THE EFFECTS OF VACANT LOT GREENING AND THEIMPACT OF LAND USE AND BUSINESS VIBRANCY
By Jesse Cui and Shane T. Jensen
University of Pennsylvania
We examine the ongoing Philadelphia LandCare (PLC) vacantlot greening initiative and evaluate the association between this builtenvironment intervention and changes in crime incidence. We developa propensity score matching analysis that estimates the eﬀect of va-cant lot greening on diﬀerent types of crime while accounting forsubstantial diﬀerences between greened and ungreened lots in termsof their surrounding demographic, economic, land use and businessvibrancy characteristics. Within these matched pairs of greened vs.ungreened vacant lots, we estimate larger and more signiﬁcant ben-eﬁcial eﬀects of greening for reducing violent, non-violent and totalcrime compared to comparisons of greened vs. ungreened lots with-out matching. We also investigate the impact of land use zoning andbusiness vibrancy and ﬁnd that the eﬀect of vacant lot greening on to-tal crime is substantially aﬀected by particular types of surroundingland use zoning and the presence of certain business types.
The recent availability of urban data gives us the op-portunity to investigate urban environments at a higher resolution than everbefore and quantitatively test urban design principles of the past century.Our focus is the evaluation which characteristics of the built environmentare most eﬀective in promoting the safety of local neighborhoods in our largeurban centers. We are particularly interested in the extent to which the im-pact on crime of built environment interventions are inﬂuenced by diﬀerenttypes of surrounding land use and business vibrancy.There are many theories in urban planning and criminology that hypoth-esize associations between aspects of the built environment, human activityand safety. A prominent idea by Jane Jacobs in her 1961 book, The Deathand Life of Great American Cities, was the concept of eyes on the street,which summarized her viewpoint that safer and more vibrant neighborhoodswere those that had many people engaging in activities (either commercial orresidential) on the street level at diﬀerent times of the day (Jacobs, 1961).This viewpoint is also encapsulated in the theory of natural surveillance(Deutsch, 2016): to the extent that criminal acts are a decision that may
Keywords and phrases: vacant lot greening; crime; urban analytics; land use; businessvibrancy a r X i v : . [ s t a t . O T ] J u l be impacted by the environment, we should be able to design spaces thatdiscourage crime because people feel they are being monitored.Historically, these theories have been diﬃcult to test because our abilityto experiment with changes in urban environments is severely limited asit is challenging to alter the city environment and impose treatments on ahuman population. Still, some experimentation with the built environmentof cities does occur. MacDonald (2015) and MacDonald et al. (2019) re-views previous research on the built environment and safety, where manyquasi-experimental studies have shown that changes in green space, housing,zoning and public transit does have association with crime.We focus on a particular type of built environment intervention: an ongo-ing vacant lot greening initiative by the Pennsylvania Horticulture SocietysPhiladelphia LandCare (PLC) program. In this ongoing intervention, thou-sands of vacant lots in Philadelphia have been cleaned up and turned intosmall public spaces in an eﬀort to improve the surrounding area. This pro-gram provides a unique opportunity to evaluate the eﬀects of an interventionon the built environment of neighborhoods across the city, while also inves-tigating the impact of surrounding businesses and land use zoning on theeﬀects of greening initiatives in urban environments.Branas et al. (2011) examined the early years of this PLC vacant lotgreening initiative using a diﬀerence-in-diﬀerences analysis of the impact ofvacant lot greening in Philadelphia and found reductions in gun assaults andvandalism. More recently, a randomized-control trial of the PLC greeningprogram showed reductions in serious crime, fear of crime, and shootings(Branas et al., 2018; Moyer et al., 2019). Vacant lot greening programs havealso been associated with reductions in violent crimes in Youngstown, Ohioand Flint, Michigan as well as drug crimes in New Orleans, Louisiana (Kondoet al., 2016, 2018; Heinze et al., 2018).However, when comparing crime rates around greened versus ungreenedvacant lots, we must be extra careful to take into account the surroundingcontext of these vacant lots. First, systematic diﬀerences between greenedand ungreened vacant lots in terms of their surrounding neighborhood char-acteristics can confound our comparisons such that observed diﬀerences incrime can not be attributed to the greening intervention. We will see inSection 3 that the areas surrounding greened vacant lots are substantiallydiﬀerent from the areas surrounding ungreened vacant lots in terms of theirdemographic, economic and land use characteristics.The diﬀerence-in-diﬀerences analysis of Branas et al. (2011) does not ac-count for these systematic diﬀerences in surrounding context, though theydo examine separate eﬀects for diﬀerent regions of the city. The randomized ACANT LOT GREENING IN PHILADELPHIA control design studied in Branas et al. (2018) better addresses the possibilityof confounding but this study involves a smaller subset of lots and thus maynot be representative of all neighborhood contexts present in Philadelphia.We address the systematic diﬀerences in surrounding context betweengreened and ungreened lots by a careful matching of individual greenedlots with individual ungreened lots based on their surrouding neighborhoodcontext. We use propensity score matching (Rosenbaum and Rubin, 1983) tocreate matched pairs of vacant lots where each pair of consists of one greenedlot and one ungreened lot that have highly similar demographic, economicand land use characteristics. These matched pairs allow us to make morebalanced comparisons of crime rates between greened and ungreened lots inorder to provide a comprehensive evaluation of the eﬀects of this vacant lotgreening intervention in Philadelphia.Beyond this overall evaluation, we are particularly interested in the impactof diﬀerent types of surrounding land use and business vibrancy on theeﬀect of vacant lot greening on crime. Previous studies of this PLC program(Branas et al., 2011, 2018) do not examine whether these aspects of thesurrounding context modiﬁes the eﬀects of vacant lot greening.In addition to demographic and economic data from the U.S. Census Bu-reau, we use zoning data from the City of Philadelphia to create detailedmeasures of land use around each greened and ungreened vacant lot. We alsoincorporate detailed business location data from Humphrey et al. (2020) intoour analysis. All of these data sources will be involved in the creation of ourmatching procedure in order to ensure our matched pairs of greened vs. un-greened lots are highly similar on many diﬀerent aspects of their surroundingarea.These matched pairs also facilitate our investigation of the inﬂuence ofnearby land use and business vibrancy on the eﬀects of vacant lot greeningon crime. Since each matched pair will contain two lots with highly similarland use and business vibrancy characteristics, we can subset our pairs inorder to explore whether the eﬀect of vacant lot greening on crime changesbetween pairs that diﬀer substantially on aspects of their surrounding landuse or business locations.In summary, we harness sophisticated matching methodology and avail-able data for local areas in Philadelphia to investigate the eﬀects of vacantlot greening at a high resolution while also exploring the impact of land useand business activity around the locations of these greening interventions.We describe our available data on criminal activity, demographic, eco-nomic characteristics, land use and business vibrancy for Philadelphia in Sec-tion 2. In Section 3, we compare the set of greened and ungreened vacant lots and observe systematic diﬀerences in their surrounding demographic, eco-nomic and land use characteristics. We address these systematic diﬀerenceswith a careful matching of greened vs. ungreened vacant lots in Section 4and use our matched pairs to evaluate the eﬀect of vacant lot greening oncrime in Section 4.2. In Section 5, we use diﬀerent subsets of these matchedpairs to investigate the impact of diﬀerent types of land use and businessactivity on the eﬀect of vacant lot greening on crime. We conclude witha brief summary and discussion in Section 6. The code repository for dataprocessing and analysis can be viewed at https://github.com/jessecui/WSII-Urban-Analytics-Business-Vibrancy.
2. Urban Data in Philadelphia.
Our analyses will be based on pub-licly available data on crime, economic and demographic neighborhood char-acteristics, and land use zoning as well as a comprehensive database onbusiness locations and open hours for the city of Philadelphia that has beencompiled by our research group. We have the location and type of every re-ported crime over the past decade from the Philadelphia Police Department.We also have detailed data on neighborhood-level income, poverty, race andpopulation density from the U.S. Census Bureau. The city of Philadelphiaprovides zoning designation for the approximately half million lots in the citywhich we have used to summarize the land use around vacant lots. Belowwe provide additional details about the processing of each data source andcreation of quantitative measures of the surrounding area for each vacantlot in the city of Philadelphia.2.1.
Vacant Lots Data.
We have data on the location (and date of green-ing) for each vacant lot greened through the Pennsylvania Horticultural So-ciety’s Land Care program. We focus on 4651 vacant lots greened in theperiod from 9/01/2007 - 9/01/2017, which is the time period for which wealso have high resolution crime data (Section 2.2). We also have data on thelocation of ungreened vacant lots over the same time period from the City ofPhiladelphia’s Licenses and Inspections Oﬃce. We ﬁltered this data to onlyinclude vacant property (non-building) violations and removed duplicate vi-olations at the same location by only including the ﬁrst violation instanceat each vacant lot location. After this ﬁltering, we have the locations of ≈ Crime Data.
We retrieved crime data for the city of Philadelphiathat is made available by the Philadelphia Police Department on the open-dataphilly.org data portal. Our dataset contains the date, time and GPSlocation of each reported crime from 2007 to 2019, as well as the type of
ACANT LOT GREENING IN PHILADELPHIA crime (e.g. homicide, aggravated assault, etc). We ﬁltered out some minorcrime types that are unlikely to be related to the use of public spaces (suchas fraud and embezzlement). After this ﬁltering, we have ≈ . violent crimes , which contain homicides, rapes, rob-beries, and aggravated assaults versus non-violent crimes , which containsburglary, thefts, vehicle thefts, other assaults, arson, vandalism, oﬀensesagainst family and children, public drunkenness, disorderly conduct, andvagrancy/loitering.2.3. Demographic Data.
Population demographic data for Philadelphiawas obtained from the U.S. Census Bureau website (Table SF1 P5 in theirdata portal). The raw demographic data gives the population count by racefrom the 2010 census in each of the 18,872 census blocks in Philadelphia.We use this data to calculate the population count and racial proportions(black, white, hispanic, and asian) surrounding each greened and ungreenedvacant lot.2.4.
Economic data for Philadelphia was obtained fromthe 2015 American Community Survey (Tables B19301 and C17002 in theU.S. Census Bureau data portal). This data contains the per-capita meanincome and the proportion of households in seven diﬀerent “poverty” brack-ets based on the ratio of income to poverty line for each of the 18,872 censusblocks in Philadelphia.2.5.
Land Use Zoning Data.
Zoning data is made available by the Cityof Philadelphia through the opendataphilly.org data portal. This data con-sists of a shapeﬁle that provides the area and registered land use zoningdesignation for the ≈ ,
000 lots in the city. We aggregated these zon-ing designations into eight primary types: Residential, Commercial, Indus-trial, Civic/Institution, Transportation, Cultural/Park, Water, Vacant, andOther.We use these registered zoning designations to create several quantitativemeasures of the land use around each of the greened and ungreened vacantlot locations in Philadelphia. Speciﬁcally, for the area in a 200 meter radiusaround each vacant lot location, we calculate the proportion of that areathat is designated as each of those eight zoning types, i.e. the proportion ofresidential land use, proportion of commercial land use, etc.2.6.
Business Vibrancy Data.
Our research group manually assembled adatabase of Philadelphia business locations by scraping three diﬀerent web resources (Google Places, Yelp, and Foursquare). Each business is catego-rized into one or more of eight business types: Cafe, Convenience, Gym,Institution, Liquor, Lodging, Nightlife, Pharmacy, Restaurant, and Retail.This data is described in more detail in Humphrey et al. (2020).From this database, we create measures of business vibrancy around eachof the greened and ungreened vacant lot locations in Philadelphia. For eachvacant lot, we create a set of eight binary variables (for our eight busi-ness types) that indicate whether there is a business of that particular typepresent within 200 meters of that vacant lot. For each vacant lot, we alsotabulate the total number of businesses located within 200 meters of thatvacant lot.
3. Exploratory Comparison of Greened vs. Ungreened VacantLots.
Before proceeding with our primary matching analysis, we will ﬁrstcompare greened and ungreened vacant lots in Philadelphia in terms of crimerates as well as surrounding demographic, economic and land use charac-teristics. The substantive diﬀerences that we will observe in surroundingcontext will motivate the need for a careful matching analysis of greenedand ungreened vacant lots in Section 4.Previous evaluations of this PLC vacant lot greening initiative (e.g. Branaset al. (2011) have employed a diﬀerence-of-diﬀerences (DoD) analysis ofcrime rates. In this approach, a diﬀerence in crime is calculated for eachgreened vacant lot as the crime rate in a time period after the greeningintervention minus the crime rate in a time period before the greening inter-vention for that lot. In our version of this DOD analysis, the time periodsare 6 to 18 months before the lot was greened versus 6 to 18 months after thelot was greened. If we can calculate a corresponding after-before diﬀerencein crime rates for each ungreened vacant lot, then the diﬀerence of thesegreened vs. ungreened diﬀerences is the DoD estimate of the eﬀect of vacantlot greening on crime rates.However, an immediate issue with this approach is that although eachgreened (“treatment”) vacant lot has a well-deﬁned greening interventiondate upon which we can center the before vs. after time periods, there is nocorresponding intervention date for each of the ungreened (“control”) lotsin Philadelphia. We will address this time period issue with our matchinganalysis in Section 4, but for this preliminary comparison we chose to centerour before vs. after time periods for the ungreened vacant lots on October30th, 2012, which is the average intervention dates of the greened vacantlots.Figure S1 in our supplementary materials compares the distribution of
ACANT LOT GREENING IN PHILADELPHIA crime counts around greened lots vs. ungreened lots. The average after-before reduction in total crimes is -16.19 for greened lots and -20.14 forungreened lots, which results in a DoD estimate of an increase of 3.95 totalcrimes for the eﬀect of vacant lot greening. This DOD analysis clearly doesnot provide evidence for beneﬁcial eﬀects of vacant lot greening on crime.However, this simple diﬀerence-in-diﬀerences comparison of crime ratesdoes not address the possibility that greened vacant lots could diﬀer greatlyfrom ungreened vacant lots in terms of their surrounding context, and thatthese diﬀerences would confound our attempts to attribute crime diﬀerencesto the greening intervention. Indeed, we observe substantial imbalance be-tween greened and ungreened vacant lots in terms of their surrounding de-mographic, economic, land use and business vibrancy characteristics.Figure 1 gives side-by-side boxplots that compare greened vacant lotsto ungreened vacant lots on several demographic and economic measures.We see that greened and ungreened lots diﬀer substantially in terms oftheir surrounding racial proportions, per capita income and proportion ofhouseholds in each poverty bracket. Fig 1: Comparing greened versus ungreened vacant lots in terms of population count(top left), racial proportions (top right), per capita income (bottom left) and povertybrackets (bottom right). . Speciﬁcally, we observe substantially lower proportions of hispanics andwhites and substantially higher proportions of blacks in the neighborhoodssurrounding greened lots compared to the neighborhoods surrounding un-greened lots. The neighborhoods surrounding greened lots also tend to have lower per capita income and a larger proportion of households in the highpoverty brackets. We also see diﬀerences between greened and ungreenedvacant lots in terms of the surrounding land use zoning as well as num-ber and types of businesses. Details on this comparison are given in oursupplementary materials.These substantial diﬀerences in surrounding context for greened versusungreened vacant lots make it diﬃcult to attribute any observed diﬀerencesin crime rates to the greening intervention itself. This problematic imbalanceon surrounding characteristics is in addition to the issue that our collectionof ungreened (control) lots do not have a well-deﬁned intervention date forestablishing a before-after comparison of crime rates. Both of these problemswill be addressed by our matching analysis in the next section.
4. Matched Pairs Comparison of Greened and Ungreened Va-cant Lots.
We can address the imbalance in surrounding context betweengreened and ungreened vacant lots by performing a careful matching of eachgreened (treatment) vacant lot with an ungreened (control) vacant lot thathas highly similar surrounding characteristics. By creating matched pairs ofindividual greened and ungreened lots, we can better attribute any observedwithin-pair diﬀerence in crime rates to the greening intervention.In addition, our matched pair analysis addresses the issue that we do nothave intervention dates for our ungreened (control) vacant lots. Once we havepaired up an individual greened vacant lot with a highly similar ungreenedvacant lot, we can use the intervention date of that greened lot for its pairedungreened vacant lot. This ensures that we are comparing changes in crimeover the same time period within each of our matched pairs.4.1.
Propensity Score Matching.
In order to create matched pairs ofhighly similar greened vs. ungreened vacant lots, we must ﬁrst choose somemeasure of the similarity between the surrounding characteristics of anypair of greened and ungreened vacant lots. We base our matching upon the propensity score (Rosenbaum and Rubin, 1983) which is deﬁned as the prob-ability that a particular unit (vacant lot) receives the treatment (greening)based on their surrounding neighborhood context.Two vacant lots with highly similar surrounding characteristics shouldhave highly similar propensity scores. Thus, for every greened vacant lot inPhiladelphia, we will create a matched pair by ﬁnding an ungreened lot thathas the closest propensity score to that greened vacant lot.We use a logistic regression model to calculate these propensity scores foreach greened and ungreened vacant lot in our data. In this logistic regressionmodel, each unit i is a vacant lot in the city of Philadelphia with outcome ACANT LOT GREENING IN PHILADELPHIA Y i = 1 if vacant lot i was greened or Y i = 0 if vacant lot i was ungreened.The probability P ( Y i = 1) for each vacant lot i is modeled as a function of itssurrounding characteristics X i which includes our demographic, economic,land use and business vibrancy measures outlined in Sections 2.3-2.6.Details of our ﬁtted logistic regression model and signiﬁcance of coeﬃ-cients are given in our supplementary materials. From those details, we notethat the logistic regression model that uses all diﬀerent type of surround-ing characteristics (demographic, economic, land use and business vibrancy)provides the best ﬁt to the data compared to models that exclude one ormore of these diﬀerent data types. As expected, our ﬁtted logistic regressionmodel has highly signiﬁcant coeﬃcients for the surrounding characteristicswith large observed diﬀerences between greened and ungreened lots in Fig-ure 1, such as per capita income and the racial proportions.For each vacant lot i in our data, our ﬁtted logistic regression model pro-duces ˆ p i which is the predicted probability of greening for that lot, otherwiseknown as the propensity score for that lot. We use these propensity scores tocreate matched pairs of vacant lots where each greened vacant lot is matchedup with the ungreened vacant lot that has the closest propensity score tothat greened vacant lot.In Figure 2, we evaluate the eﬀectiveness of our propensity score match-ing procedure in terms of improving the balance in surrounding contextbetween greened and ungreened lots. Speciﬁcally, we compare the standard-ized mean diﬀerence in each surrounding area measure between all greenedand ungreened vacant lots before matching (gray dots) with the standard-ized mean diﬀerence in each surrounding area measure within our matchedpairs (black triangles).We see that our matched pairs of greened and ungreened lots have smallerdiﬀerences on almost all surrounding area measures compared to the pop-ulation of greened and ungreened lots before matching. The reduction indiﬀerences by matching is most dramatic for the racial proportions, incomeper capita, and land use proportions where we saw the greatest imbalancein Figure 1. By reducing the diﬀerences in the surrounding characteristicsbetween greened and ungreened vacant lots, our matching procedures al-lows us to better isolate the eﬀect of the greening intervention on changesin crime rates.4.2. Matched Pairs Evaluation of Eﬀect of Greening on Crime.
Our eval-uation of the eﬀect of the PHS Landcare greening intervention in Section 3was based on a diﬀerence-of-diﬀerences (DoD) estimate where ﬁrst a af-ter vs. before intervention diﬀerence in crime rates was calculated around Fig 2: Comparing the standardized mean diﬀerences in each measure of the sur-rounding area between greened and ungreened vacant lots before matching (graydots) and after matching (black triangles). each vacant lot and then the average diﬀerence of those diﬀerences wascalculated between the greened vacant lots and the ungreened vacant lots.As discussed above, this estimate suﬀers from two major issues: imbalancebetween greened and ungreened vacant lots in terms of surrounding charac-teristics and the lack of a comparable intervention date for the ungreenedvacant lots.We now correct both of these issues by calculating the diﬀerence-of-diﬀerences (DoD) in crime rates within each matched pair , which ensuresthat we are only comparing vacant lots that have highly similar surroundingcharacteristics as well as an identical time frame for the comparison (sincewe use the same intervention date for the greened and ungreened lots withineach pair). Our overall estimate of the eﬀect of greening is the average ofthese within-pair DoD values across all matched pairs.In Table 1, we present within-pair DoD estimates of the eﬀect of vacantlot greening on crime rates. We present estimates for all crimes as well asjust violent and just non-violent crimes. Crime rate calculations were basedon a 200 meter radius around each vacant lot (just as in Section 3) thoughwe see similar trends when using 100 meter or 500 meter radii.The negative signs on the diﬀerence-of-diﬀerence estimates imply that thedecrease in crime rates (after - before intervention date) was larger aroundthe greened vacant lots than their matched ungreened vacant lots. Theseestimates are signﬁcantly diﬀerent from zero for both violent and non-violent
ACANT LOT GREENING IN PHILADELPHIA Table 1
Within-pair diﬀerence-of-diﬀerence estimates of the eﬀect of vacant lot greening on crimerates.
Crime type Estimate Standard Error T-stat p-value
Total -4.57 0.82 -5.58 2.61E-08Non-violent -4.02 0.72 -5.61 2.14E-08Violent -0.55 0.24 -2.25 0.025 crime types, though the eﬀect is much larger for non-violent crimes whichare also much more frequent than violent crimes as seen in SupplementaryFigure S1.Overall, we see a beneﬁcial eﬀect of vacant lot greening on surroundingrates of both violent and non-violent crimes. The size of these eﬀects can beinterpreted as the diﬀerence between greened versus ungreened vacant lots inthe change in number of crimes in a one year period around the interventiondate. In other words, greened vacant lots showed an additional decrease of4.57 crimes per year compared to ungreened vacant lots over the same timeperiod. In the next section, we use our matched pairs to investigate furtherhow the eﬀect of vacant lot greening on crime is potentially moderated orinﬂuenced by diﬀerent types of surrounding land use and amounts of businessvibrancy.
5. Impact of Land Use Zoning and Business Vibrancy on Ef-fects of Vacant Lot Greening.
Our analysis in Section 4 was based oncreating matched pairs of individual greened and ungreened vacant lot lo-cations that share highly similar surrounding demographic, economic, landuse and business vibrancy characteristics. We can also use these matchedpairs to explore whether diﬀerent types of surrounding land use or businessvibrancy are associated with larger or smaller crime eﬀects. We investigatethese associations by subsetting our set of matched of pairs into high vs. lowlevels of certain types of land use zoning or presence vs. absence of certainbusiness types and then comparing the DoD estimates of vacant lot greeningon crime between these subsets of matched pairs.5.1.
Inﬂuence of Surrounding Land Use Zoning.
We ﬁrst investigatewhether diﬀerent types of land use zoning surrounding vacant lots has animpact on the eﬀect of vacant lot greening on crime. Speciﬁcally, for a par-ticular type of land use zoning designation such as “commercial”, we iden-tify the subsets of our matched pairs with the largest (top 75%) and thesmallest (bottom 25%) proportions of commercial zoning. We then calcu-late diﬀerence-of-diﬀerence estimates of the eﬀect of vacant lot greening on crime separately within these two subsets of matched pairs. Our tabulationsof crime are based on a 200 meter radius around each vacant lot, just as inSection 4In Figure 3 (left), we compare our estimated eﬀect of vacant lot greeningon total crime between the subsets of matched pairs with the largest (blackpoints) and the smallest proportions (gray points) of each diﬀerent land usezoning designation described in Section 2.5. For each estimated eﬀect, wealso indicate which of these estimated eﬀects are signiﬁcantly diﬀerent fromzero (solid points). For reference, the black vertical line represents the eﬀectof vacant lot greening on total crime across all matched pairs (from Table 1).We see that several types of land use zoning seem to have a substantialinﬂuence on the eﬀect of vacant lot greening on total crime. Greening ofvacant lots in high residential areas showed an increased reduction on crimecompared to areas with low residential proportions. Correspondingly, thegreening of vacant lots in low commercial areas showed an increased reduc-tion on crime compared to areas with high commercial proportions. We alsosee that the greening of vacant lots was associated with greater reductionsin total crime in areas with high proportions of civic/institutions and lowproportions of transportation . The signiﬁcance of these results suggest thatthe surrounding land use around vacant lots has a substantial impact on theeﬀect of greening on total crime.5.2. Inﬂuence of Surrounding Business Vibrancy.
We now investigatewhether diﬀerent types of surrounding businesses has an impact on the eﬀectof vacant lot greening on crime. Speciﬁcally, for a particular type of businesssuch as cafes or restaurants, we identify the subsets of our matched pairswhere that type of business is present within a 200 meter radius versus thesubsets of matched pairs where that business is absent within a 200 meterradius. We then calculate diﬀerence-of-diﬀerence estimates of the eﬀect ofvacant lot greening on crime separately within these two subsets of matchedpairs. We also compared estimates of the eﬀect of vacant lot greening oncrime between subsets of matched pairs with the largest (top 75%) and thesmallest (bottom 25%) total number of businesses in a 200 meter radius.In Figure 3 (right), we compare our estimated eﬀect of vacant lot greeningon total crime between the subsets of matched pairs the presence (blackpoints) versus absence (gray points) of each diﬀerent business type describedin Section 2.6. For each estimated eﬀect, we also indicate which of theseestimated eﬀects are signiﬁcantly diﬀerent from zero (solid points). The blackvertical line again represents the eﬀect of vacant lot greening on total crimeacross all matched pairs (from Table 1).
ACANT LOT GREENING IN PHILADELPHIA We see that the presence (or absence) of many business types seem to havea substantial inﬂuence on the eﬀect of vacant lot greening on total crime.The presence of convenience stores and pharmacies was associated with anincreased reduction in crime around greened vacant lots, whereas the absenceof cafes, gyms and restaurants was associated with an increased reductionin crime around greened vacant lots. Interestingly, the eﬀects of vacant lotgreening on total crime are not signiﬁcantly diﬀerent from zero amongst thesubsets of vacant lots that have the largest and smallest numbers of totalbusinesses. As we discuss in Section 6, additional research and new datasources are needed to further investigate these potential associations.
Estimated eﬀect of vacant lot greening on total crime between matchedpairs with the largest (black points) and smallest (gray points) proportions of eachland use zoning designation.
Right : Estimated eﬀect of vacant lot greening on totalcrime between matched pairs with the presence (black points) vs. absence (graypoints) of each business type. In both plots, signiﬁcant eﬀects are indicated by solidpoints and the black vertical line is the eﬀect of vacant lot greening on total crimeacross all matched pairs.
The recent explosion in data collection on so many as-pects of city life gives us the opportunity to investigate urban environmentsat a higher resolution than ever before. In this big data age, we can har-ness many diﬀerent types of data to evaluate associations between the builtenvironment and the health and safety of neighborhoods. We focus on aparticular built environment intervention: the ongoing vacant lot greeninginitiative undertaken by the Pennsylvania Horticulture Society through theirPhiladelphia LandCare (PLC) program.We develop a sophisticated propensity score matching analysis that es-timates the eﬀect of vacant lot greening on diﬀerent types of crime whileaccounting for systematic diﬀerences between greened and ungreened lotsin terms of their surrounding demographic, economic, land use and business vibrancy characteristics. By creating matched pairs of individual greenedvs. ungreened vacant lots that share highly similar surrounding characteris-tics, we can better isolate the eﬀect of the greening intervention on crime.Our matched pair design also addresses the issue that our ungreened vacantlots (control group) do not have a natural intervention date around whichto examine changes in crime. Within our matched pairs, we estimate largerand more signiﬁcant beneﬁcial eﬀects of vacant lot greening on violent, non-violent and total crime than simpler comparisons of greened vs. ungreenedlots without matching.We also used our matched pairs to investigate the impact of land usezoning and business vibrancy around the locations of these greening inter-ventions by comparing subsets of matched pairs that diﬀer substantially ontheir land use or business characteristics. We ﬁnd that the eﬀect of vacantlot greening on total crime is substantially aﬀected by certain types of sur-rounding land use zoning and the presence or absence of certain businesstypes.In particular, the eﬀects of vacant lot greening seem most beneﬁcial inareas with high residential and civic/institution zoning, as well as in loca-tions where convenience stores and pharmacies are present. Interestingly,the eﬀects of vacant lot greening on total crime are not signiﬁcantly diﬀer-ent from zero amongst the subsets of vacant lots that have the largest andsmallest numbers of total businesses, which perhaps suggests that a moder-ate number of businesses is more ideal in terms of the most beneﬁcial eﬀectsof vacant lot greening.However, further research is needed in order to conﬁrm these potentialassociations and investigate the underlying mechanisms by which the builtenvironment impacts safety. We also need data that more closely reﬂects theextent of human activity around greened and ungreened vacant lot locationsand changes in public space usage due to greening interventions. In particu-lar, direct measures of foot traﬃc around greened and ungreened vacant lotswould provide a higher resolution picture of public usage of these spaces.There is also a need for more research on the impact of vacant lot greeninginitiatives on outcomes beyond crime and safety. Heckert and Mennis havefound that property values surrounding greened vacant lots had a greaterincrease in value than properties surrounding non-greened vacant lots (Heck-ert and Mennis, 2012). Branas et al. (2011) vacant lot greening was associ-ated with residents reporting less stress and more exercise in certain areasof Philadelphia. South et al. (2015) found that views of a greened vacantlot was associated with a signiﬁcant reduction in heart rate and concludedthat remediating neighborhood blight may reduce stress and improve health. ACANT LOT GREENING IN PHILADELPHIA Feelings of depression and self-reported poor mental health were reduced inparticipants living near greened vacant lots (South et al., 2018).Nevertheless, our matching analyses indicate promising results that high-light how the PHS Landcare greening intervention is associated with signif-icant reductions in crime. This research can be used to better inform thePennsylvania Horticultural Society as well as other interested parties ongreening practices to provide the greatest beneﬁt to the safety of local ur-ban areas in Philadelphia. The code repository for data processing and anal-ysis can be viewed at https://github.com/jessecui/WSII-Urban-Analytics-Business-Vibrancy.
We thank John MacDonald, Ajjit Narayananand Park Sincharisi for helpful contributions to this research. We are alsograteful to the Wharton Social Impact Initiative for their generous supportof our work.
Branas CC, Cheney RA, MacDonald JM, Tam VW, Jackson TD and Ten Have TR (2011)A Diﬀerence-in-Diﬀerences Analysis of Health, Safety, and Greening Vacant UrbanSpace. American Journal of Epidemiology 174(11): 1296–1306. . URL https://doi.org/10.1093/aje/kwr273 .Branas CC, South E, Kondo MC, Hohl BC, Bourgois P, Wiebe DJ and Macdonald JM(2018) Citywide cluster randomized trial to restore blighted vacant land and its eﬀectson violence, crime, and fear. Proceedings of the National Academy of Sciences 115:2946–2951.Deutsch W (2016) Crime prevention through environmental design. The Balance .Heckert M and Mennis J (2012) The economic impact of greening urban vacant land: Aspatial diﬀerence-in-diﬀerences analysis. Environment and Planning A: Economy andSpace 44(12): 3010–3027. . URL https://doi.org/10.1068/a4595 .Heinze JE, Krusky-Morey A, Vagi KJ, Reischl TM, Franzen S, Pruett NK and ZimmermanMA (2018) Busy streets theory: The eﬀects of community-engaged greening on violence.American Journal of Community Psychology 62: 101–109.Humphrey C, Jensen ST, Small DS and Thurston R (2020) Urban vibrancy and safetyin philadelphia. Environment and Planning B: Urban Analytics and City Science :2399808319830403.Jacobs J (1961) The Death and Life of Great American Cities. Random House, New York.Kondo M, Hohl B, Han S and Branas C (2016) Eﬀects of greening and community reuseof vacant lots on crime. Urban Studies 53: 3279–3295.Kondo M, Morrison C, Jacoby SF, Elliott L, Poche A, Theall KP and Branas CC (2018)Blight abatement of vacant land and crime in new orleans. Public Health Reports 133:650–657.MacDonald J (2015) Community design and crime: The impact of housing and the builtenvironment. Crime and Justice 44(1): 333–383.MacDonald J, Branas C and Stokes R (2019) Changing Places: The Science and Art ofNew Urban Planning. Princeton University Press. Moyer R, MacDonald JM, Ridgeway G and Branas CC (2019) Eﬀect of remediatingblighted vacant land on shootings: A citywide cluster randomized trial. AmericanJournal of Public Health 109: 140–144.Rosenbaum PR and Rubin DB (1983) The central role of the propensity score in obser-vational studies for causal eﬀects. Biometrika 70: 41–55.South EC, Hohl BC, Kondo MC, MacDonald JM and Branas CC (2018) Eﬀect of GreeningVacant Land on Mental Health of Community-Dwelling Adults: A Cluster RandomizedTrial. JAMA Network Open 1(3): e180298–e180298. . URL https://doi.org/10.1001/jamanetworkopen.2018.0298 .South EC, Kondo MC, Cheney RA and Branas CC (2015) Neighborhood blight, stress,and health: A walking trial of urban greening and ambulatory heart rate. AmericanJournal of Public Health 105(5): 909–913. . URL https://doi.org/10.2105/AJPH.2014.302526 . PMID: 25790382.ACANT LOT GREENING IN PHILADELPHIA Supplementary Materials for “The Eﬀectsof Vacant Lot Greening and the Impact ofLand Use and Business Vibrancy”
8. Crime Comparisons between Unmatched Greened and Un-greened Vacant Lots.
Figure S1 compares the distribution of crimecounts in a 200 meter radius around greened lots vs. ungreened lots in thetime periods before and after either the greening intervention for that lot(in the case of greened lots) or October 30th, 2012 (in the case of ungreenedlots). We observe similar patterns for greened and ungreened vacant lots: adecrease in crime counts in the “after” time period compared to the “before”crime period. The average after-before reduction in total crimes is -16.19 forgreened lots and -20.14 for ungreened lots, which gives a DoD estimate of anincrease of 3.95 total crimes for the eﬀect of vacant lot greening. This DODanalysis clearly does not provide evidence for beneﬁcial eﬀects of vacant lotgreening on crime.
Fig S1: Distribution of violent and non-violent crime counts within 200m radius ofall greened and ungreened vacant lot locations in Philadelphia
9. Land Use Zone Comparisons Between Greened and UngreenedVacant Lots.
Figure S2 displays side-by-side boxplots comparing greenedvs. ungreened vacant lots in terms of land use proportions that we createdfrom the City of Philadelphia zoning data. . We observe that greened va- Fig S2: Land use proportions (based on zoning data) for greened versus ungreenedvacant lots cant lots tend to occur in neighborhoods with less commercial and industrialzones but more cultural/park zones compared to ungreened vacant lots. Thevariance of transportation land use around greened lots is smaller than thevariance of transportation land use around ungreened lots. We also observethat greened lots tend to have a higher proportion of surrounding vacantland use than ungreened vacant lots.
10. Business Vibrancy Comparisons Between Greened and Un-greened Vacant Lots.
Figure S3 displays side-by-side boxplots compar-ing greened vs. ungreened vacant lots in terms of the number of businesses ina 200 meter radius. We observe a smaller number of businesses surroundinggreened lots compared to the number of businesses surrounding ungreenedvacant lots, which is also evident in the lower commercial proportion aroundgreened lots seen in Figure S2.Figure S4 gives barplots that compare greened and ungreened vacant lotsin terms of the proportions of each business type within a 200 meter radius.We observed that greened vacant lots have a lower proportion of conveniencestores, gyms, liquor stores compared to ungreened vacant lots.
11. Logistic Regression Model for Propensity Scores.
We used alogistic regression model to calculate the propensity scores for each greenedand ungreened vacant lot in our data. In this logistic regression model, eachunit i is a vacant lot in the city of Philadelphia with outcome Y i = 1 if vacant ACANT LOT GREENING IN PHILADELPHIA Fig S3: Number of Businesses surrounding greened vs. ungreened vacant lotsFig S4: Proportions of each business type surrounding of greened and non-greenedvacant lots lot i was greened or Y i = 0 if vacant lot i was ungreened. The probability P ( Y i = 1) for each vacant lot i is modeled as a function of its surroundingcharacteristics X i which includes our demographic, economic, land use andbusiness vibrancy measures outlined in the data section of our paper.The Other land use zoning proportion and one poverty bracket (income topoverty line above 2.00) were removed from the model due to high collinear-ity with the other surrounding characteristics. We also removed the indica-tors for retail business from the model since almost all vacant lots had atleast one of this type of business in their surrounding area.In Table S1, we provide several common evaluation metrics of our ﬁt-ted logistic regression model. For the “Accuracy” and “Balanced Accuracy” metrics, we chose the decision boundary to be the proportion ( p = 0 .
22) ofall vacant lots in our dataset that are greened. We compare the ﬁtted logis-tic regression model that uses all available measures of the surrounding area(“All”) to ﬁtted logistic regression models that only use one set of measures(e.g. “Economic” vs. “Demographic”).
Evaluation Metrics of Logistic Regression Models for diﬀerent sets of includedsurrounding characteristics.
Metric All Economic Demographic Land Use BusinessROC AUC
Pos Pred Value
Neg Pred Value
We can see that the logistic regression model that uses all surroundingcharacteristics has the best ﬁt to the data by all evaluation metrics, whichsuggests that each data type is making a signiﬁcant contribution to themodel. Amongst the models based just on single data type, the land usezoning characteristics seem to provide the best ﬁt to the data.These observations are conﬁrmed when we compare the ROC (ReceivingOperating Characteristic) curves for these diﬀerent ﬁtted logistic regressionmodels in Figure S5. We see that the model using “All” surrounding charac-teristics has the best ROC curve, followed by the model that only uses thesurrounding land use zoning proportions.In Table S2, we examine the estimated coeﬃcients from the ﬁtted logis-tic regression model that uses all surrounding characteristics. We see thatcoeﬃcients with the largest (in magnitude) statistics are income per capita,the racial proportions, population count and indicators for several of thebusiness types.
Jesse CuiWharton Social Impact InitiativeThe Wharton SchoolUniversity of Pennsylvania400 Jon M. Huntsman Hall3730 Walnut StreetPhiladelphia, PA 19104Email: [email protected] Shane T. JensenDepartment of StatisticsThe Wharton SchoolUniversity of Pennsylvania400 Jon M. Huntsman Hall3730 Walnut StreetPhiladelphia, PA 19104Email: [email protected]
ACANT LOT GREENING IN PHILADELPHIA Fig S5: ROC Curves for Diﬀerent Covariate Groups Predicting Greening in Lots Table S2
Summary of Coeﬃcients from Logistic Regression Model using all surroundingcharacteristics
Coeﬃcient Estimate Std. Error Stat P-value(Intercept) -2007.954719 1132.887511 -1.772421975 0.076324537
Residential block total count white percent black percent asian percent hispanic percent block per capita income -1.466296008 0.103165642 -14.21302652 7.61E-46 income to poverty under .50 -1.65123502 0.270758785 -6.098546426 1.07E-09 income to poverty .50 to .99 income to poverty 1.00 to 1.24 -2.243585887 0.336208549 -6.673197026 2.50E-11 income to poverty 1.25 to 1.49 -0.69900368 0.330699879 -2.11371012 0.034540034 income to poverty 1.50 to 1.84 income to poverty 1.85 to 1.99 cafe convenience -0.057606872 0.048074951 -1.198272096 0.230811106 gym -0.511807752 0.043691912 -11.71401592 1.08E-31 institution liquor -0.293541954 0.046507647 -6.311692183 2.76E-10 lodging -0.018729199 0.043165373 -0.433894063 0.664365371 nightlife pharmacy restaurant retail business countbusiness count