Urban forest biodiversity and cardiovascular disease: Potential health benefits from California’s street trees

We first assess the sensitivity of our results to our biodiversity measure and spatial source of identifying variation. Column two of Table 3 uses the Shannon index for biodiversity at the species, instead of genus, level. The effect is slightly smaller but similar to the main estimate. Column three removes fixed effects at the 3-digit zip code level and thus identifies the correla- tion off of variation across California. Removing the 3-digit zip code fixed effects fails to change the estimate substantially. Next, we use a blunt approach to assessing selection bias in our sample. Column four restricts the sample to zip codes with at least one thousand trees sampled in our dataset from private arborists. As shown in S1 Fig, there is a positive correlation between the Shannon Index and trees per 100,000 individuals. By imposing this threshold, we likely lose variation in the Shannon Index, reducing power. The estimate is still negative for Heart Disease, but it is near zero for Stroke and, unsurprisingly, both estimates are imprecisely estimated. This sug- gests that the zip codes with few trees are important when estimating our effect. This could be because of a threshold effect. In the non-linear specification in columns two and five of Table 2, the coefficients for Shannon II, III, and IV are not very different from one another. By removing the zip codes with the fewest trees, it is akin to comparing across these quartiles. We also lose power, having likely limited variation in the Shannon Index. Alternatively, given that

the arborist sampling frame is not randomly assigned, the change in the coefficient estimate could reflect omitted variable bias if arborists are less likely to collect tree inventory data in zip codes with high mortality rates. Another explanation of the difference in the trimmed and full sample would be heteroge- neous treatment effects. Comparing OLS and weighted least squares helps assess if this is likely. Column five weights the OLS regression estimate by population, giving greater weight to zip codes with larger populations. This decreases the coefficient estimates by about half, though they remain negative and precisely estimated. As noted, differences in weighted least squares and ordinary least squares estimates may indicate model misspecification [52,53]. In our case, the direction of the effect and the precision of the estimate is similar across the two specifica- tions, suggesting that our estimates are not sensitive to assumptions in the basic OLS model. However, we find that magnitude of the effect differs, suggesting some heterogeneity in the effect of biodiversity on the mortality rate. To address omitted variable bias in our sample, columns six and seven add socioeconomic and pollution burden controls. The controls in column six include population density, median household income, mean commute time, and the share of the population that were Black, Hispanic, or Asian in the 2010 census. Adding these controls increases the magnitude of the estimate and slightly decreases the precision of the heart disease estimate. Column seven adds

the same controls as column six as well as the zip code’s Pollution Burden Index. Adding the Pollution Burden Index slightly reduces the magnitude of the coefficients from that of column six. Finally, column eight and S2 Table addresses spatial autocorrelation and S3 Table addresses aggregation bias. Column eight reports the total effect from a linear spatial autoregression model using data for 2018. The model includes spatial lags on the Shannon Index and error term and the total effect is the linear combination of the direct effect on mortality in a given zip code and indirect effects on correlated zip codes. The coefficient is similar in magnitude and precision to our main specification. Results for each year are very similar and are reported in S2 Table in Supporting Information. We also subsampled our data by two-digit zip code and report estimates using our main specification in S3 Table. We lose power by restricting the sample, and the estimates are less precise, particularly for Stroke. For Heart Disease, all but one of coefficients is negative and the association seems stable across two-digit zip code areas. PLOS ONE Biodiversity and cardiovascular disease PLOS ONE | https://doi.org/10.1371/journal.pone.0254973 November 3, 2021 12 / 23