Short-term impact of tropical cyclones in Madagascar: evidence from nightlight data

ABSTRACT

This paper explores the short-term effect of tropical cyclones on economic activity at a local level in Madagascar. To do so, we combine high-resolution spatial data about nightlight brightness and exposure to tropical cyclones with geographic information at the smallest administrative level in Madagascar, namely the Fokontany. Our findings reveal that exposure to tropical cyclones leads to an ambiguous economic response as proxied by nocturnal brightness during the first year after the shock. However, during the second year, nightlights clearly increase, leading to an overall beneficial effect of 5% just 2 years after the tropical cyclone. We then provide a finer analysis by interacting wind speed exposure with variables that capture many heterogenous dimensions of our data. This analysis shows that for Fokontany sharing specific characteristics, the short-run effect of tropical cyclones is negative. However, the positive effect of exposure in the second year emerges as a regular pattern in our analysis. Our empirical study is in line with economic mechanisms suggesting that after a period of contraction, the impacted economy rebounds beyond the counterfactual trend that would otherwise be observed in the absence of a shock.

KEYWORDS:

• Tropical cyclones
• nightlight data
• madagascar
• high-resolution spatial data

JEL CLASSIFICATION:

• O44
• Q54

Acknowledgments

We gratefully acknowledge the financial support from the Fond Européen de Développement Régional (FEDER), the Région Réunion, and the Observatoire des Sociétés de l’Océan Indien (OSOI). For his insightful comments and research assistance, we thank Pascal Mouquet who helped us collect the nightlight data for Madagascar. We are also grateful to Thomas Vogt for providing us with an updated version of TCE-DAT for Madagascar.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ Many studies use economic and human damages related to disasters from the EM-DAT dataset. However, the use of this dataset has at least two limitations. First, data on economic damages are collected from different sources, while the quality of reporting changes over time (Strobl 2012). Second, monetary damages are likely to be correlated with output, namely the dependent variable in growth regression models (Felbermayr and Gröschl 2014).

² For more details about the economic mechanisms at play and the potential economic path followed after catastrophic natural events, the interested reader can refer to Hsiang and Jina (2014) or Heger and Neumayer (2019).

³ In Madagascar, the spatial unit of the Fokontany broadly corresponds to a ‘village’ or hamlet. Throughout the paper, we will interchangeably employ the terms of village and Fokontany.

⁴ “Black marble’ ’ products are freely available on adsweb.modaps.eosdis.nasa.gov.

⁵ In particular, we use the ‘BND_BRDF-Corrected_NL’ nightlight data from the VNP46A2 collection. For this data, the data producer classifies each value for each pixel to be of ‘high quality’, ‘poor quality’, and ‘no retrieval’. To compute our monthly average, we only keep high-quality pictures. Relying on all available pictures increases the variability of the time series of nightlight data, which can be imputed to data quality. We thank Pascal Mouquet for his research assistance during this step.

⁶ To compute the mean nightlight brightness for the pixels of a given Fokontany, we use the R function exact_extract from the package exactextractr.

⁷ For a detailed description of the data construction, the interested reader can refer to Kummu et al. (2018).

⁸ We also checked for the existence of a positive relationship between GDP and nightlight at a local level without taking log into account. Furthermore, we also considered a less spatially disaggregated measure, namely the ‘commune’ and found a similar positive association with a correlation of about 0.7.

⁹ More specifically, the dependent variable of the regression is the log of local GDP and the independent variable is the log of the nightlight radiance value.

¹⁰ For the global sample, we find an elasticity of local GDP to nightlight of 1.36 with a R-squared of 0.39. Complete regression results associated with this application for all samples are available upon request.

¹¹ For the sake of simplicity, we do not add an index to designate pixels.

¹² The initial version of the dataset is referenced as Geiger et al. (2017) and is available in its initial form at https://dataservices.gfz-potsdam.de/pik/showshort.php?id=escidoc:2387904

¹³ As recalled by Dell et al. (2014), in a pooling model context, increasing the number of control variables does not necessarily reduce the omitted variable bias.

¹⁴ Details about the ‘fixest’ package can be found via the following link: https://cran.r-project.org/web/packages/fixest/vignettes/fixest_walkthrough.html.

¹⁵ To further strengthen this choice, we ran the Lagrange multiplier test. Specifically, we tested for the presence of village, time, and both kinds of fixed effects. Each time, the tests indicates the presence of fixed effects. As the presence of the Fokontany and time fixed effects is not rejected, the model with both of them is our preferred specification.

¹⁶ At this stage, it should be mentioned that we have no information about the true origin of the positive boosts in nightlight brightness unveiled by our regressions. The latter could come from either the private or the public sector (through specific reconstruction policy programs after the adverse shock) or from international aid.

¹⁷ Examples include Strobl (2011), Bertinelli and Strobl (2013), Elliott et al. (2015), and Mohan and Strobl (2017)

¹⁸ Implicitly, the first bin ${\tilde{w}}_{\mathit{t}}^0=1(W_t\in [0;88])$ serves as the reference in the regression.

¹⁹ We also conducted a similar analysis based on other months. This had no impact on the qualitative or quantitative results.