Economic impact of natural disasters, spillovers, and role of human development: case of Indonesia

Abstract

We investigate how natural disaster damages affect regional economic growth and their spatial spillover effect in various human development (HDI) level settings by utilising a panel annual district-level dataset from Indonesia. We employ Spatial Durbin Model to achieve these objectives. We find that disaster damage to houses has a negative effect on district-level per capita output growth. Meanwhile, the effect of disaster damage to people on economic growth depends on their HDI level, with low-HDI districts affected negatively and high-HDI districts positively. Our analysis also finds spatial spillover effects on neighbouring regions’ growth, and heterogeneous effects across economic sectors.

Appendices

Appendix A Disaster data summary statistics

See Appendix Table 3.

Table 3 People and houses-related damages by disaster type (median and range)

Appendix B Selection bias test

Selection bias in disaster damage reports might occur when districts with particular characteristics (such as higher level of development) are more likely to report any damages than others. One of the methods of testing this is to check if there exists any correlation between disaster damage reporting probability and district characteristics after controlling for actual disaster intensity (Felbermayr and Gröschl 2014). Because comprehensive district-level disaster intensity data in Indonesia is not available, we use disaster damage in adjacent districts as its proxy. Furthermore, following Felbermayr and Gröschl (2014), we choose GDP per capita (in natural logarithm form) as the district characteristic of interest.

Table 4 presents the estimation result. The dependent variable is a dummy variable which value is 1 if there is any reported disaster damage and 0 if otherwise. There is no correlation between log GDP per capita and damage reporting probability after controlling for the spatial lag of the weighted number of people affected. Thus, we do not find any evidence of selection bias in the DIBI dataset.

Table 4 Selection bias test result

Appendix C Variable imputation

We employ linear extrapolation to impute outliers and missing values in variables with the time trend. For those without time trend, we impute them with the median value of the nearest aggregation level available (i.e. using provincial median value, then national median value). We define outliers as any value outside 0 to 100 range for share variables (in %) and outside $-30$ to 30 for output growth variables (in %). The percentage of imputed observations can be seen in Table 5.

Table 5 Percentage of imputed obervations