consumers, regardless of real estate agency chosen. In this case, prices go up very slightly,

but consumers choose products with other characteristics, that overall utility and welfare go

up by 18 per person on average.

These results open several avenues of future research. First, there could be heterogeneous

eﬀects by borrower characteristics; perhaps ﬁrst-time, minority, and/or female borrowers

are impacted diﬀerently than other borrowers. Second, understanding how agencies recom-

mend lenders, independent of common ownership is important to understanding this market.

Third, local market conditions may make change the ability of conglomerate lenders to ad-

just their prices, and thus change the penalty impacted borrowers pay. Extensions to the

model would incorporate limited consideration sets for buyers, or more complex or dynamic

pricing decisions by lenders. I leave these analyses to future papers. Finally, the data set I

constructed for this project are useful for other projects which may want to link loans and

the associated home purchase, or which want to exploit variation in corporate ownership

structure of real estate ﬁrms over time.

References

Agarwal, Sumit, Gene Amromin, Itzhak Ben-David, and Douglas D Evanoﬀ,

“Loan product steering in mortgage markets,” Technical Report, National Bureau of Eco-

nomic Research 2016.

Akg¨un, U˘gur, Cristina Caﬀarra, Federico Etro, and Robert Stillman, “On the

welfare impact of mergers of complements: Raising rivals’ costs versus elimination of

double marginalization,” Economics Letters, 2020, 195, 109429.

Bartlett, Robert, Adair Morse, Richard Stanton, and Nancy Wallace, “Consumer-

lending discrimination in the FinTech era,” Journal of Financial Economics, 2021.

Barwick, Panle Jia, Parag A Pathak, and Maisy Wong, “Conﬂicts of interest and

steering in residential brokerage,” American Economic Journal: Applied Economics, 2017,

9 (3), 191–222.

Berger, Allen N, Rebecca S Demsetz, and Philip E Strahan, “The consolidation

of the ﬁnancial services industry: Causes, consequences, and implications for the future,”

Journal of Banking & Finance, 1999, 23 (2), 135–194.

Bhutta, Neil and Aurel Hizmo, “Do minorities pay more for mortgages?,” The Review

of Financial Studies, 2021, 34 (2), 763–789.

, Andreas Fuster, and Aurel Hizmo, “Paying too much? Price dispersion in the US

mortgage market,” 2020.

Buchak, Greg, Gregor Matvos, Tomasz Piskorski, and Amit Seru, “Fintech, Regu-

latory Arbitrage, and the Rise of Shadow Banks,” Journal of Financial Economics, 2018,

130 (3), 453–483.

Choi, Jay Pil, “Mergers With Bundling In Complementary Markets,” The Journal of

Industrial Economics, 2008, 56 (3), 553–577.

Ershov, Daniel, Jean-William Lalibert´e, and Scott Orr, “Mergers in a model with

complementarity,” 2018.

Grunewald, Andreas, Jonathan A Lanning, David C Low, and Tobias Salz, “Auto

Dealer Loan Intermediation: Consumer Behavior and Competitive Eﬀects,” Working Pa-

per 28136, National Bureau of Economic Research November 2020.

Guiso, Luigi, Andrea Pozzi, Anton Tsoy, Leonardo Gambacorta, and

Paolo Emilio Mistrulli, “The cost of steering in ﬁnancial markets: Evidence from the

mortgage market,” Journal of Financial Economics, 2021.

Lopez, Luis A, Shawn McCoy, and Vivek Sah, “Steering Consumers to Aﬃliated

Financial Services: Evidence from Mortgage Referrals and Cost of Credit,” Available at

SSRN 3365347, 2019.

Robles-Garcia, Claudia, “Competition and incentives in mortgage markets: The role of

brokers,” Unpublished working paper, 2019.

Stroebel, Johannes, “Asymmetric Information about Collateral Values,” The Journal of

Finance, 2016, 71 (3), 1071–1112.

Sun, Liyang and Sarah Abraham, “Estimating dynamic treatment eﬀects in event stud-

ies with heterogeneous treatment eﬀects,” Journal of Econometrics, 2021, 225 (2), 175–

199. Themed Issue: Treatment Eﬀect 1.

Villas-Boas, Soﬁa Berto, “Vertical Relationships between Manufacturers and Retailers:

Inference with Limited Data,” The Review of Economic Studies, 2007, 74 (2), 625–652.

Woodward, Susan E and Robert E Hall, “Diagnosing consumer confusion and sub-

optimal shopping eﬀort: Theory and mortgage-market evidence,” American Economic

Review, 2012, 102 (7), 3249–76.

A Matching the Data

A.1 MLS to CoreLogic Property Data

I begin by matching the CoreLogic MLS data to the CoreLogic Property data. I ﬁrst merge

on APN and county FIPS code. However, there are a number of properties in the MLS

which do not have APN numbers. Thus, I take all unmerged properties in both data sets,

and attempt to mere them on address.

This gives me a data set of basic property characteristics merged to the MLS data. Now,

I can merge the data to the CoreLogic Mortgage data. This merge is an imperfect merge

on APN number, FIPS, and tax account number. However, as this is not a unique match, I

restrict this sample to cases where the ownership data match.

Now, I have a data set of MLS listings matched to property records, including the mortgage

provider. I now clean the names of the real estate agencies and lenders. This is a nontrivial

process, requiring much hand checking. However, without this, it is not possible to have

consistent ﬁrm names.

A.2 Merging in the LLMA Data

There are very few unique identiﬁers in the LLMA data that are shared with my MLS-

CoreLogic Property data set. Thus, this is a very inexact match. I merge ﬁrst on origination

year-month, loan amount, and zip code. Then, I use a random number generator to keep

one observation of each of the observations matched on these three variables.

I next take the unmatched LLMA and MLS-CoreLogic Property data, and merge on just

year-month and zip code. Again, I use a random number generator to keep a unique obser-

vation from each match.

After merging in the LLMA Originations data, it is easy to merge in the events data. There

is a unique loan-level identiﬁer which can be matched.

A.3 Merging in the HMDA Data

Similar to the other inexact matching steps, this process is an iterative match. Data are ﬁrst

matched between the CoreLogic Property Records and HMDA on tract, exact loan amount,

and lender name. Then, a random number generator is used to keep one observation from

each match. Next, the data are matched just on tract and exact loan amount, and a unique

observation from each match is kept. Next, the loan amount is allowed to vary sequentially

more each iteration from ±$1, 000 up to ±$10, 000 in increments of 1,000. At each iteration,

the previous two steps are repeated. That is, ﬁrst tract, loan amount, and lender name are

used. Then, after a unique observation is kept lender name is dropped and the process is

repeated with just tract and loan amount before moving on to the next iteration with a

higher loan range.

B Model Results

Below are the ﬁrst stage results for the instrumental variables in the logit demand. I use one

instruments: the 10-year Treasury Rate interacted with the number of lenders in the market.

The results of this estimation are reported below. The coeﬃcient on Merged Pair is positive

but insigniﬁcant, which is consistent with the reduced form results. Furthermore, the

Table 11: First Stage

Interest Rate

Treasury Rate * Num. Lenders 0.004

∗∗∗

(0.0002)

Jointly Owned Pair 0.016

(0.016)

Time to Close −0.001

∗∗∗

(0.0002)

FE Lender

F-Statistic 40

Observations 162,667

0.200

* 0.10 ** 0.05 *** 0.01

Notes: Results of ﬁrst stage regressions on a 25% random sample of data. Unit of observation is a loan

product j in a market t. Standard errors calculated on 500 bootstraps of random sample at market level.

C Results for All Agent Preferred Loans

Below, I report the same borrower and loan-level outcomes I report in the main paper, but

keeping all loans I ﬂag as agent preferred. In general, the results are robust to this change in

speciﬁcation: there is little to no impact on time to close, ex-ante borrower characteristics,

or ex-post loan performance.

The one set of results which do substantively change are those on interest rate in Table

C. Here, the coeﬃcient on T reat ∗ P ost

jlt

is positive and signiﬁcant in my main sample, but

not when looking at all loans. This is consistent with the fact that the full sample includes

distribution channels that are unlikely to be subject to price eﬀects from the merger, such

as mortgage brokers.

Table 12: Time to Close

Days 30+ Days 45+ Days 60+ Days

Mean 42.0 0.72 0.35 0.16

Treat*Post -2.0*** -0.034*** -0.032* -0.014

(0.54) (0.013) (0.019) (0.011)

Treat 1.6** 0.035** 0.025 -0.0012

(0.61) (0.014) (0.022) (0.013)

FE CBSA,Year CBSA,Year CBSA,Year CBSA,Year

Lender,Agency Lender,Agency Lender,Agency Lender,Agency

R-squared 0.29 0.22 0.22 0.20

N 1,132,834 1,132,834 1,132,834 1,132,834

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the number of days between a home under contract and the home closing

(column (1)), or indicator for if it took more than 45 days (column (2)), 60 days (column (3)), or 75 days

(column (4)) to close. Unit of observation is a loan matched to a home purchase. Treat*Post represents

the coeﬃcient of interest and equals one if the lender and agency are a merged pair. Treat equals one

for any lender-agency pairs which ever merge. Standard errors are clustered at the agency level. ALl

columns include CBSA, Year, Lender, and Agency ﬁxed eﬀects.

Table 13: Borrower Characteristics

Amount LTV FICO 30 Days 60 Days 90 Days

Mean 254,484 87.00 736 14.2% 7.8% 5.8%

Treat*Post -8,213 -0.55 3.1* -0.0072 0.0051 0.0012

(5,923) (0.71) (1.8) (0.011) (0.0062) (0.0043)

Treat 2,692 0.28 -2.1 0.016 -0.0067 -0.0076

(6,431) (0.74) (2.1) (0.013) (0.0084) (0.0062)

FICO*LTV -0.000025*** -0.000022*** -0.000019***

(5.3e-07) (4.1e-07) (3.6e-07)

FICO Score 0.00053*** 0.00087*** 0.00083***

(0.000047) (0.000036) (0.000031)

LTV 0.020*** 0.017*** 0.015***

(0.00041) (0.00032) (0.00028)

FE CBSA,Year CBSA,Year CBSA,Year CBSA,Year CBSA,Year CBSA,Year

Lender,Agency Lender,Agency Lender,Agency Lender,Agency Lender,Agency Lender,Agency

R-squared 0.52 0.20 0.18 0.16 0.14 0.13

N 846,338 1,189,050 1,096,155 1,095,983 1,095,983 1,095,983

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is loan amount (column (1)), loan-to-value ratio (column (2)), FICO (credit)

score at origination (column (3)), or indicator for if loan is ever 30 days delinquent (column (4)), 60 days

delinquent (column (5)), or 90 days delinquent (column (6)). Unit of observation is a loan matched to a

home purchase. Treat*Post represents the coeﬃcient of interest and equals one if the lender and agency

are a merged pair. Treat equals one for any lender-agency pairs which ever merge. Standard errors are

clustered at the agency level.

Table 14: Interest Rate (Mean = 4.09%)

(1) (2)

Treat*Post 0.026 0.054*

(0.032) (0.030)

Treat -0.021 -0.041*

(0.033) (0.025)

FICO*LTV 0.000042***

(6.7e-07)

FICO Score -0.005***

(0.0001)

LTV -0.030***

(0.001)

Sun&Abraham(2020) 0.067*** -0.032

(0.009) (0.034)

Sample All All

FE CBSA,Year CBSA,Year

Lender,Agency Lender,Agency

R-squared 0.41 0.43

N 1,162,386 1,162,386

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is loan interest rate. Unit of observation is a loan matched to a home purchase

where loan is originated through the retail channel. Treat*Post represents the coeﬃcient of interest and

equals one if the lender and agency are a merged pair. Treat equals one for any lender-agency pairs

which ever merge. Standard errors are clustered at the agency level.

D Robustness Checks

D.1 Placebo Test

In Table 5, I show that the eﬀect is statistically signiﬁcant for buyers who use the retail dis-

tribution channel. This is consistent with the story that buyers who use the retail channel

are most captured by the one-stop shopping model and thus charged the higher rate.

By similar logic, buyers who use a mortgage broker should be the least captured by a merged

pair, and thus, I would not expect to see an eﬀect on the interest rate. To that end, running

my speciﬁcation on buyers who use the mortgage broker distribution channel is a placebo

test for my results.

I report that speciﬁcation in Table 15. As can be seen, the result is insigniﬁcant. Even

if it were signiﬁcant, the point estimate is negative, suggesting that buyers who use mort-

gage brokers pay less after the merger. These results are consistent with the story I have

put forth.

D.2 Points Paid

One way that my identifying assumption would be violated is if borrowers not exposed to

a merged lender-agent pair behave diﬀerently than the borrowers not exposed to a merged

lender-agent pair. I have shown in Section 7.5 that on observable borrower characteristics

that are plausibly determined prior to the loan contract (namely, loan amount, FICO score,

and loan-to-value ratio), borrowers do not diﬀer substantially in most respects. However,

it is possible that borrowers change their behavior within the loan contract. Speciﬁcally, in

deciding to pay discount points. Borrowers can decide to pay points at origination, where

they trade an upfront fee (called ”points”) in exchange for lowering the interest rate on the

Table 15: Interest Rate for Mortgage Broker Intermediated Loans

(1)

Treat*Post -0.144

(0.200)

Treat -0.211

(0.178)

Sun&Abraham(2020) -0.44

(0.057)

Sample Mtg.Broker

FE CBSA,Year

Lender,Agent

R-squared 0.63

N 7,032

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the loan interest rate. Unit of observation is a loan matched to home

purchase. Treat*Post represents the coeﬃcient of interest and equals one if the lender and agent are a

merged pair. Treat equals one for any lender-agent pairs which ever merge. Standard errors are clustered

at the agency level.

loan. Bhutta and Hizmo (2021) document that the apparent gap in interest rates paid by

white vs. minority borrowers can be explained by the fact that white borrowers tend to pay

more points in exchange for a lower interest rate.

If borrowers who use merged lender-agent pairs after merger are paying fewer points than

their pre-merger counterparts, that would appear as higher interest rates post-merger, which

is not truly a consequence of the joint ownership.

Unfortunately, the LLMA data do not include points paid. However, due to a rule change,

points paid are reported in HMDA data beginning in 2018. Thus, for the last two years of my

data, I can merge in HMDA and observe both the interest rate and points paid. Due to the

small nature of this sample, I am not able to restrict to just agent referred buyers, however

earlier versions of my results were robust to either sample, so that is not driving these results.

In Table 16, I make points paid the dependent variable in columns (1) through (3). The

ﬁrst two columns are for all loans, while the third column restricts to just retail channel

loans. Beginning with the ﬁrst column, we see that the coeﬃcient on T reat ∗ P ost is both

very small in magnitude and insigniﬁcant. This holds with the inclusion of credit metrics in

column (2), and when restricting the sample to just retail loans in column (3). All in all,

points paid cannot explain my main results.

Table 16: Discount Points

(1) (2) (3) (4)

Treat*Post 1.2e-08 1.4e-08 -2.3e-09 -6.0e-09

(4.8e-08) (5.6e-08) (5.8e-08) (5.8e-08)

Treat -1.9e-08 -2.3e-08 -7.4e-09 -2.4e-09

(4.7e-08) (5.5e-08) (5.8e-08) (5.8e-08)

FICO*LTV 6.4e-12*** 9.3e-12***

(5.7e-13) (6.5e-13)

FICO -7.6e-10*** -9.4e-10***

(5.1e-11) (5.8e-11)

LTV -6.0e-09*** -8.1e-09***

(4.3e-10) (5.0e-10)

Sample All All Retail Retail

FE CBSA, Year, CBSA, Year, CBSA, Year, CBSA, Year,

Lender, Agency Lender, Agency Lender, Agency Lender, Agency

R-squared 0.14 0.15 0.14 0.14

N 690,167 651,645 469,841 469,293

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the amount of discount points paid at origination. Unit of observation is

a loan matched to home purchase. Treat*Post represents the coeﬃcient of interest and equals one if the

lender and agent are a merged pair. Treat equals one for all lender-agency pairs which merge at any

point in time. Standard errors are clustered at the agency level.

D.3 Total Origination Costs

Similarly to the discussion above, it is possible that after the merger, buyers pay lower orig-

ination costs in exchange for higher interest rates. While the LLMA data do not include

origination costs, the same rule change which required lenders to report any discount points

paid also required lenders to report total origination costs. Thus, for the last two years of my

data I can test for diﬀerences in total origination costs by merged and unmerged lender-agent

pairs.

I report these results in Table 17. In columns (1) and (2) I include all loans while column

(3) considers only retail channel loans. Beginning with column (1), there is no relationship

between origination costs and merged status. This continues to be true when including credit

score, loan-to-value ratio, and their interaction in column (2) and subsetting to retail channel

loans in column (3). In all three speciﬁcations the coeﬃcient on T reat ∗ P ost is insigniﬁcant.

Furthermore, the coeﬃcient on T reat is also insigniﬁcant, suggesting that even before the

merger buyers using eventually merged lender-agent pairs were not paying diﬀerent origina-

tion costs. In short, origination costs do not explain the interest rate eﬀects I ﬁnd in Table 5.

Table 17: Total Origination Costs

(1) (2) (3) (4)

Treat*Post 251 84 185 39

(317) (379) (459) (455)

Treat -223 -82 -204 -94

(308) (364) (440) (433)

FICO*LTV -0.19*** -0.22***

(0.0049) (0.0057)

FICO 12*** 16***

(0.42) (0.48)

LTV 165*** 188***

(3.7) (4.3)

Sample All All Retail Retail

FE CBSA, Year, CBSA, Year, CBSA, Year, CBSA, Year,

Lender, Agency Lender, Agency Lender, Agency Lender, Agency

R-squared 0.29 0.33 0.33 0.35

N 714,475 674,194 481,761 481,177

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the total origination costs. Unit of observation is a loan matched to home

purchase. Treat*Post represents the coeﬃcient of interest and equals one if the lender and agent are a

merged pair. Treat equals one for all lender-agency pairs which merge at any point in time. Standard

errors are clustered at the agency level.

D.4 Only Agency-Agency Mergers

One concern about my identiﬁcation strategy is that mergers are not exogenous. Speciﬁcally,

residential real estate agencies and mortgage lenders will strategically decide to merge. In

other words, mergers not occur randomly.

First, I do not believe this is a major issue, as more than half of the mergers in my data

set are actually horizontal between two residential real estate agencies where one happens

to have a lending arm. Thus, it seems that the exposure to the lender is occurring quasi-

randomly. However, to be sure, I re-run my main analysis keeping only those mergers which

are agency-agency, removing all observations which are aﬀected by an agency-lender merger.

These results can be found in Table 18. As can be seen, if anything the results are stronger

than my main speciﬁcation. Thus, this cannot explain the results.

Table 18: Only Exogenous Mergers

(1) (2) (3) (4)

Lender CBSA Share Within Agency Share Interest Rate Interest Rate, Retail Only

Treat*Post 0.0072** 0.14*** 0.086*** 0.089***

(0.0013) (0.011) (0.0091) (0.014)

Treat -0.0096 -0.093*** -0.090***

(0.0082) (0.014) (0.016)

FICO*LTV 0.000041*** 0.000041***

(6.7e-07) (7.5e-07)

FICO -0.0049*** -0.0049***

(0.000059) (0.000067)

LTV -0.030*** -0.030***

(0.00052) (0.00058)

Sample All Retail

FE CBSA, Year, CBSA, Year, CBSA, Year, CBSA, Year,

Lender Lender, Agency Lender, Agency Lender, Agency

R-squared 0.62 0.61 0.43 0.45

N 414,265 3,175,875 1,056,558 796,412

* 0.10 ** 0.05 *** 0.01

Notes: Treat*Post represents the coeﬃcient of interest and equals one if the lender is merged (column

(1)) or if a lender and agent are a merged pair (columns (2) through (4). Treat equals one for any

lender-agent pairs which ever merge. Standard errors are clustered at the lender level in column (1) and

at the agency level in columns (2) through (4).

D.5 Stacked Event Study

A common way to study mergers retrospectively is to construct a control group for each

merger, generate a merger-speciﬁc ID, and then stack each of these mergers in an event

study using the merger-IDs as ﬁxed eﬀects. This way, each merger is considered individu-

ally, and the issues about staggered DiD do not apply. This is similar to the approach taken

in prior merger retrospective papers. Because I observe over 100 mergers, I can exploit a

similar approach in my data, and I include it as a robustness test.

For each merger, I construct a control group of all unmerged lender-agent pairs at time

t, and who come from a county where I also observe the merging lender-agent pair. In other

words, for each merger, my sample for that merger consists of all observations for a given

merger as well as all observations from that county that are never treated. In this way, each

merger-ID compares the eﬀects of that merger to observations never treated by a merger that

creates sibling lender-agent pairs. This speciﬁcation avoids the issues of staggered diﬀerence-

in-diﬀerence by giving each merger its own control group. I report the results of this for each

of my main regression speciﬁcations in Table 19. Due to sample size, I omit the mortgage

broker results.

As can be seen below, the results are largely consistent with the main speciﬁcation, with

the exception of the lender’s CBSA market share. Now, this speciﬁcation ﬁnds that lenders

that merge with residential real estate agencies lose 0.53 percentage points of market share

after merging, in contrast to the 0.54 percentage point gain found in my main speciﬁcation.

However, the result in column (2) that a lender’s within-agency market share increases by

0.20 percentage points when they merge with a residential real estate agency is consistent

with the main ﬁnding of 0.15 percentage points. Likewise, the positive but insigniﬁcant 5

basis point eﬀect from using a merged lender-agent pair in column (3) and the signiﬁcant

eﬀect of 11 basis points when restricting to only buyers going directly to the lender in column

(4) mirror the results in the main section of the paper.

With the exception of the lender CBSA market shares, the results in Table 19 are stronger

than my main results. However, as this is dependent on the particular control group used, I

choose to keep this as a robustness check instead of as the main speciﬁcation.

Table 19: Stacked Event Study

(1) (2) (3) (4)

Dependent Variable CBSA Share Agency Share Rate Rate

Treat*Post -0.0053* 0.20*** 0.055 0.11***

(0.0030) (0.013) (0.047) (0.031)

Treat 0.015** -0.047 -0.091***

(0.0069) (0.046) (0.026)

FICO*LTV 0.000043***

(7.4e-07)

FICO -0.0050***

(0.000066)

LTV -0.030***

(0.00056)

Sample All All All Retail

FE CBSA, Year, CBSA, Year, CBSA, Year, CBSA, Year,

Lender Lender, Agent Lender, Agent Lender, Agent

R-squared 0.41 0.72 0.45 0.47

N 2,465,198 24,701,215 14,777,831 11,301,873

* 0.10 ** 0.05 *** 0.01

Notes: Treat*Post represents the coeﬃcient of interest and equals one if the lender is merged (column

(1)) or if a lender and agent are a merged pair (columns (2) through (4). Treat equals one for any

lender-agent pairs which ever merge. Standard errors are clustered at the lender level in column (1) and

at the agency level in columns (2) through (4).

D.6 Additional Fixed Eﬀects

Below, I report the interest rate result using CBSAxYear ﬁxed eﬀects, FIPSxYear ﬁxed

eﬀects, and Zip-Year ﬁxed eﬀects. In general, the results are robust to the ﬁxed eﬀects

speciﬁcation I choose, and of similar magnitude.

Table 20: FIPS-Year Fixed Eﬀects

(1) (2) (3) (4)

Treat*Post 0.021 0.037 0.069** 0.079***

(0.039) (0.036) (0.028) (0.025)

Treat -0.0085 -0.019 -0.040 -0.052**

(0.039) (0.037) (0.029) (0.026)

FICO*LTV -0.000012*** 0.00004***

(1.5e-06) (7.4e-07)

FICO 0.00018 -0.0049***

(0.00013) (0.00007)

LTV 0.011*** -0.030***

(0.0011) (0.0006)

Sample All All Retail Retail

R-squared 0.42 0.47 0.44 0.46

N 1,188,913 1,095,474 829,938 824,220

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the loan interest rate. Treat*Post represents the coeﬃcient of interest and

equals one if the lender and agent are a merged pair. Treat equals one for any lender-agent pairs which

ever merge. Standard errors are clustered at the agency level. Fixed eﬀects included in all columns are

FIPSxYear, FIPS, Year, Lender, Agency.

Table 21: CBSAxYear Fixed Eﬀects

(1) (2) (3) (4)

Treat*Post 0.020 0.044 0.071** 0.081***

(0.040) (0.034) (0.028) (0.024)

Treat -0.0088 -0.024 -0.041 -0.053**

(0.040) (0.035) (0.029) (0.026)

FICO*LTV 0.000041*** 0.000042***

(6.5e-07) (7.4e-07)

FICO -0.0049*** -0.0049***

(0.000057) (0.000066)

LTV -0.030*** -0.030***

(0.00050) (0.00057)

Sample All All Retail Retail

R-squared 0.41 0.44 0.44 0.46

N 1,189,092 1,095,646 830,157 824,442

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the loan interest rate. Treat*Post represents the coeﬃcient of interest

and equals one if the lender and agent are a merged pair. Treat equals one for any lender-agent pairs

which ever merge. Standard errors are clustered at the agency level. CBSAxYear, CBSA, Year, Lender,

Agency.

Table 22: ZipxYear Fixed Eﬀects

(1) (2) (3) (4)

Treat*Post 0.036 0.053* 0.074*** 0.087***

(0.042) (0.031) (0.019) (0.015)

Treat -0.033 -0.050 -0.056** -0.067***

(0.044) (0.034) (0.028) (0.025)

FICO*LTV 0.000039*** 0.000039***

(9.2e-07) (1.0e-06)

FICO -0.0047*** -0.0047***

(0.000081) (0.000091)

LTV -0.028*** -0.028***

(0.00070) (0.00078)

Sample All All Retail Retail

R-squared 0.44 0.46 0.46 0.48

N 622,278 572,097 428,953 425,511

* 0.10 ** 0.05 *** 0.01

Notes: Dependent variable is the loan interest rate. Treat*Post represents the coeﬃcient of interest and

equals one if the lender and agent are a merged pair. Treat equals one for any lender-agent pairs which

ever merge. Standard errors are clustered at the agency level. ZIPxYear, ZIP, Year, Lender, Agency.