Figure 1. Nonfinancial corporate debt, loans and debt securities (percent of GDP).Source: IMF Global Debt Database and authors calculations.Note: Consistent sample of 70 countries across time: advanced economies (35) and emerging market and developing economies (35). Dotted upper(lower) lines depict the 75th (25th) percentiles of the overall distribution. Solid lines depict the median.

Table 1. Contribution of leverage to the average investment response following recessions

Figure 2. Evolution of (log) investment following a recession.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{is}^{k}+\gamma _{nq}^{k}+\sum _{j=-\mathrm{k}}^{4}\beta _{j}^{k}R_{i,t-j}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession (and is the variable of interest), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{is}^{k}$ are country-sector fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\beta _{0}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure 3. Differential effect of recession on (log) investment for high-debt companies relative to low-debt companies.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure 4. Robustness: exclude recessions years when computing high-debt dummy.Note: Robustness check where companies are classified into high or low debt based on whether average leverage in all non-recession years is above or below median of their industry. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage in non-recession years within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure 5. Robustness: include interaction of macro variables with high-debt dummy.Note: Robustness check where interaction of macro variables with high-debt dummy are included as controls. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum \nu _{j}^{k}M_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage in non-recession years within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. Interaction of various macro variable $M_{i,t-j}$ with the high-debt dummy are included as controls. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure 6. Robustness: include interaction of recession dummy with firm characteristics.Note: Robustness check where interaction of dummy variables for various firm characteristics and the recession variable are included as controls. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum \nu _{j}^{k}R_{i,t-j}*F_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. $F_{n}$ is a dummy variable defined analogously to the high-debt dummy but for different firm characteristics i.e. firm classified as 1 if it is above the median as per the characteristic when averaged across the entire sample. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure 7. Differential effect of recession on (log) investment for high-debt companies relative to low-debt companies—controlling for profitability.Note: impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid black line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. The solid red line indicates the baseline response reported in Figure 3. Standard errors are clustered at two-way at the firm and country-time level.

Figure 8. Differential effect of recession on (log) investment for high-debt companies relative to low-debt companies—role of bankruptcy efficiency.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha_{ist}^{k}+\gamma_{nq}^{k} + \sum_{j=-\mathrm{k}}^{4}\mu^{k}_{j} R_{i,t-j}\ast D_{n} + \sum_{j=-k}^{4} \sigma_{j}^{k} R_{i,t-j} \ast D_{n} \ast Z_{i,t-1} + \sum\delta^{k}_{j} D_{n}*Z_{i,t-1}+\sum^{4}_{j=1} \theta^{k}_{j} y_{n,i,t-j} + \varepsilon^{k}_{n,i,t}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company $n$ is a high-debt company (above median leverage within industry), $Z_{i,t-1}$ is an indicator of bankruptcy efficiency, $\lambda^{k}_{nd}$ are firm-quarters fixed effects, and $\alpha^{k}_{ist}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid black line shows the point estimate for $\mu^{k}_{o}$ for different horizons k, while the dotted lines are the 68 percent and 90 percent confidence intervals. The solid red line indicates the baseline response reported in Figure 3. Standard errors are clustered at two-way at the firm and country-time level.

Figure 9. Differential effect of recession on (log) investment for high-debt companies relative to low-debt companies—global vs. domestic recessions.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid black line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. The solid red line indicates the baseline response reported in Figure 3. Standard errors are clustered at two-way at the firm and country-time level.

Figure 10. Differential effect of recession on (log) investment for high-debt companies relative to low-debt companies—global vs. domestic recessions and export dependence.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=-k}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=-k}^{4}\sigma _{j}^{k}R_{i,t-j}*D_{n}*M_{i,s,t-1}+\sum \delta _{j}^{k}D_{n}*Z_{i,s,t-1}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $M_{i,s,t-1}$ is a dummy for high export dependence, $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid black line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. The solid red line indicates the baseline response reported in Figure 3. Standard errors are clustered at two-way at the firm and country-time level.

Figure 11. Differential effect of recession on (log) investment for high-debt companies relative to low-debt companies—firms characteristics.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on equation (3) for different horizons ‘k’. The solid line shows the point estimate for $\nu _{o}^{k}$, the coefficient on the triple interaction term, for different horizons k, while the dotted lines are the 68% and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Table A1. Sample of 75 countries by region

Table A2. Number of firms and observations by country

Table A3. Number of firms and observations by sector

Table A4. Summary statistics by leverage

Table A5. Summary statistics of firm-level database

Table A6. Summary statistics of macroeconomic variables

Table A7. Correlation of investment between Capital IQ and World Economic Outlook data

Figure A1. Bankruptcy efficiency in advanced economies (AE) and emerging markets (EM).Source: World Bank.

Figure B1 Robustness: excluding GFC and COVID-19.Note: Left panel excludes the year 2009 from the estimation while the right panel excludes the year 2020. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B2. Robustness: excluding countries with most observations.Note: Each impulse response checks robustness to excluding one country from the baseline regression—the 10 countries with the most observations in the sample are excluded one at a time. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B3. Robustness: excluding one region at a time.Note: Each impulse response checks robustness to excluding one region at a time. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B4. Robustness: excluding one sector at a time.Note: Each impulse response checks robustness to excluding one 2-digit sector at a time. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B5 Robustness: different winsorizing schemes.Note: Left panel winsorizes 0.05 percent tails of the dependent variable while the right panel winsorizes 5% tails of the dependent variable (compared to the baseline which winsorizes 1 percent tails). Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B6. Robustness: different dependent variables.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B7. Robustness: alternate recessions variables.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 to denote a recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level. Periods of banking crises are identified as the quarters during which a banking crisis-led recession occurred. Periods of peaks and troughs in economic activity are identified using the Harding-Pagan algorithm applied to real GDP. Periods of negative growth correspond to those where quarterly real GDP growth was negative.

Figure B8. Robustness: alternate debt dummies.Note: Impulse response functions based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 to denote a recession, $D_{n}$ is a dummy variable if company n is a high-debt company. A high-debt company is identified if its leverage is above the median leverage within country (left-hand graph), or above the median leverage within countries that are advanced economies or EMDEs respectively (right-hand graph). $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B9. Robustness: alternate specifications for high-debt dummies II.Note: Impulse response functions based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 to denote a recession, $D_{n}$ is a dummy variable if company n is a high-debt company. A high-debt company is identified if its leverage is above the highest tercile (quartile) leverage within industry. $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B10. Alternative high-debt specification: companies in the 2nd quartiles vs companies in the 4th quartile.Note: Impulse response functions based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 to denote a recession, $D_{n}$ is a dummy variable if company n is a high-debt company. A high (low) debt company is identified if its leverage is above (below) the highest (second) quartile leverage within industry. $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B11. Robustness: additional firm-level controls.Note: Robustness to adding firm level controls to the baseline. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\sum _{j=1}^{4}\theta _{j}^{k}x_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. $x_{n,i,t}$ are additional firm level controls. Left panel includes lags of firm size (ln assets), return on assets, and a measure of liquidity (current assets minus liabilities as a share of assets) as additional controls. The right panel also adds log revenues and the firms age as controls. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B12. Unconditional response of log (investment) to macro variables.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{is}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\beta _{j}^{k}M_{i,t-j}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $M_{i,t-j}$ is different macro variable in different panels, $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{is}^{k}$ are country-sector fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\beta _{0}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B13. Coefficients on interaction of macro variables with high-debt dummy as controls.Note: Robustness check where interaction of macro variables with high-debt dummy are included as controls. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum \nu _{j}^{k}M_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage in non-recession years within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. Interaction of various macro variable $M_{i,t-j}$ with the high-debt dummy are included as controls. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\nu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B14. Robustness: coefficient on interaction of recession dummy with firm characteristic dummies.Note: Robustness check where interaction of dummy variables for various firm characteristics and the recession variable are included as controls. Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum \nu _{j}^{k}R_{i,t-j}*F_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is a dummy which takes value 1 at the start of a technical recession, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. $F_{n}$ is a dummy variable defined analogously to the high-debt dummy but for different firm characteristics i.e. firm classified as 1 if it is above the median as per the characteristic when averaged across the entire sample. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\nu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B15. Robustness: placebo test.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on the regression $y_{n,i,t+k}=\alpha _{ist}^{k}+\gamma _{nq}^{k}+\sum _{j=0}^{4}\mu _{j}^{k}R_{i,t-j}*D_{n}+\sum _{j=1}^{4}\theta _{j}^{k}y_{n,i,t-j}+\varepsilon _{n,i,t}^{k}$ for different horizons ‘k’, where $y_{n,i,t+k}$ is the log change in capital expenditure of firm n in country i at time t over the next k quarters, $R_{i,t-j}$ is our random recession dummy, $D_{n}$ is a dummy variable if company n is a high-debt company (above median leverage within industry), $\gamma _{nq}^{k}$ are firm-quarters fixed effects, and $\alpha _{ist}^{k}$ are country-sector-time fixed effects. The regression is estimated separately for different horizons k (for up to 12 quarters). The solid line shows the point estimate for $\mu _{o}^{k}$ for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.

Figure B16. Nonlinear effects with triple interaction using smooth transition functions.Note: Impulse response function based on local projection methods following Jordà (2005) using firm-level quarterly data from 75 countries for the period 2001Q1 to 2020Q4. Estimates based on equation 3 for different horizons ‘k’, where the firm level characteristic $X_{n,i,t}$ is transformed using a smooth transition function $G\left(X_{n,i,t}\right)=\frac{1}{1+\exp \left(-X_{n,i,t}\right)}$ before interacting with the debt and recession dummies. The solid line shows the point estimate for $\nu _{o}^{k}$, the coefficient on the triple interaction term, for different horizons k, while the dotted lines are the 68 and 90% confidence intervals. Standard errors are clustered at two-way at the firm and country-time level.