Project Motivation

We believe that property crime in Atlanta does not happen everywhere at once it piles up in a handful of predictable places. This project asks a simple question: when someone’s car is broken into or a package goes missing, what kind of place is it happening in, and how stable are those patterns over time? We follow each recorded property crime from the street address where it was reported to the tax parcel where it most likely occurred. That lets us talk about incidents not just as points on a map, but as events happening at specific types of parcels: apartments, single-family homes, commercial strips, parking lots, and tax-exempt or institutional sites.

By aggregating these incidents to Census tracts and joining neighborhood indicators like poverty and renter share, we move from individual addresses to broader questions about stability, investment, and risk. A BeltLine buffer also lets us ask whether land around the trail carries a different mix of property crime than the rest of the city. In short, the goal is to connect three levels of analysis: parcels (land use and function), neighborhoods (social and economic conditions), and corridors like the BeltLine where new investment and heavy activity might change opportunities for crime.

Research Questions

Guided by that motivation, we focus on three main questions:

  1. Land-use patterns
    Do commercial, mixed-use, and parking-heavy parcels see more property crime than mostly residential parcels?

  2. Neighborhood conditions
    When we convert incidents into tract-level rates, how do characteristics like poverty, renter share, and ownership relate to property crime risk?

  3. Stability over time
    Across pre-COVID, COVID, and post-COVID phases, do hot spots move around the city, or do the same tracts stay high or low from year to year?

These questions frame everything that follows: the way we process the data, the spatial statistics we use, and how we interpret the maps and charts.

Theoretical Framework

Our interpretation leans on two ideas: routine activity theory and neighborhood effects.

Routine activity theory says crime is more likely when three elements line up: a motivated offender, a suitable target, and weak guardianship. In our context, that often looks like:

Neighborhood effects add a slower layer: long-term conditions such as lower incomes, older multifamily housing, and a high share of renters can shape where property crime becomes normal rather than exceptional. If those forces are strong, we expect tract-level crime rates to show clear spatial clustering: high-crime tracts next to other high-crime tracts, and low-crime tracts forming their own clusters.

The BeltLine adds an important corridor lens. If parcels near the trail repeatedly show higher shares of commercial and multifamily incidents than the rest of the city, it suggests that new amenities, parking, and nightlife around the BeltLine create a durable environment for opportunistic property crime—alongside all the benefits of investment and activity.

Data and Processing Pipeline

Step 1: Building the incident–parcel file

Our starting point is a set of APD incident tables filtered to property crime categories and a 2025 parcel dataset with classcd, assessed values, and neighborhood attributes.
To make them work together, we:

  1. Standardized incident addresses
    We cleaned the raw address string into components (house number, street name, suffix, and direction) and removed extra punctuation so that incident addresses and parcel Site Address fields share a common format.

  2. Matched incidents to parcels
    Using a combination of fuzzy string matching and spatial distance, we linked each incident to nearby parcels and recorded the best candidate.
    The key variable here is match_distance, a score that measures how close the incident address is to the parcel address; smaller values indicate a better match.

  3. Filtered to high-confidence matches
    We kept only records with match_distance <= 0.25, producing the file property_crimes_with_parcels.
    This file is the backbone of the project and is stored as both an .rds and a .csv in the data_intermediate folder.

Within this file, we use the first character of classcd to build location_type_calc, grouping parcels into categories like “Residential”, “Commercial / Business”, “Industrial”, and “Institutional / Exempt.”
These categorization allows us to turn a long list of codes into a small set of environments we can talk about in plain language.

Step 2: From incidents to neighborhoods

Next, we aggregate the incident–parcel file to larger units:

  • At the parcel level, we tabulate the number of incidents by location_type_calc to see whether risk concentrates in certain land-use environments.
  • At the tract level, we join each incident to a Census tract and compute crime_rate_per_10k as incidents per 10,000 residents, then attach tract attributes such as median_income, renter_share, and ownership share.
Preview: property_points_unique.csv
Incident_ID datetime year month hour phase Crime_Type Address Location_Type Neighborhood NPU Lat Long ss Report_Date Day_Occur Time_Occur Zone OBJECTID Day_Week Event_Watch Report_Num
72932 2018-01-02 00:00:00 2018 1 0 Pre-COVID LARCENY-NON VEHICLE 183 MOUNT ZION RD SE Other / Unknown NA NPU Z 33.67268 -84.38507 180020741 1/2/2018 Tuesday 00:00:00 3 72932 Tuesday Evening Watch NA
72929 2018-01-02 01:20:00 2018 1 1 Pre-COVID AUTO THEFT 3238 PACES FERRY PL NW Residential NA NPU B 33.84325 -84.38202 180021478 1/2/2018 Sunday 01:20:00 2 72929 Sunday Evening Watch NA
72395 2018-01-02 04:44:00 2018 1 4 Pre-COVID LARCENY-NON VEHICLE 1977 MARTIN L KING JR DR NW Residential NA NPU I 33.75110 -84.44790 180020324 1/2/2018 Tuesday 04:44:00 1 72395 Tuesday Morning Watch NA
72919 2018-01-02 05:57:00 2018 1 5 Pre-COVID AUTO THEFT 285 CENTRAL AVE Other / Unknown NA NPU M 33.74809 -84.39201 180020375 1/2/2018 Tuesday 05:57:00 5 72919 Tuesday Morning Watch NA
72404 2018-01-02 08:00:00 2018 1 8 Pre-COVID LARCENY-NON VEHICLE 2766 METROPOLITAN PKWY SW Other / Unknown NA NPU X 33.67967 -84.40736 180020496 1/2/2018 Tuesday 08:00:00 3 72404 Tuesday Morning Watch NA
72923 2018-01-02 08:00:00 2018 1 8 Pre-COVID AUTO THEFT 3530 ZIP INDUSTRIAL BLVD SE Other / Unknown NA NPU Z 33.65859 -84.38330 180021149 1/2/2018 Saturday 08:00:00 3 72923 Saturday Morning Watch NA
72928 2018-01-02 08:00:00 2018 1 8 Pre-COVID AUTO THEFT 3730 DUMBARTON RD NW Retail / Commercial NA NPU A 33.85739 -84.41687 180020976 1/2/2018 Tuesday 08:00:00 2 72928 Tuesday Morning Watch NA
72944 2018-01-02 08:30:00 2018 1 8 Pre-COVID LARCENY-FROM VEHICLE 3550 CAMPBELLTON RD SW Other / Unknown NA NPU P 33.68984 -84.50491 180020569 1/2/2018 Tuesday 08:30:00 4 72944 Tuesday Morning Watch NA
72237 2018-01-02 09:00:00 2018 1 9 Pre-COVID LARCENY-FROM VEHICLE 1046 RIDGE AVE SW Other / Unknown NA NPU V 33.72631 -84.39189 180020682 1/2/2018 Thursday 09:00:00 3 72237 Thursday Morning Watch NA
72939 2018-01-02 09:00:00 2018 1 9 Pre-COVID LARCENY-FROM VEHICLE 1058 NORTHSIDE DR NW Residential NA NPU E 33.78382 -84.40782 180021130 1/2/2018 Tuesday 09:00:00 5 72939 Tuesday Morning Watch NA

This pipeline lets us move smoothly between detailed questions like “What happens in parking parcels?” and a broader context like “How does crime_rate_per_10k vary with median_income across the city?”

Workflow

Exploratory Data Analysis

knitr::include_graphics(boxplot_path)

Most tracts fall near the lower end of the crime rate and incident count, but the upper tail stretches far beyond the interquartile range. Several tracts report crime rates above 4,000–6,000 per 10k residents, and a few exceed 10,000, placing them many standard deviations above the city median. Incident counts follow the same pattern: while many tracts record fewer than 300–400 incidents, the outliers reach 1,000+, confirming that a small share of tracts accounts for a large share of all reported events.

For the predictors, the distributions are less extreme but still uneven. Poverty and renter share show long right tails—most tracts fall in mid-range values, but a minority reach 40–75% poverty or 80–100% renters. Retail and commercial activity also vary, with a small group of tracts carrying much higher retail share.

Turning to histograms, these together clarifies the data clearer:

As for the scatter matrix, the data seems solidfying the idea of disadvantage and instability make it harder to maintain informal control, while strong employment centers can provide some protective effect.

Spatial Methods and results

We measure clustering with Global Moran’s I (citywide autocorrelation) and Local Indicators of Spatial Association (LISA) to flag specific hot and cold spots. Queen contiguity defines tract neighbors, and standardized weights keep influence balanced. Spatial regression addresses dependence in residuals when relating tract crime rates to neighborhood covariates.

Global Moran’s I Results

Global Moran’s I for property crime rate by phase
variable I expected variance z_score p_value
prop_rate_10k 0.5587873 -0.0030769 0.0010675 17.19673 0
pov_rate 0.5132679 -0.0030769 0.0010979 15.58334 0
rent_rate 0.4023548 -0.0030864 0.0011165 12.13375 0

Global Moran’s I remains consistently high across all phases, with values tightly grouped around 0.56. The expectation under spatial randomness is near 0, so the observed statistic sits far above what chance would generate. The associated p-values fall below 10⁻⁶⁵ in every phase, confirming that the spatial pattern is extremely unlikely to be random.While the near-identical values before, during, and after COVID indicate that the spatial structure of property crime did not loosen or reorganize; the same parts of the city continue to form the core high-crime and low-crime clusters.

The Moran scatterplot shows how this plays out at the tract level. The fitted line rises sharply, with most points falling in the lower-left (low rate, low spatial lag) and upper-right (high rate, high spatial lag) quadrants. Roughly speaking, tracts with rates under 1,000–2,000 per 10k cluster together, while tracts above 4,000 per 10k tend to sit next to neighbors with averages in the 3,000–5,000 range. A small group of extreme tracts—those reaching 6,000–10,000 per 10k—extend the upper tail and reinforce the slope instead of weakening it. Only a modest number of tracts fall into mixed quadrants, which shows that high-crime tracts rarely sit next to low-crime neighbors.

LISA Cluster Patterns

LISA results show a clear split in where property crime concentrates. In the post-COVID map, high-high clusters make up the majority of significant tracts, with roughly 20–25 tracts forming a single, continuous block around the urban core. These tracts sit at the upper end of the distribution—many above 4,000–6,000 incidents per 10k—and almost all are bordered by neighbors with similarly elevated levels. The size and contiguity of this block indicate that the core is not a loose set of hotspots but a tightly connected zone of high activity.

Low-low clusters account for the next largest group, concentrated mainly in northern single-family areas. These clusters involve about 10–15 tracts, most with rates near the bottom decile of the distribution. They form one extended low-crime zone rather than isolated pockets, showing strong internal consistency and little disruption across phases.

Cross-type outliers remain limited. High-low tracts appear only in small numbers, typically 1–3 tracts at the perimeter of the high-high core. These boundary cases show rates substantially higher than their neighbors but well below the extreme upper tail, marking areas where local conditions diverge from the broader corridor. Low-high cases are nearly absent, confirming that sharp jumps from low-crime neighborhoods into isolated high-crime tracts are rare.

Model Interpretation

Spatial regression (post-COVID): coefficients and rate ratios
term estimate std.error statistic p.value conf.low conf.high rate_ratio pct_change_per_sd
(Intercept) -2.5499899 0.1530040 -16.666169 0.0000000 -2.8356311 -2.2343497 0.0780825 -92.19175
z_pov_rate 0.3186855 0.1883688 1.691817 0.0906809 -0.0957225 0.7700866 1.3753187 37.53187
z_rent_rate 0.5886624 0.1780191 3.306736 0.0009439 0.1919294 1.0009033 1.8015770 80.15770
z_total_jobs -0.2221681 0.1641781 -1.353214 0.1759872 -0.4889266 0.0579544 0.8007807 -19.92193

The model results align with the exploratory maps:

Property Type Additional Evidence

When we look at the LISA map, many of the high–high clusters of property crime appear in and around the BeltLine corridor. That pattern motivates a closer look at what kinds of places along the BeltLine are driving those hot spots.

The incident–parcel file property_crimes_with_parcels gives us a clear way to talk about where property crime happens in terms of land use. The summary table type_overall shows the distribution of incidents by location_type_calc across the entire city.
Here, “Commercial / Business” location types sit at the top, followed by “Residential” parcels, confirming that retail, and apartment contexts dominate the property crime landscape.

All matched incidents by parcel location type
location_type n pct
Commercial / Business 77228 49.7
Residential 59975 38.6
Other / Unknown 7735 5.0
Institutional / Exempt 6302 4.1
Industrial 3645 2.3
Historic / Special 442 0.3
Utility / Infrastructure 83 0.1
Top parcel class codes (overall)
classcd n
R3 59610
C3 40768
C4 23982
C5 8565
C1 3400
E1 2458
I3 2114
E2 1374
I4 1234
E6 991
Incidents with BELTLINE in the address (proxy for BeltLine-adjacent sites)
location_type n pct
Commercial / Business 83 84.7
Industrial 10 10.2
Residential 5 5.1
Incidents within 400m of BeltLine overlay (by location type)
beltline_buffer location_type n pct
FALSE Commercial / Business 42356 51.1
TRUE Commercial / Business 34872 48.1
FALSE Residential 30307 36.6
TRUE Residential 29668 40.9
FALSE Other / Unknown 5152 6.2
FALSE Institutional / Exempt 3490 4.2
TRUE Institutional / Exempt 2812 3.9
TRUE Other / Unknown 2583 3.6
TRUE Industrial 2213 3.1
FALSE Industrial 1432 1.7
TRUE Historic / Special 333 0.5
FALSE Historic / Special 109 0.1
FALSE Utility / Infrastructure 43 0.1
TRUE Utility / Infrastructure 40 0.1

The top_class table lists the most frequent parcel classcd values associated with incidents, making it possible to see which specific property classes—such as particular commercial or multifamily codes—are most affected.

The associated bar plots show that, within the BeltLine buffer, incidents are even more concentrated in “Commercial / Business” and multifamily “Residential” categories than they are citywide. Outside the buffer, incidents spread more into commercial parcel types.

This combination of tables and plots strengthens the conclusion that property crime in Atlanta is closely tied to specific land-use environments rather than being evenly distributed across all parcels.

Key Takeaways

To answer our questions:

  1. Stable hot spots. Hot spots are remarkably stable. Pre-COVID, COVID, and post-COVID maps all show similar clusters around the urban core, major mixed-use corridors, and Beltline-adjacent areas.Shocks change levels, not the map. Overall incident counts fluctuate across years, but the location of high- and low-crime tracts hardly moves—suggesting that underlying land use and neighborhood structure drive the long-run pattern.

  2. Land use and neighborhood structure matter. Commercial, mixed-use, and parking parcels carry the heaviest load of property crime. Incidents concentrate around garages, surface lots, retail strips, and mixed commercial–residential parcels. Residential risk is highest in multifamily contexts. Apartments and other higher-density residential parcels show more incidents than single-family parcels, especially near major corridors and the Beltline

  3. Neighborhood conditions. Poverty and renting go hand-in-hand with higher risk. Tracts with higher poverty rates and larger renter shares consistently have higher property-crime rates. Ownership and affluence are protective. Areas with more homeowners and higher property values tend to have lower rates, even when they are not far from hot-spot corridors.

Limitations

Several limitations shape how we read these results:

Conclusion and Planning Implications

Looking across parcels, tracts, and spatial models, a consistent story emerges: Atlanta’s property crime patterns are anchored less by short-term shocks and more by the city’s land-use structure and neighborhood stability.

Commercial corridors, parking facilities, and multifamily areas, especially near the BeltLine, repeatedly appear as environments where property crime clusters. Tracts with more renters and lower incomes show higher rates, while higher ownership and property values are associated with lower risk.

For planners and policymakers, this points to a two part strategy:

By tying incidents to parcels and neighborhoods, this project shows how urban analytics can move from “where the dots are” to which environments and corridors truly structure property crime risk in Atlanta.