3.1 Study period and the cyclones covered in the study

The study period for analyzing the impact of cyclones on the sales and assets of Indian manufacturing firms is April 2008 to March 2009 (hereafter written as 2008) to 2019–20 (2019). Data for the years 2013–14 to 2020–21 are used to analyze the impact of cyclones on the survival probability of manufacturing plants. Twenty-three cyclones that occurred during 2008–2019 are included in the study. Details are provided in online appendix A1.

Over the last two decades, there has been an upward trend in India's landfalling (tropical) cyclones. During 2002–2007, there were six landfalling cyclones in India; during 2018–2022, the corresponding number was 17. Analyzing the districts impacted by the 23 cyclones that occurred from 2008 to 2019 that were selected for the study, we find that the districts repeatedly affected are concentrated in the eastern coastal states of India, mainly Andhra Pradesh, Odisha, and Tamil Nadu. However, a shift is underway towards an increasing number of cyclones on the west coast (Deshpande et al., Reference Deshpande, Singh, Ganadhi, Roxy, Emmanue and Kumar2021), which might be connected with climate change.

3.2 Incorporating wind speed as an explanatory variable in the econometric model

If an analysis of the impact of cyclones is undertaken at the establishment level as, for instance, in Elliott et al. (Reference Elliott, Liu, Strobl and Tong2019) and Juárez-Torres and Puigvert (Reference Juárez-Torres and Puigvert2021), then there is a one-to-one correspondence between wind speed and the establishment-level performance variable. Hence, the wind speed variable or a measure of firm exposure to high-speed wind based on wind speed data may be included as an explanatory variable in the econometric models estimated. However, problems arise when one uses data at the firm level, and the firms have multiple establishments located in different parts of the country. The one-to-one correspondence between wind speed and the outcome (performance) variable does not hold. If a firm has n establishments, taking an average wind speed at the n establishments as an explanatory variable in the econometric model is not correct. Forming a measure of exposure to high-speed wind at each of the different establishments and then aggregating these measures at the firm level by simple summation is also problematic.

We address the above estimation issue in two ways. First, we confine our analysis to firms with at most three plants or establishments (see online appendix A2 which explains in detail the rationale for this choice). By restricting the analysis to firms with three or fewer plants, we ensure greater homogeneity in the sample of firms included in the study. Second, for most of the analysis presented in the paper, we capture, for each year, the effect of cyclones by using a dummy variable, C, which represents whether a firm has a plant that was exposed to a cyclone. For each cyclone included in the study, we identify the districts that have been affected and the firms that have one or more plants in those districts. We have collected data on districts affected by different cyclones from diverse sources. These include various annual reports titled Disastrous Weather Events, compiled by the India Meteorological Department, Pune, Government of India. In addition, the Government of Andhra Pradesh provides information on the districts affected by various cyclones (https://apsdma.ap.gov.in/history_of_storms.php). The Government of Odisha also provides information on the estimates of the extent of damage caused by cyclones in different districts (available in various Memorandum on the Very Severe Cyclonic Storm, Government of Odisha). This compiled information helps identify the districts affected. We have also considered the newspaper and media reports on the cyclones in India as additional sources of information. For constructing the aforementioned dummy variable (C) for our econometric analysis, we consider only the districts affected by cyclones and leave out the districts partially affected.

The dummy variable C is constructed at the firm level. The value of this dummy variable for firm i in year t (denoted by ${C_{it}}$) is one if any of the firm's plants are located in a district that was affected by a tropical cyclone in year t. The variable is assigned the value of zero otherwise. Although this measure is crude, it has the advantage of simplicity and also has precedence.Footnote ⁶

Our approach based on a dummy variable for cyclones is a major deviation from the previous studies on the effect of cyclones on firms, which have commonly used wind speed data. Since the econometric estimates based on our approach need verification, we have made alternate estimates based on wind speed.

The International Best Track Archive for Climate Stewardship (IBTrACS) data have commonly been used to measure wind speed to undertake studies on the impact of TCs. We have used a different dataset, namely ERA5, for our analysis based on wind speed data, on several considerations. To our knowledge, this is the first study in which a reanalysis dataset for meteorological variables (ERA5) has been used to study the impact of TCs on the performance of industrial firms.

For constructing the explanatory variable representing TCs, we take the surface wind component at a 10-meter elevation. The analysis is done only for the single-plant firms so that a district is identified for each firm/plant. The maximum wind speed during a cyclone is computed for different districts of the country. If there were two cyclones in a year, the maximum wind speed for the two cyclones is compared for each district, and the higher of the two figures is taken (similar treatment for three cyclones in a year).

The ERA5 dataset has the advantage of having a high spatial (up to 0.25 degrees) and temporal (hourly) resolution, ensuring detailed regional and local wind speed variations. It provides information at various vertical levels, enabling a comprehensive analysis of the vertical profile of wind speed. Its reanalysis nature leads to a more consistent and homogeneous dataset compared to IBTrACS, which relies on best-track information and may have coarser resolutions, especially in data-sparse regions.

A disadvantage in using ERA5 data is that it tends to understate/under-represent the intensity of the TCs (Xiong et al., Reference Xiong, Yu, Fu, Dong and Liu2022; Dulac et al., Reference Dulac, Cattiaux, Chauvin, Bourdin and Fromang2024). To address this issue, the equation for computing firm exposure to wind which we adopt is a modified version of that in Pelli et al. (Reference Pelli, Tschopp, Bezmaternykh and Eklou2023). It may be written as

(1)\begin{equation}{x_{dh}} = \frac{{{{({w_{dh}} - \phi )}^2}}}{{{{({w^{max}} - \phi )}^2}}}\;\textrm{if}\;{w_{dh}} > \phi ;\,{x_{dh}} = 0\;\textrm{otherwise}.\end{equation}

In this equation, ${w_{dh}}$ is the (maximum) wind speed for cyclone h in district d. ${w^{\textrm{max}}}$ is the maximum wind speed observed in the entire sample (all cyclones and districts). If there were two cyclones in a year, the values of ${x_{dh}}$ are computed for the two cyclones, and the higher of the two figures is taken (similar treatment for three cyclones).

The value of $\phi$ is not specified but allowed to be determined by the data. Define ${w^{Q3}}$ as the third quartile of wind speed observed in the entire sample. The lower limit of $\phi$ is taken as ${w^{Q3}}$. Then, various integer values of $\phi$ (in meters per second) higher than ${w^{Q3}}$ are taken as possible alternative values of $\phi$, and the econometric models explaining real sales and real capital stock are estimated for each of the values of $\phi$. The value of $\phi$ for which the results for the firm exposure to wind variable is found to be the best in terms of statistical significance (the t-value of the estimated coefficient is the highest) has been finally chosen as the right estimate of $\phi$.

4.1 Data sources

The basic data source for the analysis of the impact of cyclones on the sales and total assets of manufacturing firms is the Prowess database of the Centre for Monitoring Indian Economy, Mumbai, India, combined with data on cyclones and wind speed explained above. Our analysis is confined to manufacturing firms, covering 2008 to 2019.

Prowess is an extensive database on the financial performance of companies. It includes listed companies, unlisted public companies, and private companies of different sizes and ownership groups. The data are procured from the companies' audited annual reports and financial statements. This data source has been widely used for research on Indian manufacturing firms.

Manufacturing firms covered in the Prowess database account for about 60–70 per cent of the economic activity of organized manufacturing in India and most of the capital stock of organized manufacturing. This makes an analysis based on Prowess useful for understanding how cyclones impact Indian manufacturing (see online appendix A2). As stated above, we consider only firms with a single establishment (i.e., a plant) or, at most, three establishments.

Prowess provides information (latest) on the district location of plants belonging to manufacturing companies. Using the information on the districts affected by the cyclones and the district location of the establishments of companies, we have constructed the previously mentioned dummy variable representing the incidence of cyclones, which is subsequently used for the econometric analysis. One econometric model is estimated to explain real sales, and another model is estimated to explain the real value of total assets. In the estimated econometric models, the choice of model specification is mainly driven by the consideration of parsimony. Thus, only a few fixed effects have been considered and included.

For one component of the analysis in section 6, the firm-exposure-to-wind variable based on wind speed associated with cyclones is used instead of the dummy variable for cyclones. The purpose of this analysis is to ascertain if the key findings of the analysis based on the cyclone dummy are borne out by the econometric results based on wind speed data.

4.2 Key variables and the specification of econometric models

4.2.2 Model specification

To analyze the impact of cyclones on real sales and the real value of total assets of manufacturing firms, two different specifications of the econometric model have been used. The model estimated for the econometric analysis of real sales, in terms of its core elements, may be written as

(2)\begin{equation}ln{S_{it}} = {a_i} + {b_t} + {\beta _S}{C_{it}} + \mathop \sum \limits_u {\gamma ^u}Y_{it}^u + {\xi _{it}}.\end{equation}

In this equation, S denotes the real (i.e., deflated) value of sales. The subscript i is for firm, and subscript t is for time (year). The variable C is a dummy variable to represent the incidence of cyclones (explained earlier). ${C_{it}}$ takes the value of one if the plant (or one of the plants) of firm i is in a district that was affected by a cyclone (or cyclones) in year t; it takes the value of zero otherwise. Y is a set of variables representing firm characteristics that are introduced in the model as controls. The term ${\xi _{it}}$ is the error term. The model is estimated by applying the fixed-effects model.

For firms with a single plant and for firms with all their plants in a single state, a better specification has been used in which state-by-year dummy variables are introduced. The model may thus be written as

(3)\begin{equation}ln{S_{irt}} = {a_{ir}} + {\theta _{rt}} + {\beta _S}{C_{irt}} + \mathop \sum \limits_u {\gamma ^u}Y_{irt}^u + {\xi _{irt}}.\end{equation}

In this equation, subscript r is for State/Union Territory (UT). The term ${\theta _{rt}}$ represents state-by-year dummy variables. Other notations are the same as in equation (2). S_irt denotes the real value of sales of firm i, belonging to state/UT r, in year t. Equations (2) and (3) are estimated by applying the fixed-effects model. For both single-plant firms and for firms with multiple plants in the same state, the specification in equation (3) has been used. For a firm having multiple plants in different states, the estimates are based on equation (2).

A dynamic specification is used to explain total assets (deflated). Since changes in K are determined by investments made by firms, which are likely to be influenced by the level of capital stock prevailing at the end of the previous year, a dynamic specification seems more appropriate. The model may be written as

(4)\begin{equation}ln{K_{it}} = {a_i} + \lambda t + \; \phi ln{K_{i,t - 1}} + {\beta _K}{C_{it}} + \mathop \sum \limits_u {\delta ^u}W_{it}^u + {\zeta _{it}}.\end{equation}

In this equation, K denotes the real value of total assets. W is a set of variables representing firm characteristics that are introduced in the model as controls. A time trend variable is included whose coefficient is λ. Other notations are similar to that in equations (2) and (3). Since the lagged dependent variable is included in the model as an explanatory variable, the fixed-effects model cannot be applied. Hence, a different approach has been taken for its estimation.

To estimate equation (4), it is transformed into a difference form. Thus, the dependent variable is $\Delta ln{K_t}$.Footnote ⁸ The explanatory variables include the lagged dependent variable $\Delta ln{K_{t - 1}}$. To estimate the model, $\Delta ln{K_{t - 1}}$ has been instrumentedFootnote ⁹ by $\Delta ln{K_{t - 2}}$ along with another instrument, namely the one-period lagged and two-period lagged values of the expenditure on repair and maintenance and insurance, which is deflated by a price index for fixed assets (using the implicit deflator for gross fixed capital formation in manufacturing which has been computed using data from National Accounts Statistics). Another point to note is that two-period-lagged values of the explanatory variables have been used. Thus, the equation estimated may be written as

(5)\begin{align}\Delta ln{K_{it}} & = \lambda + \; \phi \Delta ln{K_{i,t - 1}} + {\beta _{K1}}\Delta {C_{it}} + {\beta _{K2}}\Delta {C_{i,t - 1}} + \mathop \sum \limits_u {\phi ^u}\Delta W_{it}^u \nonumber \\ & \quad +{+} \mathop \sum \limits_u {\pi ^u}\Delta W_{i,t - 1}^u + \varphi {\; _{it}}.\end{align}

The control variables (Y) used for estimating equations (2) and (3) are: (i) the age of the firm, based on year of incorporation; (ii) OFDI (outward foreign direct investment) intensity, that is, investment outside India divided by total assets; (iii) export intensity (ratio of exports to sales), one year lagged; and (iv) labour productivityFootnote ¹⁰ (one year lagged). The control variables (W) used for estimating equation (5) are (i) debt-equity ratio (one year lagged), (ii) profitability (PBDITA – profit before depreciation, interest, tax and amortization – divided by sales), and (iii) capital intensity or capital-labour ratio (total assets at constant prices divided by employment), one year lagged.

6.1 Impact on sales

The estimates of equations (2) and (3) are shown in table 1. Regression-1 (hereafter Reg-1) is for single-plant firms. In Reg-2, firms with up to three plants are included. In Reg-1 and Reg-3, equation (3) is estimated, allowing for state-by-year effects using dummy variables. In Reg-2, equation (2) is estimated. The last two columns of table 1 present the results when Reg-3 is re-estimated separately for the bottom 75 per cent of the firms (observations) and the top 25 per cent of the firms (observations) in terms of the real value of total assets. The aim is to find out if the size of the firm makes a difference in the impact of cyclones on sales.

Table 1.

Regression results, explaining real sales, fixed-effects model

Notes: Robust standard errors clustered at the district level for Reg-1, state level for Reg-3, Reg-4, and Reg-5, and at the firm level for Reg-2, shown in parentheses. Reg-1, Reg-3, Reg-4 and Reg-5 include state-by-year dummies. Reg-2 includes only year dummies. Some firms are included in Reg-4 for some years and in Reg-5 for others (because of increased asset size). For the joint test of coefficients, the null hypothesis is that all are equal to zero.

Source: Authors' computations based on the Prowess database.

The signs of the estimated coefficients of the firm characteristics variables are along expected lines, and therefore, the results regarding the control variables are not discussed.Footnote ¹² Instead, the focus is on the coefficient of the dummy variable representing cyclones (hereafter called the cyclone dummy variable). The estimated coefficient of the cyclone dummy variable is negative and statistically significant in Reg-1, which is estimated for single-plant firms, as well as in Reg-3, a regression equation estimated for firms that have 1–3 plants, all in the same state/UT. Based on the results obtained, it may be inferred that cyclones reduce sales of single-plant manufacturing firms and firms with up to three plants having all plants in the same state/UT by about six per cent. This finding is broadly consistent with the finding of Pelli et al. (Reference Pelli, Tschopp, Bezmaternykh and Eklou2023).

When the firms with 1–3 plants are considered for the analysis (estimating equation (2), shown in Reg-2), whether or not they have their plants in the same state/UT, the coefficient of the cyclone dummy variable is found to be statistically insignificant.

The firms that have multiple plants are bigger in size than the firms with a single plant (see table A2 in online appendix A2). Firms with two or three plants located in different states are generally bigger than firms with two or three plants in the same state. The finding of a statistically insignificant coefficient of the dummy variables for cyclones in Reg-2 indicates that the sales of big-sized firms are less impacted, or not impacted at all, by cyclones. This claim is supported by the results presented in the last two columns of the table. The results show that the impact of cyclones on sales is negative and statistically significant for the bottom 75 per cent of firms (relatively smaller in size), but the impact is statistically insignificant for the top 25 per cent of firms in terms of total assets. Thus, it may be inferred that for small-sized firms, there is a significant loss in sales due to cyclones, but for relatively much bigger multi-plant companies, especially those with plants in different states, the effect of cyclones on their sales might be small or even marginal (corroborative empirical evidence is presented in online appendix A4).

6.2 Impact on total assets

Table 2 presents the estimates of equation (5), which pertains to the impact of cyclones on total assets. The estimates have been made by applying the FD-IV regression. For the explanatory variables, other than the lagged dependent variable, the current difference and the lagged difference are included. Observations in which $\Delta lnK$ is more than 2 or less than (−2) have been dropped, as these are exceptionally high increases or decreases in capital stock in a year, although this does not make much difference to the results.

Table 2.

Estimates of the model explaining real capital stock, first-differenced instrument variable (FD-IV) regression

Notes: Robust standard errors clustered at the district level in Reg-6, and at the firm level in Reg-7, shown in parentheses. Observations in which $\Delta lnK$ is more than 2 or less than (–2) have been dropped.

Variable notation:lnK = log of the real value of total assets; C = Cyclone dummy; DE = Debt-equity ratio (lagged); PR = Profit (PBDITA) by sales; $lnKlr$ = Capital-labour ratio (in logarithms, one year lagged).

Source: Authors' computations based on the Prowess database.

Reg-6 is for single-plant firms, and Reg-7 is for firms having 1–3 plants, whether or not the plants are in the same state/UT.Footnote ¹³ It is seen in the results that the coefficient of profitability is positive and statistically significant and that of the debt-equity ratio is negative and statistically significant. The signs of these coefficients are as expected.

The coefficient of the lagged capital-labour ratio is negative and statistically significant. A possible interpretation of this finding is that if the ratio of capital to labour was relatively high in the previous year, there is a tendency for a downward adjustment in the current year.

The coefficient of the dummy variable for cyclones is negative and statistically significant in both Reg-6 and Reg-7. The results suggest that cyclones reduce the capital stock by about 1.5 per cent (corroborative evidence is provided in online appendix A5 based on the difference-in-difference estimator). The finding that cyclones cause a fall in total assets is in accord with the finding of Pelli et al. (Reference Pelli, Tschopp, Bezmaternykh and Eklou2023). Considering the impact in the current year and the next year, the total comes to about 2.5 per cent, which is close to the estimate (2.2 per cent) of Pelli et al. (Reference Pelli, Tschopp, Bezmaternykh and Eklou2023).

The estimated model for sales presented above indicated that cyclones adversely affect sales in medium-size and relatively smaller firms, which may not be present among relatively bigger plants. By contrast, in the case of total assets, the results indicate that cyclones cause a reduction in total assets among single-plant firms and relatively bigger multi-plant firms. The effect on bigger firms' assets does not appear to be any less than that for small firms.

6.3 Lagged impact of cyclones on sales and total assets

An interesting question is whether cyclones have a lagged impact on sales and total assets beyond the effect in the current year. This issue has been analyzed, and the results are presented in online appendix A6. For this purpose, equation (3) has been extended by adding two lagged terms of ${C_{irt}}$, the cyclone dummy variable (i.e., ${C_{ir,t - 1}}$ and ${C_{ir,t - 2}}$ are introduced in the regression equation). Similarly, equation (5) has been extended by adding $\Delta {C_{i,t - 2}}$ and $\Delta {C_{i,t - 3}}$.

When the extended version of equation (5) is estimated, which explains the variation in total assets, the coefficients of $\Delta {C_{i,t - 2}}$ and $\Delta {C_{i,t - 3}}$ are found to be statistically insignificant. When the extended version of equation (3) is estimated, which explains variations in sales, the coefficient for ${C_{ir}}{,_{t - 2}}$ is found to be statistically insignificant. The coefficient for ${C_{ir}}{,_{t - 1}}$ is negative and statistically significant in the estimated regression equation for firms with 1–3 plants in the same state/UT, but it is statistically insignificant in the regression for single-plant firms (which dominate our sample). The results disagree, making it difficult to draw a definite conclusion regarding the lagged impact of cyclones on sales. Hence, the inference that may be drawn based on the results reported in online appendix A6 is that there is either no lagged impact of cyclones or, maybe, one period lagged impact of cyclones on sales and assets.

6.4 Influence of firms' trade-technology orientation on the cyclone impact

The next question investigated is whether the effect of cyclones on manufacturing firms is influenced or conditioned by their trade and technology orientation. This aspect was discussed in section 2, and some arguments were given as to why a firm with a high trade-technology orientation will have greater resilience against cyclones.

We have formed an index of firms' trade-technology orientation to investigate the above issue. For this purpose, four variables are considered: R&D intensity, technology import intensity, capital goods import intensity, and materials import intensity. Then, by applying principal component analysis, an index is formed, representing the trade-technology orientation of different firms (see online appendix A7 for details). This measure is hereafter called the trade-technology orientation index.

Considering the nature of the distribution of the trade-technology orientation index (L-shaped), the firms (or, to be more truthful to the procedure adopted, the observations) have been divided into two sets or groups. The 67^th percentile has been taken as the cut-off. Thus, the bottom two-thirds of the firms (or observations) in terms of the index have been included in one set, and the upper one-third have been taken in the other. Then, equations (3) and (5) have been estimated separately for these two groups of firms (observations). The results of this exercise for sales are shown in table 3, and those for capital stock are presented and discussed in online appendix A8.

Table 3.

Regression results, explaining real sales, fixed-effects model, distinguishing according to trade-technology orientation

Note: Robust standard errors clustered at the district level in Reg-8 and Reg-9 and at the state level in Reg-10 and Reg-11, shown in parentheses. State-by-year dummy variables are included. The measure of trade and technology orientation is discussed in the text. The two groups are formed by taking the bottom 67% and the top 33%.

Source: Authors' computations based on the Prowess database.

In table 3, separate estimates are presented for single-plant firms and multi-plant single-state firms. The coefficient of the cyclone dummy variable is negative and statistically significant in Reg-8 and Reg-10, matching the results in table 1. However, the coefficient is not statistically significant in Reg-9 and Reg-11, which are for highly trade-technology-oriented firms. It appears from these results, therefore, that cyclones lead to a fall in sales of manufacturing firms that do not have much involvement in international trade and technology acquisition activities. On the other hand, firms that are highly trade-technology-oriented may not experience a fall in their sales after a cyclone. A similar pattern is seen when an analysis is undertaken of the effect of cyclones on capital stock; there is an effect for less trade-technology oriented firms, but not for high trade-technology oriented firms (see online appendix A8).Footnote ¹⁴