The ILO nowcasting model: Using high-frequency data to track the impact of the COVID-19 pandemic on labour markets1

2.1Target variable

The first step in providing estimates tracking the labour market disruption caused by COVID-19 is finding a suitable indicator for this purpose. We focused the search for indicators on the most common source of labour statistics, labour force surveys. Labour force surveys are one of the primary national household surveys conducted by countries. They are designed to produce official national statistics on the labour force, employment and unemployment for monitoring and planning purposes. Labour force surveys are the main source behind headline indicators of the labour market for short-term monitoring as well as more structural information on the number and characteristics of the employed, their jobs and working conditions, the job search activities of those without work, etc. In the current study, the term labour force survey is understood to include household surveys that broadly follow standards set by the International Conference of Labour Statisticians (ICLS).

Within the substantial number of indicators that can be derived from labour force surveys, the commonly used indicators for assessing labour market trends related to economic activity such as unemployment and employment, present severe limitations in capturing the effects of the pandemic on the labour market. The limitations of unemployment as a labour underutilization indicator have been long established. Critically it is prone to undercount the actual degree of labour underutilisation. More comprehensive measures of labour underutilization, which capture those in time-related underemployment or those that are in the potential labour force22 were proposed in the 19th ICLS [3]. The pandemic has made these even more clear.33 In contrast, comparability problems of employment as a work activity indicator are to a high extent specifically linked to the pandemic. Following the shock that COVID-19 containment measures produced, institutional factors largely determined to which degree the economic shock impacted employment. For instance, in 2020 the European Union experienced a loss of employment of 1.8 per cent and the United States of 6.7 per cent (Source Eurostat and ILOSTAT, population aged 15–64 in employment). The reason for this disparity is not the lower impact of the pandemic on economic activity (GDP declined more during 2020 in the European Union than in the United States), rather the cause is the implementation of job-retention schemes in European countries, which the United States did not apply at scale.

As the main purpose of the exercise is producing global and regional aggregates concerning work activity disruption, this and other differences between countries would greatly reduce the comparability of the data and hence complicate the econometric modelling of missing observations. Moreover, the objective of capturing disruptions in the world of work requires capturing declines in work activity caused by the pandemic and public health containment measures, even if they were offset or mitigated by public policy.

Hours actually worked, in contrast to employment or unemployment, are expected to capture the effect on work activity in a more comprehensive manner. Declines in hours can certainly result from job losses, but they can also derive from being employed and not at work (furlough schemes) or from a reduction in average weekly hours worked. Similarly, a fall in hours will capture diminished work activity of those who lost their job – regardless of whether they become unemployed or transition out of the labour force, see ILO Monitor [4]. The difference between a person who is unemployed and one outside of the labour force is defined by the answers to behavioural questions of a labour force survey. For instance, searching for a job is generally required to consider someone unemployed. In the context of the COVID-19 pandemic, the difference between the two might have become more ambiguous, as depending on public health measures job search would not have been possible, even if in other circumstances it would have occurred.

2.2 Hours worked: Data collection and treatment

The main source of hours worked data is ILOSTAT (https://www.ilo.org/ilostat/), the ILO’s repository of international labour statistics.44 There are several measures of hours worked captured in labour force surveys, we focus on hours actually worked. International standards, [5], define hours actually worked as the time persons spend in activities that contribute to the production of goods and services. Other measures, such as hours usually worked, would not be as useful in capturing short-term fluctuations in work activity. We include in our target variable only hours actually worked in the main job. Hours worked in secondary jobs tend to be much lower than hours spent in the main job (by an order of magnitude) and are not as widely available.

In the estimates used for the 9th edition of the ILO Monitor on the world of work – published on the 23d of May – [6] – we have available data up to 2021Q3 for 58 countries and territories. And for 67 countries there was at least one observation available in the 2020Q1-2021Q4 period. This represents a sizeable sample, nonetheless, it leaves substantial international data gaps, as the target number of countries and territories for estimation is 189. These countries and territories are the ones that the ILO modelled estimates collection targets. They account for a very large share of the world’s population, albeit they do exclude some countries with relatively low populations.

Data availability is heavily skewed towards high-income countries. Table 1 summarizes the information. Whereas high-income countries only account for 30 per cent of the target countries, they account for 60 per cent of the actual observations. As national income decreases, so does the actual data coverage. At the lowest end of the spectrum, we find low-income countries, which account for 1 per cent of observations even if they represent 14 per cent of the countries. More recent data is also scarcer. The quarter with the most observations is 2020Q1, with 66 observations. In contrast, only 3 countries had some partial (monthly) data available in 2022Q1.

To enhance comparability across countries and time, we divide hours worked by population. Given that countries present significantly different population growth rates, tracking activity without accounting for changes in the population would limit comparability. Countries with high fertility rates would tend to show higher growth rates of hours worked (or lower losses of hours worked) than countries with lower fertility rates. A similar argument can be made for the change in population in a given country over time. To properly capture work activity, changes in working hours need to account for this change to ensure that the level increase in population is not driving growth in hours worked (for the same reason, employment is often adjusted by population, using the employment-to-population ratio indicator). Population aged 15–64 is used in the normalization, as persons aged 15 to 64 tend to have much higher participation rates in the labour market than those 65 or older. Hence for each quarter, total hours worked are divided by population aged 15–64. The source of the population data is the United Nations World Population Prospects (UN WPP), 2019 edition. To avoid discontinuities population at the quarterly frequency based on the UN WPP annual data is estimated. These estimates simply assume a constant quarterly growth rate that is consistent with the annual one.

Given the objective of tracking the disruption caused by COVID-19 on work activity, we focus on relative changes with respect to a pre-crisis benchmark, which we define as 2019Q4. Hence the target of estimation for a country of interest is:

Where t indexes each quarter. It is convenient to also define the level of hours worked (divided by population aged 15–64, denoted henceforth as P) at any given quarter as:

In some instances, to simplify the text the “divided by population aged 15–64” is omitted, and population is used as a shorthand for population aged 15–64.

2.3High-frequency indicators

For nowcasting estimation, we require indicators that are available on a timelier basis than the target variable. In the present study, we use the shorthand high-frequency to refer to this circumstance.55 We can distinguish two different types of high-frequency indicators based on the temporal horizon of availability.

The first type of indicator is available during the pre-pandemic period, prior to 2020Q1. These variables, which include amongst others retail sales, household sentiment and business confidence, purchasing managers index (PMI), and labour market indicators have been commonly used in economic nowcasting, including during the pandemic, [7, 8]. The ILO compiled approximately 4’000 time series from various online repositories,66 roughly three-quarters of them available at the monthly frequency and the rest at the quarterly frequency. Table 2 presents the main types of indicators included in this repository. These indicators are highly unevenly distributed across countries. Around half of the target countries have only five or fewer high-frequency indicators, while 30 per cent of countries have more than 10 indicators.

The second type of indicator used only became available after the onset of the COVID-19 pandemic. The indicators in this category covered three broad areas: mobility data from mobile phone users, stringency of COVID-19 containment measures taken by governments, and COVID-19 epidemiological data. These three indicators are expected to contain information about the evolution of work activity. Larger declines in mobility, more strict containment measures and a worse epidemiological situation are expected to be associated with larger declines in work activity. All three indicators are available at a high frequency, either daily or weekly, and we average the data to the quarterly frequency.

The source of mobility data is the Google Community Mobility Reports. These reports contain aggregate changes in mobility compared to a pre-pandemic baseline. We take an average of the workplace mobility index and the retail and recreation index as input for the model.77 Other indices available in this dataset are related to behavioural changes during the pandemic and more indirectly related to economic activity and hence are expected to contain less information about the labour market and are not included in the model. As a proxy for stringency of COVID-19 containment measures taken by governments, we use the COVID-19 Government Response Stringency Index (hereafter “Oxford Stringency Index”). The source of the index is the Oxford Covid-19 Government Response Tracker (OxCGRT) database [9]. The index is a composite of restrictions across several dimensions, such as workplace closures and stay-at-home requirements. The indicator aims to capture quantitatively the degree of stringency of public health measures. As a proxy of a country’s COVID-19 epidemiological situation we use deceased patients per population. The source for the data is [10]. We expect this measure to be a more accurate reflection of the public health situation than COVID-19 cases, due to testing practices, the difference in vaccination rates and other potential comparability problems.

3.1 Empirical strategy

Nowcasting hours worked at the global and regional level presents a substantial variation with respect to standard nowcasting applications. As described in [8], the nowcasting problem typically consists of predicting a variable of interest at time t using an information set at time v, Ωv. The information set is more timely than the target variable, for which we will only have a time series up to a time prior to period t, meaning that it has information about a later time than the variable of interest. Additionally, the information set is usually available or at least updated at a higher frequency than the target variable.

In our set-up, the information set is similar to the description above. In contrast, we do not have any observation for the target variable. Knowing the evolution of global88 hours worked requires having complete time series available for every country up to time t. As discussed above, international labour statistics are scarce, hence hours worked are missing in many countries in each time period. Even when focusing on yearly data, instead of scarcer quarterly frequency, sizeable data gaps are present. For 2019, hours worked data are available for half of the target countries. Moreover, for 22 per cent of countries, no data on hours worked are available from 2005 to 2019. Hence, a projection based on a direct statistical relationship between the target and the information set, Ωv, is not possible, and we must take another approach.

We want to produce global aggregates of the evolution of relative hours worked considering population changes. For this purpose, we need to aggregate across the target countries indexed by i for all quarters of interest t:

Notice that the population in the denominator does not correspond to the benchmark period 2019Q4, and rather it refers to the target period t. This is to avoid biasing (downwards) the indicator in the context of population growth, as even constant per capita hours worked result in an increasing trend in the overall level of hours worked. The target variable can be interpreted as the relative ratio of hours worked during quarter t to the counterfactual level of hours worked if the per capita hours remained at the 2019Q4 level but the population stood at the level of quarter t.

The change in hours can also be reframed as a subtraction instead of a division:

Which in this case indicates the difference in the level of hours worked. This formulation is used to compute full-time job equivalents, which can be obtained by dividing by a pre-set workweek (such as a 48-hour workweek). This full-time job equivalent is widely used in the ILO Monitor publication series.

Given that we have complete population data, our estimation target requires either having observations or estimates of {hti,h2019⁢Q⁢4i} for each country (which we can label country-based approach) or having estimates of the overall summations (aggregate-based approach). We take the country-based approach for several reasons. First, the main aim of the exercise is tracking the evolution of hours worked globally, regionally and by various country groupings of interest. Producing country-level estimates enables ex-post aggregation for country groupings without the need for re-estimation. This allows for cost-efficient production and ensures that each regional aggregate will be consistent with the rest of the aggregates produced. Second, the country-based approach allows for judgemental analysis of the estimation results of each target country. This procedure is a standard quality control practice for estimates published in ILOSTAT.

The problem of estimating {hti,h2019⁢Q⁢4i} can be divided into two types of estimation problems, according to the data availability situation of each country.99 The first would constitute a “classic” nowcasting problem. In these countries, a quarterly series of the total hours worked divided by population aged 15–64 is available at time t with a delay of k. Typically, the delay will be of one or more than one quarter, which underpins the necessity for nowcasting. Moreover, countries in this category also have high-frequency indicators with a record prior to the pandemic (described in the previous section). Hence, the estimation problem is producing projections for the target variable during the period when data are missing, i.e. obtain {h^τ-k+1i,…,h^τi} or simply h^τi if k=1, where τ indicates the latest quarter targeted for the nowcasting exercise. The estimation procedure of these countries follows a time-series/panel approach.

The second type of estimation problem is applied to countries for which the quarterly time series of the target variable is not available. Moreover, the information sets of these countries do not have sufficient (or any) high-frequency indicators with a pre-pandemic record. Hence, the estimation target is the entire sequence since the pre-pandemic reference period, i.e. {h^2019⁢Q⁢4i,…,h^τi} for each country in this category. Moreover, only high-frequency indicators covering the pandemic period (and not before) can be used. The estimation procedure of these countries follows a cross-sectional approach.

3.2 Time-series and panel approach, “direct nowcast”

The direct nowcast approach exploits the historical relationship between the target variable, hours worked per population ht, and observed high-frequency data, to predict movements in hours worked in the most recent periods where labour force survey data is not yet available. The approach uses time series techniques. As there needs to be a temporal overlap between the target variable and the high-frequency indicators, in this section only indicators with an observation record prior to the pandemic are used.

The relationships can be estimated at the country level or within a panel structure. Models estimated at the country level allow the use of country-specific high-frequency data, and also the estimation of country-specific coefficients. Panel techniques require the use of common high-frequency data, i.e. the time series have to represent the same variable across countries. However, they also have a number of advantages. First, an estimate of the change in working hours can be made for countries with available high-frequency data, but without a historical time series of hours worked. Second, differences in the estimated change in hours worked are easily traced to differences in high-frequency data. Third, pooling many countries in a panel reduces the risk of overfitting the model. Our approach relies on both country-specific and panel regressions.

3.3Cross-sectional approach, “indirect nowcast”

This section describes how estimates for those countries with more limited data availability are produced. This approach is based on the cross-sectional variation and is labelled “indirect nowcast”. From the direct nowcast, we have obtained a series of estimates of hours worked per population up to the target quarter. To simplify the notation, let estimates represent the actual observations whenever these are available. Hence, we can obtain the relative change with respect to 2019 produced via the direct nowcast:

Where the super script ∼ indicates that the relative change is derived from either actual observations or the direct nowcast. The main objective of the indirect nowcast is producing estimates of the relative change of hours worked H^ti based on the available estimates from the direct nowcast. The estimation of h^2019⁢Q⁢4i, the pre-crisis benchmark, will be discussed at the end of the section.

The imputation of these countries aims to palliate bias in the sample of countries included in the direct nowcast, see [19, 20]. Indeed, if those were representative of the 189 target countries, we could simply produce global results from that subset of countries. Nevertheless, there are important differences between the countries with sufficient information to create a direct nowcast and those without, for instance in terms of national income. Even more relevant are the country differences in the evolution of the COVID-19 pandemic and the containment measures taken by governments. Quarterly hours worked respond strongly and rapidly to public health measures, hence pandemic evolution information can be exploited for estimation in missing data countries.

Given the wide availability of high frequency data on mobility (Google Mobility Reports data) and the Oxford Stringency Index, described above, we use those indicators as a proxy for the social and economic disruption caused by COVID-19. We expect (and do observe for countries with data) that larger declines in mobility and increases in stringency lead to larger declines in hours worked. Because each variable is expected to have substantial noise, in the baseline specification we combine both variables into a single index using Principal Component Analysis to capture the common variation. We refer to this index as Iti.1212 We then estimate the statistical relationship between hours worked and this proxy of COVID-19 disruption.

The relationship is heavily dependent on each quarter. For instance, during 2020Q2, the period when the strictest lockdowns occurred, the reaction of hours to the proxy of COVID-19 economic disruption was very strong. In contrast, by the end of 2020 a similar change in disruption was associated with much lower hour responses. This is to be expected, there are multiple structural factors that plausibly change the response of the economy to the disruption caused by the pandemic. For instance, governments and economic agents adapt their behaviour, cycles of contraction and recovery can have different elasticities of economic activity to changes in restrictions, public health measures taken evolve and so does the virus. Given this, we estimate the econometric model independently at each quarter.

Our preferred specification is the following simple linear regression model:

Where α indicates the constant term, β the slope coefficient and εi⁢t the error term. We estimate the parameters of interest using OLS. We do not have an observation of Iti for all countries; hence, we need complementary regression models to generate a prediction for the dependent variable for all 189 target countries. We implement a linear regression model with the following explanatory variables in a recursive manner (only if the previous is not available we apply the next). First, a model with Google Mobility Reports Data. Second, a model based on COVID-19 incidence and regional fixed effects. Third, a model based only on regional fixed effects is used. The last two models are only used for 11 and 2 territories respectively in the model update corresponding to the 9th edition of the ILO Monitor on the World of Work. During 2021, additionally, a dummy variable for developed countries was incorporated into the models to account for differential impacts of mobility and stringency variables on hours worked during the recovery phase. Finally, the model for 2022Q1 was further augmented to include GDP growth forecasts and to take into account the time-series properties of the recovery process. This time period presented a turning point in terms of global activity indicators, and only 3 actual observations of hours worked were available. At the same time, as many countries lifted completely COVID-19 restrictions we expect that both mobility and stringency indicators lose economic relevance. For this reason, we increased the information set used.

Based on these approaches we can produce an estimate for all those target countries for which we do not have either reported observations or a direct nowcast estimate. Hence, we can produce an estimated value of the relative change of hours worked divided by population aged 15–64, H^ti for each target country i and each quarter of the target horizon t. To simplify the notation, let H^ti indicate three different sources of data: reported data, direct nowcast estimates, and indirect nowcast estimates (with the variable reflecting the data of these sources in this order of preference).

3.4Aggregation

Albeit we have an estimate for the relative change with respect to 2019Q4 of each quarter of interest based on the indirect nowcast, we cannot yet derive from it the change in hours worked at the global level, Htg. As Eq. (3) makes explicit, our interest in global and regional aggregates requires estimates of the starting level of hours worked, not just the relative change. This is intuitive, as we need a weight to represent the relative importance of the change in each country. Hence an estimate of the initial level is necessary for tacking global hours worked.

The estimate of the benchmark of hours worked per population aged 15–64, h^2019⁢Q⁢4i, is obtained in two steps. First, as annual data is more prevalent than quarterly data, we estimate hours worked in 2019 per population aged 15–64. This estimate follows a linear prediction based on the employment-to-population ratio, time-related underemployment, and the share of population aged 15–64. If a country has any observation available a country fixed effect is added, and if it does not a region and income-group fixed are added. The entire 2005–2019 period is estimated using this approach. In a second step, the observed difference (typically very small) of the seasonally adjusted hours worked in 2019Q4 and the 2019 value is regressed against a regional and a national income control to estimate an adjustment factor.

With an estimate for the benchmark of hours worked h^2019⁢Q⁢4i it is straightforward the computation of the estimate of the global change in hours worked based on H^ti and Eq. (3):

It is worth highlighting that systematic errors in the country estimate of the benchmark of hours worked, h^2019⁢Q⁢4i, will tend to have a much lower impact on the global relative change than errors in the country-level relative change in hours, H^ti.1313 Another interesting metric to analyse is the estimate of the global hours worked divided by population, computed as follows:1414

3.5Pseudo-out of sample evaluation

The practice of evaluating a model prediction based on pseudo out-of-sample error simulations is commonplace in econometrics. Whereas in many applications the focus is on comparing competing models, [21], the pseudo-out of sample error contains useful information to assess the uncertainty surrounding the prediction [22]. The empirical strategy of this type of evaluation reserves a certain subset of the data by the model. The predictions of the model for this subset are then assessed against the actual observations, by the use of a loss function. The label “pseudo” indicates that the econometrician has access to the information, even if the model does not use it.

Because the main objective of the ILO nowcasting model is producing global estimates of the evolution of hours worked, Htg, we target in our out-of-sample simulation an aggregate metric. For this purpose, we identify the countries in our sample with a complete time series of hours worked for the period 2020Q1-2021Q3 (50 countries in total). Then we exclude 30 per cent of them from the model input, for the entire time period. Hence, this evaluation concerns the indirect nowcast exercise. We then estimate the model on the remaining 70 per cent of countries and produce estimates for the missing countries. Following a formula analogous to equation 9, we compute the estimated aggregate corresponding to these countries. This estimate can be compared with the aggregate resulting from actual observations. We repeat the procedure 700 times, by sampling randomly the subset of the 30 per cent excluded countries each time. We compute the root mean squared error (RMSE) and the average hours worked in the excluded set of countries across simulations.

Even if our focus of interest is global and regional aggregates, simulating the nowcasting error at the country level is highly informative – as country errors are the drivers of aggregate ones. We perform two exercises to analyse errors at the country level. One to evaluate the country-level direct nowcast error and another one to evaluate the indirect nowcast error. For the direct nowcast evaluation, we perform a one-step out-of-sample simulation exercise. Hence, for countries with a sufficient time series, we discard the last data point in the time series and predict its value following a rolling window approach. The direct nowcasting model is mostly used to do one step ahead predictions, hence we focus on this metric. We then compare this prediction with the actual data. To simulate the error at the country level for the indirect nowcast, we follow the same procedure as the one used for aggregate error simulation, described just above. Nonetheless, instead of aggregating the estimation from the 30 per cent of excluded countries and then computing the error – as in the aggregate error procedure – we now compute directly the root mean square error at the country level.

Table 3 presents the results of the aggregate exercise. The results show that the model performs reasonably well, particularly during 2020. In 2020Q2 the average simulation result implied a decline in hours worked of 23.1 percent (with respect to 2019Q4), whereas the error estimate was 2.9 per cent. By 2021Q3, the performance deteriorates as the average simulation projects a decline in hours of 5.1 per cent with an error estimate of 3 per cent.

Table 4 presents the results of the direct nowcast exercise at the country level. The model performance is reasonable – albeit errors suggest sizeable uncertainty. The average one-step ahead RMSE is 3.1 per cent, whereas the median RMSE is 1.8 per cent. The 10th and 90th percentiles show sizeable dispersion, from 0.4 to 5.5 per cent. These results are not directly comparable to the indirect nowcast error simulations, given that the indirect nowcast has access to partial information about the current period, in contrast to the direct nowcast in the one step ahead prediction.

Table 5 presents the results of the indirect nowcast exercise at the country level. The average RMSE across countries and the country level median RMSE are generally not far from the aggregate error. Albeit the average and median RMSE tend to be lower than the aggregate error, this is not true in every period. The 10th and 90th percentiles of RMSE by country for each quarter are also presented. These show a considerable spread (with differences around 1 order of magnitude), hence using country specific results is subject to a high degree of uncertainty.

Figure 1.

Estimates of global weekly hours worked (divided by 15–64 population). Results of ILO Nowcasting model for the period 2005-2022Q1 at the global level. The 2005–2019 period is only estimated at the annual level, whereas from 2020Q1 and onwards quarterly estimates are available.

4.1 A global view of the labour market during the pandemic

The latest estimates from the ILO nowcasting model, featured in the 9th edition of the ILO Monitor on the World of Work published on the 23d of May of 2022, confirm the unprecedented disruption caused by the pandemic on the world of work (see Fig. 1). In the pre-crisis benchmark period, 2019Q4, global weekly hours worked per person aged 15–64 stood at 27.3. In 2020Q2, as lockdowns and other strict containment measures were implemented by countries across the globe, weekly hours declined to 22.3, a 5-hour difference. In comparison, the decline in hours worked registered between 2007 and 2009, at the height of the impact of the global financial crisis, was 0.7 hours per person aged 15–64. This reflects that the COVID-19 pandemic produced both steeper declines in work-activity within countries and a wider impact across countries.

As countries gradually lifted the strictest containment measures, hours worked recovered strongly. As a result, weekly hours increased to 25.4 in 2020Q3. By the end of the year, weekly hours worked per person aged 15–64 stood at 26.2, 1.2 hours below the pre-crisis benchmark. Both the initial decline and rebound were uncharacteristically fast, as labour markets tend to react with substantial delay in economic activity. In this instance the reaction was simultaneous1515 due to the nature of the economic shock. Nonetheless, the recovery in hours has been partial, and from 2020Q4 up to 2022Q1 hours have essentially stagnated. This left a gap of 1.1 weekly hours worked per person aged 15–64 remaining in 2022Q1. It is critical to highlight that given that hours are divided by population, it might well be the case that a recovery to the 2019Q4 level is never attained. Up to 2019, there was a slight but persistent declining trend in hours worked per person aged 15–64. If the trend persists into the future, the hours worked per capita would remain below the last quarter of 2019, even at very long-time horizons.

Figure 2.

Estimates of weekly hours worked (divided by 15–64 population) by income level, indexed data (2019Q4 = 100). Results of ILO Nowcasting model for the period 2005-2022Q1 at the global level. The 2005–2019 period is only estimated at the annual level, whereas from 2020Q1 and onwards quarterly estimates are available.

A key driver of the slow pace of recovery since 2020Q4 is the divergence between higher and lower-income countries (see Fig. 2). High-income countries experienced a strong recovery during 2021, becoming in 2022Q1 the income group with the levels of hours worked closest to the pre-crisis benchmark. Upper-middle income countries presented a sizeable recovery during 2021, which nonetheless was entirely offset by the lockdowns related to the Omicron variant of COVID-19 in China during early 2022. Low-income and lower-middle-income countries, in contrast, have presented a stagnant level of hours since 2020Q4. Uneven access to vaccination and lack of fiscal space to carry out stimulus programs at scale seems to have severely limited the recovery in low and lower-middle-income countries as shown in [23].

4.2 Timely analysis based on the ILO nowcasting model

Since its inception, the ILO Monitor on the World of Work has aimed producing timely and policy-relevant analysis. The work related to the ILO nowcasting provided two valuable sources of data (beyond the production of aggregate hours worked): an up-to-date, internationally comparable repository of hours worked, and timely country-level estimates. In this section, we highlight analysis in three key areas, based on these two data sources and published in the ILO Monitor publication series.

The first area of analysis was the effect of testing and tracing on labour market disruption, which was published in the early phase of the COVID-19 pandemic, [24]. At that stage, no actual observation for 2020Q2 was available, as the quarter had not yet elapsed, and labour force surveys present sizeable publication delays. The World Health Organization (WHO)1616 had recently re-iterated the importance of case finding, testing, contact tracing, isolation and care (which we shorten to testing and tracing). At that time, providing estimates of the potential economic effects of a policy recommended for public health reasons was useful to design policy measures for facilitating a safer return to work. Using the only data available for hours worked, estimates from the ILO nowcasting model, it was shown that following a strategy of intensive testing and tracing was associated with very substantial reductions in the loss of hours worked. The association suggested that moving from the lowest end of testing and tracing intensity to the highest end could reduce losses in working hours by 50 per cent.

The second area studied concerns the effects of fiscal stimulus on hours lost. In the 6th edition of the ILO Monitor, [25], the imbalance in fiscal stimulus between developing and advanced economies was analysed. This gap was concerning, as it could lead to sizeable divergence in economic outcomes. Having timely data on hours worked from the nowcast provided an opportunity to gauge to what extent the lack of fiscal space could materialize in lower hours worked in the developing world. The estimates showed that a 1 per cent increase in fiscal stimulus (in terms of annual GDP) reduced hour losses by 0.8 percentage points in 2020Q2 – the analysis accounted for differences in the economic disruption caused by public health restrictions. The magnitude of this association can be illustrated as follows. A country with a sufficiently larger fiscal stimulus package could expect to mitigate 60 per cent of hours losses compared to a country that did not implement any stimulus at all. This conclusion was later corroborated by using reported observations instead of modelled estimates in [23].

The third topic that we want to highlight is the effect of vaccination on the pace of recovery in hours worked. In the 8t⁢h edition of the ILO Monitor, [23], the imbalances in vaccine access and deployment across advanced and developing economies were analysed. Similar to the case of fiscal stimulus, the input data used in the ILO nowcasting model was leveraged to quantify the short-run effect of vaccination on hours worked recovery. Using hours worked based on labour force survey data the analysis showed that for the 28 countries in the sample, growth between 2021Q1 and 2021Q2 was strongly associated with vaccination. For each 14 fully vaccinated persons, hours worked increased by an amount equivalent to one full-time job. This strong association held after including relevant control factors including the previous speed of recovery, national income, growth of COVID-19 cases during the period, and the share of the elderly population.