Print / Save as PDF

Statistical appendices accompanying the paper. Use the table of contents on the right to jump to a specific appendix.

Appendix A. Ordered logistic regression with full index​

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0

Appendix B. Sensitivity analysis for selection bias​

B.1 Purpose

Our dataset includes 686 utility-scale solar plants in the United States that reached operation between January 2022 and November 2023. We selected operational plants because permitting jurisdiction can be coded more consistently once projects complete the permitting. This choice excludes projects that were canceled or withdrawn before operation, and it can bias estimates of conflict prevalence downward if non-operational projects are, on average, more conflict-prone (which they probably are).

This appendix reports an assumption-driven sensitivity analysis using date compiled from existing research to illustrate the possible magnitude of that selection bias. Our study does not observe whether conflict leads to delay, redesign, or mitigation. Other research examines the sources and consequences of opposition and conflict.

B.2 Analysis and notation

Consider 100 proposed utility-scale solar projects. Let r denote the share of proposed projects that reach operation. Then (1 - r) is the share that do not reach operation (e.g. withdrawn or canceled). Let H_op denote the share of operational projects that fall into the high conflict-attention category. In our operational sample, H_op = 0.19 (19%). Let H_nonop denote the share of non-operational projects that would fall into the high conflict-attention category, if we could observe comparable conflict-attention measures for them. The overall prevalence of high conflict-attention among all proposed projects can be written as a weighted average:

H_all = r * H_op + (1 - r) * H_nonop.

This expression adds high-conflict projects among those that reach operation to high-conflict projects among those that do not, weighted by the size of each group, to estimate the overall share of proposed projects that would fall into the high conflict-attention category.

B.3 External evidence used to inform an illustrative value for H_nonop

We cannot estimate H_nonop because we do not know the true population of proposed projects, so we use existing research on developer-reported survey evidence to explore a scenario. In a survey of utility-scale wind and solar developers by Nilson et al. 2024 [7], respondents reported causes of project delays and cancellations. For cancellation questions, respondents could select primary causes, and the analysis weights causes by the number selected (for example, if three causes are selected, each is counted as one-third).

In the solar cancellation results from that survey, community opposition accounts for roughly 24% of the weighted primary-cause attributions. This statistic is not a project-level probability (it reflects weighted cause attribution, and multiple causes can apply). We therefore use it as a proxy for the fraction of non-operational projects that plausibly involve high conflict, not as a definitive estimate of the share of cancellations caused by opposition.

B.4 Conservative scenario and worked examples

We set H_op = 0.19 based on our operational sample, and we use H_nonop = 0.24 as an illustrative proxy grounded in the developer-reported cancellation survey evidence [7]. Under this scenario: H_all ≈ r * 0.19 + (1 - r) * 0.24.

Calculated examples (out of 100 proposed projects for ease of interpretation):

The final row uses r = 0.19 as an illustrative benchmark drawn from a different analysis of interconnection-related withdrawal (see Queued Up: 2024 by Rand et al. [24]). It is included to show how the calculation changes when overall project completion rates are low due to any cause. We interpret the analysis to imply that under this scenario, incorporating non-operational projects increases the implied prevalence of high conflict-attention from 19% to roughly 20% to 23%, depending on the completion rate plugged into the calculation. This is a meaningful difference but one that we argue does not invalidate this study’s findings and approach.

B.5 Assumptions and limitations

Conflict-attention is an observational measure of attention and conflict language in media coverage; it does not reveal what the precise relationship between conflict-attention and various outcomes are e.g. delays, mitigations, etc. Future research could expand the range of project outcomes from cancellation to operation.

Appendix C. Decomposition of conflict–attention index by category​

Appendix D. Conflict-only robustness analyses​

Appendix D reports conflict-only robustness checks requested by reviewer; results are consistent with the main models. Tables generated from Stata binary logit models and descriptives.

Pearson chi2(1) = 317.06, p < 0.001; Cramer's V = 0.68; For these robustness checks we construct binary indicators for the presence of attention and conflict: AttentionAny = 1 if the attention sub-score >0 (else 0) and ConflictAny = 1 if the conflict sub-score >0 (else 0).

Note: 0% of projects without attention had conflict detected. 80.5% of projects with attention had conflict detected. Only 19.5% of projects with attention had NO conflict.

Note: Values represent percentage point changes in probability. Standard errors in parentheses. * p<0.10, ** p<0.05, *** p<0.01 Model 3 estimated only among projects with attention detected.

Appendix E. News article count distribution​

Table E1 presents the distribution of news article counts across the 686 solar projects in our sample.

The distribution is highly right-skewed (skewness = 5.5), with most projects receiving little to no media coverage while a small number received extensive attention. Twenty-seven percent of projects (n=184) had no news coverage identified through our search protocol. Conflict-term totals are computed via automated case-insensitive string matching in article titles and bodies for the ten-word lexicon (e.g., “protest” matches “protesters”; “concern” matches “concerned”). Articles are linked to projects using project-specific search terms/plant names; matches are intentionally inclusive to reduce false negatives. As a result, term totals should be interpreted as approximate indicators of the volume of conflict-language detected in the corpus rather than precise linguistic token counts or a direct measure of conflict intensity. Importantly, the conflict sub-score used in the index is based on the presence of lexicon terms/platform signals rather than the exact number of matched strings; the counts in this table are provided for descriptive context.

Note: The conflict word count stats reflect terms detected in news articles only. Three projects (0.4%) received conflict scores based solely on social media presence.

Appendix F. Distribution of conflict terms​

Table F1 presents summary statistics for conflict term mentions across the 686 solar projects.

Note: Projects may contain multiple conflict terms; percentages do not sum to 100%.

Appendix G. Bivariate associations with conflict-attention category​

Note: Panel A reports means and one-way ANOVA F-statistics. Panel B reports percentages (% with characteristic=1) and Pearson chi-square statistics.

*** p<0.001, ** p<0.01, * p<0.05

Sample sizes by category

No Conflict: 172, Low: 213, Medium: 171, High: 130