Interpolating Resident Attitudes Toward Exurban Roadside Forest Management

Context: Knowledge about spatial patterns of human dimensions data within landscape ecology is nascent despite its importance in natural resources management decision-making. We explored this topic within the context of utility roadside forest vegetation management, a complex situation involving ecological, cultural, and aesthetic aspects of forests and reliable power. Objectives We applied spatial interpolation to investigate patterns of human attitudes toward exurban roadside vegetation management data across an exurban landscape. Methods Mail surveys (n = 1962) were used to collect social science data from residents in four areas of Connecticut, USA. For each area, three attitudes variables were evaluated for spatial autocorrelation using Moran’s I statistic. Based on identied autocorrelation distance or scale, attitudes were interpolated using inverse distance weighting. Model validation of interpolated surfaces was completed using root mean square error. Results: Statistically signicant spatial autocorrelation was present for ve of 12 study area-attitude pairings at variable distances. Accuracy of interpolations also varied among study areas, suggesting that the choice of spatial scale of analysis inuenced model results. Conclusions: Social processes within the exurban landscape were spatially heterogeneous and multiscalar for the same variables in different locations, exemplifying the complexity of social processes within exurban land use. Interpolation assumptions often applied toward ecological studies did not work well for social processes studied in this analysis. Results demonstrated the importance of understanding spatial dimensions at which social processes operate and, therefore, may inuence ecological outcomes of the roadside forest within the context of state-level natural resources management and policy.

Natural resource decision-making at the landscape level often requires spatially explicit empirical data, which can be expensive and time-intensive to obtain (Ohmann and Gregory 2002;Bell et al. 2015). For social science, incomplete coverage, lack of longitudinal data, uncertainty of scale, and inconsistent data quality create additional challenges for integrating social science data within broader landscape analyses (e.g., Redman et al. 2002;Collins et al. 2011;Rounsevell et al. 2012; Bowen et al. 2016;Elsawah et al. 2020). When empirical data are lacking, tools and techniques such as spatial interpolation allow managers to construct inferences for sample-based data (Azpurua and Ramos 2010); biophysical examples include predicting forest composition across regions (Ohmann and Gregory 2002), interpreting changes in seasonal rain patterns (Camera et al. 2014), and assessing distribution of groundwater contamination (Gong et al. 2014) and temperature gradients across regions (Kim et al. 2010). The few applications of such tools and techniques to explore social science connections to ecological processes at the landscape level include assessing where human-perceived and physically measured ecological values overlap in socioecological hotspots (Alessa et al. 2008), and understanding visitor movement for improving park and protected area management (Beeco and Brown 2013). Despite this knowledge gap, such examples demonstrate potential for addressing social data challenges to strengthen bidirectional linkages between social science and landscape ecology. In this study, we focused on the context of utility vegetation management to explore spatial characteristics of human dimensions of roadside forest management at the landscape level. Studies focused on social dynamics of vegetation management have indicated preferences for taller street trees (Schroeder 1989), shorter trees to decrease possible powerline obstructions (Flowers and Gerhold 2000), removing trees deemed hazardous to homes (Conway 2016), and perception that utility pruning harms aesthetics of roadside trees (Kuhns and Reiter 2007). Human dimensions studies of vegetation management to the landscape context have suggested that attitudes toward vegetation management and its effects on the roadside forest were more likely to be in uenced by social-psychological variables than residential context characteristics (Hale and Morzillo 2020; Kloster et al. 2021), and that factors in uencing attitudes vary across study locations (DiFalco and Morzillo 2021). Therefore, human dimensions data provide information and context for potential scalar alignment of social and ecological processes associated with vegetation management.
Our objective was to investigate spatial patterns of attitudes toward vegetation management with speci c focus on spatial proximity and scale. To do this, we used spatial interpolation, a technique facilitated by spatial autocorrelation and based on the fundamental geographic principle that points, or phenomena at points, closer together are more related than those further away (Tobler 1970). Previous human dimensions research has supported this principle, such that individuals located closer together are more likely to hold more-similar attitudes regarding a natural resource issue than those located further away (Berenguer et al. 2005 Results advance our understanding of the patterns of social processes across multiple spatial scales, and their interplay with associated short-and long-term roadside forest management planning goals.

Study context
Connecticut is a small northeastern US state (14,357 km 2 ) that has experienced rapid population growth since the 1950s, much of which has occurred in exurban areas outside of cities (Brown et al. 2005). The integration of exurban development, extensive forest cover (72.6% of the state, Nowak

Data collection
Four geographically distinct study areas in Connecticut (Northeast, Southwest, Northwest, Southeast; The sampling frame included a list of all residential street addresses within the four study areas, and the sampling unit was the individual household. Addresses used for mailing were purchased from Marketing Systems Group (Horsham, PA), which uses US Postal Service delivery routes to generate address lists. To focus sampling on residents most likely involved in tree management decisions at the property level, an effort was made to select single-family owner-occupied addresses. Post o ce boxes, seasonal homes, mail drops, and vacant homes were excluded from the sample. Based on expected response rate and a desired sampling error of α = 0.05 (95% con dence interval; Sheskin 1985), 1800 surveys were mailed to each study area. To ensure coverage across the urban-to-rural gradient, surveys were sent to an equal number of urban and rural respondents, as determined by the 2010 Census classi cation of urban and rural (U.S. Census Bureau 2011).
A modi cation of the Tailored Design Method was applied to data collection (Dillman et al. 2009). Multiple mailings were used as an effort to increase response rate, which included a: 1) pre-notice postcard to introduce the project, 2) packet containing a cover letter, survey and pre-paid return envelope, 3) reminder/thank you postcard, and 4) second survey packet to those who had not yet responded. To evaluate potential for non-response bias, non-respondents to the original survey received a short followup mail survey focusing on ten key items from the original survey. The University of Connecticut Institutional Review Board (IRB) granted permission for use of human subjects (IRB #H16-007).

Variables
Attitudes measure favor or disfavor toward a person, object, event, or situation (Fazio et al. 1982). Previous topics related to this study have included attitudes related to urban tree maintenance ( To assess attitudes toward vegetation management, we measured respondent agreement with a series of attitude statements on the survey. Responses were coded using a ve-point Likert scale measuring level of agreement (5 = strongly agree; 1 = strongly disagree). Principle component analysis (PCA) with varimax rotation was used for data reduction to identify attitude statements that factored together. Cronbach's alpha (α) was used to test the internal reliability of groups of statements that factored together (Cortina 1993  To describe respondents, data were collected for ve sociodemographic variables (Table 1), as previous research suggested these variables may in uence attitudes toward natural resources (e.g., Morzillo et al. 2010; Keener-Eck et al. 2020). Data collected for each respondent included: respondent sex (Sex; male or female), year they were born (Age in years), and length of time lived at their current address (Tenure in years). Education was represented as the highest education level selected among seven formal education levels: (a) Less than high school, (b) High school or equivalent (e.g., GED), (c) Some college, (d) Vocational or trade school, (e) College degree (2-year or certi cate), (f) College degree (Bachelor's), or (g) Graduate or professional degree; Education was represented as the highest education level selected. For household income (Income), respondents selected from among ve income groups ranging from <$25,000 to ≥$100,000. The rst step of this analysis was to identify the scale or distance at which each attitude variable was spatially autocorrelated for each study area. For each of the 12 attitude-study area pairings, we performed incremental spatial autocorrelation (ISA) in ArcGIS 10.7.1 at 100 m increments, ranging from 100 m to 3000 m, to identify the distance band of maximum spatial autocorrelation (Fig. 2). ISA measured spatial autocorrelation at each distance and computed a Moran's I and z-score for each distance (Carter et al., 2014). In this analysis, we used a 95% con dence interval to identify signi cant spatial autocorrelation, such that z-scores > 1.96 represented clustering of attitudes, and z-score < -1.96 represented a dispersed pattern of attitudes. Only pairings for which ISA produced a spatial autocorrelation value were assessed in subsequent steps.

<Fig. 2>
For the second step of this analysis, inverse distance weighting (IDW) interpolations were completed in ArcGIS 10.7.1 using a xed distance band for each attitude identi ed as having spatial autocorrelation ( Table 2). The xed distance band was equal to the distance of spatial autocorrelation for each attitudestudy area pairing. Raster output was set at the 30m 2 cell size to match the National Land Cover Database (NLCD) cell size (Homer et al. 2020). One-fourth of the survey respondents from each paired analysis was reserved for use as an independent data set for model validation (Verbyla and Litvaitis 1989). The "Geostatistical Analysis Layer to Points" tool in ArcGIS, which extracts a predicted attitude score for each respondent in the independent dataset based on the IDW interpolation, was used to validate the interpolated surface. Model performance was measured using root mean square error  where p i is the predicted attitude score for the respondent at location i, a i is the actual attitude score for the respondent at location i, and n is the number of respondents in the subset. Spatial analyses were completed using ESRI ArcGIS 10.8 and Python 2.7 (ArcPy module).  √ ISA analysis indicated the presence of spatial autocorrelation for ve of the 12 attitude-study area pairings tested (Table 2). For those ve pairings, distances of maximum autocorrelation varied by study area. Maximum spatial autocorrelation for AttProfessional was identi ed in the Northeast study area at 2400 m. AttSafety reached maximum spatial autocorrelation in the Northwest at 200 m, and in the Northeast at 300 m. ISA identi ed two different distances of signi cant spatial autocorrelation for AttTradeoff in the Southeast study area (600 m and 1500 m), as well as maximum spatial autocorrelation for AttTradeoff in the Southwest study area at 2200 m. Two of the ve pairings showed negative spatial autocorrelation (i.e., dissimilar attitude scores are clustering together): AttProfessional-Northeast (Moran's I = -0.022), and AttTradeoff-Southwest (Moran's I = -0.044, Table 2).
Interpolated surfaces indicated the spatial variation of attitudes scores across study areas (Figs. 3 & 4). IDW surfaces varied in spatial coverage depending on the distance band used for interpolation. For the Northeast, maximum autocorrelation distance bands for AttProfessional and AttSafety occurred at different distances; thus, those two attitude variables had different spatial coverages for the same location ( Fig. 4). At short distances (i.e., 200 m and 300 m) attitudes did not interpolate extensively across the study area extents for the Northwest and Northeast, respectively (Fig. 4).
<Fig. 3> <Fig. 4> Model validation indicated differences in interpolation accuracy among attitude variables and study area pairings ( Table 2). Only independent data points located inside of interpolated surface were included in RMSE calculations, which varied by distance band used in the interpolation. For the Northwest, 14.7% of independent data points were included in model validation at 200 m. For both Northeast at 2400 m and Southwest at 2200 m, all 100% of independent data points were included (Table 2). RMSE scores ranged from 2.66-5.48, which also indicated variation in interpolation accuracy.

Discussion
We applied spatial interpolation to understand spatial patterns of human dimensions data, with focus on proximity and scale of attitudes metrics within the context of utility vegetation management of roadside forests. Complex levels of spatial heterogeneity existed, such that attitudes toward vegetation management varied across space, and likely are associated with location-speci c characteristics. Preceding analyses suggested that respondents had generally favorable attitudes toward vegetation management across all study areas, interplay of social and landscape factors affected individual attitudes (Hale and Morzillo 2020), and variation in social variables in uenced attitudes among study areas (DiFalco and Morzillo 2021). In our analysis, maximum spatial autocorrelation distances for the three attitudes variables among the four study areas supports multiscalar spatial variation of people's attitudes toward vegetation management, which may hinder success of a state-level one-size-ts-all vegetation management policy. To focus the discussion, we describe two underlying social phenomena that may contribute to the observed spatial heterogeneity among locations and offer direction for further analysis.
Supporting our rst hypothesis that social phenomena closer together are more likely to be related than those further away, clustering of similar attitude scores existed for three of the 12 attitude-study area pairings (AttSafety-Northwest, AttSafety-Northeast, and AttTradeoff-Southeast). Contrasting the same hypothesis were two pairings with negative autocorrelation (AttProfessional-Northeast, and AttTradeoff-Southwest) and seven with no autocorrelation. Spatial clustering of favorable attitudes toward natural resources has been observed in other natural resources contexts. Attitudes toward tigers in Nepal were clustered based on human cultural factors, educational achievement, and experience with tiger attacks (Carter et al. 2014). Attitudes toward the desert were spatially clustered within neighborhoods of similar social and landscape characteristics; more favorable attitudes existed in high-income areas closer to preserved desert parks ( However, ancillary evidence from elsewhere in our analysis (DiFalco and Morzillo 2021) did not suggest clustering based on self-selected urban-rural residential designation (LocReside; Table 1), or landscape characteristics surrounding respondent homes, except for AttTradeoff-Northwest which was positively associated with a greater percentage of tree cover (DiFalco and Morzillo 2021). Areas of negative spatial autocorrelation of attitudes demonstrates diversity of attitudes among proximate individuals, and potentially that processes observed in one area are in uenced by neighboring areas (Gri th and Arbia 2010). It is possible that vegetation management actions completed by homeowners may affect and be affected by attitudes of their neighbors (e.g., Belaire et al. 2016), but support for this conclusion is beyond the scope of our data. Collectively, the mixed results from our study exempli ed the heterogeneity and complexity of social processes that occur within exurban land use (e.g.,  ) suggested that municipaland state-level land-use zoning affected household and neighborhood vegetation structure in the city of Baltimore, where a policy was implemented to increase urban tree canopy. Findings from that study suggested that most tree plantings would need to occur on private residential properties, thus relying on property-scale decision-making to achieve municipal goals, and demonstrating the importance of multiscale processes in policy outcomes (Chowdhury et al. 2011).
Besides zoning, other governance structures such as town ordinances present additional layers of complexity to governance and therefore spatial distribution of tree plantings and removals on both public and private properties (Johnson et al. 2020). Comments from our survey suggested overall confusion about jurisdictional coordination of vegetation management: We are perplexed by the randomness of activity in roadside tree removal/trimming. Is there an overall state or town plan for a comprehensive and methodical approach?
In Connecticut, some towns have speci c ordinances in place that delay or prohibit implementation of utility practices related to state vegetation management guidelines. For example in the Northeast study area, the towns of Mans eld and Coventry encourage maintaining a closed forest canopy to preserve the aesthetic quality of forested scenic roadways (Town of Mans eld 1995; Town of Coventry 1997).
However, in the same study area, the adjacent towns of Bolton and Andover do not have analogous regulations. Another commonly mentioned suggestion is to bury powerlines underground, which is now Diversity in respondent attitudes toward vegetation management revealed by interpolation results also may re ect differences in regional culture across the state. Others have reported direct relationships between greater household incomes and a greater likelihood to plant trees in neighborhoods that sustain greater tree cover (Conway et al. 2011;Nitoslawski et al. 2016 Tree management to reduce power outages is important to us since we are one of the last areas to have power restored [after storm events] because of our sparse population. We live in a rural setting. This management should be done in an environmentally responsible way.
Maintaining the aesthetic character of an area also was expressed by respondents as an important outcome and existed concurrently with respondent understanding of the necessity for vegetation management: Trees are so important but tree health is just as important. The woods, land preserved and forests should remain untouched. Trees in residential areas should be maintained for health and resident safety. I would happily allow the utilities to remove the tall trees in my front yard that stand tall enough to fall on the lines but also appreciate the other trees around my property.
This parallels previous research from New England on the topic of developing strategies for maintaining the rural character of a town while balancing economic growth and development (Zabik and Prytherch 2013). Our results also suggested opportunity for integrating homeowner preferences into management plans, such as incorporating desired visual outcomes of tree trimming and potentially replace taller trees with shorter statured species less likely to interfere with powerlines, echoing ndings by Flowers and Gerhold (2000).
Limitations in this study offer opportunities for better integration of social and ecological data in landscape ecology. First, the geographic assumption that points closer together are more similar (Tobler 1970; Differences in model performance among attitude-study area pairing also alluded to opportunities to improve landscape analytics for assessing social phenomena. For example, Gong et al. (2014) found that interpolations to assess contamination of groundwater wells across Texas were more accurate when data were divided into regional districts (i.e. speci c aquifers or areas of the state) rather than pooled for the whole state -i.e., adjustments made to level of analysis. Bhowmik (2012) concluded that performance between climate prediction models and actual weather conditions improved as additional meteorological stations were added over subsequent years through increasing the number of data points. In our analysis, interpolated surfaces were more likely to predict the actual attitude score in portions of our study areas where the distance between respondents was smaller, suggesting that additional data points or oversampling among locations where households are further apart may be useful in areas of heterogeneous land use and development density, such as exurban landscapes. Future studies may also consider experimenting with mismatches between scale of data collection and social processes (Robinson et al. 2019) to better align methodological design of data collection to the spatial scale of the social process being assessed.
Results of this analysis supported the creation of multi-scalar strategies for roadside vegetation management that consider not only landscape-level decision-making but also local-level stakeholder heterogeneity outreach strategies and protocols that re ect and respond to both regional and local variation.  Output from incremental spatial autocorrelation for the ve attitude-study area pairings with spatial autocorrelation: A) AttProfessional-Northeast, B) AttSafety-Northwest, C) AttSafety-Northeast, D) AttTradeoff-Southeast, and E) AttTradeoff-Southwest. Black circles denote distances of spatial autocorrelation determined by Moran's I statistic. Open circles are distances without identi ed spatial autocorrelation.