A. Firm-level data

Our firm-level data comes from S&P Capital IQ. The database provides extensive balance sheet and income statement information at the firm-level and at the quarterly frequency. It covers a large, unbalanced sample of 150 countries from 1950Q1 to 2021Q2; in order to reduce significant gaps in the time series, we restrict the sample to 2001Q1 onwards. This leaves us with a sample of 75 countries. Details regarding the sample of countries used in the analysis, by geographic region, are available in Table A1.1 of Annex 1. The data is restricted to non-financial corporations and cleaned to remove firms which had negative values for assets or debt in any year, and observations with the incorrect sign for revenue, capital expenditure, cash, tangible assets, and interest expenditure were set to missing (see Kim et al. 2020, and Arbatli-Saxegaard et al. 2022, for details). We further restrict the sample to exclude real estate and insurance companies. Tables A1.2 and A1.3 display the number of firms across countries and 20 economic sectors.

For our revenue measure, we use total revenue (IQ_TOTAL_REV), Table A1.4 displays the summary statistics for this variable. For our intangibles measure, we use intangible assets (IQ_GW_INTAN) as a share of total assets (IQ_TOTAL_ASSETS). All firm-level variables have been winsorized at the 1^st and 99^th percentiles to eliminate outliers.

B. Recessions and other macroeconomic data

Our baseline measure for recessions is the start of a technical recession, defined by two consecutive quarters of negative GDP growth. Quarterly real gross domestic product growth from Haver Analytics is the main source used to construct this variable, but we complement it with World Economic Outlook (WEO) data for countries that have limited data in Haver. For these countries, we replace the full country series with WEO data. We then define a dummy where the first observation for each country’s recession episode is set to 1. This leaves us with a total of 231 recessions for advanced economies and 336 for emerging market and developing economies.

We also use two alternative measures to complement our recession indicators: banking and currency crisis dummies from the Global Crises Data of Reinhart and Rogoff (2009). Recession data summary statistics are reported in Table A1.5.

C. Digitalization measures

There is not yet a generally accepted definition and measure of digitalization, though substantial work is ongoing to agree on a consistent international approach (OECD 2021a). This reflects the pervasive nature of the technologies and the activities involving them, from physical computers to data streaming to real-world transactions mediated through e-commerce platforms (e.g., taxi rides and room rentals). We therefore adopt a pluralist approach, drawing on a range of sources to capture different facets of digitalization. In this section we outline a series of alternative digitalization measures D^m which we use in the main specification and the robustness checks.

First, our baseline measure $D_r^C$ is an industry-wise digitalization quartile, constructed by Calvino et al. (2018) using data on ICT input shares, the number of robots per employee, the share of ICT specialists in total employment, and the share of turnover from online sales. This ‘off-the-shelf’ measure has several advantages: in addition to being time-invariant, and therefore exogenous to recessions, it is independently constructed and draws on a wide range of data sources to capture and synthesize several facets of digitalization—namely purchases of digital tools themselves, investment in the human capital required to embed them in production, and the exploitation of digital channels for transacting with customers. This composite and relative approach, based on digitalization quartiles rather than absolute values, also produces sector-wise estimates that remain accurate over time. For instance, Calvino et al. (2018) find that only 17 percent of sectors would change their relative digitalization quartile, if it were recalculated separately in 2001–03 and 2013–15. This suggests that the approach produces relevant estimates across the whole timespan of our study, reflecting fundamentals rather than volatile sub-categories of digitalization technologies.

Nonetheless, there remain limitations to this measure. Given that several different metrics are used to construct the composite, the Calvino et al. (2018) method is relatively demanding on data availability. It only uses data from 12 developed countries—so does not take into account various country specificities when constructing the industry-wise average. We therefore complement this measure by constructing our own country-industry-specific measures using information from harmonized input-output tables. First, we construct a measure $D_{r,i}^{ICIO}={\Sigma }_l{\alpha }_{rli}\cdot D_l$ based on direct ICT input shares, where α_rli is the input share of industry l in industry r country i, and D_l = 1 for “Computer, electronic and optical products” and “IT and other information activities”.⁴ To construct this measure, we use consolidated data from 66 countries via the OECD Inter-Country Input-Output Tables (OECD 2021b).⁵ Specifically, this measures the share of each industry’s direct inputs that come from narrowly defined ICT industries. Second, we note that digitalization can also be a property of broader supply chains, where what matters is the extent to which suppliers and their suppliers in turn also use digital inputs. We therefore calculate a second measure $D_{r,i}^{TiVA}={\Sigma }_l{\zeta }_{rli}\cdot D_l$ where ζ_rliu is the share of total input value added of industry r country i that comes from industry l. ⁶ This uses data from the OECD Trade in Value Added database (OECD 2021c), again for 66 countries. $D_{r,i}^{TiVA}$ therefore repeats $D_{r,i}^{ICIO}$ but drawing on digital inputs in all stages in production, not just in direct inputs. Intuitively, it measures the total dependence of industry r on narrowly defined ICT industries, both as direct inputs to production, and as inputs to other inputs to production, and so on.

Together, these input-output table measures provide a good gauge of the relative dependence of sectors on both physical IT hardware and digital services. The set of countries used to construct them, however, is still not identical to that for which we have firm revenue data in Capital IQ. We therefore construct a further measure $D_{r,i}^{Int}$ which is the median share of intangible assets among firms within a country-industry cell, calculated using the same Capital IQ balance sheets.⁷ While intangibles as listed on balance sheets are an imperfect measure of digitalization, since they contain other elements such as copyrights, licensing agreements and post-merger goodwill (see, for instance, Haskel and Westlake 2018), this provides an additional robustness check for a larger set of countries.

Finally, for the pandemic period we use recent data from LinkedIn profiles to create an alternative digitalization measure focused on human capital. We define:

where TechSkillsInTop50₀ is the number of skills categorized as ‘tech skills’ appearing in the top 50 skills listed by workers in each occupation o and industry-country ri, and w_o,r,i is the relative weight of each occupation in the total employment of each industry-country ri. Intuitively, this measures the relative intensity of digital skills among all those skills used by each country and industry pair.⁸ Digitalization measures summary statistics are reported in Table A1.6.

Table 1 shows the cross-correlations between these various measures of digitalization, and Table A1.7 shows the most and least digital sectors according to each of the rankings. In general, the measures are strongly positively correlated and rank sectors in an intuitive manner. Between them they capture the major facets of digitalization and reassure us that our conclusions are not driven by a narrow range of technologies (for instance, industrial robotics or mobile internet) but instead accurately reflect the widespread impacts of the broad concept of digitalization.

A. Historical Recessions

Figure 3 presents the evolution of log revenue following a recession episode. The solid line displays the average estimated response, while dotted and dashed regions denote the 90 and 68 percent confidence bands respectively. We find that recessions are associated with persistent effects on the level of revenue relative to pre-recession trends. In particular, the average recession in our sample is associated with a reduction in the level of revenue by more than 10 percent four quarters after the recession and by 5 percent after 8 quarters.

Figure 4 is analogous to Figure 3 but reports the differential response of revenue to recessions between a firm in a relatively digitalized industry and a firm in an average industry. The figure shows that firms in industries that are one standard deviation more digitalized than the mean experience 1 percent lower revenue losses after two years. The difference is larger (about 2 percent) and highly statistically significant four years after the recession, highlighting that companies with higher digitalization levels are more resilient to economic shocks over the medium term.

B. The COVID-19 Pandemic

Digitalization could be even more important during recessions caused by the COVID-19 pandemic. Governments, firms, and individuals had to adapt to lockdowns and social distancing measures that have had deep and lasting effects on work and consumption practices. Digitalization became an important channel to mitigate harms from pandemic-response measures, in particular by facilitating remote work and contactless sales.

To examine the impact of COVID-19 period recessions on firms’ revenue, we use our baseline recession dummy—the start of two periods of negative growth—and artificially restrict it to be equal to one only for those recessions begun on or after 2019Q4.¹⁰ We also restrict the overall sample to 2016Q1 onwards so that our fixed effects are only estimated from a relevant timespan. Figure 5 presents the evolution of log revenue following a COVID-19 recession episode. COVID-19 recessions are associated with persistent effects on the level of revenue, of up to 20 percent after five quarters—a magnitude almost four times that associated with a typical recession, reflecting the severity and the unprecedented nature of the pandemic.

Figure 6 reports the differential response of revenue to COVID-19 recessions for a firm in an industry with one standard deviation higher digitalization than the average. Revenue in such is almost 4 percent higher four quarters after a recession and increases to close to 5 percent after five quarters. This effect is statistically significant and precisely estimated, and more than four times larger than the corresponding effect of an average historical recession after five quarters. Thus, the data strongly support the idea that digitalization has been especially important during the COVID-19 pandemic recession, given the unprecedented measures introduced to reduce mobility and social contact. This higher revenue growth may also have allowed firms in more digitalized industries to expand relative to other firms. Figure 7 shows the differential effect of COVID-19 recessions on growth in hiring rates and total transitions by workers into and out of each industry, using industry-level outcome variables from LinkedIn. Panel A shows that hiring rates in highly digitalized industries grew almost 3 percent faster than in average industries in the year after the COVID-19 recession, and almost 4 percent faster after two years. Similarly, Panels B and Panel C show that the inflow (outflow) of workers to (from) highly digitalized industries grew more than 3 percent faster (1 percent slower) than in other industries in the two years after the shock.

A. Alternative digitalization measures

While the effects of digitalization on our professional and personal lives—and on broader society and the economy—are clearly visible, its measurement is not straightforward. There is not yet a clear consensus on how to define and measure the concept (OECD, 2021). To account for this, in this paper we take a pluralist approach and, as a first robustness check, we examine the sensitivity of our results to several measures of digitalization as described in Section II.

Figure 8 shows the differential effect of recessions on revenue growth using digitalization measures defined at the country-industry level. Panel A shows results using $D_{r,i}^{ICIO}$, the share of direct digital inputs, constructed from OECD Inter-Country Input-Output tables (OECD 2021b). Panel B shows results using $D_{r,i}^{TiVA}$, the digital share of total input value added, constructed from the OECD Trade in Value Added database (OECD 2021c). Panel C shows results using $D_{r,i}^{Int}$, the average share of intangible assets, constructed using Capital IQ balance sheet data. Results are qualitatively robust, with firms in more digital country-industries facing 1–2 percent less revenue losses four years after recessions, though the differential effect is less precisely estimated.

Since the OECD input-output tables are calculated every year for a consistent and balanced panel of country-industries, we can also construct time-varying versions of these measures. Figure 9 shows the results, first using the contemporaneous values of $D_{r,i,t}^{ICIO}and{D}_{r,i,t}^{TiVA}$, then using lagged versions $D_{r,i,t-1}^{ICIO}and{D}_{r,i,t-1}^{TiVA}$ to mitigate potential endogeneity.¹¹ Our main findings are qualitatively robust in all cases.

Lastly, we check that our pandemic-specific results also hold when using alternative digitalization measures. For this shorter period, we can also use the additional digitalization measure $D_{r,i}^{TechSkills}$ calculated from the skills listed on LinkedIn profiles, as described in Section II. Figure 10 shows that our main findings are robust to these checks—except for Panel C, using the intangibles share, which we speculate may reflect the particular pressures facing firms with a high share of intangibles (which includes post-merger goodwill and potential goodwill impairment) in the boom/bust turmoil of pandemic-era merger activity.

B. Alternative recession dummies

We next look at the sensitivity of our results to the definition of recessions. While our baseline approach follows the standard technical definition of recessions, we also consider alternative versions, specifically the start of a currency crisis, or the start of a banking and currency crisis. The results produced using these variables are very similar to those in the baseline specification, with marginally larger effects over four years (Figure 11). This similar response to these alternative recessions emphasizes the generality of the role played by digitalization in fostering resilience.

C. Alternative samples

We next examine the sensitivity of our results to alternative samples. A noted above, the restrictions imposed by countries in response to the COVID-19 pandemic triggered an economic crisis with peculiar characteristics. Digitalization and telework both played an important role in mitigating the effects of this crises among companies, since it allowed them to continue their activity by reducing the disruption to their work. Therefore, we first check whether our results from historical recessions are driven by this specific crisis episode. Figure A2.1 Panel A in Annex 2 shows our baseline results excluding the year 2020 from our regression. Our results remain similar to our baseline, confirming that digitalization already mitigated the effects of recessions even before the COVID-19 pandemic.

Second, we look at whether the results are driven by other major crisis episodes, such as the 2008 Global Financial Crisis or other systemic banking crises. The results reported in Figure A2.1 panels B and C confirm that the differential response of revenue for firms in more digitalized industries remains similar to, and not statistically different from, the baseline.

Third, we check whether the results are driven by specific countries or groups of countries. To this end, we re-estimate equation (2) but excluding one country at a time or one region at a time. The results, reported in Figure A2.2 and A2.3 of Annex 2, suggest that our baseline results are not dependent on any specific country or group of countries. Similarly, we repeat this test for industries, excluding one 2-digit sector at a time, with the same result (Figure A2.4 of Annex 2).

Fourth we also check whether the results change depending on countries’ income levels. Figure A2.5 of Annex 2shows the differential effect of recessions between highly digitalized firms and the average firm restricting the sample to Advanced (left panel) and Emerging Economies (right panel). The results show almost no difference between the overall effect and the effect depending on countries’ income levels, revealing that digitalization serves as a buffer for these economic shocks independently of countries’ income characteristics.

Finally, we investigate whether our results are sensitive to our choice of winsorization threshold. We therefore winsorize 0.05 and 5 percent of the tails of the distribution of our dependent variable, allowing in turn a greater or lesser role for extreme observations than in our baseline. The results obtained are again similar to the baseline (Figure A2.6).

D. Alternative methodology

In a final robustness check, we repeat our pandemic results using an alternative difference-in-differences specification following Duval et al. (2020). This focuses on changes in firm and industry behavior around a single specified date, which in our case is the onset of the global pandemic in 2020Q1, rather than allowing for country-specific variation in the timing of pandemic recessions R_i,t, as in our local projection specifications.

For the Capital IQ firms, we calculate the growth in average quarterly revenue between the two years pre-pandemic and the first year post-pandemic, and how it varies with the degree of digitalization of the industry:

where Δy_n,i is firm n’s log average quarterly revenue in the post-pandemic period (2020Q2–2021Q1) minus log average quarterly revenue in the pre-pandemic period (2017–2019), $D_r^C$ is the standardized Calvino et al. (2018) measure of digitalization in industry r, and α_i and α_s are country and sector fixed effects respectively. For the alternative industry-wise outcome variables from LinkedIn, we run a similar specification but drop the sector fixed effects due to the less granular industry classification available in the LinkedIn data:

where y_r,i is now the log of average hiring rates or transitions into or out of the industry. The results for this exercise are shown in Figure A2.7 of Annex 2. We again find similar effects to the baseline, with an approximately 2 percentage point rise in revenue growth post-pandemic, along with 3 percentage point higher growth in transitions into more digitalized industries, and roughly 1 percentage point lower growth in transitions out of such industries.