COPY THY NEIGHBOR: THE ROLE OF SOCIAL NORMS ON THE ADOPTION OF SOLAR PANELS Hester Huisman

(1)

COPY THY NEIGHBOR: THE ROLE OF SOCIAL NORMS

ON THE ADOPTION OF SOLAR PANELS

Hester Huisman

S2718006

24-08-2020

Supervisors:

Prof. dr. J.E. Wieringa

Prof. dr. M. Mulder

University of Groningen

Faculty of Economics and Business

MSc Marketing

(2)

ABSTRACT

(3)

INTRODUCTION

To reduce greenhouse gas emissions, the European Union (EU) strives to produce 32% of EU energy from renewables in 2030. EU member countries have committed to binding national targets to raise their share of renewables in their energy consumption under the Renewable Energy Directive (EC, 2008). Currently, the Netherlands is the furthest away to reach its national objective of all member states (Eurostat Press Office, 2019).

One way for the Netherlands to reach this objective by incentivizing consumers to install photovoltaic (PV) panels on residential rooftops. This will convert consumers into ‘prosumers’, as they generate their own electricity and are able to export their residuals. The generated electricity can directly be consumed by the producing households. Consequently, the PV panels help to reduce the volatility on the transmission lines often caused by renewable energy sources (Razmara et al., 2017). Moreover, increasing voluntary adoption of PV panels by residents will contribute to reaching the Netherlands’ target of the Renewable Energy Directive, without being heavily dependent on government subsidies.

Current research has examined various incentives to motivate consumers to switch to renewable energy sources, including promoting the adoption of PV panels. Much research has focused on individual consumer characteristics, such as income (Batley, Colbourne, Fleming, & Urwin, 2001), social status (Batley et al., 2001), education level (Conte & Jacobsen, 2016; Tabi, Hille, & Wüstenhagen, 2014), stated climate concern (Arkesteijn & Oerlemans, 2005; Hobman & Frederiks, 2014; Newton, Tsarenko, Ferraro, & Sands, 2015; Tabi et al., 2014), and green perceived values (Y. S. Chen & Chang, 2012). These individual consumer characteristics are often found to be powerful to predict behavioral intention of switching to PV panels, but not to predict the execution of the behavior (Hobman & Frederiks, 2014).

(5)

influenced by the attitude towards the behavior, the subjective norm, and perceived behavioral control. Though the model is highly recognized (Nosek et al., 2010), there has also been criticism. The model provides accurate predictions for behavioral intention, but not for the actual behavior (Ajzen, 2011). When applied to behavior related to renewable energy (e.g. the adoption PV panels) the model shows a major gap between the behavioral intentions and the overt behavior of consumers (Claudy, Peterson, & O’Driscoll, 2013; Gupta & Ogden, Denise, 2006; Hobman & Frederiks, 2014). Psychological research refers to this as the value-action gap (Blake, 1999; Boulstridge & Carrigan, 2000; Flynn, Bellaby, & Ricci, 2009; Kennedy, Beckley, McFarlane, & Nadeau, 2009), the attitude-action gap (Boulstridge & Carrigan, 2000), or the intentions-action gap (Sheeran, 2002; Sheeran & Abraham, 2003). Because of this gap, only extreme attitudes are found to affect overt behavior (van Doorn, Verhoef, & Bijmolt, 2007).

(6)

of the effects of social norms on their behavior (Nolan et al., 2008). This explains the conflicting results with prior research, which did not find significant effects of self-reported social norms, e.g. Chen & Hung (2016). Therefore, it is important to measure consumer behavior in a social context.

There are similarities between the behavior of energy conservation and the adoption of PV panels as they are both related to similar pro-environmental norms. Therefore, the behavior of peers is likely to be a good predictor for the adoption of PV panels. However, the adoption of PV panels requires higher behavioral costs than energy conservation. The installation of PV panels involves a higher (initial) financial investment, higher risks due to higher switching costs and is not suitable for all residence types (such as rental houses). The effect of the energy conservation can therefore not be generalized to the adoption of PV-panels. This research will test if the social effects can predict the adoption of PV panels. To fully capture the effect of the social pressure, the adoption of PV panels should not be measured on a micro-level as it is too narrow to allow for a proper identification of the effect of social pressure. Instead the adoption of PV panels should be measured on a meso-level including the social system. To better understand the role of social pressure on a meso-level, this research will answer the question how do social dynamics affect the adoption of PV panels on a meso-level?

(7)

be perceived as ambiguous. In the same vein, Schaffer and Brun (2015) observed spatial contagion for the adoption of PV panels in Germany. They found an initially small spatial autocorrelation, which increased over time. They suggest drivers as uneven spatial diffusion of specified craft skills and intermediary agents such as solar initiatives for this autocorrelation, but leave this question open for future research. Balta-Ozkan et al. (2015) used spatial econometrics to examine the patterns of PV panel adoption among British residents. They suggest that the spatial effects could be related to contextual factors on similarities and the coordination of local policies by enterprise-partnerships or non-governmental organizations. They propose for future research to take a more local scope to investigate the nature of the spatial effects to determine whether these are peer effects or such contextual effects. Hence, a local scope is necessary to explore the drivers of the spatial autocorrelation found in prior research.

This is what Graziano and Gillingam (2015) did in their study on diffusion patterns of PV panels in Connecticut, USA. These show that smaller centers contribute to adoption more than larger urban areas, in a wave-like centrifugal pattern. The effect of nearby systems diminishes with distance and time, suggesting a spatial neighbor effect conveyed through social interaction and visibility. Thereby, the authors support findings of prior research of peer influence on PV installations (Bollinger & Gillingham, 2012; Müller & Rode, 2013; Rode & Weber, 2016).

(8)

local organizations promoting PV panels to be the most important factors explaining high local PV panel diffusion rates. The peer effects influenced the residential PV diffusion both actively through direct communication between peers and passively through observing PV systems in the area. The research was limited to five municipalities with high early adoption rates. The findings should be extended to regions with average or low PV panel adoption rates.

This research will contribute to the current research stream of the adoption of PV panels by combining observed behavior on a meso-level with underlying drivers. The research will do so by analyzing the diffusion of PV panels over time in a local scope. These behavioral findings are combined with survey data on the social dynamics which should explain the diffusion effect. Prior research focused on either the diffusion of PV panels without measuring the motivations of residents, or the underlying motivations without measuring the actual (and not just intended) behavior of residents. This research combines the outcomes and the antecedents to measure both the pattern and the process of the diffusion of PV panels in a community, and therefore gain more insights in the motivations leading to actual behavior, even when the behavior has high behavioral costs. This will help to understand the underlying mechanisms of the contagion effect and help to distinguish it from other potential vindications. The combination of the pattern and the process of the adoption of PV panel is distinct from previous research and contributes to the existing literature. Moreover, this approach adheres to the request for a more a multi-disciplinary approach in energy studies (Sovacool, 2014; Stern, 2014).

(9)

behavior, measurement on a micro-level might be biased. The focus on a community instead of individual consumers might provide new insights for other research on product diffusion. Moreover, the study shows how spatial econometrics can be used in marketing research. The use of spatial-econometrics is under-exploited in marketing research and may be useful in various sub-fields of marketing research, such as research on product diffusion, targeting strategies, or social network influences. The findings on the PV panels might provide insights to various other sustainable technologies with high behavioral costs, such as electric vehicles.

Third, the study provides important implications for marketing managers of energy producers and for policy makers. Better understanding the underlying mechanisms of spatial contagion will provide insights on how to better target consumers and how to provide more persuasive instruments for consumers to adopt PV panels. This research can help marketing managers and policy makers to understand the importance of local dynamics, which provides insights on whether to invest in local or national campaigns. This will enhance the effectiveness of marketing strategies and policy instruments to motivate consumers to switch to PV panels, thus increasing the share of renewable energy consumption, without heavily relying on subsidies.

In the following sections, I will explain the theoretical background and form my hypotheses. After, I will describe my choice of sample and the spatial-econometric methods to analyze the spatial contagion effects. Next, I will present my findings and discuss the implications of the findings and the limitations of the research.

THEORY

(10)

This indicates social contagion, the interpersonal copying effect of a behavior within social networks (Iyengar, van den Bulte, & Valente, 2011). This effect is positively moderated by the social norms and the social cohesion in the community. Moreover, local energy initiatives can further increase the impact of the social contagion.

Figure 1. Conceptual Model

(11)

perform than the overt action. This indicates the presence of a situational threshold to translate intentions into overt behavior. For many consumers, their stated preference for renewable energy is not substantial enough to lead to the actual behavior, while it is substantial enough to lead to the easy action of a stated intention. Therefore, consumers’ values and beliefs have only strong predictive power over intentions and behaviors related to low behavioral costs, such as willingness to change or policy acceptance (e.g. Harland, Staats, & Wilke, 2007; Nordlund & Garvill, 2003; Steg, Dreijerink, & Abrahamse, 2005; Stern, Dietz, Abel, Guagnano, & Kalof, 1999; Wall, Devine-wright, & Mill, 2007), while having less predictive power over behaviors with high related behavioral costs in terms of effort, inconvenience, time, or money (Abrahamse, Steg, Vlek, & Rothengatter, 2005; Bamberg & Schmidt, 1999; Hunecke, Blöbaum, Matthies, & Höger, 2001).

(12)

Reducing greenhouse gas emissions by the adoption of PV panels cannot be achieved by an individual consumer. Hence, consumers are co-dependent on the adoption by other consumers. When consumers a region in a community adopt PV panels, the barrier related to PV panels being an impure public good diminishes for neighboring regions, as consumers see others do not free-ride on their investment. Simultaneously, the benefits increase as the utilities are co-dependent.

Because the utilities of the adoption of PV-panels are co-dependent, it makes more sense to analyze the adoption of PV panels on a meso level than on a micro level. The adoption of PV panels can be analyzed on a meso level by considering the patterns of the distribution of PV panels in communities. Analyzing distribution patterns over space and time can reveal whether the adoption of PV panels is contingent on the adoption of PV panels by neighboring regions. Graziano and Gillingham (2015) revealed a centrifugal dispersion pattern of the adoption of PV panels, as they found the adoption of PV panels to be strongly related to the number of prior installed nearby PV panels. This relationship diminishes with distance, suggesting a spatial neighbor effect conveyed through social interaction and visibility. This leads to the following hypothesis:

Hypothesis 1: The adoption of PV panels in a region is affected by the adoption rate in neighboring regions, where the effect of neighbors’ adoption diminishes with distance.

(13)

found that that normative information spurred people to conserve more energy than any of the standard appeals that are often used to stimulate energy conservation, such as protecting the environment, being socially responsible, or even saving money. Hence, social influence can predict behavior, even when the costs of the behavior are high. Curtius et al. (2018) tested the effects of descriptive and injunctive norms on the adoption of PV panels. They found that descriptive norms positively affect the intention of the adoption of PV panels. An increased exposure of PV panels in a community sends a positive signal to adopters, which reduces the uncertainty to invest in PV. Moreover, a higher descriptive norm makes consumers more mindful about the advantages of the adoption of PV panels. The visibility of the PV panel therefore serves as advertisement for neighbors. Curtius et al. (2018) also found a positive effect of injunctive norms, as consumers who think that adopting PV panels is desired by their peers is more likely to adopt PV panels than consumers who do not perceive this desired behavior. This lead to the following hypotheses:

Hypothesis 2a: A descriptive norm towards the adoption of PV panels in the community positively affects the rate of self-generated electricity in a region.

Hypothesis 2b: An injunctive norm towards the adoption of PV panels in the community positively affects the rate of self-generated electricity in a region. y.

(14)

behavior of neighbors who do install PV panels. Hence, a higher descriptive norm is expected to strengthen the social contagion between regions. This leads to the following hypothesis:

Hypothesis 3a: The relation between the adoption of PV panels in a region and the adoption rate in neighboring regions is positively moderated by the descriptive norm towards the adoption of PV panels in the community.

When residents perceive that adopting PV panels is desired by their neighbors (i.e. the injunctive norm is high), the adoption of PV panels by others may provide social benefits. The adoption of PV panels signals self-identity, which is an important aspect of making pro-societal decisions (Axsen & Kurani, 2012). The public visibility of pro-societal behavior leads consumers to desire green products (Griskevicius, Tybur, & Van den Bergh, 2010). The adoption of PV panels can enhance consumers’ status as they appear more pro-societal rather than pro-self. When a consumer feels that this pro-societal behavior is desired by the community, the consumer is motivated to copy the neighbor’s behavior as this will improve the consumer’s status. When the consumer does not feel that the behavior is desired by the community, the consumer will be less likely to copy neighbor’s behavior to adopt PV panels. Therefore, a higher injunctive norm is expected to strengthen the social contagion in a region. This leads to the following hypothesis:

Hypothesis 3b: The relation between the adoption of PV panels in a region and the adoption rate in neighboring regions is positively moderated by the injunctive norm towards the adoption of PV panels in the community.

(15)

Therefore, they will be less focused on personal benefits and more focused on benefits of the community. Residents will be less concerned with the adoption of PV panels being a party public good and will be less afraid of paying the costs for the community. Hence, residents in a community with high social cohesion will perceive less behavioral costs to the adoption of PV panels. This leads to the following hypothesis:

Hypothesis 4: The social cohesion in the community positively affects the rate of self-generated electricity in a region.

Moreover, the social cohesion moderates the effect of social cohesion on the adoption of PV panels in a community. The adoption of a product by consumers depends on the existence of preceding adopters in their social system (Strang & Soule, 1998). For example, the consumers’ social network has been found to encourage the adoption of electric vehicles among initially hesitant consumers (Eppstein, Grover, Marshall, & Rizzo, 2011). According to social network analysis, not all peers are equally influential, but primary those in consumers’ social network (Rogers & Kincaid, 1981). When residents in a community are part of each other’s social network, they are more likely to be affected by each other. In a community with a high degree of social cohesion, the residents are part of each other’s social network and therefore more likely to be affected by each other’s behavior. When the degree of social cohesion in a community is low, the consumers are less inclined to copy each other’s behavior. Hence, a higher degree of social cohesion is expected to strengthen the social contagion in a region. This leads to the following hypothesis:

(16)

(17)

have an advantage over incumbent energy corporations to convert consumers’ intentions into action for the adoption of PV panels.

Hypothesis 6: The launch of a local energy initiative in the community positively affects the rate of self-generated electricity in a region.

The local initiatives can increase the effect of social contagion on the adoption of PV panels in a community. The launch of a local initiative in the community increases the knowledge and awareness of residents on the adoption of PV panels. As residents become more acquainted with the process, costs, and outcomes of the adoption of PV panels, they might become more inclined to talk about the possibilities with other residents in their community. When the topic of conversations more often concerns the adoption of PV panels, consumers may perceive the purchase less as an individual decision and more as a community decision. Hence, the launch of a local energy initiative may increase the influence of neighbors on the adoption of PV panels. This leads to the following hypothesis:

Hypothesis 7: The relation between the adoption of PV panels in a region and the adoption rate in neighboring regions is positively moderated by the launch of a local energy initiative in the community.

To test these hypotheses, a series of analyses are conducted which are described in the next section.

METHODS Data set

(18)

sources: Enexis, the Kadaster, and a survey research conducted by Jacob Dijkstra and Fleur Goedkoop (scholars at the University of Groningen).

Enexis provides open source data on the yearly residential usage of electricity and gas. Due to privacy constraints, the information cannot be provided on a household level, but on a six-digit postcode level (PC6). To safeguard that the data cannot be linked to an individual household, all PC6 regions are clustered in such that every cluster contains at least 10 connections to the grid. The clusters are made alphabetically. For example, 1234AA with 2 connections, 1234AB with 3 connections, 1234AC with 5 connections are clustered into one observation which contains 10 connections of PC6 1234AA to 1234AC. From a spatial perspective, this is a reasonable way of clustering as regions with alphabetically sequential letters are located close together, given they have the same PC4 code. Because of this clustering, some observations are at PC6 level and some observations are in PC6 groups. The observations are collected from all five communities on a yearly level for the years 2010-2019.

The Kadaster provides the basic registration for addresses and buildings (BAG) including the geographic location and information on all official buildings, including residences. The information from the BAG is used to determine the spatial distances between PC6 groups and additional information on the buildings. The dataset is cross-sectional, with the data extracted on 23-05-2020. The data is aggregated from a residence level to a PC6-level. The data is filtered to only include buildings with the status to be in use and with a residence function.

(19)

and a prepaid response envelope. The purpose of this letter was to inform residents about the upcoming research, derive initial insights regarding the interest in renewable energy and CEIs, and ask participants whether they were willing to complete the main questionnaire shortly after. Respondents had the option to either answer the questionnaire online in a Qualtrics survey or request a paper version by mail with a prepaid response envelope. The researchers conducted additional door-to-door follow-ups for the main questionnaire for a random sample of households (randomly selecting streets) which did not respond to the short questionnaire to increase response rates. The main questionnaire was dropped off in cases where no residents were home. All main questionnaires were distributed between three to five months after the short questionnaire and two weeks after the initial invitations reminders were sent for the online main questionnaires. This resulted into an overall response rate of 29%, with response rates of 26%, 26%, 30%, 31%, and 39% in community 1 to 5. Respondents were always asked to sign an informed consent form before answering the main questionnaire. The results of the survey research were shared anonymously. Hence, they the responses can be connected to a community, but not to a location within the community. Therefore, the results of the survey cannot be linked to a PC6 group and are averaged to a community level.

To summarize, the total dataset includes 416 PC6 groups, located in 5 communities, with data for the years 2010-2019.

Dependent variable

(20)

Research, 2020). The self-generated electricity measure is estimated by subtracting the delivery rate from 100, providing the percentage not generated by Enexis. The data is provided annually per PC6 group because of privacy restrictions. When more residents in a PC6 group adopt solar panels, the delivery rate of electricity delivered by Enexis is lower, and the self-generated electricity higher. Hence, an increase of self-generated electricity rate in a PC6 group indicates an increase in the adoption of PV panels.

This estimation of the adoption of PV panels is based on a research by NRC (Poort, 2020) and differs from those used in prior academic research. Most studies measure the number residential PV installations per region (Balta-Ozkan et al., 2015; Bollinger & Gillingham, 2012; Graziano & Gillingham, 2015; Jayaweera, Jayasinghe, & Weerasinghe, 2018; Kwan, 2012; Rode & Weber, 2016). The number of PV installations does not consider how much electricity is generated by these installations and to which extent consumers depend on other PV panels as their source for energy consumption. The self-generated electricity rate measures how much of the total electricity usage is self-generated, and therefore to which extent consumers have switched to solar energy. The self-generated electricity percentage is similar to the measure used by Schaffer & Brun (2015), who measured the installed capacity from small-scale installations per square kilometer in a region.

Explanatory variables

(21)

of the variables is derived from the Enexis dataset, as described before. The Euclidean distances are estimated with the geographic information from the BAG dataset. For every PC6 group, the centroid is estimated of all residential locations in that PC6 group. Hence, the centroid is not the center of an area, but the center of the residences within the area. The distances are estimated between the centroids of these areas.

The social norms of the community include the descriptive norms and the injunctive norms of a community. The descriptive norms are measured by the number of people a resident knows with solar panels and by the norm of reciprocity, how many neighbors the respondent expects to participate in the PV-panel initiative. The first variable is measured as an open question, in which respondents could fill in the number of people they know. The second variable is measured on a 5-point Likert scale from almost no-one to almost all. This provides an indication of to what extent residents in the community perceive their neighbors to have adopted PV panels. The injunctive norms are measured by three questions indicating the community sustainable energy motivation, in which the respondents state how important sustainable energy consumption is to their neighbors. The answers are given on a seven-point Likert scale from completely disagree to completely agree. This provides an indication of to what extent residents in the community perceive their neighbors to approve the adoption of PV panels.

(22)

themselves with their community with four items on a seven-point Likert scale. The perceived interpersonal interaction captures the extent of interaction among neighbors. It is measured with two items on a seven-point Likert scale. Together, these three constructs measure the connectedness and solidarity among neighbors in a community.

Finally, initiative indicates the launch of a local energy initiatives in the community. The variable is a binary variable whose value is zero in the years before the launch of the local energy initiative, and whose value is one when the local energy initiative is launched in the community. The initiatives started in 2014, 2014, 2015, 2018, and 2017 in community 1 to 5 respectively.

Control variables

To capture the effect of the social contagion, it is important to control for other factors which might explain the adoption of PV panels (Bernerth & Aguinis, 2016). This helps to distinguish the imitation effect from individual factors influencing the adoption of PV panels.

(23)

According to the theory of planned behavior, ability is one of the most important factors to predict behavior (Ajzen, 1991; Fishbein & Ajzen, 1975). The suitability of PV panels might be affected by the construction year of a building. Therefore, the median construction year of all buildings is estimated for every PC6 group. The surface of the parcel might also indicate the suitability of PV panels. Therefore, the median surface of the parcels in a PC6 group is estimated.

Moreover, there are various community characteristics which might influence the percentage of self-generated electricity. The survey provides insight in various demographic characteristics of the residents of a community. Though these variables only represent the characteristics of the respondents, and not the entire community, the sample provides some understanding of the differences among the communities. For example, characteristics as Age, Level of education, Employment, Income, and the Years lived in community provide valuable information. Prior research shows that a resident’s age is an important predictor of PV panels (Kwan, 2012), as consumers between the age 35–45 have a larger purchasing power for green products than those of other ages (Pickett-Baker & Ozaki, 2008). The level of education is found to be a predictor of the adoption of PV panels as well (Balta-Ozkan et al., 2015; Davidson, Drury, Lopez, Elmore, & Margolis, 2014; Jager, 2006; Keirstead, 2007; Kwan, 2012). Employed residents with higher incomes might have more resources to install PV panels. Households who already live for a long time in the community might be more influenced by the social norms of the community and might be more likely to participate in community initiatives than households who recently moved in the community.

(24)

are higher. Hence, the PV panels become more valuable for the consumer. The Distributive Justice provides an indication of how different parties benefit from the renewable energy project as the adoption of PV panels can be perceived as a partly public good (Kotchen, 2006; Wiser, 1998). As these variables are measured on a community level and cannot be linked to a specific PC6 group, they are controlled for by adding a community variable to the analyses indicating which community the PC6 group belongs to.

Community variables

The survey variables provide information on various characteristics of the communities. To reduce the dimensions of the survey questions, a factor analysis is performed, which aggregates the sub-items to one construct. All survey questions are provided in Appendix I. An overview of the steps in the data reduction is provided in Appendix II.

(25)

to a better environment by the adoption of PV panels, efficacy production measures the extent to which the consumer is able to produce its own electricity, and efficacy environment measures the extent to which to resident can contribute to the environment by the adoption of PV panels. Age measures the age of the resident who filled in the questionnaire. Income measures the monthly household income after taxes. The years living in community indicates for how many years the resident has lived in the community. Education groups captures the highest level of education of the respondent. Man measures the percentage of male respondents in the community, employed measures the percentage of employed respondents in the community, and member indicates the percentage of respondents in a community who were at least member of one club.

As the data of the respondents is anonymous, the answers cannot be linked to a certain PC6 group. Therefore, the data is aggregated to a community-level by taking the mean of all variables for each community. The dummy variables are converted to the percentage of responses being equal to 1 in the community. Furthermore, for the variables distributive norm solar and years living in community the communities are estimated using the median instead of the mean as the variables show many outliers. An overview of the variables per community is provided in Table 1. The table shows there is some variance in the variables among the communities. It is important to note that the demographic variables only refer the demographic characteristics of the respondents, and not of the entire community. The high average age of the respondents explains the low employment rate in community 4 and 5, as many respondents are already retired. Though only view survey variables are used in the analysis, the other variables help to understand the differences between the communities.

Table 1. Survey variables per community

Com 1 Com 2 Com 3 Com 4 Com 5

Interpersonal Contact 3.23 2.82 2.87 3.02 2.75

(26)

Perceived interpersonal

interaction 4.50 4.46 3.93 4.18 3.71

Community sustainable

energy motivation 4.47 4.57 4.69 4.00 3.93

Distributive fairness others 3.03 2.98 2.85 3.15 3.49

Distributive fairness self 3.76 3.42 3.47 3.33 3.64

Distributive norm solar 5.00 10.00 3.24 5.00 4.00

Norm of reciprocity 2.58 2.09 2.40 2.22 2.52 Number of clubs 1.48 2.13 0.96 1.33 0.88 Efficacy contribution 5.52 5.12 4.83 5.10 5.65 Efficacy production 4.25 - - 3.44 3.54 Efficacy environment 4.68 4.61 5.26 5.03 5.40 Age 59.30 59.00 52.02 64.34 62.75 Income 3.42 3.09 3.14 3.36 2.92

Years living in community 20.00 21.00 16.00 23.00 21.00

Education groups 2.18 2.33 2.90 2.52 2.25 Man 0.71 0.58 0.50 0.49 0.55 Employed 0.52 0.49 0.56 0.33 0.34 Member 0.68 0.80 0.36 0.63 0.35 Number of respondents 126 51 129 80 65 Variable overview

All observations which include missing variables are omitted. These variables include years of certain PC6 groups prior to the existence of these groups. For example, if the postcode 1234AB did not exist before 2016, observation 2010-2015 of this postcode are omitted. This decreases the dataset from 4160 observations to 4092 observations.

Table 2 provides an overview of the descriptive variables of the dependent variable, the explanatory variables, and the control variables. A further investigation in the distribution of these variables shows that the variables self-generated electricity, number of connections, annual consumption, and surface are highly skewed (skew = 4.34, 1.89, 2.46, and 3.36 respectively). Hence, in the analysis these variables will be estimated using the natural logarithm. The outliers of all variables are retained as they do not indicate measurement errors.

Table 2. Descriptive variables

n mean sd median min max

Self-generated electricity 4092 4.10 9.65 0 0 100

(27)

Norm of reciprocity 4092 2.43 0.14 2.52 2.09 2.54 Community sustainable energy

motivation 4092 4.27 0.33 4 3.93 4.69

Interpersonal Contact 4092 2.92 0.18 2.87 2.75 3.23

Identification with community 4092 4.49 0.14 4.42 4.35 4.72

Perceived interpersonal interaction 4092 4.04 0.31 3.93 3.71 4.50 Initiatives 4092 0.41 0.49 0 0 1 Number of connections 4092 20.15 7.21 20 10 72 Annual consumption 4092 3412.39 2649.62 2980.43 1.15 22856 Build year 4092 1953.21 34.85 1968 1840 2018 Surface 4092 118.96 51.15 107 24 617

(28)

buildings in a PC6 region and is time-invariant. The surface indicates the mean surface of the parcels of the residential buildings in a PC6 region, which is time-invariant as well.

It is important to check for multicollinearity problems prior to the analyses. The VIF scores are checked for the variables number of connections, annual consumption, build year, surface, initiative and community. Only build year and community showed a VIF score higher than the critical value of 4, indicating collinearity problems. Hence, the control variable build year will not be used in the analyses across communities. The correlations are presented in Table 3. Only the community level variables show high correlations (> 0.50), which is expected as they are equal for all PC6 groups in a community. All variables are standardized prior to the analyses.

Table 3. Correlation matrix

1 2 3 4 5 6 7 8 9 10 11 1. Distributive norm solar - 2. Norm of reciprocity -0.51 - 3. Community sustainable energy motivation 0.05 -0.11 - 4. Interpersonal Contact 0.20 0.12 0.38 - 5. Identification with community 0.38 0.10 0.73 0.83 - 6. Perceived interpersonal interaction 0.59 -0.18 0.48 0.89 0.90 - 7. Initiatives 0.03 -0.01 0.36 0.23 0.32 0.24 - 8. Number of connections -0.15 0.15 -0.24 -0.25 -0.28 -0.30 -0.09 - 9. Annual consumption 0.19 0.01 0.01 0.31 0.25 0.32 -0.10 -0.07 - 10. Build year 0.12 0.18 -0.66 -0.04 -0.28 -0.10 -0.21 0.26 -0.08 - 11. Surface 0.38 -0.07 -0.02 0.39 0.32 0.45 0.03 -0.11 0.50 -0.03 - Model specification

(29)

community, and therefore shows whether the distribution pattern in the community is clustered or dispersed.

A Moran’s I is not sufficient to estimate the interdependence in the adoption of PV panels. Therefore, a fixed effects model should provide more information about the size of the copying effect while correcting for other explanatory variables. There are various models which account for this spatial-autocorrelation (Figure 2). Elhorst (2010) proposes to use a general-to-specific approach to select the most appropriate econometric model (right to left in Figure 2). Hence, begin with an ordinary least squares (OLS) regression and make the model more specific with every model. The most specific model is the Manski Model (equation 1). The model is described in matrix notation, with N regions and K explanatory variables, Y represents a (N×1) vector of dependent variable observations, X a (N×K) matrix of observations of explanatory variables with a (K×1) vector of associated regression coefficients β. Moreover, ρ is a spatial autoregressive variable which measures the degree of interdependence across regions and indicates the spatial lag of dependent variable and θ provides the spatial autocorrelation in the explanatory variables. W indicates a inversed-distance matrix, discounting the effect of the dependent variable of the neighbors by the distance between the focal observation and the other neighbors. Furthermore, u denotes an independently and identically distributed error term with a mean of zero and a constant variance of σ2_{. The autocorrelation in the error term is indicated}

by λ.

𝑌 = 𝜌𝑊𝑌 + 𝑋𝛽 + 𝑊𝑋𝜃 + 𝑢 𝑢 = 𝜆𝑊𝑢 + 𝜀

With this model, Manski (1993) distinguishes three ways in which the adoption of a product in one location is affected the adoption in another location: endogenous interactions, contextual interactions, or correlated effects. Endogenous interactions describe the tendency of a region to behave in some way which varies with the behavior of other regions. In the model, this effect

(30)

is provided by ρWY. Contextual interactions show the tendency of a region to behave in a way which varies with exogenous characteristics influencing the behavior of other regions. Hence, the similarity of explanatory variables explaining the adoption of PV panels in neighboring regions. This effect is provided in the model by WXθ. Finally, the correlated effects describe how similar unobserved correlated effects might influence similar decisions of regions, such as similarities in income, education, and household size. The model includes this effect with the spatially autocorrelated error term u = λWu+ε.

Figure 2. Econometrics models for spatial dependence, specific to general

Source: Elhorst (2010)

(31)

preferred over the spatial lag model. When θ = 0 and θ = −ρβ both do not hold, the model considers a spatial lag in the explanatory variables. This is the case for the spatial Durbin model. Both the spatial lag and the spatial error model are nested in the Durbin model. This allows for comparison of the models through the likelihood-ratio test. If there is no spatial autocorrelation in the explanatory variables, but there is autocorrelation in the dependent variable and the error term, the Kelejian-Pucha model is preferred.

The effect of the spatial autocorrelation on the adoption of PV panels, indicated by ρ, is expected to variate between the five communities and depend on the social norms, the social cohesion and the local energy initiatives in a community. Interaction effects are added to the selected model to assess these moderating effects

RESULTS Pattern of distribution

(32)

of PV panels in another region, the diffusion of PV-panels is expected to become more clustered over time. Hence, the values of Moran’s I are expected to increase over the years. For each PC6-group, the centroid is calculated by taking the Euclidean center of all residential locations in the PC6-group. A weight matrix Wij is created which provides the Euclidean inversed distance between the centroid of PC6-group i and the centroid of PC6-group j. The weight matrix is estimated with the inverse distance as communities with smaller distances are expected to be more similar than communities with large distances. The inversed distance is estimated by 1 𝐷# !", where Dij is a Euclidean distance matrix between point i and j.

𝐼 = 𝑁 × ∑ ∑ [𝑊!"× (𝑦!− 𝑦5 )]

$ "%& $ !%&

∑$!%&∑$"%&𝑊!"× ∑$!%&(𝑦!− 𝑦5 )']

Equation 2 presents the estimation of Moran’s I, where I represents the Moran’s I, N the number of regions, Wij a the corresponding value in the weight matrix of value i and j, 𝑦 the log of the percentage of electricity generated by PV panels in region i, and 𝑦% the average percentage of the log of electricity generated by PV panels in the community. The results of the Moran’s I are presented in Table 4.

Table 4. Moran’s I EV 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 Com 1 -0.012 0.016* 0.016* 0.029*** -0.001 0.033* 0.064*** 0.061*** 0.052*** 0.057*** 0.047** Com 2 -0.043 -0.059 -0.086 -0.051 -0.037 -0.042 -0.023 0.055* 0.003 0.043* 0.065* Com 3 -0.010 -0.012 -0.012 -0.012 -0.014 -0.015 -0.013 -0.013 -0.018 -0.008 -0.009 Com 4 -0.016 -0.014 -0.014 0.000 -0.008 -0.020 0.002 0.010 0.018 0.003* 0.002† Com 5 -0.007 -0.006 -0.003 0.008 0.004 0.016 0.010† 0.631*** 0.575***

Note: EV = expected value, bold indicates initiative has started, † indicates p < .1, * indicates p < .05, ** indicates p < .01, *** indicates p < .001, where p indicates the pseudo-significance.

Table 4 presents the Moran’s I of each community for each year. The expected value of the Moran’s I serves as a benchmark as it indicates the Moran’s I at random dispersion and is unique for every community. The expected Moran’s I is negative by default, as it is defined by

#$

%#$. This is close to zero for large sample sizes, but not for the sample sizes of the communities.

(33)

The Moran’s I differs over time and across communities. A non-significant Moran’s I indicates a random pattern which is nor clustered nor dispersed. For community 1, in most years the Moran’s I is small, but positive and significant. There is a slight increase in the value of the Moran’s I over the years and a small decline after 2016. The positive significant Moran’s I indicates a small spatial clustering effect of the adoption of PV panels among regions. Community 2 shows an increase in the significant positive Moran’s I over the years. This indicates an increase in the degree of spatial clustering. In community 3, however, all Moran’s I values are insignificant, indicating no spatial clustering in the adoption of PV panels. Community 4 shows a small but positive Moran’s I in 2018 and 2019, indicating a small spatial clustering effect. Finally, community 5 shows a strong spatial clustering effect from 2018 and 2019. The values of 2010 and 2011 could not be estimated for community 5 as there was no variance among the adoption of PV panels (there was no self-generated electricity in all regions).

(34)

Effect of social contagion

A pooled OLS regression is performed to estimate the model as presented in equation 3. The higher-level variables are not included because of multicollinearity problems. The results of the pooled OLS estimation are reported in Table 5. A Breusch-Pagan test revealed no heteroscedasticity problems in the error term with BP = 940.59 and p = 0.000.

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦)!( = 𝛽)+ 𝛽&ln (𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() + 𝛽'ln (𝐶𝑜𝑛𝑠𝑢𝑝𝑚𝑡𝑖𝑜𝑛!() + 𝛽*ln (𝑆𝑢𝑟𝑓𝑎𝑐𝑒!) + 𝛽+𝐶𝑜𝑚𝑚𝑢𝑛𝑖𝑡𝑦!+ 𝛽,𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝜀!( Where: i = 1, … , N: PC6 group j = 1, …, N-1 and i≠j: PC6 group t = 1, …, T: year m = 1, … , M: community

Self-generated electricityit = the percentage electricity consumption self-generated by PC6 group i in year t

Connectionsit = number of connections of PC6 group i in year t

Consumptionit = annual electricity consumption of PC6 group i in year t

Surfacei = average surface of residence in PC6 group i

Communityi = community of PC6 group i

Initiativemt = dummy variable indicating if an initiative has launched in community m in year t

To investigate the contagion effect of the adoption of PV-panels, a fixed effect regression is executed using the model in equation 4. Pooled OLS does not allow for the inclusion of spatially correlated variables. Therefore, a fixed effects regression is employed as it enables to regress panel data with the inclusion of spatial autocorrelation effects. A random effects model would be preferred over a fixed effects model as it allows for the inclusion of time-invariant variables (surface and community), which are omitted in a fixed effects model. A random effects model can therefore be estimated using the model in equation 3. However, the random-effects model has the strong assumption that the explanatory variables cannot be correlated to any unobserved time-invariant observation specific variables. To test if this assumption holds, a Hausman test is performed. The Hausman tests shows whether the differences in the coefficients in the time-varying variables in the fixed-effects model and the random effects model are significantly different. The Hausman test shows a Chi2_{(3) = 42.954 with p = 0.000.}

(35)

Hence, the coefficients are significantly different in the models and the assumption for a random effects model does not hold. A fixed effects model should be employed.

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦)!(= 𝛽)!+ 𝛽&ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() + 𝛽'ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!() +𝛽*𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝜀!( Where: i = 1, … , N: PC6 group j = 1, …, N-1 and i≠j: PC6 group t = 1, …, T: year m = 1, … , M: community

Self-generated electricity = the percentage electricity consumption self-generated by PC6 group i in year t Connections = number of connections of PC6 group i in year t

Consumption = annual electricity consumption of PC6 group i in year t

Initiative = dummy variable indicating if an initiative has launched in community m in year t The spatial fixed effects models are estimated on panel data with the splm package in R using maximum likelihood estimations. Model 3 in Table 5 is estimated with spatial autocorrelation in the dependent variable (equation 5). Model 4 is estimated including only autocorrelation in the error term (equation 6). Model 5 is estimated including both autocorrelation in the dependent variable and in the error term (equation 7). Model 6 is estimated using OLS and includes spatially lagged explanatory variables (equation 8). Model 7 includes autocorrelation of the dependent variable and the explanatory variables (equation 9). All equations are compared to find the best fit to model the log of the rate of self-generated electricity.

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦)!(= 𝛽)!+ 𝛽&ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() + 𝛽'ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!() +𝛽*𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝜌 L 𝑤!"𝑦"( $.& !%& + 𝜀!( Where: i = 1, … , N: PC6 group j = 1, …, N-1 and i≠j: PC6 group t = 1, …, T: year m = 1, … , M: community

wij = inversed Euclidean distance between i and j

Initiative = dummy variable indicating if an initiative has launched in community m in year t

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦)_𝛽 !(= 𝛽)!+ 𝛽&ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() +

'ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!() + 𝛽*𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝑢!(

(4)

(5)

(36)

with 𝑢!(= 𝜆𝑤!"𝑢!(+ 𝜀!( Where: i = 1, … , N: PC6 group j = 1, …, N-1 and i≠j: PC6 group t = 1, …, T: year m = 1, … , M: community

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦)!(= 𝛽)!+ 𝛽&ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() + 𝛽'ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!() + 𝛽*𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝜌 L 𝑤!"𝑦"( $.& !%& + 𝑢!( With 𝑢!(= 𝜆𝑤!"𝑢!(+ 𝜀!( Where: i = 1, … , N: PC6 group j = 1, …, N-1 and i≠j: PC6 group t = 1, …, T: year m = 1, … , M: community

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦)!(= 𝛽)!+ 𝛽&ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() + 𝛽'ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!() + 𝛽*𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝜃&𝑤!"× ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() + 𝜃'𝑤!"× ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!() +𝜃'𝑤!"× 𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝜀!( Where: i = 1, … , N: PC6 group j = 1, …, N-1 and i≠j: PC6 group t = 1, …, T: year m = 1, … , M: community

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦)!(= 𝛽)!+ 𝛽&ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() + 𝛽'ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!()

+ 𝛽*𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝜃&𝑤!"× ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!() + 𝜃'𝑤!"× ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!()

(7)

(8)

(37)

+𝜃'𝑤!"× 𝐼𝑛𝑛𝑖𝑡𝑎𝑡𝑖𝑣𝑒-(+ 𝜌 L 𝑤!"𝑦"( $.& !%& + 𝜀!( Where: i = 1, … , N: PC6 group j = 1, …, N-1 and i≠j: PC6 group t = 1, …, T: year m = 1, … , M: community

Table 5. Regression results Model 1 P-OLS Model 2 FE Model 3 SLM Model 4 SEM Model 5 KPM Model 6 SLX Model 7 SDM Constant 0.057 Log Connections 0.016 -0.120** -0.122*** -0.121*** -1.122*** -0.123** -0.122*** Log Consumption -0.053*** -0.068*** -0.070*** -0.070*** -0.070*** -0.071*** -0.071*** Initiative 1.170*** 1.166*** 1.152*** 1.154*** 1.152*** 1.140*** 1.144*** Surface 0.338*** Community 2 0.569*** Community 3 0.049 Community 4 0.329*** Community 5 0.172*** W *Connections -0.00001 -0.00001 W*Consumption 0.001 0.001 W*Initiative 0.039*** 0.022* N 4020 4020 4020 4020 4020 4020 4020 WY 0.007*** 0.006* 0.004* Wu 0.008*** 0.004 R2 _0.360 _0.365 _0.403 Adjusted R2 _0.358 _0.294 _0.336 Pseudo R2 _0.597 _0.589 _0.596 _0.597 Log likelihood -15348.47 -15351.56 -8998.88 -15346.25 AIC 10593.6 9618.9 31510.94 31517.12 18813.75 9558.9 31512.50 BIC 10656.5 12176.3 34074.65 34080.83 21383.76 12135.2 34095.11

Note: * indicates p < .05, ** indicates p < .01, *** indicates p < .001, WY indicates the spatial autocorrelation in the dependent variable, Wu indicates the spatial autocorrelation in the error term, P-OLS = pooled OLS, FE = fixed effects, SLM = spatial lag model, SEM = spatial error model, KPM = Kelejian-Pucha model, SLX = spatial lagged X model, SDM = spatial Durbin model

(38)

variable connections becomes significant while it was insignificant in the pooled OLS model. The fixed effects model does not outperform the pooled OLS model, as the adjusted R2_does

not increase. This is most likely due to the loss of the time-invariant variables. Still, the fixed effects model will be used as it allows to include the spatial autocorrelation in panel data.

The spatial lag model shows a small but significant and positive effect of the adoption of PV panels by neighboring regions with ρ = 0.007 and p = 0.000, implying a significant positive spatial autocorrelation in the dependent variable. The spatial error model shows a small but significant and positive effect in the spatial autocorrelation in the unobserved variables with λ = 0.008 and p = 0.000, indicating a significant positive spatial autocorrelation in the error term. The Kelejian-Pucha model shows a small but significant and positive spatial autocorrelation with ρ = 0.006 and p = 0.022, indicating a significant spatial autocorrelation in the dependent variable. The error term, however, is no longer significantly spatially autocorrelated with λ = 0.004 and p = 0.314.

The spatial lagged X model is estimated with OLS and shows that only the explanatory variable initiative is spatially autocorrelated with θ = 0.039 and p = 0.000. Adding a spatial component to this explanatory variable makes conceptually no sense, as the variable is equal to all PC6 groups in the community regardless of the distance to the other PC6 groups. Hence, the effect of initiatives in a region on the log of the rate of self-generated electricity is not spatially dependent on presence of an initiative in neighboring regions in a community, as all regions in a community have the same value by default. Therefore, the spatially lagged X model is not preferred over the fixed effects model, regardless of the higher adjusted R2_{. For the spatial}

(39)

The log likelihood, AIC, BIC, and pseudo-R2_{all provide indications of the fit of the models}

estimated through maximum likelihood. The Kelejian-Pucha provides a substantially higher log likelihood value than the other models. This is confirmed by the likelihood-ratio tests, which show that the Kelejian-Pucha model significantly outperforms the spatial lag model (LR = 12699.41, p = 0.000), the spatial error model (LR = 12705.17, p = 0.000), and the spatial Durbin model (LR = 12694.55, p = 0.000). The pseudo R2_{of the Kelejian-Pucha model is not}

the largest, but is very close to the values of the other models. The AIC and BIC allow comparison with the models estimated with OLS. Again the AIC and BIC of the Kelejian- Pucha model outperforms the spatial lag model, spatial error model and the spatial Durbin model, as its value is much lower. However, the model does not outperform the pooled OLS model, the fixed effects model or the spatial lagged X model. The pooled OLS model and the fixed effects model both do not contain a spatial autoregressive element. As shown in the other models, the PC6 groups are spatially interdependent. Hence, the model specification should include this spatial interdependence. Therefore, the Kelejian-Pucha model is preferred over the pooled OLS and the fixed effects model, regardless of the higher AIC and BIC. The spatial lagged X is conceptually not meaningful, as it only includes the spatial lag of the explanatory variable initiative. Therefore, the Kelejian-Pucha model remains the preferred model to estimate the log of the rate of self-generated electricity.

(40)

spatial autoregressive term in the dependent variable. The Lagrange multiplier shows that ρ ≠ 0 when the model already contains a spatial autoregressive error term, with LM = 62.49 and p = 0.000. Moreover, the Lagrange multiplier shows λ ≠ 0 when the model contains a spatial autoregressive dependent variable, with LM = 5.39 and p = 0.020. Hence, the model should include both an error term in the dependent variable and in the error term by estimating a model containing both these spatial terms, such as the Kelejian-Pucha model.

The Kelejian-Pucha model shows that the log of the number of connections is significantly negatively related to the log of the self-generated electricity rate with β = -1.122 and p = 0.001. Moreover, the annual consumption is as well significantly negatively related to the log of the rate of self-generated electricity with β = -0.070 and p = 0.000. The local energy initiatives are significantly positively related to the log of the rate of self-generated electricity with β = 1.152 and p = 0.000. As the effect between the local energy initiatives and the adoption of PV panels is positive and significant, Hypothesis 6 is supported. Moreover, the spatial autocorrelation in the dependent variable shows a significant positive effect with ρ = 0.006 and p = 0.022. The spatial autocorrelation in the error term does not show a significant effect with λ = 0.004 and p = 0.314.

The Kelejian-Pucha model shows a small, but positive and significant spatial autocorrelation effect in the dependent variable. This indicates that the adoption of PV panels in one region is affected by the adoption rate in neighboring regions. Hence, the model finds support for Hypothesis 1.

Community moderators

(41)

different communities to define the spatial autocorrelation in the log of the percentage of self-generated electricity for each community. The results are presented in Table 6.

Table 6. Kelejian-Pucha model per community

Com 1 Com 2 Com 3 Com 4 Com 5

Log Connections -0.072 -0.063 0.048 -0.274** -0.011 Log Consumption -0.063 -0.056 -0.108 -0.181 -0.148*** Initiative 1.724*** 2.852*** 1.261*** 2.109*** 1.123*** WY 0.789 -4.884 -1.121* -0.576 0.007*** Wu 4.213*** 8.382*** 1.594*** 4.285*** 0.003 N 830 210 960 610 1410 Pseudo R2 _0.592 _0.551 _0.481 _0.588 _0.536 Log likelihood -1812.45 -473.79 -1986.27 -1330.13 -2971.58 AIC 3804.454 1001.94 4176.53 2794.27 6237.15 BIC 4223.11 1092.31 4672.96 3089.97 7009.10

Note: * indicates p < .05, ** indicates p < .01, *** indicates p < .001, WY indicates the spatial autocorrelation in the dependent variable, Wu indicates the spatial autocorrelation in the error term.

The models in Table 6 show large differences in coefficients across the different communities. Only community 5 shows a significant positive spatial autocorrelation in the log of the rate of self-generated electricity rate. The other communities show a positive spatial autocorrelation in the error term. Community 3 has a strong negative spatial autocorrelation in the log for the rate of self-generated electricity rate. This indicates that the log of the rate of self-generated electricity in a PC6 group in community 3 decreases when the log of the self-generated electricity rate of neighbors increases. The global Moran’s I of community 3 already showed that there is no clustering in community 3. Hence, the negative relation could be coincidental correlation, or the presence of a distant influential area (the spatial autocorrelation increases with high distance, instead of low distance), for example because it is at a busy cross road with high exposure.

(42)

community-level data cannot be linked to specific PC6 regions. Therefore, the PC6 groups’ survey scores are given by the average of all responses in the community. This implies that within communities, all PC6 groups have the same score on the community-level variables, which leads to multicollinearity problems. Including all variables in one model leads to omitted variables due to singularity problems. Therefore, two different models are created: one to measure the impact of social norms and one to measure the impact of social cohesion. Panel data requires the use of a fixed effects model, which omits the community time-invariant variables. Reducing the dataset to only 2019 data allows the inclusion of time-invariant data. The year 2019 is chosen as this year has the highest rate of self-generated electricity. To examine the effect of social norms, the model as specified by equation 10 is employed:

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦!) = 𝛽)+ 𝜌 ∑$.&!%& 𝑤!"𝑦"+ 𝛽&ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!) +

𝛽'ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!) + 𝛽*ln(𝑆𝑢𝑟𝑓𝑎𝑐𝑒!) + 𝛽+𝐵𝑢𝑖𝑙𝑑𝑦𝑒𝑎𝑟!+ 𝛽,𝐷𝑒𝑠𝑐𝑟𝑖𝑝𝑡𝑖𝑣𝑒𝑁𝑜𝑟𝑚-+

𝛽/𝑅𝑒𝑐𝑖𝑝𝑟𝑜𝑐𝑖𝑡𝑦-+ 𝛽0𝐼𝑛𝑗𝑢𝑛𝑐𝑡𝑖𝑣𝑒𝑁𝑜𝑟𝑚-+ ∑!%&$.&𝑤!"𝑦"+ 𝛽1𝐷𝑒𝑠𝑐𝑟𝑖𝑝𝑡𝑖𝑣𝑒𝑁𝑜𝑟𝑚-× ∑$.&!%& 𝑤!"𝑦"+

𝛽2𝑅𝑒𝑐𝑖𝑝𝑟𝑜𝑐𝑖𝑡𝑦-× ∑$.&!%& 𝑤!"𝑦"+ 𝛽&)𝐼𝑛𝑗𝑢𝑛𝑐𝑡𝑖𝑣𝑒𝑁𝑜𝑟𝑚-× ∑$.&!%& 𝑤!"𝑦" + 𝑢!;

With 𝑢!= 𝜆𝑤!"𝑢!+ 𝜀!

Where:

i = 1, … , N: PC6 group

j = 1, …, N-1 and i≠j: PC6 group

m = 1, … , M: community

Self-generated electricityi = the percentage electricity consumption self-generated by PC6 group i

Connectionsi = number of connections of PC6 group i

Consumptioni = annual electricity consumption of PC6 group i

Buildyeari = mean year the residences are built in PC6 group i

DescriptiveNormm = median people known with PV panels in community m

Reciprocitym = average people expected to participate in the initiative in community m

InjunctiveNormm = average sustainable energy motivation in community m

(43)

containing the values of the log of the rate of self-generated electricity rates for all PC6 groups. As the weight of wii is 0, this sum product equals ∑%#$𝑤_!"𝑦_"

!&$ . To include the spatial

autocorrelation in the dependent variable and in the error term, a spatial error model is estimated including the spatial autocorrelation of the dependent variable in the explanatory variables. This way, the model has the same specification as a Kelejian-Pucha model.

First, the model is estimated without interaction effects (Model 8). Next, the interaction effects are included to the Kelejian-Pucha model (Model 9). A spatial Breusch-Pagan test shows no heteroskedasticity in the error terms for both models with BP = 13.93 with p = 0.084 for Model 8 and BP = 19.51 and p = 0.053 for Model 9. The p-value of the Breusch-Pagan test of Model 9 is only slightly larger than the threshold of 0.05. This implies that Model 9 might not have equal variance among the error terms. This is likely the result of spatial dependence among the communities. However, it is important remain careful with the interpretation of the results as the test statistics are no longer distributed as defined by the usual laws (Floch & Le Saout, 2018). All explanatory variables are standardized prior to the estimation. Table 7 provides an overview of the results.

Table 7. Interaction effects social norms

(44)

The results show that the log of the average surface significantly increases the log of the rate of self-generated electricity, with (β = 0.622 and p = 0.000). The descriptive norm does not significantly affect the log of rate of self-generated electricity (β = 0.021 and p = 0.780). Surprisingly, the norm of reciprocity significantly decreases the log of the rate of self-generated electricity (β = -0.189 and p = 0.009), while the injunctive norm significantly increases the log of the rate of self-generated electricity in a PC6 group (β = 0.550 and p = 0.025). Therefore, Hypothesis 2a is rejected and Hypothesis 2b is supported. Moreover, the spatial autocorrelation in the log of the rate of self-generated electricity is positive and significant with ρ = 0.010 and p = 0.000, indicating a significant spatial autocorrelation of the log of the self-generated electricity rate. The spatial autocorrelation in the error term is not significant with λ = 0.003 and p = 0.959.

(45)

BIC of Model 9. This shows that the increased complexity of the interaction effects do not lead to better performance of the model.

To test whether the social cohesion in a community affects the social contagion of the adoption of PV panels, the model is as specified by equation 11 is employed:

ln(𝑆𝑒𝑙𝑓 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑒𝑑 𝑒𝑙𝑒𝑐𝑡𝑟𝑖𝑐𝑖𝑡𝑦!) = 𝛽)+ 𝜌 L 𝑤!"𝑦" $.& !%& + 𝛽&ln(𝐶𝑜𝑛𝑛𝑒𝑐𝑡𝑖𝑜𝑛𝑠!) + 𝛽'ln(𝐶𝑜𝑛𝑠𝑢𝑝𝑡𝑖𝑜𝑛!) + 𝛽*ln(𝑆𝑢𝑟𝑓𝑎𝑐𝑒) + 𝛽+𝐵𝑢𝑖𝑙𝑑𝑦𝑒𝑎𝑟 + 𝛽,𝐶𝑜𝑛𝑡𝑎𝑐𝑡-+ 𝛽/𝐼𝑑𝑒𝑛𝑡𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛 -+ 𝛽0𝐼𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛-+ L 𝑤!"𝑦" $.& !%& + 𝛽1𝐼𝑛𝑡𝑒𝑟𝑝𝐶𝑜𝑛𝑡𝑎𝑐𝑡-× L 𝑤!"𝑦" $.& !%& + 𝛽2𝐼𝑑𝑒𝑛𝑡𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛-× L 𝑤!"𝑦" $.& !%& + 𝛽&)𝐼𝑛𝑡𝑒𝑟𝑎𝑐𝑡𝑖𝑜𝑛-× L 𝑤!"𝑦" $.& !%& + 𝑢! With 𝑢! = 𝜆𝑤!"𝑢!+ 𝜀! Where: i = 1, … , N: PC6 group j = 1, …, N-1 and i≠j: PC6 group m = 1, … , M: community

Self-generated electricityi = the percentage electricity consumption self-generated by PC6 group i

Connectionsi = number of connections of PC6 group i

Consumptioni = annual electricity consumption of PC6 group i

Buildyeari = average year the residences are built in PC6 group i

Contactm = average interpersonal contact in community m

Identificationm = average identification with the community in community m

Interactionm = average interpersonal interaction among residents in community m

(46)

model. A spatial Breusch-Pagan test shows heteroskedasticity in the error term of Model 10 with BP = 25.71 and p = 0.001 and in Model 11 with BP = 25.11 and p = 0.009. This is likely the result of spatial dependence among the communities. However, it is important to remain careful with the interpretation of the results as the test statistics are no longer distributed as defined by the usual laws (Floch & Le Saout, 2018). All explanatory variables are standardized prior to the estimation. Table 8 provides an overview of the results.

Table 8. Interaction effects social cohesion

Model 10 Model 11 Constant 1.667*** _1.437*** Log Connections -0.052 _-0.043 Log Consumption 0.495 _0.677 Surface 0.556*** 0.593*** Build year 0.019 _0.057 Contact -0.333* _-0.043 Identification -0.068 _-0.277 Interaction 0.475** 0.283 WY * Contact -1.413 WY *Identification 0.567 WY*Interaction 1.181 WY 0.010*** _0.482 Wu -0.002 -0.002 N 402 402 Log likelihood -628.23 _-627.07 AIC 1278.46 _1282.14 BIC 1322.42 1338.09

COPY THY NEIGHBOR: THE ROLE OF SOCIAL NORMS ON THE ADOPTION OF SOLAR PANELS Hester Huisman