For the full panel of tracts and years, I ran nested OLS models with the distance to the nearest garden as the dependent variable and independent variables of percent in poverty, median monthly housing costs, percent with a college degree, percent Black, percent Hispanic, percent Asian or Pacific Islander, and percent foreign born. I then added the distance to downtown variable, and finally the variable for year of measurement to assess and control for any change in overall accessibility over time. Quantile-quantile plots show that, as expected, controlling for distance to downtown greatly reduced the right-skew of the residuals resulting from the consistent under-service of far-flung areas. However, due to the clustering of high-poverty neighborhoods relatively close to downtown, controlling for distance to downtown also reversed the coefficients for the effect of percent in poverty. Therefore I chose to continue nesting models with and without distance to downtown in order to gain an accurate picture of garden accessibility for low-income residents. Results showed that adding year to the model generally did not have a strong impact , rolling benches hydroponics but did yield slight improvements in model fit, so I retained year as an independent variable in subsequent models.
To determine whether correlations between percent foreign born and racial composition would interfere with modeling, I looked at variance inflation factors and tested the impact of removing percent foreign born, percent Hispanic, and percent Asian or Pacific Islander. I ran separate models for each city due to differences in their immigrant composition. In the full model, variance inflation factors were consistently below 10 for all racial groups and percent foreign born, indicating that multicollinearity was likely not a problem. Running the models with each variable removed, I confirmed that one was not masking the effect of another in any of the cities. Finally, in order to assess whether OLS or spatial models would be more appropriate, I ran Lagrange multiplier tests following Anselin et al. . Lagrange multiplier tests consistently indicated that spatial modeling would be more accurate than OLS modeling. When the variable for distance to downtown was included to correct for some of the spatial autocorrelation in the dependent variable, the Lagrange multiplier test results showed more significance for a spatial error model than for a spatial lag model. Spatial lag models measure the spillover effects of dependent variable values from adjacent tracts, while spatial error models treat these effects as random error. In accordance with the Lagrange multiplier results, I chose to run spatial error models and to include distance to downtown.
For this study, the relationships of interest are between the independent variables and the dependent variables , so any spillover effects in garden accessibility between adjacent tracts can reasonably be treated as a nuisance to control for—as with spatial error modeling—rather than being measured and reported as with spatial lag models.Overall, spatial error models indicate that the gardens of each program are closer to neighborhoods with higher proportions of Black and Hispanic residents, as shown in Table 1. Models for the three cities are less consistent regarding the accessibility of gardens for Asian and Pacific Islander communities, immigrants, and people in poverty. In Milwaukee, the spatial error model predicts that with all other factors held constant for a neighborhood, the nearest garden will be about 8 meters closer for every 1% increase in the share of the tract’s residents who are Black, 13 meters closer for every 1% increase in the share who are Hispanic, and 46 meters closer for every 1% increase in the share who are Asian and Pacific Islander. In Milwaukee, regression results also suggest that low-income neighborhoods have been relatively well served by the city’s garden programs7. The model predicts that the nearest garden will be about 10 meters closer for every 1% increase in a tract’s poverty rate, with all other factors held constant. The opposite is true for the city’s immigrant population, however; a 1% increase in percent foreign born is associated with an increase of about 24 meters to the nearest garden. Other factors thought to influence the creation and maintenance of gardens—that is, real estate values and levels of cultural capital—do not appear to have a strong effect on garden locations in Milwaukee, as neither is significant at the p<0.05 level.
The model does indicate that the reach of the city’s main garden programs has changed over time: on average, with other factors held constant, every additional year is associated with about 14 meters more distance to the nearest garden. This suggests that Milwaukee’s main garden programs have not been able to increase their gardens’ proximity to residents over time, and in fact the garden distribution is becoming less accessible overall. Given that the policy victories achieved in Milwaukee have done the least to secure permanent land tenure, this finding makes sense. In contrast, Philadelphia’s garden distribution appears to be more accessible over time, with a decrease of nearly 31 meters to the nearest garden on average for every year, with other factors held constant. Similar to Milwaukee, neighborhoods with more Black and Hispanic residents tend to have closer gardens . The proportion of Asian and Pacific Islander residents and the percent foreign born residents are not significantly associated with garden proximity in the spatial error model for Philadelphia. While low-income neighborhoods in Milwaukee appear to have gardens closer to them, in Philadelphia higher poverty rates are associated with the nearest garden being further away—about 18 meters further away for every 1% increase in the poverty rate, with all other variables held constant. As noted in the previous section, the association between poverty rates and garden proximity in Philadelphia reverses direction when the model controls for distance to downtown. Many of the city’s low-income neighborhoods are relatively close to Center City, which may help explain the reversal; however, the clustering of poverty near downtown does not entirely explain the relationships observed. In Philadelphia, the Pearson correlation between poverty rate and distance to downtown is negative , but at -0.013, this correlation is quite weak. Examining maps of poverty rates and garden locations in Philadelphia over time shows that several high poverty tracts further from Center City have never had gardens nearby, cannabis indoor grow system while the gardens near Center City are often just as close to low-poverty tracts as they are to high-poverty ones. In general, the spatial error model suggests that Philadelphia Green’s garden development efforts did not maximize the benefits of garden access for low-income communities, and in fact may have better served neighborhoods with lower poverty rates. In Seattle, poverty rates are not significantly associated with distance to the nearest garden, suggesting that gardens are equally likely to be found in high-income and low-income neighborhoods. As in Milwaukee and Philadelphia, a higher share of Black or Hispanic residents is associated with closer garden proximity—about 11 meters for every 1% increase in Black residents and 22 meters for every 1% increase in Hispanic residents, with all other factors held constant. Similar to Philadelphia, associations with percent Asian and Pacific Islander and percent foreign born are not significant at the p<0.05 level. Other factors thought to be associated with garden development and maintenance, the neighborhood’s real estate values and cultural capital, do appear to be significant in Seattle. Every $1 increase in median monthly housing costs is associated with an increase of 0.75 meters to the nearest garden, meaning that a tract with median rent $100 more than an otherwise identical tract is predicted to be 75 meters further away from the nearest garden. For every 1% increase in the proportion of residents with a college degree, distance to the nearest garden is predicted to decrease by about 19 meters with all other factors held constant. This association suggests that cultural capital has been a factor in the development of P-Patch gardens, which is plausible given that much of the program’s expansion has relied upon local residents organizing to demonstrate demand and seek grants to help build new gardens.To understand the impact of organizational framing and decision-making on the equity in garden distributions, we should consider the possibility that proximity to various communities of interest may have changed over time.
Expanding the models with interaction terms for year and the main independent variables , I assessed ways in which organizational priorities may have impacted garden distributions as they changed over time.Table 2 shows the results for spatial error regression models including interaction terms for year and the main independent variables. For all three cities, adding in the interaction terms produced better fitting models as shown by their lower Akaike Inference Coefficients. In general, the results are consistent with the original spatial error models, though the interaction terms appear to dilute some relationships while revealing others. In the case of Milwaukee, none of the interaction coefficients are significant at the p<0.05 level. This suggests that the development and preservation of gardens has not followed a clear trajectory in terms of targeting specific communities more deliberately over time. Milwaukee’s main gardening programs have highlighted the value of urban agriculture for lowincome people, and throughout my interviews advocates referred to the Near North Side, where poverty is highest and where the city’s Black residents are concentrated, as the area most “in need” of the benefits that community gardens can bring. On the one hand, a relatively static pattern of garden accessibility in Milwaukee may suggest the city’s main programs were succeeding in their goals from the outset. Indeed, maps of Milwaukee’s gardens over time show that the Near North Side has consistently hosted a substantial share of the gardens developed by Shoots n Roots from the 1970s to the 1990s and Milwaukee Urban Gardens from 2000 onward. The spatial error model without interaction terms suggests that Black and poor communities have been well served by the program, with significant negative coefficients for both percent in poverty and percent Black, as described above. On the other hand, a lack of significant interaction terms may suggest that program leaders have been unable to undertake a concerted effort at reaching potentially underserved groups, such as immigrants.In Philadelphia, program leaders did undertake a concerted effort to bring the benefits of gardens to specific communities: the Greene Country Townes developed in the 1980s and early 1990s. However, as explained in Chapter 2, the Greene Country Townes initiative was as much about demonstrating the potential public benefit of a greening intervention as it was about improving the lives of poor and marginalized people. With interaction terms added to the spatial error model, higher poverty rates are still significantly associated with the Philadelphia Green program’s nearest garden being further away. There is not a significant effect of year on the association between poverty rates and garden proximity. Adding the interaction terms does reveal more about the relationship between garden locations and housing values. This relationship was not significant in the original spatial error model, but it is in the interaction model, which estimates that every $1 increase in housing costs in 1980 was associated with a decrease in 1.658 meters to the nearest garden. In other words, a neighborhood with $100 more in median monthly housing costs would be predicted to have a garden about 166 meters closer than an otherwise identical neighborhood. The interaction term for housing costs is also significant, and it suggests some attenuation of the underlying relationship over time. For every additional year after 1980, the model predicts that every $100 increase in median monthly housing costs would yield an increase of about 4 meters to the nearest garden. Taken together, the significant coefficients for housing cost and housing costs’ interaction with year suggest that the association between housing costs and garden proximity has gradually gone away over the last 40 years, to the point where housing costs should have no impact on distance to the nearest garden in 2020. Given that the Pennsylvania Horticultural Society did little to resist garden sales and lot redevelopment as real estate values rapidly appreciated in some neighborhoods—including former Greene Country Townes—the changing relationship over time makes sense. While gardens were established closer to high rent neighborhoods, they remained largely vulnerable to the development pressures of a growth machine continuously seeking higher rents and denser land use, and many have been replaced with housing over time.