Government policies and manufacturing production during the COVID-19 pandemic

Políticas gubernamentales y producción manufacturera durante la pandemia de COVID-19

Víctor Manuel Cuevas Ahumada vmca@azc.uam.mx

Universidad Autónoma Metropolitana, Mexico

Cuauhtémoc Calderón Villarreal

Colegio de la Frontera Norte, Mexico

This paper investigates the effects of government support and containment measures on the manufacturing production cycle of 39 nations, including Mexico, during the COVID-19 pandemic. To capture the impact of closure and containment measures as well as vaccination endeavors on the manufacturing output cycle, we make use of the containment and health index (CHI). Such an index comprises the actions undertaken by governments in 13 different areas, such as school and workplace closing, restrictions on traveling and internal movements, testing and contact tracing policies, vaccination funding, and emergency health care investments (Hale et al., 2021). Moreover, government support measures are captured by three key indicators:

1. 1. The income support index, which reflects the extent to which the government of a given country is reacting to the pandemic by: 1.1) Making direct cash payments to the unemployed and other people in need, 1.2) Granting temporary subsidies to private enterprises to prevent job losses, and 1.3) Providing financial aid to households or firms.
2. 2. The index of debt/contract relief for households. This index captures to what degree a government is implementing temporary measures such as: 2.1) Delaying tax payments, 2.2) Freezing mortgages and other loan repayments to private banks, 2.3) Stopping evictions, and 2.4) Ensuring the provision of services like water and electricity regardless of payment (Government Response Tracker (OxCGRT) Coding Interpretation Guide, 2022).
3. 3. To reflect the impact of monetary policy actions on the manufacturing production cycle (MPC, hereafter), we make use of the real policy-related interest rate.
4. 4. To account for the external sector, we employ the real effective (or multilateral) exchange rate based on consumer price indices.¹

To the best of our knowledge, the impact of this set of variables on the MPC has not been studied, so we attempt to fill this gap in the empirical literature. The analysis of the MPC and its determinants is relevant not only because the manufacturing sector is a major source of job creation, but also because it generates more learning by doing activities (LBD), greater positive externalities, and better prospects for export expansion and for attracting foreign direct investment (FDI) than other sectors of the economy (Sachs and Warner, 1995; Gylfason and Zoega, 2006). Our econometric work relies on a dynamic panel data model, which comprises 39 nations and monthly data from January 2020 through December 2021. The number of cross-section units (or nations) is restricted because some countries provide insufficient data (or not data at all) for at least one of the variables of the model. Given that the econometric methods employed here require a sufficiently large number of cross-section units (i.e., countries), for reasons of data availability our 39-nation group consists of 28 high-income countries and 11 upper-middle-income countries. Mexico, of course, belongs to the upper-middle-income country subgroup. Appendix 1 contains the complete list of nations.

To increase the robustness of the findings, we make use of two versions of the Generalized Method of Moments (GMMs) to estimate the panel data model: 1) The Arellano-Bond GMMs (1991), and 2) The Arellano-Bover GMMs (1995). When the number of cross-section units (N) exceeds the number of time periods (T), the Arellano-Bond GMMs is not only consistent but also robust to endogeneity problems, and these properties hold regardless of the behavior of the error term (Arellano and Bond, 1991; Ahn and Schmidt, 1995; Baltagi, 2008, pp. 147-155). So, when N>T and the sample is large, the Arellano-Bond GMMs estimator basically requires no information as to the behavior of the model residuals. In this study, N(=39)>T(=24) and there is a fairly large number of observations: 858 for each variable after adjusting for the loss of data caused by the presence of a lagged dependent variable and a set of instrumental variables. Nonetheless, we also resort to the Arellano-Bover GMMs estimator, which in relatively small samples tends to yield smaller standard deviations and even lower biases than the Arellano-Bond GMMs estimator (Hayakawa, 2009). Moreover, to further improve the efficiency of the estimations in a finite-sample setting, we include some remedial measures for residual heteroscedasticity and correlation.

As we shall see, the evidence is very consistent across econometric methods and remedial measures. Broadly speaking, the evidence shows that closure and containment measures are recessionary, just as expected. Moreover, income support measures, debt relief measures, and expansionary monetary policies are effective in boosting manufacturing production. Lastly, real exchange rate depreciation may not encourage manufacturing production and could even discourage it in the short term. Put differently, it could even exert a negative influence on the MPC. As we shall see, there are four reasons behind this finding, one of them being that real currency depreciation raises the domestic currency costs of a range of imported intermediate inputs, which are used intensively in manufacturing production.

The rest of this paper is organized as follows. Section 2 is a brief review of the empirical literature. Section 3 describes the empirical model, the estimation methods, and the variables involved. Section 4 presents the empirical evidence. Finally, as part of the conclusions, we summarize the findings and discuss their policy implications.

The next task is to specify the determinants of the MPC, describe the variables involved and the treatment of the data, and explain the pros and cons of the econometric methodologies employed. The selection of explanatory variables is consistent with economic theory and allows us to assess the role played by closure and containment measures, fiscal and monetary policy actions, and real multilateral exchange rate fluctuations. However, the larger the number of countries included in the panel, the smaller the set of variables for which complete statistical series are available. Given this tradeoff and the fact that the dependent variable is not manufacturing production itself but the manufacturing production cycle, we put forward the following model:

where the subscripts i and t denote the country and the year, and $M{PC}_{it}$ is the manufacturing production cycle. The MPC was calculated by applying the Hodrick-Prescott (1997) filter to the manufacturing production index, going several years back to properly derive the manufacturing output gap for the period January 2020-December 2021. The model is dynamic as it includes a lagged dependent variable, $MP{C}_{it-1}$, which means that the impact of the other regressors depends on the past behavior of the manufacturing production cycle. Under the standard specification suggested by Arellano and Bond (1991, p. 280) and Baltagi (2008, p. 147), all the regressors are current variables except for the lagged dependent variable, which is $MP{C}_{it-1}$ in this case. The rationale behind this is to capture the predetermined or inertial component of the dependent variable, so that one can assess to what degree the other explanatory variables can change its current trajectory. Regarding the rest of regressors, we point out that:

1. 1. $CH{I}_{it}$ is the containment and health index. Such an index measures the extent to which the different countries have been enforcing the following 13 policy areas to cope with the COVID-19 pandemic at different points in time: on-line teaching rather than in-person teaching, remote work, bans on public events, limitations on the number of people that can gather, public transportation stoppage, stay-at-home mandates, domestic travel protocols and limitations, international travel protocols and limitations, public information efforts to create awareness, massive testing campaigns, contact tracing endeavors to break the virus propagation channels, mandatory use of masks, and vaccination campaigns (Hale et al., 2021, p. 530). In this context, the CHI gauges the intensity and scope of the government interventions in these 13 areas.
2. 2. $I{S}_{it}$ stands for the income support index, which is designed to reflect the degree to which governments are counteracting the effects of the pandemic through cash transfers to unemployed people and informal workers (such as roadside vendors affected by the lockdown), or through food rations to vulnerable population groups. This index also covers temporary financial support to either households or private sector enterprises. In the case of firms, the financial aid is usually aimed at preventing firms from firing workers during the pandemic (OxCGRT Coding Interpretation Guide, 2022; Hale et al., 2021).⁴
3. 3. $D{R}_{it}$ is the index of debt/contract relief for households, which records the government’s efforts to alleviate the financial situation of homeowners by: 3.1) Delaying tax payments, 3.2) Temporarily lowering or freezing residential mortgage payments to private banks, 3.3) Extending soft loans backed by the government, 3.4) Banning evictions for a given period, and 3.5) Ensuring the provision of basic services such as water, electricity, and primary health care (OxCGRT Coding Interpretation Guide, 2022; Hale et al., 2021).
4. 4. $R_{it}$ is the real policy-related interest rate, which was calculated by subtracting the inflation rate (based on the consumer price index) from the nominal policy-related interest rate. In the case of only three countries,⁵ we made use of the real lending rate as there was little or no data regarding the policy interest rate. The rationale behind this is that policy interest rates are supposed to influence other interest rates, such as the lending rate.
5. 5. $Q_{it}$ is the real effective (or multilateral) exchange rate based on consumer price indices. This variable is incorporated to estimate how real exchange rate fluctuations affect the manufacturing production cycle in times of pandemic.
6. 6. Finally, ${\mu }_i$ denotes a cross-section disturbance term that captures the heterogeneity of nations, whereas $v_{it}$ is a combined disturbance term changing from one nation to another and over time.

In this context, Equation (1) is estimated by two econometric methods: 1) The Arellano-Bond GMMs (1991), and the Arellano-Bover GMMs (1995). Broadly speaking, the Arellano-Bond GMMs takes care of the following potential problems: First, the dependent variable $\left(M{PC}_{it}\right)$ is influenced by the cross-section error term $\left({\mu }_i\right)$, which means that the lagged dependent variable $\left(MP{C}_{it-1}\right)$ and ${\mu }_i$ could be correlated. Secondly, the lagged dependent variable and the combined error term $\left(v_{it}\right)$ could be correlated as well. Thirdly, any of the explanatory variables ($CH{I}_{it}$, $I{S}_{it}$,$D{R}_{it}$, $R_{it}$, and $Q_{it}$) could be correlated with the combined error term due to endogeneity problems. All these three problems give rise to biased estimations and thus must be solved. The first step to address these problems is to first-difference all the variables of Equation (1):

By first-differencing the data, the cross-section error term $\left({\mu }_i\right)$ is eliminated and, therefore, its potential correlation with the lagged dependent variable. The potential correlation between the regressors of Equation (2) and the new combined error term $\left(\Delta v_{it}\right)$ is solved by “internal” instrumental variables estimation. The instruments are internal because they are provided by the appropriate lags of the independent variables “in levels”; that is, by the appropriate lags of $M{C}_{it-1}$, $CH{I}_{it}$, $I{S}_{it}$,$D{R}_{it}$, $R_{it}$, and $Q_{it}$.⁶ The instruments generated in this fashion fulfill two requirements: 1) They are highly correlated with the regressors of Equation (2), which are in first differences, and 2) They are uncorrelated with the combined error term, so that the endogeneity problem is properly solved (Arellano and Bond, 1991).

The next step is to address serial correlation and heteroscedasticity by combining this set of instruments with a sequential Generalized Least Squares (GLS) procedure, which gives rise to the Arellano-Bond first-, second-, and n-step consistent estimators. In this context, it has been shown that when the number of cross-section units (N) in the panel is greater than the number of periods (T), the Arellano-Bond GMMs has three advantages: 1) It is free of endogeneity problems, 2) It is consistent, and 3) It is reliable regardless of the initial behavior of the error term (Arellano and Bond, 1991; Ahn and Schmidt, 1995; Baltagi, 2008, pp. 147-155). In fact, the first- and second-step consistent estimators are equivalent in large samples when, from the start, the model residuals are homoscedastic and uncorrelated (Baltagi, 2008, p. 150).

As we have shown, the Arellano-Bond (1991) GMMs transforms the data by taking first differences. In contrast, the Arellano-Bover (1995) GMMs makes use of what is known as forward orthogonal deviations (FODs). Both procedures are useful to eliminate the cross-section error term $\left({\mu }_i\right)$, which as we explained tends to be correlated with the lagged dependent variable $\left(M{PC}_{it-1}\right)$, thereby generating biased estimations. As with the first-differencing procedure, the FODs procedure applies to all the variables of the model, including the two disturbance terms (${\mu }_i$ and $v_{it}$). Let us assume that ${\widehat{X}}_{it}$ is the variable transformed by the FODs procedure and that $X_{it}$ is the original variable in levels. As explained by Abdul-Karim, Azman-Saini, and Abdul-Karim (2011), the relationship between ${\widehat{X}}_{it}$ and $X_{it}$ is given by the following mathematical expression:

where $T_{it}$ denotes the number of periods for the ith country. Put differently, first we get the difference between a given observation and the mean of all future observations, and then we multiply that difference by the scale factor ${\left(\frac{T_{it}}{T_{it}+1}\right)}^{1/2}$. Using Monte Carlo simulation, Hayakawa (2009) and Phillips (2019) study the performance of the Arellano-Bond GMMs, based on first differencing, and the Arellano-Bover GMMs, based on FODs. In this context, both authors conclude that the Arellano-Bover GMMs tends to yield more efficient and less biased estimations when working with small samples. Along these lines, although the panel data models estimated here involve a fairly large number of observations, we decided to employ both estimation procedures.

For each variable of the model, we gathered monthly data from January 2020 to December 2021 for the 39 countries of the sample. As stated before, this means that there are 858 observations for each variable, once we have adjusted for the loss of data due to the presence of a lagged dependent variable and a set of instrumental variables.Appendix 1 contains the list of the 39 the countries involved in the analysis, whereas Appendix 2 explains the measurement units and data sources for each variable.

As is well known, when working with time-series equations, there are basically two types of problems with the model’s residuals that compromise efficiency: correlation and heteroscedasticity across time. However, when working with panel equations there are four types of problems: time-series correlation and heteroscedasticity, on the one hand, and cross-section correlation and heteroscedasticity, on the other. We deal with these four problems in three ways:

1. 1. By making use of two robust estimation methods: the Arellano-Bond GMMs and the Arellano-Bover GMMs. When N>T and the sample is large, as in this case, the Arellano-Bond GMMs is consistent, robust to endogeneity problems, and basically requires no information concerning the behavior of the residuals (Arellano and Bond, 1991; Ahn and Schmidt, 1995; Baltagi, 2008, pp. 147-155). Considering the value of T and N, and the loss of data due to the use of a lagged dependent variable and a set of instrumental variables, we have 858 observations for each variable, which is quite reasonable. Nonetheless, we also make use of the Arellano-Bover GMMs, which in relatively small samples is proven to yield even lower standard deviations and biases (Hayakawa, 2009).
2. 2. We make use of a white period instrument weighting matrix under both econometric methods: the Arellano-Bond GMMs and the Arellano-Bover GMMs. The white period instrument weighting matrix is a good remedial measure for serial correlation.
3. 3. We also resort to coefficients’ white period standard errors and covariance, which is a coefficients variance-covariance method that is robust not only to serial correlation, but also to heteroscedasticity across time. This coefficients covariance method is useful to further enhance efficiency and is available for the Arellano-Bover as well as for the Arellano-Bond estimator.

In this context, we estimate Equation (1) by the two econometric methodologies: the Arellano-Bond GMMs and the Arellano-Bover GMMs. In addition, we make use of the remedial measures previously mentioned: 1) The white period instrument weighting matrix, and 2) The white period standard errors and covariance. Table 1 displays the estimations results:

Table 1
Dynamic panel data models

Asterisks *, **, and *** denote statistical significance at the 10%, 5% and 1% levels, respectively.

GMMs= Generalized Method of Moments.

∆ is the first difference operator, whereas the hat ˄ denotes the forward orthogonal deviations operator.

d.f.= Degrees of freedom.

Source: Authors’ estimations based on information from the data sources depicted in Appendix 2 and the use of EViews 12.

The second column of Table 1 shows the estimation results under the Arellano-Bond GMMs estimator, whereas the fourth column shows the estimation results under the Arellano-Bover GMMs estimator. Before interpreting the signs and statistical significance of the parameter estimates, we will display the outcome of the Arellano-Bond test for serial correlation, which is useful only to detect first- and second-order autocorrelation. According to Arellano and Bond (1991), if the error term in Equation (1) is identically independently distributed, then the error term in Equation (2) should show two things: 1) Negative first-order autocorrelation, and 2) No second order autocorrelation. Along these lines, Table 2 shows that the probability values for the null hypotheses of first- and second-order autocorrelation are 0.1025 and 0.6046, respectively. Therefore, there is no second-order autocorrelation, but we fail to establish -in the margin, given that the probability value is slightly higher than 0.10- the presence of negative first-order autocorrelation in the underlying residuals as Arellano and Bond require. Although these results are reasonably good, we deemed appropriate to implement the remedial measures previously mentioned.

Table 2
Arellano-Bond (AB) test for autocorrelation

Note: The AB autocorrelation test applies only to first- and second-order serial correlation. Source: Authors’ estimations based on information from the data sources depicted in Appendix 2 and the use of EViews 12.

Another important test available for the Arellano-Bond and the Arellano-Bover estimators is the test of exogeneity of the instruments. Put differently, in addition to being correlated with the regressors, the instruments must be uncorrelated with the error term (i.e., they must be exogenous). Table 3 shows the results of the test of exogeneity of the instruments under both econometric methods.

Table 3
J-test of exogeneity of the instruments

Note: GMMs= Generalized Method of Moments. Source: Authors’ estimations based on information from the data sources depicted in Appendix 2 and the use of EViews 12.

According to Table 3, the probability values for the null hypothesis that “all the instruments are uncorrelated with the error term” are 0.330510 for the Arellano-Bond GMMs and 0.278180 for the Arellano-Bover GMMs. This means that such a null hypothesis cannot be rejected and, therefore, the instruments are valid. In this context, the next step is to briefly summarize the findings displayed in Table 1:

1. 1. The coefficient of the lagged dependent variable is statistically significant at the 1% level in the two cases. This means that the past behavior of the manufacturing production cycle (MPC) is a key driver of the current behavior. Therefore, the MPC involves an inertial (or predetermined) component. The stronger this inertial component is, the more difficult is for fiscal and monetary policies to influence the MPC.
2. 2. The containment and health index (CHI) has a negative estimated coefficient, which is statistically significant at the 1% level under both econometric methods. As stated earlier, this is a comprehensive index encompassing 13 policy areas. So, broadly speaking, a negative coefficient in this case reflects the recessionary effects of the closure and containment measures previously described.
3. 3. The parameter estimate associated with income support measures (IS) is positive and statistically significant at the 1% level in both regressions. And the same applies to the estimated coefficient of debt contract relief (DR). Therefore, countercyclical fiscal policies are effective in encouraging manufacturing output.
4. 4. The estimated parameter of the real policy-related interest rate (R) is negative and statistically significant at the 1% level in both cases, which means that lower interest rates would boost manufacturing production.
5. 5. The coefficient of the real multilateral exchange rate (Q) is negative and statistically significant at the 1% level under the Arellano-Bover GMMs and at the 10% level under the Arellano-Bond GMMs. Thus, as we argue below, real exchange rate depreciation lowers manufacturing output growth, at least in the short term.

This paper relies on a dynamic panel data model involving 39 countries and monthly observations from January 2020 to December 2021. To enhance the robustness of the evidence, the model is estimated through two econometric methods. The first is the Arellano-Bond GMMs, which for N>T is consistent, free of endogeneity problems and basically requires no information regarding residual behavior (Arellano and Bond, 1991; Ahn and Schmidt, 1995; Baltagi, 2008, pp. 147-155). Although the number of observations for each variable is reasonably large (858), we also resort to the Arellano-Bover GMMs, which in finite samples performs even better in terms of standard deviations and biases (Hayakawa, 2009). Using the tools at hand for these econometric methods,⁷ we further improve efficiency by implementing remedial measures for serial correlation and heteroscedasticity across time.

The evidence across econometric methods is very consistent in indicating that:

1. 1. The MPC has a strong predetermined or inertial component, given the statistical significance of the coefficient linked to the lagged dependent variable. This means that a dynamic panel data model is a suitable specification and the effects of the exogenous regressors is contingent upon the past behavior of the MPC. Moreover, the stronger the inertial component of the MPC is, the more difficult is for fiscal and monetary policies to alter the course of the MPC.
2. 2. Closure and containment measures, which are captured by the CHI, have been clearly recessionary. The deep negative impact on global economic activity must be explained not only by the nature and severity of these measures, but also by their worldwide implementation. Therefore, to normalize global production chains, reactivate consumption and investment, and generate positive spillover effects through international trade, there must also be a coordination of national economic policies (Susskind and Vines, 2020; McKibbin and Vines, 2020) addressing the demand- as well as the supply-side factors behind the crisis (Maital and Barzani, 2020). As the reader may recall, those factors were analyzed in the literature review.
3. 3. Income support measures seem to be effective in raising manufacturing production, and the same conclusion applies to debt contract relief for households. Put differently, in times of crisis, enhancing consumption through higher income (i.e., direct financial support to the unemployed and other people in need) can be effective in raising manufacturing production and creating jobs. On the other hand, direct cash transfers to MSMEs and financial assistance to big corporations is key to preventing bankruptcies and thus preserving production capacities, so that workers can remain in the same job or industry. In this manner, the accumulated work experience can be protected insofar as workers won’t have to shift from one job to another that requires a different set of skills (Akbulaev, Mammadov, and Aliyev, 2020).
4. 4. An expansionary monetary policy in the form of interest rate cuts has clearly a positive impact on the MPC. This highlights the importance of cheaper credit to consumers and investors to stimulate manufacturing production. One caveat to this conclusion is that the implementation of expansionary fiscal and monetary policies must go through very gradual stages given the time it takes for global production chains to normalize even when proper supply-side policies are in place.

The real multilateral exchange rate is negatively related to the manufacturing production cycle, regardless of the econometric method employed. Put differently, our evidence is consistent in indicating that real currency depreciation discourages manufacturing output. This result holds at the 1% significance level under the Arellano-Bover estimator, and at the 10% significance level under the Arellano-Bond estimator. Nonetheless, as previously stated the Arellano-Bover estimator is more reliable when the sample size is not too large. Conventional economic theory suggests that exchange rate depreciation raises manufacturing output by boosting manufacturing sales abroad. Therefore, a plausible interpretation of our finding is that real currency depreciation influences manufacturing production through two channels working in opposite directions. The demand-side channel of currency depreciation stimulates manufacturing output by making manufacturing products abroad cheaper, whereas the supply-side channel of currency depreciation discourages manufacturing production by making imported intermediate inputs more expensive. In this manner, our result underlines the prevalence of the supply-side channel of currency depreciation over the demand-side channel, at least in the short term. There are several reasons behind this conclusion:

1. 1. The increasing role of international production sharing has made manufacturing firms across the world more dependent on imported parts, components, and materials. Therefore, currency depreciation increases the local currency costs of all those imported intermediate inputs, which in turn tends to deteriorate manufacturing export performance. Along these lines, the higher the import content of manufacturing exports is, the stronger the supply-side effect of currency depreciation becomes.
2. 2. By reducing manufacturing exports and production, exchange rate depreciation could result in scale contraction for many manufacturing firms, thereby reducing productivity levels. In this regard, Fung and Liu (2009) analyze the effect of exchange rate depreciation on manufacturing domestic and foreign sales, value-added and even productivity. However, those authors highlight the demand-side effect of currency depreciation on Taiwanese manufacturers.
3. 3. Closure and containment measures disrupted trade flows, investments, and international production sharing, giving rise to bottlenecks in the production of numerous critical parts and components for the manufacturing sector. In this context, during the great lockdown, currency depreciation interacted with production shortfalls to make all those inputs even more costly.
4. 4. Exchange rate depreciation often entails higher exchange rate volatility, which also worsens the export growth prospects of manufacturing firms (Zia and Mahmood, 2013). This negative impact on export performance may be stronger for those manufacturing firms holding net liabilities in foreign currency, in addition to producing goods with a high import content. Higher exchange rate volatility could also reduce investment by raising uncertainty.
5. 5. Regardless of the direction of minor exchange rate movements, manufacturing exports could hardly increase amid falling global demand and recession.

Disentangling the transmission channel through which real exchange rate depreciation could lower manufacturing output is beyond the scope of this investigation. However, the increasing role of global supply chains and processing trade is consistent with the view that the supply-side effect of exchange rate fluctuations is gaining force vis-à-vis the demand-side effect. In this context, weakening the domestic currency may not be the best alternative to raise manufacturing output, at least over the short-term horizon.