Estimating Residential Displacement in the Central Puget Sound Region using Household Survey Data

This paper proposes a new methodology using a Bayesian spatiotemporal model and household survey data to estimate sub-county residential displacement rates in the Central Puget Sound Region, revealing geographic variations and a temporary moderation in displacement during 2020-2021.

The Gist

Imagine you are trying to understand why people are being forced out of their homes in the Seattle-Tacoma-Bellevue area. You want to know: How many people are being displaced? Where is it happening? And who is getting hit the hardest?

The problem is that nobody has a perfect map. It's like trying to draw a detailed picture of a storm using only a few scattered raindrops.

This paper is about how the authors built a super-powered magnifying glass to see the whole storm clearly, even when the data is messy and incomplete. Here is how they did it, explained simply:

1. The Three-Legged Stool (The Data Problem)

The authors needed three different types of information to get the job done, but each one had a major flaw on its own. They combined them like a sturdy three-legged stool:

• Leg 1: The Local Travel Survey (HTS). This is like a detailed diary kept by a small group of locals. It asks, "Why did you move?" and "Where did you go?"
  • The Flaw: Only a tiny fraction of people fill out this diary. In some neighborhoods, they only have a handful of answers. If you tried to guess the whole neighborhood's story based on just three people, you'd be guessing wildly.
• Leg 2: The Census Survey (ACS). This is like a massive, reliable headcount of the entire region. It tells you exactly how many people live in each neighborhood, their income, and their race.
  • The Flaw: It's great at counting people, but it doesn't ask why they moved. It's like knowing how many people are in a room, but not knowing if they left because they were happy or because they were kicked out.
• Leg 3: The Housing Survey (AHS). This is like a national report card that asks big cities, "What percentage of people were forced out?"
  • The Flaw: It gives a great answer for the whole city, but it's too blurry to tell you what's happening in specific small neighborhoods.

2. The Magic Recipe (The Method)

Since they couldn't rely on just one leg, they cooked up a recipe called Multilevel Regression with Poststratification (MRP). Think of this as a smart translator that turns the "diary" answers into a "headcount" map.

Here is the step-by-step process:

1. The Detective Work (The Model): They took the tiny "diary" (HTS) and taught a computer to look for patterns. They asked: "Do renters move more often than owners? Do poor families move more than rich ones? Does it happen more in the city center or the suburbs?" The computer built a rulebook based on these patterns.
2. The Translation (Poststratification): Now, they took that rulebook and applied it to the massive "headcount" (ACS). They said, "Okay, we know from our rulebook that renters in low-income areas are likely to be displaced. Since we know exactly how many renters live in Neighborhood X, we can estimate how many of them were displaced."
   • Analogy: Imagine you have a recipe for a cake that says "1 cup of flour makes 10 cookies." You don't have enough flour to bake the whole batch, but you know exactly how much flour the whole city has. You use the recipe to calculate how many cookies the whole city could make.
3. The Reality Check (Benchmarking): Finally, they compared their new, detailed map against the "national report card" (AHS). If their map said 50% of people were displaced, but the national report said 20%, they knew something was wrong. They adjusted their map until the total numbers matched the national report. This ensures their detailed map is accurate for the whole region.

3. What They Found (The Results)

Once they built this map, some interesting patterns emerged:

• The East-West Divide: There is a clear line. The West (closer to the city center and water) has high displacement rates. The East (suburbs and rural areas) has much lower rates. It's like a wave pushing people out of the city center and toward the edges.
• The Pandemic Pause: During 2020-2021 (the height of the pandemic), the displacement rate actually dropped. It's like the storm took a brief nap. This might be because of eviction bans or because people were too scared to move during uncertain times.
• Who Gets Hit Hardest:
  • Renters are much more likely to be displaced than homeowners (about twice as likely).
  • Low-income families are at the highest risk.
  • Large families are more likely to be pushed out than single people.
  • Interestingly, Race wasn't the biggest factor when you accounted for income and housing type. The main driver was money and whether you owned your home.

4. Why This Matters

Before this study, policymakers were flying blind. They knew displacement was happening, but they didn't know where to send help.

• Old Way: Guessing based on general trends or looking at "gentrifying" neighborhoods only.
• New Way: A high-resolution map that shows exactly which small neighborhoods are losing their residents.

This allows city planners to say, "We know Neighborhood A is losing 25% of its renters, so we need to put our housing assistance money there right now."

The Bottom Line

The authors took three imperfect sources of information, mixed them together with a clever statistical recipe, and created a clear, reliable map of who is being forced out of their homes in the Puget Sound region. It's a tool that turns a blurry, confusing picture into a sharp, actionable plan for helping vulnerable communities.

Technical Summary

Here is a detailed technical summary of the paper "Estimating Residential Displacement in the Central Puget Sound Region using Household Survey Data."

1. Problem Statement

Residential displacement is a critical housing stability issue, yet it lacks robust, granular measurement tools for local policymakers.

• Measurement Gap: Standard definitions (e.g., Grier and Grier, 1978) define displacement as forced moves due to uncontrollable conditions (e.g., rent hikes, eviction, demolition). However, major national datasets like the American Community Survey (ACS) identify movers but do not capture reasons for moving.
• Data Limitations: The American Housing Survey (AHS) asks about reasons for moving but only reports at the Metropolitan Statistical Area (MSA) level, lacking the sub-county granularity (Public Use Microdata Areas or PUMAs) required for local planning.
• Local Survey Constraints: The Puget Sound Regional Council Household Travel Survey (HTS) contains specific questions on why households moved and PUMA-level identifiers. However, the sample size of recent movers within specific PUMAs is often too small for direct statistical estimation, leading to high variance and unrepresentative results.

Goal: To estimate the proportion of displaced households in every PUMA within the Central Puget Sound Region (King, Pierce, and Snohomish counties) for four consecutive two-year move cohorts (2016–2023) with quantified uncertainty.

2. Methodology

The authors propose a three-stage Multilevel Regression with Poststratification (MRP) workflow integrated within a Bayesian framework. This approach synthesizes data from three sources: HTS (displacement indicators), ACS (population weights), and AHS (regional benchmarks).

Stage 1: Bayesian Spatiotemporal Logistic Regression

The authors model the log-odds of displacement ($Y_i$) for individual movers in the HTS data using a mixed-effects logistic regression:
$$\text{logit}(p_i) = \beta_0 + X_i^\top \beta + \alpha_{t[i]} + \omega_{s[i], t[i]} + \gamma_{w[i]}$$

• Fixed Effects ($X_i \beta$): Demographic and housing covariates including household size, income, vehicle ownership, children, tenure (rent/own), and race.
• Random Effects:
  • Temporal ($\alpha_t$): A first-order random walk to smooth trends across move cohorts.
  • Spatial ($\omega_{s,t}$): A BYM2 (Besag, York, and Mollié) model. This captures spatial dependence by pooling information between adjacent PUMAs (structured effect) while allowing for idiosyncratic heterogeneity (unstructured effect). This is crucial for handling PUMAs with zero or very few samples.
  • Survey Effect ($\gamma_w$): An independent random effect to account for systematic biases between HTS survey waves (2019, 2021, 2023).

Stage 2: Poststratification

To ensure estimates represent the actual population composition:

• The model generates displacement probabilities for 900 unique "cells" (combinations of covariates) within each PUMA and move cohort.
• These cell-level estimates are aggregated using weights derived from the ACS. The ACS provides the population count ($N_{j,s,t}$) for each cell within a PUMA.
• The final PUMA-level estimate is a weighted average of the cell probabilities, ensuring the result reflects the demographic reality of recent movers in that area.

Stage 3: Benchmarking

To ensure consistency with high-quality national data, the authors apply Bayesian benchmarking (Okonek and Wakefield, 2024):

• The MRP estimates are constrained to match the regional displacement rates derived directly from the AHS.
• A Gaussian likelihood term is added to the posterior distribution, penalizing deviations from the AHS aggregate estimates.
• A rejection sampler is used to filter posterior draws that are inconsistent with the AHS benchmarks, resulting in a refined set of posterior samples.

3. Key Contributions

1. Novel Data Integration: This is the first study to integrate HTS, ACS, and AHS data specifically for residential displacement estimation. It treats the three surveys as a "three-legged stool," where HTS provides local granularity, ACS provides representativeness, and AHS ensures regional consistency.
2. Methodological Advancement: The application of Spatiotemporal MRP with BYM2 spatial effects to a small-area displacement problem allows for robust estimation in areas with sparse data, overcoming the limitations of standard small-area estimation models.
3. Public Resource: The authors provide publicly available, high-resolution estimates (PUMA-level) and code, enabling stakeholders to monitor displacement trends over space and time.

4. Key Results

• Spatial Patterns: There is a distinct east-west gradient in displacement. Rates are significantly higher in the urban west (Seattle and immediate suburbs) and lower in the suburban/rural east. Within King County, southwestern PUMAs show elevated displacement, consistent with displacement from the urban core to the south.
• Temporal Trends:
  • 2020-2021 Cohort: A temporary moderation (flattening) of displacement rates was observed during the height of the pandemic. This may be attributed to eviction moratoriums, income support, or behavioral resistance to moving during turbulent times.
  • Uncertainty: The 2018-2019 and 2022-2023 cohorts showed higher uncertainty (wider credible intervals), likely due to sample size limitations and post-pandemic movement volatility.
• Covariate Effects (Odds Ratios):
  • Tenure: Renters have 2.17 times higher odds of displacement compared to owners.
  • Income: Lower-income households face significantly higher risks, with risk scaling inversely with income.
  • Household Size: Larger households (3+ people) are at higher risk than individuals or couples.
  • Race: When controlling for other variables, race showed limited association, with only the "Other" category showing a statistically significant increase in risk.
• Benchmarking Impact: Benchmarking against AHS data did not drastically shift the point estimates (medians) but significantly reduced the posterior standard deviations (increased precision), particularly for cohorts with high uncertainty.

5. Significance and Limitations

Significance:

• Policy Utility: Provides local officials with actionable, granular data to identify vulnerable neighborhoods and target interventions (e.g., tenant protections, affordable housing) without relying on proxy risk indices.
• Transferability: The methodology is transportable to other jurisdictions with similar data structures (local travel/housing surveys + ACS + AHS).
• Measurement: Moves beyond "risk of displacement" proxies to direct measurement of actual displacement events.

Limitations:

• Sample Size: The primary constraint remains the small number of movers in specific PUMAs, necessitating the heavy reliance on MRP and spatial smoothing.
• Definition of Displacement: The study relies on self-reported reasons for moving. It may miss "social displacement" (moving due to neighborhood cultural changes rather than financial coercion).
• Directionality: The data captures where displaced households arrive, not where they were displaced from, limiting the ability to trace specific origin-destination flows of displacement.
• Uncertainty Sources: The model does not explicitly propagate uncertainty from the ACS population estimates, potentially underestimating total uncertainty.

In conclusion, this paper establishes a rigorous statistical framework for estimating residential displacement at a sub-county level, offering a vital tool for evidence-based housing policy in the Central Puget Sound Region and beyond.