Assessing the Impact of Offshore Wind Power Deployment on Fishery: A Synthetic Control Approach

Abstract

Global efforts to decarbonize the energy sector are necessary to curb global warming, and many countries plan to achieve this through the rapid deployment of offshore wind power. However, the deployment will likely decelerate if local fishers plying their trade near maritime areas likely to host offshore wind farms oppose the deployments due to anticipated fishery disruptions. To date, there is no body of knowledge on the causal impact of offshore wind farm installation on local fisheries. Using fishery production panel data at the municipality level in Japan, this study applies a synthetic control method to measure causal impacts. The results suggest that offshore wind farms currently installed in Japan are unlikely to disrupt local fisheries. We find no statistically significant effects either on aggregated fishery production or on production categorized by fishery type. Moreover, we find no spatial spillover effects. Although the generalization of our findings requires caution because they are based on small-scale wind farms, our results imply that such moderate-size wind projects may not harm local fisheries.

Appendices

Appendix A: Methodological Details

1.1 A.1 Objective Function

1.1.1 A.1.1 Synthetic Control Method

Let $\pmb {X}_0=\left( \pmb {X}_2, \ldots , \pmb {X}_N \right)$ denote a $T_0\times N$ matrix of the pre-treatment outcomes of municipalities without offshore wind farms. Similarly, let $\pmb {Z}_0=\left( \pmb {Z}_2, \ldots , \pmb {Z}_N \right)$ represent a $K\times N$ matrix of municipality covariates of controlled municipalities. In addition, define $\pmb {M}_i=(\pmb {X}_i', \pmb {Z}_i')'$ and $\pmb {M}_0=\left( \pmb {M}_2, \ldots , \pmb {M}_N \right)$. Then, the weights of the conventional SCM in Eq. (2), $\pmb {w}$, can be obtained by solving the following problem:

$$\begin{aligned}&\mathop {\mathrm{arg~min}}\limits _{\pmb {w}} \left\Vert \pmb {M}_1-\pmb {M}_0\pmb {w}\right\Vert , \nonumber \\ \text {s.t. } \forall i\in&\{2,\ldots ,N+1\},w_i\ge 0 \wedge \sum _{i=2}^{N+1}w_i = 1. \end{aligned}$$

(A.1)

The norm in Eq. (A.1) is constructed to account for the importance of each variable as follows:

$$\begin{aligned} \left\Vert \pmb {M}_1-\pmb {M}_0\pmb {w}\right\Vert = \left( \sum _{k=1}^{T_0+K} v_k \left( M_{k1}-\sum _{i=2}^{N+1}w_i M_{k i}\right) ^2 \right) ^{1/2}, \end{aligned}$$

where $\pmb {v}=(v_1,\ldots ,v_K)'$ is a vector of weights that reflects the relative importance of each variable. The optimal weight $\pmb {w}^*$, which is the solution of Eq. (A.1), can be obtained at a fixed value of $\pmb {v}$ or by updating its value within an algorithm. See Abadie (2021) for a more detailed description of the algorithms.

1.1.2 A.1.2 Augmented Synthetic Control Method

Ben-Michael et al. (2021) propose Ridge ASCM, where weights in Eq. (3) can be obtained as the solution of the following problem:

$$\begin{aligned} \mathop {\mathrm{arg~min}}\limits _{\eta _0, \pmb {\eta }_x, \pmb {\eta }_z}\frac{1}{2}\sum _{i=2}^{N+1} \left( Y_i - \left( \eta _0+\pmb {X}_i'\pmb {\eta }_x+\pmb {Z}_i'\pmb {\eta }_z \right) \right) ^2 + \lambda _x \left\Vert \pmb {\eta }_x\right\Vert ^2_2 + \lambda _z \left\Vert \pmb {\eta }_z\right\Vert ^2_2, \end{aligned}$$

(A.2)

where $\lambda _x$ and $\lambda _z$ are hyperparameters that reflect the degree of penalty on extrapolation. When only pre-treatment outcomes are used for approximation, Ben-Michael et al. (2021) show that Eq. (A.2) can be written as

$$\begin{aligned} \mathop {\mathrm{arg~min}}\limits _{\pmb {\eta }\text { s.t. }\sum _{i=2}^{N+1}\eta _i=1}\frac{1}{2\lambda _x}\left\Vert \pmb {X}_1-\pmb {X}_0\pmb {\eta }\right\Vert ^2_2 + \frac{1}{2}\left\Vert \pmb {\eta }-\pmb {w}\right\Vert ^2_2, \end{aligned}$$

(A.3)

and the ASCM estimator of $Y_{it}(0)$ is:

$$\begin{aligned} {\hat{Y}}_{1t}(0)=\sum _{i=2}^{N+1} \eta _i Y_{it}. \end{aligned}$$

Equations (A.2) and (A.3) allow the weights to take negative values; therefore, extrapolation is allowed.

1.2 A.2 Inference

Conventional SCM conducts statistical inference using permutation methods (Abadie et al. 2010). The treatment effects are estimated by iteratively assigning the treatment to units that do not receive it. The placebo effects are then compared to the impact estimated using the unit with the actual treatment. If the unit of interest has a significant impact, it should be larger than the placebo effects.

For statistical inference on the ASCM, Ben-Michael et al. (2021) proposed the conformal inference approach developed by Chernozhukov et al. (2021). The approach first specified a null hypothesis, $\tau _{it}=\tau _0$ (typically $\tau _{0}=0$). Then, the post-treatment outcome was adjusted under the null hypothesis: $\tilde{Y}_{1t}=Y_{1t}-\tau _0$. The ASCM weights $\tilde{\pmb {\eta }}(\tau _0)$ are re-estimated after including $\tilde{Y}_{it}$ in the original data. The p-value is obtained by calculating how the adjusted residual for $t>T_0$, $Y_{1t}-\tau _0-\sum _{i=2}^{N+1}{\tilde{\eta }}_i(\tau _0)Y_{it}$, conforms to the residuals for $t\le T_0$. See Ben-Michael et al. (2021) and Chernozhukov et al. (2021) for more details.

Appendix B: Additional Figures

See Figs. 11, 12, 13, 14, 15, 16, 17 and 18.

Fig. 11

ASCM estimation results for each fishery type (Choshi). Notes: The shaded areas show 95% confidence intervals

Fig. 12

ASCM estimation results for each fishery type (Kitakyushu). Notes: The shaded areas show 95% confidence intervals

Fig. 13

ASCM estimation results for each fishery type (Goto). Notes: The shaded areas show 95% confidence intervals

Fig. 14

ASCM estimation results for each fish species (Choshi). Notes: The shaded areas show 95% confidence intervals

Fig. 15

ASCM estimation results for each fish species (Kitakyushu). Notes: The shaded areas show 95% confidence intervals

Fig. 16

ASCM estimation results for each fish species (Goto). Notes: The shaded areas show 95% confidence intervals

Fig. 17

ASCM estimation results for marine aquacultural production. Notes: The shaded areas show 95% confidence intervals

Fig. 18

ASCM estimation results at the prefectural level for coastal fishery production. Notes: The shaded areas show 95% confidence intervals. a–c are the results obtained by using the same sample period as in the text. d–f represent the results obtained by expanding the sample period from 2004–2018 to 1984–2018. The results in d–f were derived using only the shares of fishery types as the covariates

Appendix C: Additional Tables

See Tables 4, 5, 6, 7, 8, 9 and 10.

Table 4 Pooled OLS of total fishery production

Table 5 Pooled OLS of coastal fishery production

Table 6 Difference-in-difference estimation of total fishery production

Table 7 Difference-in-difference estimation of coastal fishery production

Table 8 Productions of major fish species by fishery types

Table 9 Aquacultural production

Table 10 Comparison of covariates over treatment and control groups