Social norms, pro-environmental identity, and finances: what motivates households to participate in energy efficiency programs?

Abstract

Municipal governments, often in collaboration with utilities, have implemented a range of energy efficiency programs to encourage homeowners and businesses to adopt energy efficiency upgrades. Energy efficiency holds promise to reduce energy consumption, reduce greenhouse gas emissions, improve public health, and reduce energy bills. However, these programs often suffer from poor participation and have typically had limited success. In this analysis, we use novel data to understand the relationship between social norms, pro-environmental identity, and household finances to understand program participation and retrofit decision-making. We find that the variables that predict retrofit decision-making do not explain a household’s initial decision to contact an energy efficiency program. We suggest that the processes that drive households to contact energy efficiency programs—a necessary first step in improving energy efficiency—are different from the processes that explain why households decide to upgrade their homes.

Appendices

Appendix 1. Descriptive statistics and factor analysis

Social norms

Fig. 2

Distribution of social norms questions

Table 5 Factor analysis for social norms

Pro-environmental identity

Fig. 3

Distribution of pro-environmental identity items

Table 6 Factor analysis for pro-environmental identity

Debt aversion and debt acceptance

Fig. 4

Distribution of debt acceptance and debt aversion items

Table 7 Factor analysis for debt acceptance and debt aversion items

Appendix 2. Average marginal effects for specific retrofits

Fig. 5

Average marginal effects for specific retrofits. Note: dependent variables are coded 0,1. Estimates were derived by re-running the regression model specifications from Table 2 and 3 with each outcome

We also asked respondents to report on what upgrades they made to their home. Response categories included: rooftop solar, heating upgrades, air conditioning upgrades, window and door sealing, and insulation. The question also included and “other” category and the option for respondents to write their “other” retrofit into a text box. The responses for the “other” category were too different from each other to be grouped effectively and hence we did not use the data for the “other” response category.

We estimated a series of binary logistic regression models for each outcome, using the model specifications from Table 2 and 3 (that is, the same combination of predictor variables). For each model, we calculated average marginal effects for the predictor of interest, like our approach in the main text. We provide these average marginal effects in Fig. 5 above. Overall, our results imply that, for many specific retrofits, the predictors are not statistically significant and have substantively small effects. However, awareness of consequences does predict improvements in insulation and the adoption of rooftop solar.

Appendix 3. Robustness checks

As shown in our regression models, some variables that were statistically significant in other papers (using other data, of course) were not consistently statistically significant in our models. Yet, compared to some work, our sample sizes are smaller. For instance, pro-environmental identity was statistically significant in multiple papers (e.g., Gatersleben et al., 2014; Schleich et al., 2021; Whitmarsh and O’Neill, 2010). To determine if a difference in sample size explains the divergence between our work and prior research, we conducted a series of simulations wherein we increased the size of our dataset by duplicating observations and then re-running the models in Table 3 and 4 for both the participation and the retrofit dependent variables. Appendix Table 8 shows the results of these simulations. The simulations suggest that, for program participation, pro-environmental identity was not statistically significant even when the sample size is much larger. The null effect is robust to a larger sample size. On the other hand, our indicator for emergency repair would cross the alpha = 0.05 threshold at three times the current sample size (i.e., roughly 1800 cases) while the indicator for emergency repair would only become statistically significant at n*7. Overall, the simulations for program participation imply that our results may diverge from other studies because these studies used larger samples that contributed to smaller standard errors and smaller p values.

The second panel of the table shows sample size simulations for the retrofit outcome variable. For this variable, most of the predictors of interest were statistically significant, so the results of the simulations are perhaps less substantively interesting. Still, in the interest of transparency, we present these results.

Table 8 Sample size simulations

Konfound analysis

Next, we turn to the konfound method. In the current application, konfound estimates the degree of measurement error (e.g., replaced with a case with no effect, or with an effect) that would be required to invalidate an inference—that is, to render a statistically significant effect non-significant (at alpha = 0.05) and to change a non-significant effect to statistically significant (Frank & Xu, 2017; Frank et al., 2013; Xu et al., 2019). Appendix Table 9 shows the percentage of cases that would have to be measured with error to change the inference. These estimates are derived from the regression models presented in Table 3 and 4.

Table 9 Konfound analysis

For program participation, we find that social norms (which were not statistically significant) could become significant with a relatively small amount of measurement error, but the effect of pro-environmental identity could only be statistically significant if a strong majority of the cases were measured with error (83.34%)—a scenario that is dubious. The non-significant effect of rainy day funds is comparatively less robust (7.94%). Overall, the konfound analysis for program participation implies that some predictors are less robust than others, although most inferences would require a non-trivial amount of measurement error to change the inference.

For retrofits, social norms were statistically significant but not highly robust (11.01%) while pro-environmental identity was somewhat more robust (24.32%). Non-emergency repairs, which were not statistically significant, also exhibit a relatively low level of robustness to measurement error (14.18%) while awareness of consequences was slightly more robust (20.67%).

Multiverse analysis

Appendix Table 10 provides the percentage of multiverse models wherein the predictor of interest takes the same sign (i.e., positive or negative) and is statistically significant. As we noted in the main text, the nulls effects reported in Table 3 appear to be robust. That is, they do not change to non-null under alternative model specifications, and our reported models do not appear to be unusual outlier models wherein the effects are not statistically significant. Appendix Fig. 6 provides a graphical distribution of the multiverse of coefficients, and a dashed line to represent the coefficient reported in Table 3. Appendix Fig. 7 suggests that pro-environmental identity, social norms, and awareness of consequences are highly robust and exhibit strong sign stability. Further, most of the non-significant predictors from Table 4 are rarely significant in a multiverse of models.

Table 10 Results of multiverse analysis

Fig. 6

Distribution of the multiverse coefficients for program participation

Fig. 7

Distribution of multiverse coefficients for retrofit