<<

ABSTRACT

AIRBNB AND ITS EFFECTS ON : FROM CINCINNATI

by Jordan Matthew McMaster

Airbnb is a two-sided platform that connects individuals seeking short-term lodging to owners renting out their homes. For many property owners, it is possible that revenues from a short-term rental listing will surpass those of a long-term lease agreement. Therefore, it is likely that property owners are incentivized to evict their long-term tenants to list their on Airbnb. In Cincinnati, OH, the city council voted in April 2019 to enact a short-term rental policy to combat this potential issue among others. We leverage the policy intervention and study the effects of the Airbnb market on evictions in Cincinnati utilizing both Regression Discontinuity in Time and Difference-in-Difference specifications. We found that the policy led to a subsequent decrease in the total number of listings, multi-listers, and the percent of entire-unit listings. If property owners are evicting their long-term tenants to list their properties on Airbnb on a grand scale, one would expect that this dip in Airbnb activity after April 2019 would lead to fewer evictions. However, we found no significant variation in evictions during this time period. Therefore, there is little evidence of a significant relationship between Airbnb usage and evictions in Cincinnati, OH based on the data we have at this time.

i

AIRBNB AND ITS EFFECTS ON EVICTIONS: EVIDENCE FROM CINCINNATI

Thesis

Submitted to the

Faculty of Miami University

in partial fulfillment of

the requirements for the degree of

Master of Arts

by

Jordan Matthew McMaster

Miami University

Oxford, Ohio

2020

Advisor: Dr. Mark J. Tremblay

Reader: Dr. Charles C. Moul

Reader: Dr. Austin C. Smith

©2020 Jordan Matthew McMaster

This thesis titled

AIRBNB AND ITS EFFECTS ON EVICTIONS: EVIDENCE FROM CINCINNATI

by

Jordan Matthew McMaster

has been approved for publication by

Farmer Business School

and

Department of Economics

______Dr. Mark J. Tremblay

______Dr. Charles C. Moul

______Dr. Austin C. Smith

Table of Contents

List of Tables ...... iv List of Figures ...... v Chapter 1: Introduction ...... vi Chapter 2: Literature Review ...... vii Chapter 3: Data ...... viii Airbnb Data...... viii Data ...... ix Cincinnati, Ohio ...... x Cincinnati’s Short-Term Rental Policy...... xi Chapter 4: Estimation ...... xii Two-Way Fixed Effects Panel Regressions ...... xii Regression Discontinuity in Time (RDiT)...... xiii Difference-in-Differences (DiD) ...... xiv Parallel Trends Assumption for the DiD Specification ...... xiv Placebo Tests for the DiD Specification ...... xv Alternative Event-Study Design ...... xv Chapter 5: Results ...... xvi Regression Discontinuity in Time (RDiT)...... xvi Difference-In-Differences (DiD) ...... xviii Chapter 6: Conclusion...... xx References ...... xxi

iii List of Tables

Table 1: Full Sample Summary Statistics by Zip-Code Month ...... xxiv Table 2: Aggregated Annual Trends ...... xxv Table 3: Comparative Summary Statistics (2019) ...... xxvi Table 4: Reduced-Form Linear RDiT Specification (All Observations) ...... xxvii Table 5: Reduced-Form Linear RDiT Specification ( -Eight, +Four Month Bandwidths) ..... xxviii Table 6: Reduced-Form Linear RDiT (120 Day Bandwidths) ...... xxix Table 7: Pre-Enforcement Differences in Outcomes ...... xxx Table 8: Parallel Trends - DiD Specification (Zip Code/ Month) ...... xxxi Table 9: DiD Specification (Zip Code/ Month) ...... xxxii Table 10: DiD Specification (Zip Code/ Day) ...... xxxiii Table 11: DiD Specification (Zip Code/ Month - October 2018 Threshold) ...... xxxiv

iv List of Figures

Figure 1: Hamilton County Zip Codes and Boundaries...... xxxv Figure 2: Airbnb Listings Locations Across Hamilton County ...... xxxvi Figure 3: Evictions Locations Across Hamilton County ...... xxxvi Figures 4-8: Key Variable Trends in Treatment Zip Codes for RDiT ...... xxxvii Figures 9-13: Key Variable Trends in Control Zip Codes for RDiT………………………...xxxvii Figures 14-18: Event-Study Graph………………………………………………………… xxxviii Figure 19: Number of Listings Time-Disaggregated Effects by Treatment………………….xxxix Figure 20: Entire-Unit Ratio Time-Disaggregated Effects by Treatment……………………xxxix Figure 21: Evictions Time-Disaggregated Effects by Treatment………………………………..xli

v Chapter 1: Introduction

A 2015 Community Compact outlined Airbnb’s self-proclaimed commitment to work with local governments amid regulatory threats (news.airbnb.com). 1The Compact was meant to outline clear rules for short-term rentals and focused on three key aspects: taxes, community, and transparency. Burgeoning talk of regulatory action has come about in response to claims of adverse consequences attributed to the platform’s disturbance to traditional lodging through its home- sharing services. To name a few: less-reliable tax payments to cities, a contribution to wealth inequality, diminished housing affordability, and the exacerbation of housing shortages across the country. Airbnb is a two-sided platform that primarily connects individuals seeking short-term lodging to residential property owners renting out their home, whether it be a shared-room or the entire unit. The company has no stake in the listings themselves, rather, acts as an intermediary that collects commissions each time a booking takes place. 2Founded in 2008 in a San Francisco apartment, the platform has since grown to over 150 million users, 7 million global listings, and an estimated valuation of nearly 31 billion dollars leading into 2020. Airbnb’s consistent growth and impressive consumer demand have placed intense pressures on long-standing regulatory frameworks (news.aribnb.com; Samaan, 2015). 3In response, state and local governments have both proposed and enacted regulatory action against the home-sharing platform, bolstering the conflict between the sharing economy – markets for shared and idle resources - and established economic regulation. Legislation on topics ranging from land use to outright bans, has been defended by authors through alleged preservation of affordable housing and community integrity. 4Airbnb has taken on numerous initiatives and introduced direct channels of communication between the platform and the communities with which it operates, addressing the accusations head on. In September 2019, Airbnb announced its plan to invest 25 million dollars to support affordable housing in its home state of California, specifically in the San Francisco Bay Area and in Los Angeles County (Khouri, 2019). The initiative remains open-ended contingent upon its success. Furthermore, following the release of its Community Compact, Airbnb launched a Neighborhood Support feature on its website where community members can directly discuss concerns regarding local hosts and launch formal complaints. 5Data reveal that in recent years the platform has become more popular with multi-unit listers and absentee owners. This trend is a leading cause for much of the concern put forth by local governments and community leaders. Given the right location and pricing scheme, there can exist an incentive for property owners to convert long-term rentals into listings on platforms like Airbnb and HomeAway. A viral incident in 2018 took place where “This Airbnb displaced 5 people” was painted on the sidewalk outside of an Airbnb listing in New Orleans where a once long-term dwelling for a family of five is now operating as an Airbnb. (Maldonado, 2018). This

1 The Airbnb Community Compact, along with numerous other regulatory announcements, can be reviewed on news.airbnb.com. 2 These general and financial statistics come from iPropertyManagment, Yahoo! Finance, and Bloomberg. 3 The Federal Trade Commission (FTC) published a comprehensive 2015 study following the early and major regulatory discussions; “Airbnb, Rising Rent, and the Housing Crisis in Los Angeles.” Frequently updated neighborhood-level regulations can be followed on the company website: airbnb.com. 4 Reference the Airbnb Community Compact, additional announcements, and other site features like “Neighborhood Support” on news.airbnb.com. 5 The data will be explained further in section 4 with the accompanying summary statistics.

vi type of occurrence, the conversion of long-term rentals into short-term listings on a grand scale, can significantly reduce the number of available units for local renters and prospective homeowners and lead to upward price adjustments (Bibler et al., 2020; Cader-Wang, 2019; and Cosman Quintero, 2019). This mechanism requires the potential hosts to survey their surroundings, exhibit logical thinking, and put forth the effort to convert otherwise long-term units into short-term rentals. It must be believed that revenues from a short-term rental listing will outweigh those of a medium- to long-term lease agreement. Furthermore, there are numerous costs associated with the conversion of a long-term unit. If conversion costs are greater than some specified threshold unique to the individual, mainly that which will lead to lower profits overall, there will be a lower probability of conversion. Conversion costs are mainly those realized through policy actions like taxes and fees, though other property-specific maintenance costs do exist. If policy actions lead to higher costs and lower conversion rates, we would expect the rental and housing inventory to shrink at a reduced pace and evictions to decline in areas where the policy is active. The question then becomes, does this mechanism exist? This study aims to answer this question directly. Few academic studies have examined the potential displacement of long-term tenants as property owners continue to list more properties on Airbnb or other similar platforms. This study aims to uncover the causal effects of the Airbnb market on evictions using evidence from Cincinnati, OH. To do so, we use a panel of daily observations spanning a period of nearly five years between October 2014 and August 2019. Data were also aggregated at the zip code - month level and leveraged to carry out regression discontinuity in time and difference-in-differences specifications. 6The introduction and passing of a short-term rental policy specific to Cincinnati in April 2019, which aims to introduce a seven percent excise tax, regulate the number of units per building, and require licensing for all hosts is used as a policy shock, or intervention in both specifications. If property owners are evicting their long-term tenants to list their properties on Airbnb on a grand scale, one would expect this potential supply shock to lead to fewer evictions overall as Airbnb activity declines.

Chapter 2: Literature Review

Numerous studies have examined the potential consequences attributed to Airbnb, e.g., housing affordability, housing supply, and the public’s general welfare. Jefferson-Jones (2015), for example, uses a theoretical approach to examine the tension between growing patterns of home- sharing and existing regulations that aim to prohibit such behaviors. Acknowledging the potential for a positive relationship between housing prices and the presence of Airbnb, the author suggests that this increase may stem from a newfound ability of property owners to better maintain the residence. By listing properties on short-term rental platforms, owners may shift the burden of home , and in turn, increase the property’s value. On the other hand, Lee (2016), another descriptive analysis, suggests that Airbnb increases rent, reduces the housing supply, and exacerbates segregation. The author outlines clear and strict policy recommendations to prevent further costs in major U.S. metros. More recent studies have been able to estimate some causal effects. Horn and Merante (2017) study the effects of Airbnb on rental rates in Boston, Massachusetts. Aggregating data at the census tract level, the authors suggest that a one standard deviation increase in listings increases

6 The specifications and the use of the policy shock will be formally discussed in Chapter 4.

vii asking rents by 0.04 percent. Similarly, though broader scoped, Barron, Kung and Proserpio (2018) use an instrumental variable to estimate a causal effect of Airbnb on rents and housing prices across the United States. Aggregated at the zip code level, they suggest that a one standard deviation increase in listings increases rents by 0.018 percent and home prices by 0.026 percent. Furthermore, and similar to the study at hand, they show that increases in Airbnb listings lead to more short-term rentals and fewer long-term rental units. Last, Calder-Wang (2019) examines the effect of converting long-term units to short-term rentals on rent across housing types and demographics. She ultimately suggests that, overall, home-sharing is detrimental to the public’s general welfare. We deviate from the above studies in that we aim estimate the causal effect of the Airbnb market in Cincinnati on evictions by leveraging strongly balanced panel data to carry out both regression discontinuity in time and difference-in-differences specifications. We will focus on the direct displacement of long-term tenants associated with property owners listing otherwise long- term properties on short-term rental platforms, a mechanism somewhat outside of the general affordability debate. 7One of our main contributions is to shed light on whether the policy actions of many cities like Cincinnati, Ohio regarding Airbnb, specifically those that target multi-listers and un-hosted units are useful. Furthermore, we hope this study will be yet another piece of the puzzle in estimating the causal effects of home-sharing platforms on the public’s general welfare. As companies like Airbnb and Uber become larger pieces of the U.S. economy, understanding their broader consequences, either good or bad, becomes absolutely necessary.

Chapter 3: Data

Majority of the data come from AirDNA and the Hamilton County Clerk of Courts. Daily observations from both sources span a period of nearly five years and cover all of Hamilton County, Ohio. 8The inclusion of latitude and longitude in both of these datasets allows us to leverage geospatial software and aggregate observations at the zip code level, permitting zip code - day and zip code - month levels of analysis. Additional covariates such as rent, housing prices (HPI), unemployment rate, and the economic conditions index (ECI) come from Zillow, the Federal Housing Finance Agency (FHFA), and The Federal Reserve Bank of St. Louis, respectively. Figure 1 displays three key features of the study area, (1) the Hamilton County boundary, (2) the county broken up by zip code, and (3) the Cincinnati city boundary. We provide a more in-depth description of the data and its sources in the following sub-sections.

Airbnb Data

Data specific to Airbnb was retrieved from AirDNA, a third-party source that frequently scrapes airbnb.com. Similar data were used by Bibler et al. (2019) when analyzing tax enforcement in short-term rental markets and again in Bibler et al. (2020) when studying the effects of short-term rentals on housing affordability. The property-specific daily observations span a period between October 2014 and August 2019, leaving the initial and terminal years incomplete. The data cover all listings, active or otherwise, in Hamilton County and include variables such as availability, bookings, and asking and booking prices.

7 The Cincinnati STR policy will be formally discussed in the Chapter 3. 8 We used ArcGIS Pro for all of the geospatial aggregation.

viii Also included are numerous time-invariant property characteristics like latitude and longitude, the number of bedrooms, number of bathrooms, and maximum number of guests. Figure 2 displays all Airbnb listings that were active at some point within the period of study and is layered with the uniquely colored boundaries of any city with a population greater than 10,000. There are a total 3,335 unique listings that fall into this category. There are approximately 58 zip codes included in the study area. Table 1 presents summary statistics at the zip code-month level. For example, for any of the 58 zip codes in any given month during the study period, we would expect to find on average approximately 20 Airbnb listings and just over 100 bookings. Other variables in the table include the percent of multi-listers, the percent of un-hosted units, asking price, booking price, the number of bedrooms, bathrooms, and maximum guests. Table 2 presents annual trends for Hamilton County between 2014 and 2019. The year 2014 contains three months of observations while 2019 contains eight. Thus, these years must be interpreted carefully. Panel A represents the aggregated trends for all of Hamilton County, Panel B the annual trends for the area within Hamilton County but outside of the Cincinnati boundary, and Panel C the annual trends for Cincinnati, OH. The annual percent change is listed just to the right of each variable and totals are provided for those that can be summed. Regarding the overall percent change, the initial and terminal years are excluded from the calculation.

Eviction Data

Eviction related data were supplied by the Hamilton County Clerk of Courts upon request. The data spans a six-year period between January 2014 and January 2020 but was trimmed to match the availability of our short-term rental data. 9Though each observation included the address of the eviction, exact locations were converted to coordinates using Geocodio. This approach better aligned with AirDNA and allowed for the use of geospatial software to aggregate the observations at the zip code level. Hamilton County exhibits good variation in evictions. This is mainly a product of consistent and detailed record keeping. 10That said, the county is consistently about five percentage points higher than the U.S. average in its eviction rate (6.3 percent) and about ten percentage points higher in its evictions filing rate (8.7 percent). Furthermore, less than one percent (0.4 percent) of evictions filed over the study period were decided in favor of the tenant. The majority of evictions were dismissed (49.9 percent) due to negotiations and voluntary departures or were decided in favor of the landlord (47.6 percent). Any eviction filed within the period of study, no matter its outcome, is considered in our estimation. One could argue that the distribution of court rulings makes sense considering that individuals are not guaranteed representation in eviction court. Over the study period the majority of landlords had representation (88.2 percent) while the overwhelming majority

9 Geocodio allows the user to batch geocode a spreadsheet in either direction (i.e., address to coordinates, or coordinates to address) for a small fee. Once the user creates an account, they gain access to numerous attachments, such as U.S. census data. The census data in the study, however, was pulled directly from www.census.gov. 10 Additional eviction statistics were obtained from "’You are being asked to leave the premises’: A Study of Eviction in Cincinnati and Hamilton County, Ohio, 2014-2018,” a study carried out by the University of Cincinnati in 2018. Though the study was concluded prior to 2019, a year included in our analysis, the statistics are believed to be relevant and representative.

ix of tenants did not (97.5 percent). Given that landlords are statistically favored, an apparent ease of victory in eviction court in Hamilton County may lessen the cost of conversion if the property owner desires to list the property on a short-term rental platform. Nearly half of all residences (42.3 percent) in Hamilton County were renter occupied as of 2018. Between 2014 and 2019, the county filed an average of nearly 12,000 evictions each year. Summary statistics and annual eviction trends are presented in Tables 1 and 2. Figure 3 displays the location of any eviction that took place over the study period and is again layered with the uniquely colored boundaries of any city with a population greater than 10,000. Cincinnati accounted for approximately 69 to 71 percent of Hamilton County evictions throughout the study period. Just under 19 percent of Cincinnati’s evictions are attributed to the top ten organizational landlords in the area. These include the Cincinnati Metropolitan Housing Authority, Community Management Corporation, and various other LLC’s and corporations. Though the filer is not specified in our data, these statistics leave room, approximately 85 percent of the county’s evictions, for private owners to make their own conversion decision. This is our ideal variation in the dependent variable.

Cincinnati, Ohio

Cincinnati resides in southwestern Ohio on the right bank of the Ohio River. The city is the third largest in Ohio, just behind Columbus and Cleveland, and is the 64th largest city in the United States with a 2018 census population estimate of 302,605. The Cincinnati Metropolitan Statistical Area (MSA) is made up of seven counties in three states – Ohio, Kentucky, and Indiana. Data limitations restricted the sample to a single county, Hamilton County, OH, which fully encompasses the city of Cincinnati. Cincinnati is comprised of approximately 29 zip codes while the remaining 29 make up Hamilton County. Given the data restrictions, it is important to consider whether Cincinnati is a useful and representative area of study. Table 3 presents summary statistics across three cities - Cincinnati, Cleveland, and Pittsburgh – their corresponding counties - Hamilton County, Cuyahoga County, and Allegheny County - and the United States. Panel A breaks down area demographics, Panel B education, and Panel C income, labor, and housing. Both Cincinnati and Hamilton County are situated in the middle for nearly all marks included in the table. 11The two exhibit an above average labor force participation rate, a below average owner-occupied housing rate, and a slightly below average Asian population. In short, it appears there is nothing particularly unique about the city nor the county with which it resides regarding these statistics. The last row of Table 3 presents average rent for the three cities. Cincinnati, in 2019, had the highest average rent at $1,135, compared to $1,022 and $1,113 for Cleveland and Pittsburgh, respectively. These all fell below the national average of $1,601 in 2019. If we consider Cincinnati’s average Airbnb occupancy rate of 65 percent across all entire- unit listings, which accounted for 79 percent of listings in the city, we can quickly analyze the potential benefit of conversion. The average asking and booking prices in 2019 were $127 and $88, respectively. If a unit were occupied 65 percent out of the year (as the average would suggest), or approximately 19 days out of each month, a host would expect

11 Averages are calculated across states by the United Census Bureau. All data presented in Table 3 come from the U.S. Census Bureau’s Quick Facts at census.gov.

x to earn on average about $2,413 in monthly income. Given the booking price, a host would actually realize a monthly income of about $1,672, on average. For comparison, in 2019, average rent was approximately $2,500 in New York City, $2,400 in Los Angeles, and $1,600 in Portland. It is reasonable to assume that a potential host would scan asking prices on the Airbnb website when making their conversion decision, unless well versed or connected to the Airbnb market, ultimately comparing potential earnings to those in more populated cities. Furthermore, it is reasonable to assume that short-term rental regulation would disturb the potential for excess revenue above average rent. We find above that in simple averages, there is potential for the conversion mechanism in question to be carried out in Cincinnati.

Cincinnati’s Short-Term Rental Policy

The Cincinnati City Council first publicly discussed an ordinance designed to regulate short-term rentals in March of 2018. These talks originally targeted listings that were non- owner occupied or apartments that were rented out full-time. 12In other words, the concern stemmed from business-like units that could serve to degrade the character of the neighborhood, exacerbate the area’s already problematic housing shortage, and further decrease housing affordability. Councilman David Mann stated that the goal of the ordinance was “to balance the real and positive economic benefits of short-term rentals with the need to protect more vulnerable citizens from rent inflation and possible displacement.” Councilman Mann followed up in October of the same year with a more formal proposal. 13Key provisions from this version of the proposal stated that (1) owners would be limited to three un-hosted units with the exception of those grandfathered in, (2) all short-term rentals would be required to register with city administration to comply with a fee which was to be set at a later date, (3) operators of un-hosted units would need to acquire a that would be updated every three years and an engineers’ certificate that verified building, zoning, housing, and fire codes are being complied with, (4) a seven percent excise tax would be charged to all hosts and dedicated to the preservation and general benefit of affordable housing as well as to aid in the city’s eviction rate, and (5) units with three or more complaints of “nuisances” – noise, exceeding maximum capacity, criminal activity, and illegal parking - would no longer be permitted to operate. Many, council members and hosts alike, voiced concerns about the aggressiveness of the policy. Mann was quoted saying that “many hosts and elected officials say they would support a pared-down proposal with minimal requirements." The vote was then delayed. The council finally voted to pass a revised and simplified version of the previous year’s proposal in April of 2019. The legislation states that (1) each host will register with city administration and be required to display a registration number online, (2) hosts in buildings with four or more units will be required to obtain certification from an engineer or architect that the short-term rental is in compliance with codes, (3) a seven percent excise

12 Xavier University’s Community Building Institute and Local Initiatives Support Corporation in Cincinnati, Ohio has conducted numerous studies on the topic. 13 Cincinnati.com followed the council meetings closely and reported concise summaries of the minutes during the short-term rental discussions. Articles can be found on nearly every step of the process and pieced together to form a complete narrative.

xi tax will be collected from all short-term rentals – with the exception of hotels as they are already taxed - and revenues will be put into the Affordable Housing Trust Fund, and (4) regulates the number of units per building that are permitted to be operated as short-term rentals: (a) there is no limits for buildings with four or fewer dwelling units and (b) five short-term rentals for every four dwelling units in excess of four dwelling units is permitted for buildings with five or more dwelling units.

Chapter 4: Estimation

We employ multiple specifications in an attempt to estimate a causal effect of Airbnb on evictions. In the following sub-sections, we begin by outlining the ideal approach and the corresponding coefficient of interest. We then set up a simple regression discontinuity in time specification prior to introducing the DiD regressions and their underlying assumptions.

Two-Way Fixed Effects Panel Regressions

We would ideally be able to leverage the geospatial analysis and strongly balanced panel to estimate the following two-way fixed effects model,

Evictions = α + β BPL + δX + γ + 휇 + ε (1) it it it i 푡 it

where the number of evictions and bookings per listing (BPL) are our dependent and independent variables of interest. The subscripts i and t are ID and time indicators, respectively. This implies that γ, a dummy variable, with the subscript i only varies across zip code. The inclusion of this dummy would allow us to control for the time-invariant factors that only vary across zip code. On the other hand, the dummy variable 휇 with subscript t, would allow us to control for time varying factors that do not significantly vary across ID, e.g., cyclical listing and booking behaviors. In equation (1), X represents the set of additional covariates that simultaneously vary across time and zip code. This set would help us to partial out the variation in evictions unexplained by our BPL variable, negating the issue of omitted variable bias. Consider the conditional independence assumption (CIA),

{ Y0i, Y1i } ⊥ di | Xi (2)

E[ Yi| Xi, di = 1] − E[ Yi| Xi, di = 0] = E[ Y1i − Y0i| Xi] (3)

where Yi represents the outcome for individual i, in our case evictions, and di is a treatment indicator for individual i. If we believed that we could apply the CIA, we could assign the coefficient on BPL, β, a causal interpretation, i.e., a causal effect of listing and booking behavior on evictions given the conditioning set X. Otherwise, we would merely be estimating an association between the two variables of interest. In our case we are missing numerous covariates belonging to set X that would explain said variation, such as economic development, ultimately biasing the results of our coefficient, β. Thus, we must introduce alternative estimation techniques that still attempt to estimate a causal relationship between

xii listing and booking behaviors and evictions in Cincinnati. The following specifications attempt to do just that.

Regression Discontinuity in Time (RDiT)

It is our goal to estimate a difference-in-difference model given the data. Many papers’ comparison of RDiT and DiD estimates is deemed useful when an untreated group exists but its validity as a control is in doubt (Chen and Whalley 2012 and Gallego et al. 2013). Our reasons for including the RDiT specification are meant to address this concern preemptively. Our approach follows closely with directives of Hausman and Rapson (2018) and the work of Burger et al. (2014). Let t index time from {1, 2, …, T}. Then let T0 represent the policy intervention period. A Post variable then acts as a binary indicator that takes on a value of one when t ≥ T0 and zero otherwise. Thus, in our main RDiT specification we estimate the following reduced-form linear model,

ln ( y ) = α + β ( Post ) + γ + 휇 + ε (4) it it i 푡 it where the natural log of yit is the outcome for observation i at time t and β is our coefficient of interest. Here, we are estimating the difference in the conditional means prior and post policy intervention across the same group, whether it be the zip codes in Cincinnati or the of Hamilton County. The 휇푡 term acts as a month - year fixed effect. Equation (4) is leveraged for both day and month specifications with varying bandwidths – full-sample, eight months using the monthly data, and 120 days. The model lacks, however, the additional controls necessary to partial out potential bias from our estimation of β. In lieu of including the additional covariates directly, we utilize additional data from Zillow, the Federal Housing Finance Agency, and the Federal Reserve Bank of St. Louis, as well as equation (4) to show that the variables in question exhibit no significant variation around the policy threshold. Demonstrating continuity in control variables is suggested by Hausman and Rapson (2018) and according to their summary was included in 36 percent of RDiT papers. In the RDiT framework, it is not uncommon to worry about the differences in short- run and long-run estimations. Given that we only possess four months of data following the policy change, we ignore this concern. Furthermore, the short post-intervention period led us to rely solely on the reduced-form linear model, opposed to a parametric, as both unaccompanied evidence and a robustness check for the following difference-in-difference specifications. Figures 4-13 display trends for five key variables between January 2017 and August 2019. These include the number of evictions, number of listings, number of bookings, percent of multi-listing hosts, and percent of entire-unit listings. Months are averaged and binned into groups of four, i.e., the first tick on the x-axis represents January 2017 through April 2017. This organizational scheme addresses three issues, (a) we see almost two annual cycles of booking and listing behaviors, (b) we can address seasonal trends, and (c) exactly four months of observations follow intervention, which is represented by the vertical line. Figures 4-8 omit all control group observations, displaying only results for

xiii zip codes within Cincinnati. Figures 9-13 cover the remainder of Hamilton County, omitting all treatment group observations.

Difference-in-Differences (DiD)

A strongly balanced panel allows us to employ a DiD specification by leveraging a policy shock unique to Cincinnati. This specification closely follows the work and directives of Bilinski and Hatfield (2019). We have 3,422 zip code - month observations and 104,168 zip code - day observations. Each of these are broken down further into a treatment group, those zip codes that are within Cincinnati, and a control group, the remaining zip codes in Hamilton County. Let treatment be indexed by di, where di = 1 indicates that observation i is treated and di = 0 indicates that it is not. As before, we let t index time from {1, 2, …, T}. If we again let T0 represent the date that the policy was passed, t ≥ T0 then represents a post- intervention point in time. Our model then becomes,

ln ( y ) = α + ∑T β (t = j ∩ 푑 = 1) +δ ( Post ) + γ + 휇 + ε (5) it j =T0 j 푖 it i 푡 it where the natural log of yit is the outcome for observation i at time t. A Post variable once again acts as a binary indicator that takes a value of one when t ≥ T0 and zero otherwise. We let γi be a zip code fixed effect, hence its subscript i, and 휇푡 be a time fixed effect. Naturally, the panel data controls for γi, though it is important to highlight in the above equation. The εit term then captures the residuals that vary across both time and zip code. The treatment effects that we capture are the coefficients βj. These capture the differenced intervention period changes in outcomes of the treatment group relative to the control group at each point in time. The overall goal, however, is to retrieve the average treatment effect (ATE), or simply β. This is achieved by averaging the βj terms as follows:

1 T β = ∑j =T β (6) T− T0−1 0 j

In order for this to hold and the model produce meaningful results, there are various assumptions and robustness checks that we must first address.

Parallel Trends Assumption for the DiD Specification

The first traditional assumption for a DiD specification is that of parallel trends. It requires that our control group serves as an appropriate counterfactual to the trend of our treated units in the absence of treatment. That is, without treatment, the two groups would have maintained a similar path, i.e., exhibited parallel trends. A violation of this simple assumption may render the results of the DiD meaningless. Let θ represent the difference in slope of the various outcome variables between the treatment and control groups prior to T0, the initiation of intervention. Consider the following,

xiv ln ( y ) = α + ∑T β (t = j ∩ d = 1) +δ ( Post ) + θd t + γ + μ + ε (7) it j =T0 j i it i i t it where we add an interaction term, dit, to equation (5). Here, we are interested in the coefficient on this term, θ, in that we can test whether the differential slope is equal to zero. Notice that equation (5) naturally assumes θ is equal to zero, thus we have our hypothesis (H0: θ = 0). Per usual, if the p-value is greater than 0.05, we fail reject the null hypothesis and conclude that the trends are parallel. In theory, the results from the constrained equation are then fit for the standard interpretation.

Placebo Tests for the DiD Specification

We will in a sense consider two different placebo tests. The first will revert back to our RDiT specification. As previously mentioned, many papers compare RDiT and difference-in-difference estimates. This comparison is deemed useful when an untreated group exists but its validity as a control is in doubt. As our first placebo test, we will use the following model:

ln ( y ) = α + π ( Post ) + γ + μ + ε ∀ d = 0 (8) it it i t it i

The interpretation is similar to that of equation (4). When restricting the data to the control group, omitting all observations where di = 1, we allow π to capture the change in outcome in a sample that should be unaffected by the short-term rental policy within Cincinnati. As a second placebo test, we will adjust our time threshold, T0. Recall that T0 represents the date that the policy was passed, or the date intervention begins. We will again restrict our sample, though this time omit all observations where t > T0 and create a new intervention date where T0* < T0. In this case we will use the two other short-term rental policy related dates, March and October of 2018. The model is as follows:

T0-1 ln ( y ) = α + ∑ * ρ (t = j ∩ di = 1) +δ ( Post ) + γ + μ + ε (9) it j =T0 j it i t it

Similar to before, given the null-hypothesis H0: ρ = 0, if

1 푇0−1 ρ = ∗ ∑j =푇∗ ρj (10) 푇0− 푇0 −1 0

is not statistically significant, with a p-value greater than 0.05, we may again assume parallel trends and theoretically valid results.

Alternative Event-Study Design

xv In situations where non-constant treatment effects are expected, Abraham and Sum (2018) and Goodman-Bacon (2019) suggest carrying out an event-study. This approach is similar to the RDiT specification, though it allows for varying post- treatment effects in that we can measure the differential in different time periods relative to intervention. As opposed to the above DiD specification, Abraham and Sum (2018) propose a simple causal parameter that adequately accommodates treatment effect heterogeneity and allows for the missing covariates in our initial conditioning set X. Let t once more index time from {1, 2, ..., T} and T0 represent the initial intervention period. Consider the following,

Y = ∑ λ ( t - T ∈ g) + γ + μ + ε (11) it g ∈ G g 0 i t it

it where Y is the outcome for unit i at time t and γi and μt are individual and time fixed effects, respectively. Elements g ∈ G are a disjoint set of relative-period periods as defined by Abraham and Sum (2018), with which we have eight. The period g = - 6 encompasses all periods prior and period g = -1 is omitted and normalized to zero as policy is introduced. All period differentials are relative to period g = -1. The term λg represents our coefficients of interest. When applying this technique, there are ideally variations in treatment timing among groups. Due to data limitations, we are left with a single intervention. We include equation (11) estimates in the final stages of the paper but believe that future work should introduce additional metros, both treatment and control, with different short-term rental policy introduction dates.

Chapter 5: Results

In this section we present results from the specifications outlined above. We begin with the RDiT model and transition into the DiD model. We then shed light on the parallel trends assumption, the placebo tests, and present our event-study findings.

Regression Discontinuity in Time (RDiT)

Figures 4-13 as described in the Chapter 3 not only provide insight but implore additional study. When we restrict the data and omit all observations for the control group, we find graphically that the number of listings, percent of multi-listers, and percent of entire unit listings exhibit peculiar behavior around the threshold (Figures 4-8). After intervention, all three variables move downward in a manner distinct from previous patterns. As expected, the number of bookings appears to remain unaffected, suggesting the policy targets supply rather than demand. The number of evictions, however, decreases slightly, but no significant effect can be confidently inferred relying only on the graphs. No variables visually break trend when only treatment observations are omitted (Figures 9-13). We now consider equation (4), first letting i indicate zip code and t indicate the month-year. Table 4 summarizes these results. Note that here, the analysis encompasses the entire study period. There are four statistically significant coefficients when examining the treatment group (column one). These include the number of listings, number of multi-

xvi listers, number of entire unit listings, and booking price, all of which are log linearized. Between the pre- and post-intervention means at the zip code level, this specification suggests that there was an approximate 15 percent decrease in the average number of listings, a 4 percent decrease in the average number of multi-listers, a 20 percent decrease in the average number of entire-unit listings, and a 15 percent decrease in average booking price. All are significant at the 1 percent level and signed as expected, with the exception of booking price, which is significant at the 10 percent level. Restricting the sample to observations from only the control group in column two yields no statistically significant coefficients. Though not significant, the coefficient on the logged number of evictions is negatively signed and is far greater in magnitude than its treatment specific counterpart. Though this appears to run perpendicular to our hypothesis, any meaningful conclusion here will require further study. Table 4 also includes four additional covariates in an attempt to explain away exogenous discontinuities that may impact the variables of interest. These include the FRED ECI, unemployment rate, rent, and the FHFA HPI. Though not controlled for directly, in columns four through six we show that any differences in pre- and post- intervention means for these four variables are insignificant. As an additional robustness check, we next adjust the pre- and post-intervention observation bandwidths. Recalling equation (4), we again let i indicate zip code and t indicate month-year. The choice of bandwidth is outlined in Chapter 4 and results for this specification are summarized in Table 5. There are now five statistically significant coefficients when examining the treatment group in column one. This is similar to Table 4 with the addition of a significant and positive coefficient on logged bookings. Between the pre- and post-intervention means at the zip code level, this specification suggests that there was an approximate 5 percent decrease in the average number of listings, a 2 percent decrease in the average number of multi-listers, a 8 percent decrease in the average number of entire-unit listings, and a 12 percent decrease in average booking price. Though the magnitudes decrease, the ordering and significance is consistent. Once we restrict the sample to observations from only the control group in column two, the only significant coefficient is that on logged bookings, which is again positive. The coefficient on evictions is yet again negative, however, in this specification is much smaller in magnitude. Again, we find no significance in any of the four additional covariates. In both of the above specifications we find that the policy appears to have generated the expected variation in the appropriate Airbnb related variables, though its impact on evictions is unclear. That said, additional analysis is necessary if we are to build a causal case, or lack thereof, between these two variables. To ensure robust results, we again leverage equation (4) but now utilize the daily observations. We let i indicate zip code and t indicate the day. The sample is restricted to 120 days both pre- and post-intervention and results are presented in Table 6. Columns one and two summarize results for Cincinnati. Column two includes an additional day-of-week (DOW) fixed effect. Only logged listings, bookings, and asking price are significant across the two specifications with consistent signs. Listing behavior is again as expected, decreasing between its pre- and post-intervention zip code-day means, while logged bookings is significant and positive. Unique to column one, there is a significant and negative coefficient on logged evictions. This also carries over into the control group specification, however, again running perpendicular to our hypothesis that a significant

xvii drop in evictions would be observed exclusively in Cincinnati following policy intervention. As for the remaining variables in the control group specification, only logged bookings is significant and is again positive. Results from the three RDiT specifications exhibit minimal evidence that the decrease in Airbnb related activity had any significant effect on evictions in Cincinnati given the data at our disposal. In each, we find that the appropriate Airbnb related variables appear to respond to the policy as expected, generating the necessary variation for the conversion mechanism to take hold. Evictions, however, exhibit no strong response. This may stem from some additional omitted variables or too few observations following intervention. In either case, we proceed with our results from the DiD model to strengthen our case and check for consistency.

Difference-In-Differences (DiD)

We first test for pre-treatment differences in outcomes between treatment and control groups in Table 7. Column one presents full-sample averages while columns two and three present group specific means. Columns four and five display pre-enforcement differences along with their test statistics using month-year fixed effects and zip code-month-year fixed effects, respectively. As expected, each variable is statistically different at the 1 percent level with the exception of the percent of multi-listers and the percent of entire-unit listings. These are insignificant when we control for zip code-month-year fixed effects. These differences do not mean, however, that the two group cannot be compared. The DiD methodology remains valid as long there is no significant difference in pre-enforcement trends. Recall equation (7) from Chapter 4. Testing for a significant θ will shed light on whether or not we can move forward with our estimates. Table 8 presents these results in four columns across varying fixed effects which are indicated at the bottom of the table. We will proceed with interpretations of our preferred zip code-year-month fixed effect specifications. Three coefficients are significant in column four. These include logged bookings, entire-unit listings, and asking price. The pre-trend differences in booking behavior among groups is significant at the 1 percent level. This is not surprising given that one would expect demand to develop faster within the city. The positive direction and relatively small magnitude of this coefficient is not particularly alarming. The coefficient on the logged number of entire-unit listings is significant at the 10 percent level and small in magnitude. Here, we have an approximate 1 percent positive difference in the pre-trend differential among groups. It is unlikely that this would erode all validity in our DiD estimates for this particular outcome, or its potential effects on evictions. Lastly, logged asking price is significant at the 1 percent level. This is less of a concern given that there is no clear-cut theoretical prediction for the price outcomes. With these results in mind, we will move forward with our estimates. We now consider equation (5) and utilize the monthly aggregated data in an attempt to estimate β, the DiD estimator. These results are summarized in Table 9 and seven variations of fixed effects are indicated at the bottom of the table. We will proceed by interpreting the zip code-year-month fixed effects specification (column 7). First notice that the coefficient on logged listings is negative and significant at the 1 percent level. It

xviii suggests that there is an additional 17 percent decline in listings in Cincinnati relative to the remainder of Hamilton County following intervention. This result is similar to that of the RDiT treatment-specific specification. The coefficients on the logged number of bookings and multi-listers are not significant, though both are negative. The coefficient on the log of entire-unit listings is highly significant at the 1 percent level, is negative, and large in magnitude. This estimate suggests there is an additional 20 percent decline in the number of entire-unit listings in Cincinnati relative to the control group, once again similar to its RDiT counterpart. Though our estimates realize the expected variation in the Airbnb related variables, there is no significant variation in evictions. Still note that the estimate is negative and would suggest that there was an additional 5 percent decrease in evictions in Cincinnati relative to the control group following policy intervention. We again consider equation (5), but now utilize the non-aggregated daily observations. These results are summarized in Table 10 and the seven variations of fixed effects are once more indicated at the bottom of the table. We proceed with our preferred specification in column 7 which controls for zip code-month-year fixed effects. The coefficients on logged listings and entire-unit listings are both significant at the 1 percent level, negative, and nearly identical in magnitude to their coefficients in the zip code-month specification above. The coefficients on the logged number of multi-listers and the two price variables are again insignificant and carry the same sign. Specific to the non- aggregated daily data, however, the logged number of evictions estimate is slightly significant at the 10 percent level. The coefficient suggests that there is an additional 1 percent decline in evictions in Cincinnati relative to the control group after policy intervention. Though we again estimate the expected variation in the Airbnb related variables, there is no convincingly significant variation in evictions. In each of the above DiD and RDiT specifications, the logged number of evictions has been negative and similar in magnitude with only one producing a significant result at the 10 percent level. For robustness, consider a placebo test where we now analyze equation (9) using a different policy threshold, the October council meeting when a more extreme version of the policy from Councilman Mann was first introduced (see Chapter 3). Table 11 summarizes these results in the same manner as the previous DiD specification. Only three variables’ coefficients exhibit any semblance of significance. The first of these is the estimate on the logged number of listings. This coefficient is small in magnitude, only slightly significant at the 10 percent level, and positive. For these reasons, this estimate raises little concern and further supports our previous estimations. The logged number of entire-unit listings is significant at the 5 percent level but is small in magnitude and positive. Last, log booking price is slightly negative and only significant at the 10 percent level. These results all serve to support the robustness of our previous DiD estimates. We now draw our attention to equation (11) and the alternative event-study design so that we may visualize treatment effect heterogeneity across time. The main estimations are displayed in Figures 14-18. These figures present the time-disaggregated estimated effect of the Cincinnati STR policy on the natural log of our five key variables. This approach interacts a binary treatment indicator with months relative to enforcement. For example, -1 on the x-axis represents the month of April while 0 represents the initial month after policy enforcement. The vertical lines associated with each point estimate indicate the 95 percent confidence intervals. We include zip code-month-year fixed effects.

xix As expected, we realize variation in the number of listings, number of multi-listers, and number of entire-unit listings, all of which decline relative to our normalized month of April. Recalling our upward trend in the number of bookings from our previous RDiT and DiD specification, this downward shift in the number of listings should show through in a generated bookings per listing variable. Figure 17 confirms this result. Visually, we do notice a steady decline in the number of evictions on a relatively small scale. This supports previous results in that the coefficients on evictions in Cincinnati are negative, though generally insignificant and smaller in magnitude. Notice that the confidence intervals for Figures 15 and 17 are relatively less consistent than those of the other variables. We believe this is capturing the cyclical nature of listing and booking behaviors. In December, for example, the estimates are rather precise given the increase in holiday activity. Figures 19-21 further break down the time-disaggregated effects for three key variables originally presented in Figures 14-18. This break-out highlights the different behaviors for evictions, listings, and the ratio of entire unit listings between Cincinnati and non-Cincinnati zip codes. Note that Figure 19 presents the percent change in listings between the treatment and control groups after policy enforcement across months relative to April of 2019. Figure 20 presents the differences in entire unit listings and Figure 21 the differences in evictions. Distinct differences in the ratio of multi-listers are negligible, as the full sample drops in unison just after policy enforcement, thus, this figure is not presented.

Chapter 6: Conclusion

The Cincinnati city council voted in April 2019 to enact a short-term rental policy that promotes more responsible use of Airbnb. The idea was to pass legislation that balanced the risks - less- reliable tax payments to cities, a contribution to wealth inequality, diminished housing affordability, and the exacerbation of housing shortages across the country - and economic benefits associated with the platform. In this paper, we leverage the April 2019 policy intervention and study the effects of the Airbnb market on evictions in Cincinnati, OH utilizing both RDiT and DiD specifications. We found that the policy led to a subsequent decrease in the total number of listings, multi- listers, and the percent of entire-unit listings. In the monthly RDiT specification with zip code- month-year fixed effects, we estimate an approximate 15 percent decrease in the number of listings, a 4 percent decrease in the number of multi-listers, and a 20 percent decrease in the number of entire-unit listings between the pre- and post-intervention means at the zip code level. Furthermore, the monthly DiD specification with zip code-month-year fixed effects suggests that there is an additional 17 percent decline in total listings and a 20 percent decline in entire-unit listings in Cincinnati relative to the remainder of Hamilton County in the period following policy intervention. In this model, the coefficient on multi-listers is insignificant, but suggests that the variable decreases an additional 2 percent in Cincinnati relative to the control group, again similar to the RDiT specification. The daily DiD zip code-month-year fixed effects specification produces nearly identical results to its month-level counterpart. The model suggests there is an additional 17 percent decline in total listings and a 24 percent decline in entire-unit listings in Cincinnati relative to the remainder of Hamilton County following policy intervention. If property owners were evicting their long-term tenants to list their properties on Airbnb on a grand scale, one would expect that this dip in Airbnb activity after April 2019 would lead to

xx significantly fewer evictions. However, we found no significant variation in the number evictions during this time period in either specification. Therefore, there is little evidence of a significant relationship between Airbnb booking and listings behaviors and evictions in Cincinnati, OH given the data at our disposal. Given that the post-intervention is relatively short, there may be a lag in the conversion mechanism that we are simply not picking up. We suggest that future studies obtain more post-treatment observations as well as data for additional metros, both treatment and control. A flexible event-study design is better suited for heterogeneity in policy intervention timing and could produce more robust results, especially after more time has passed following intervention.

References

Abraham, S., & Sun, L. (2018). Estimating Dynamic Treatment Effects in Event Studies With

xxi Heterogeneous Treatment Effects. SSRN Electronic Journal.

Barron, K., Kung, E., and Proserpio, D. (2020). The sharing economy and housing affordability: Evidence from Airbnb. SSRN Working Paper.

Bibler, A., Teltser, K., and Tremblay, M. (2019). Inferring tax compliance from pass-through: Evidence from Airbnb tax enforcement agreements. SSRN Working Paper.

Bibler, A., Teltser, K., and Tremblay, M. (2020). Is Sharing Really Caring? The Effect of Airbnb on the Affordability of Housing. SSRN Working Paper.

Bilinski, A., & Hatfield, L. (2018). Nothing to see here? Non-inferiority approaches to parallel trends and other model assumptions. Cornell University, arxiv.org. 2020.

Burger, N. E., Kaffine, D. T., & Yu, B. (2014). Did California’s hand-held cell phone ban reduce accidents? Transportation Research Part A: Policy and Practice, 66, 162-172.

Calder-Wang, S. (2019). The distributional impact of the sharing economy on the housing market. Working Paper.

Chen, W., Wei, Z., and Xie, K. (2019). The battle for homes: How does home sharing disrupt local residential markets? SSRN Working Paper.

Cosman, J. and Quintero, L. (2019). Fewer players, fewer homes: concentration and the new dynamics of housing supply. Working Paper.

Diamond, R., McQuade, T., and Qian, F. (2019). The effects of rent control expansion on tenants, landlords, and inequality: Evidence from San Francisco. American Economic Review, 109(9):3365-94.

Farronato, C. and Fradkin, A. (2018). The welfare effects of peer entry in the accommodation market: The case of Airbnb. NBER Working Paper.

Horn, K., & Merante, M. (2017). Is home sharing driving up rents? Evidence from Airbnb in Boston. Journal of Housing Economics, 38, 14-24.

Jefferson-Jones, J. (2015). Can short-term rental arrangements increase home values? A case for Airbnb and other home sharing arrangements. Cornell Real Review, 13(1), 12-19.

Khouri, A. (2019). Airbnb pledges $25 million to support affordable housing and small business. L.A. Times. Retrieved from: https://www.latimes.com/business/story/2019-09-17/airbnb-pledges-25-million-to- support-affordable-housing-and-small-business

Lee, D. (2016) How Airbnb Short-Term Rentals Exacerbate Los Angeles’s Affordable Housing Crisis: Analysis and Policy Recommendations. Harvard & Policy

xxii Review 10: 229-253.

Maldonado, C. (2019). 'This Airbnb displaced 5 people': Here's the story behind that photo that spread on Facebook. The Lens. Retrieved from: https://thelensnola.org/2018/02/10/this-airbnb-displaced-5-people-heres-the-story- behind-that-photo-that-spread-on-facebook/

Samaan, R. (2015). Airbnb, Rising Rent, and The Housing Crisis in Los Angeles. FTC.gov.

Zervas, G., Proserpio, D., and Byers, J. W. (2017). The rise of the sharing economy: Estimating the impact of airbnb on the hotel industry. Journal of Marketing Research, 54(5):687-705.

Tables

xxiii

Table 1: Full Sample Summary Statistics by Zip-Code Month

Note : This table presents full-sample summary statistics for Hamilton, County between October 2014 and August 2019. Statistics are static and include the mean, standard deviation, 25th percentile, median, and 75th percentile. Each variable has 3,422 observations and statistics are aggregated at the zip code-month level.

xxiv

Table 2: Aggregated Annual Trends

Note: This table presents annual trends for Hamilton County between 2014 and 2019, with the initial and terminal years being incomplete. The year 2014 contains three months of observations while 2019 contains eight months and thus must be interpreted carefully. Panel A represents the aggregated trends for all of Hamilton County. Panel B represents annual trends for the area within Hamilton County but outside of the Cincinnati, Ohio city boundary. Panel C represents annual trends for Cincinnati, Ohio. Where appropriate, the annual percent change is listed for each variable.

xxv

Table 3: Comparative Summary Statistics (2019)

Note: This table presents summary statistics for 2019 in an attempt to determine whether Cincinnati and Hamilton County are relevant and useful in our analysis. Panel A displays statistics regarding area demographics, Panel B compares educational statistics, and Panel C summarizes the appropriate income, labor, and housing information. The areas of interest are Cincinnati and Hamilton County, Cleveland and Cuyahoga County, Pittsburgh and Allegheny County, and the United States. As of 2019, consider that Cincinnati is the 64th largest city in the U.S, Cleveland the 53rd, and Pittsburgh the 66th.

xxvi

Table 4: Reduced-Form Linear RDiT Specification (All Observations)

Note: This table presents results from the local linear regression discontinuity in time specification for all monthly observations. The first three columns display the coefficients on a time indexed policy enforcement dummy regressed onto the various natural logs of the listed variables. The coefficient should identify the treatment effect of the Cincinnati STR policy. Leveraging panel data, observations were broken up into Cincinnati and non-Cincinnati specific zip codes. Column 3 displays the results for the full sample. Rather than (1) shrinking the sample size and (2) using less granular data, additional covariates were handled in a similar manner to test for discontinuity around the policy date. Various fixed effects are indicated at the bottom of the table. Standard errors are robust to clustering at the zip code level.

xxvii Table 5: Reduced-Form Linear RDiT Specification ( -Eight, +Four Month Bandwidths)

Note: This table presents results from the local linear regression discontinuity in time specification using monthly data and bandwidths of four months. The first three columns display the coefficients on a time indexed policy enforcement dummy regressed onto the various natural logs of the listed variables. The coefficient should identify the treatment effect of the Cincinnati STR policy. Leveraging panel data, observations were broken up into Cincinnati and non-Cincinnati specific zip codes. Column 3 displays the results for the full sample. Rather than (1) shrinking the sample size and (2) using less granular data, additional covariates were handled in a similar manner to test for discontinuity around the policy date. Various fixed effects are indicated at the bottom of the table. Standard errors are robust to clustering at the zip code level.

xxviii Table 6: Reduced-Form Linear RDiT (120 Day Bandwidths)

Note: This table presents results from the local linear regression discontinuity in time specification using daily data and bandwidths of 120 days. The first three columns display the coefficients on a time indexed policy enforcement dummy regressed onto the various natural logs of the listed variables. The coefficient should identify the treatment effect of the Cincinnati STR policy. Leveraging panel data, observations were broken up into Cincinnati and non-Cincinnati specific zip codes. Column 3 displays the results for the full sample. Various fixed effects are indicated at the bottom of the table. Standard errors are robust to clustering at the zip code level.

xxix Table 7: Pre-Enforcement Differences in Outcomes

Note: This table presents the pre-enforcement differences in outcomes. The first three columns display the mean and standard deviations for the full, treated, and untreated samples. The last two columns display the statistical differences be tween treated and untreated zip codes in pre-enforcement months. In other words, these columns test whether or not being in a treated zip code is correlated with outcomes in pre-policy months. All regressions included in Table 7 are restricted to observations prior to April 25th, 2019.

xxx Table 8: Parallel Trends - DiD Specification (Zip Code/ Month)

Note : This table presents estimates for 휃 from equation (5). Here we attempt to defend the DiD parallel trends assumption in equation (3) by presenting whether or not the inclusion of the time-trend and treatment interaction term is statistically significant.

xxxi Table 9: DiD Specification (Zip Code/ Month)

Note: This table presents regressions of the natural log of evictions (Panel A), listings (Panel B), bookings (Panel C), multi-listers (Panel D), and entire-unit listings (Panel E) on a dummy variable for treatment and a difference-in-differences indicator (Policy x Post). The coefficient on the DiD indicator is reported above in Table 9. Panels F and G repeat the process for the average asking price (Panel F) and average booking price (Panel G). The observations are aggregated at the zip code-month level. The seven variations of fixed effects are indicated at the bottom of the table. Standard errors are robust to clustering at the zip code level.

xxxii Table 10: DiD Specification (Zip Code/ Day)

Note: This table presents regressions of the natural log of evictions (Panel A), listings (Panel B), bookings (Panel C), multi-listers (Panel D), and entire-unit listings (Panel E) on a dummy variable for treatment and a difference-in-differences indicator (Policy x Post). The coefficient on the DiD indicator is reported above in Table 10. Panels F and G repeat the process for the average asking price (Panel F) and average booking price (Panel G). The observations are aggregated at the zip code-day level. The seven variations of fixed effects are indicated at the bottom of the table. Standard errors are robust to clustering at the zip code level.

xxxiii Table 11: DiD Specification (Zip Code/ Month - October 2018 Threshold)

Note: This table presents regressions of the natural log of evictions (Panel A), listings (Panel B), bookings (Panel C), multi-listers (Panel D), and entire-unit listings (Panel E) on a dummy variable for treatment and a difference-in-differences indicator (Policy x Post) using the October 2018 policy introduction date. The coefficient on the DiD indicator is reported above in Table 10. Panels F and G repeat the process for the average asking price (Panel F) and average booking price (Panel G). The observations are aggregated at the zip code-day level. The seven variations of fixed effects are indicated at the bottom of the table. Standard errors are robust to clustering at the zip code level.

xxxiv Figures

Figure 1: Hamilton County Zip Codes and Boundaries

Cincinnati

xxxv

Figure 2: Airbnb Listings Locations Across Hamilton County

Cincinnati

Airbnb Listing Location

Figure 3: Evictions Locations Across Hamilton County

Cincinnati

Eviction Location

Note: Figures 1-3 display various snapshots of the study area. In Figure 1 we display three key features of the study area, (1) the Hamilton County boundary - dark green, (2) Hamilton County zip codes, and (3) the Cincinnati city boundary - red. Figure 2 displays all Airbnb listing locations that were active at some point in Hamilton County between October 2014 and August 2019. Figure 3 displays the location of all evictions that were filled in Hamilton County between October 2014 and August 2019. In both Figures 2 and 3, the multi-colored areas within the county boundary represent cities with a population greater than 10,000. Figures 1-3 were generated in ArcGIS Pro.

xxxvi

Figures 4-8: Key Variable Trends in Treatment Zip Codes for RDiT

Figure 4 Figure 5

Figure 6 Figure 7

Note: These figures present variable trends for Cincinnati specific zip codes across Hamilton County, Ohio between January 2017 and August 2019. The data are broken up into four-month bins and averaged across the four months. The Cincinnati STR policy was passed in the final month of the second to last bin and is represented by the vertical line. Note that the scales vary between the figures. Figures 4-8 help readers to visualize results from the RDiT Specifications.

Figure 8 xxxvii

Figures 9-13: Key Variable Trends in Control Zip Codes for RDiT

Figure 9 Figure 10

Figure 11 Figure 12

Note: These figures present variable trends for non-Cincinnati specific zip codes across Hamilton County, Ohio between January 2017 and August 2019. The data are broken up into four-month bins and averaged across the four months. The Cincinnati STR policy was passed in the final month of the second to last bin and is represented by the vertical line. Note that the scales vary between the figures. Figures 9-13 help readers to visualize results from the RDiT specifications.

Figure 13 xxxviii Figures 14-18: Event-Study Graphs

Figure 14 Figure 15

Figure 16 Figure 17

Note: These figures present the time- disaggregated estimated effect of the Cincinnati STR policy on the natural log of our five key variables. This approach interacts a binary treatment indicator with months relative to enforcement. For example, -1 on the x-axis represents the month of April while 0 represents the first month after policy enforcement. The analysis includes zip code and month-year fixed effects. Standard errors are robust to clustering at the zip code level.

Figure 18

xxxix Figure 19: Number of Listings Time-Disaggregated Effects by Treatment

Figure 20: Entire-Unit Ratio Time-Disaggregated Effects by Treatment

xl

Figure 21: Evictions Time-Disaggregated Effects by Treatment

Note: Figures 19-21 further break down the time-disaggregated effects for three key variables originally presented in Figures 14-18. This separation highlights the different behaviors for evictions, listings, and the ratio of entire unit listings between Cincinnati and non-Cincinnati zip codes. Note that Figure 19 presents the percent change in listings between the treatment and control groups after policy enforcement across months relative to April of 2019. Figure 20 presents the differences in entire unit listings and Figure 21 the differences in evictions. Distinct differences in the ratio of multi-listers are negligible, as the full sample drops just after policy enforcement.

xli