Memphis Housing Security Study

We are assisting the City of Memphis/Shelby County in administering approximately $60 million in emergency rental assistance provided under the Consolidated Appropriations Act of 2021. The City and County are randomizing whether applicant households receive assistance through a weighted lottery and, in one arm of our analysis, randomizing the assistance packages offered to households and landlords. At this time, applications are open but treatments have not been assigned.

We have several primary treatment arms:

- Pure control.
- Provision of back rent and utilities. Households will receive 100% repayment of back rent and utilities. Households are eligible for assistance on arrearages accumulated within twelve months of their application date.
- A second group of households will, in addition to back rent and utilities, also receive an offer made to their landlord on their behalf for future rent assistance. In particular, among the households who randomly receive back rent and utilities, we will randomly assign individuals to different rental assistance packages. Not all households who receive back rent will receive a future rent offer. The share of tenants who receive a future rent offer depends on the number of applications and a forthcoming City/County determination about the share of funds to be allocated to future rent. Tenants may receive up to three months of future rent. Packages are defined as shares to be contributed by the government, landlord, and tenant. For example, one package may specify that the contribution toward the next months of future rent will be: (government, landlord, tenant) = (20%, 80%, 0%), whereas another package may specify (government, landlord, tenant) = (30%, 20%, 50%).

Households that receive a future rent offer will receive this offer in addition to a back rent offer. No households receive future rent while having a (reported) rental debt. Therefore, the future rent treatment will identify the effect of receiving future rent, conditional on also receiving back rent.

Landlord contribution rate. We will test several levels of landlord write-offs from small (e.g., 10%) to large (e.g., 40%). For example, if a tenant owes $2,000 in rent for the next three months, a 10% write-off would require the landlord to forgive $200 in exchange for a government contribution. This arm will help us understand how much we can “stretch” the city’s funds further by imposing landlord write-offs.

Tenant contribution rate. We will hold the landlord’s write-off fixed but will test several levels of tenant contributions to future rent owed. The tenant contribution will range from very small (e.g., 0%) to requiring tenants to pay half of rent. This arm will help us understand how much we can “stretch” the city’s funds further by reducing, rather than fully relieving, tenants’ future rent burdens. We will randomize how much the tenant will be asked to contribute.

The precise rental assistance packages depend on the number and demographic composition of applicants, and are subject to final approval from the City/County. We therefore do not specify these here and will file an amendment at the appropriate time.

Randomization. We describe the randomization in more depth below. In waves of two to three weeks, we will randomly select a group of applicants for back and future rent. Applications for the first wave have opened. The first randomization will take place in approximately one week.

Description of lottery weights. The City/County has determined that it will conduct a “weighted” lottery, in which particularly at-risk groups receive a higher chance of treatment. The first randomization has not yet begun, and the City/County has determined that it will choose the final weights and weighting criteria once it sees the final number and composition of applicants. These weights may change in each wave; e.g., if one wave has an especially large number of weighted applicants, that may affect the weights in that wave.

Such groups may include: people with very low or low income (<30% or <50% AMI), people with disabilities, seniors, unemployed people, veterans, and households with 3 or more children. These weights may vary in each randomization. Our analysis will account for heterogeneity in treatment probabilities across households.

There may be an additional cross-randomization of provision of social services (e.g., housing counseling through a local non-profit). This cross-randomization has not yet been established. Such social services may be given to the control group.

2021-03-15

2021-12-31

Landlord take-up of assistance offer: The time period for measuring this outcome is determined by the window allowed to landlords to decide whether or not to accept. Landlord take-up will occur within roughly one month of the offer date.

Eviction filings and formal eviction judgments: We will examine weekly filings and formal eviction judgments.

We will study the number of evictions within one to twelve months after, as well as two to five years after, the date of the initial wave in which the tenant applies.

We will directly measure tenant/landlord take-up, and we will collect eviction outcomes from local housing court filings.

We have tentatively designated these as our primary outcomes not because they are necessarily the most important but because we know that we will be able to measure them. It is not necessarily the case that our treatments must affect evictions for the treatment to affect our secondary outcomes. That is, our intention in specifying these as primary outcomes is not to indicate that eviction prevention is the key mediator for causal effects on the secondary outcomes.

We also intend to study the impacts of rental assistance on several other economic outcomes. Each of these groups of outcomes could become primary outcomes if we can link them. We have identified the outcomes of interest but have not yet been able to determine whether we will be able to access and merge administrative records held by the state government, local government, or private companies to be able to study them. Once we have secured the appropriate data access, we may consider some of these to be primary tenant outcomes.

- Housing security: number of moves, characteristics of any new residential neighborhoods moved to
- Labor market outcomes: employment, wage income, hourly wages, weekly hours (as measured in Tennessee UI records).
- Transfer benefit receipt (as measured in Tennessee administrative data on, e.g., SNAP)
- Financial outcomes (as measured from credit scores from credit reporting agencies). Health and hospitalization records
- Homeless records: from the Homelessness Management Information System

We will study these outcomes within one to twelve months after, as well as two to five years after, the date of the initial wave in which the tenant applies. If surveys are administered, we anticipate that they will be during the year of treatment (i.e., 2021), but we leave open the possibility of longer-run survey outcomes.

We may also study the following landlord outcomes:
- Financial outcomes (as measured from credit reporting agencies)

Pending funding, IRB approval, and collaboration with our government and nonprofit partners, we may also administer tenant and landlord surveys to measure outcomes such as subjective probability of eviction, mental health, neighborhood and housing satisfaction, etc. These outcomes are more speculative, and we will amend the pre-registration before analyzing such data.

We will provide a more precise definition of these outcome variables in an amendment once data availability has been determined.

Tenants apply via an online portal. After several weeks of applications, they are assigned to a treatment group. Those who are not treated are automatically re-entered in the lottery. We will run multiple waves of the lottery. Using information collected at the application, we can link tenants’ outcomes to public data on evictions records and administrative datasets from governments or private companies, and follow these outcomes over time.

The City/County anticipate allocating approximately $2-3 million per lottery “wave” of applications on paying back rent. Each wave will be open for approximately two weeks of applications. Funding amounts and application windows may be revised according to funding availability and the volume of applications.

Some applicants will not be entered into the lottery, based on a determination by the City/County that they are particularly at risk. At this time, having a pending eviction filing or utility shut-off notice is a risk factor that would cause households to not be entered into the lottery and to receive immediate assistance on back rent. Similarly, the City/County is conducting bulk negotiations with large landlords to clear back rent from all delinquent tenants. Households excluded from the back rent lottery, who do receive back rent, may be entered into the future rent lottery.

After a household wins the back rent lottery, screeners hired by the City/County will confirm eligibility. Households who are not eligible may be contacted by the City/County to update their eligibility materials or removed from the lottery. They may then be re-entered in the lottery for a subsequent wave. Therefore, because winning a lottery does not perfectly determine assignment, we will use an instrumental variables design, wherein winning a lottery is an instrument for ever receiving assistance.

Households that apply but do not win the back-rent lottery will automatically be re-enrolled in a subsequent lottery. This means that people who apply early are very likely to eventually receive assistance, since they will have many chances to win a lottery. Insofar as early and late applicants may differ on observable and unobservable demographic characteristics likely to be related to our outcomes of interest, our primary specification will adjust for treatment probability, e.g. by including calendar time fixed effects for the “wave” of the lottery in which they apply. Within each wave, treatment status is randomly assigned and unrelated to observable and unobservable applicant characteristics. Because different demographic groups have different odds of winning the lottery (because of the lottery weights, explained above), primary specifications may additionally include fixed effects for these demographic groups or will be appropriately re-weighted.

At this time, we do not intend to automatically re-enroll households who lose the future rent lottery into the future rent lottery.

We first send landlords the offer for back rent. Among landlords who accept the back rent offer, we will additionally enter those tenants who indicated they will have trouble paying future rent into a weighted lottery for the future rent offers. The future rent lottery will take place several weeks after the initial back rent lottery. As above, note that any household that receives back rent (including people excluded from the back rent lottery) can be entered into the future rent lottery.

The City/County also anticipate spending approximately $500,000 to $1 million per lottery “wave” of applications on paying future rent. Therefore, approximately 20–50% of applicants who indicate they are eligible for future rent will receive a future rent offer.

Primary specification. The primary specification will use IV to assess the impact of receiving back or future rent on the desired outcome, using the lottery assignment as an instrument for receiving assistance. We will adjust for differences in treatment probability (which could vary by application wave and//or because some demographic groups will receive priority weights). In our primary or secondary specification, we may include baseline covariates selected using a robust selection procedure (e.g., double/debiased machine learning) for power.

The exclusion restriction is that the only way lottery assignment affects outcomes is via receiving back or future rent assistance. A possible confound (which we do not view as especially likely) is that receiving an offer itself changes outcomes, even if aid is not delivered. For instance, landlords who decline aid may feel guilty and forgive some (but not all) rent. For this reason, we will also present the reduced form of offering assistance on outcomes.

An alternative strategy to obtain the effect of future rent may use the specific future rent offer bundle to instrument for the landlord’s take-up decision. In particular, some landlords will randomly receive more generous future rent offers and (we conjecture) be more likely to take up. The random variation in the offer bundle can be used to instrument for the effect of future rent given take-up.

Given that households will be randomized into back rent and future rent, a specification that çaptures both forces could jointly instrument for the endogenous variables of receiving back and future rent relief, using the vector of back and future rent offers.

The effects on landlords can be measured in a similar way as for tenants. However, to study the effects on landlords, we may amend the primary specification as follows. Due to random selection, some landlords may have more tenants who receive aid. We can therefore run landlord-level regressions, using the random variation in the number or share of tenants who receive aid offers to instrument for aid receipt.

Secondary specification. Because people who enter the lottery early are very likely to win (given we will run multiple waves of the lottery), we may explore specifications where we consider the date of treatment to be randomly assigned. In these specifications, we may study the outcomes defined as of the date when they win the lottery. For instance, we may study whether households who win the lottery early have reductions in eviction filings or increases in credit scores earlier than those who win the lottery later. If almost all households ultimately win the lottery, the secondary specification may be promoted to the primary specification.

The unit of analysis in all specifications with tenant outcomes will be household-level. Because the level of random assignment is the household, we cluster standard errors for specifications with tenant outcomes at the household level.

Treatment effect heterogeneity. We are unsure of the demographic heterogeneity that we will be powered to explore, given the composition of applications. It is particularly interesting and natural to examine treatment effect heterogeneity by the size of the assistance offered. Therefore we intend to study:
- Treatment effects by back rent owed.
- Treatment effects by the rent to income ratio (or similar ratio).
- We may study whether tenant demographics affect the landlord’s propensity to accept rental assistance offers. Demographics that may affect the landlord’s propensity to accept include tenants’ gender, race, income, and age (and, for instance, whether the tenant’s demographics are the same or different from the landlord’s).
- Heterogeneity in the effects of receiving assistance by demographic characteristics is likely to be secondary. However, we may present treatment effects by gender, race, age, and household composition.

We intend to amend this discussion as data become available but before analysis.

Lottery conducted in Stata or similar statistical software by the investigators. The lottery will be conducted every few weeks (as determined by the City/County). Lottery weights will be determined by the City/County.

Household.

No

There is a budget of approximately $60 million, which the City/County intends to disperse in full. The number of treated and control households will depend on the number of applicants for rental assistance. Because the City/County plan to conduct a weighted lottery (where weights depend on household characteristics), the composition of applicants will also affect the number of treated and control households. Therefore, we cannot pre-register the number of clusters (or, equivalently, households).

Some households will receive assistance via a non-randomized “bulk settlement” process if their landlord has multiple tenants in need of rental assistance and requests funds for back rent via the bulk-settlement process. These households will be excluded from our study of back rents (though they may still be assigned to receive future rents). Other households with a pending eviction filing or utility shut-off notice will also be excluded from the lottery.

See above. The number of treated and control observations depends on the number and composition of applicants, which we cannot know at this stage.

See above.

We will file update in an amendment once we have more application data available.