1.1. Literature review

Several studies have compared the performance of countries from different income levels during the COVID-19 pandemic in various aspects such as food security [27, 28] and income shocks caused by COVID-19 pandemic [29]. Studies have also compared the impact of policies measured during the pandemic across countries in different income groups [30–34]; the impact of socio-economic variables on the spread of COVID-19 [34–38]. Erokhin and Gao [27] concluded that during the COVID-19 pandemic in lower-income countries food inflation is the major crisis, while in upper-income countries food trade restrictions and currency depreciation is more frequent. In [39] it is shown that containment and closure policies resulted in less mobility, morbidity, and mortality in high and upper-middle income countries compared to low and lower-middle income countries. In [32] economic stimulus packages have been compared over different countries using beta regression and Euclidean distance.

Authors in [36–38] found a positive correlation between income inequality and the number of COVID-19 cases and deaths across all the WB income groups. In [33] it is shown that a contextualized pandemic response plan is required for Low-and-Middle-Income Countries (LMIC). In [40] demographic factors such as age, education, family size, and income for economic hardship have been studied across five different countries, Thailand, Malaysia, the UK, Italy, and Slovenia. Generally, the most vulnerable groups to the COVID-19 pandemic were people 18–24 years of age or over 65, with lower education, with children under 18, with larger family sizes, and with flexible or no income. In [29] it is shown that infant deaths have increased by 6.8% in 2020 due to COVID-19 economic contraction. Of note, none of the above studies looks at the macroeconomic responses to the COVID-19 pandemic.

Macroeconomic changes during the COVID-19 pandemic have been analyzed in some papers, but not on a cross-country-income level. Authors in [41] have used the Phillips curve in an Auto-Regressive Distributed Lag (ARDL) model to find the effect of shocks caused by lockdowns on the monetary policies, and the economy’s response to the affected policies. They concluded that to reduce the negative impacts of the COVID-19 pandemic on the economy, countries are taking less prudent monetary policies compared to before. T. W. Abate in [42] studied unemployment and inflation rate during COVID-19 in Ethiopia and concluded that because unemployment has increased as a result of lockdowns, production has decreased. This has caused a great demand for goods and therefore, inflation has also increased at the same time.

Authors in [43] argue that with a strict lockdown, the number of infections decreases, however, economic inactivity increases. In contrast, with a loose lockdown and restriction, economic inactivity decreases, but the number of infections and deaths increases. They then use the short-run Phillips curve to build a model that describes the inverse relationship between the number of COVID-19 infections and economic inactivity. Pham and Sala [44] studied the effect of inflation and unemployment rates on cross-country connectedness for the G7 countries. They found that the connectedness of inflation is high, but unemployment is low. In general, countries with higher competitiveness have a lower connectedness and do not spread their economic shocks to other countries. Furthermore, coordination in macroeconomic policies can help reduce volatility spillovers caused by shocks.

In [6] different countries such as Germany, Norway, Japan, China, Taiwan, the UK, and different cities in the USA are compared based on GDP loss and COVID-19 deaths. Countries are classified into four different groups: (1) low GDP loss and low COVID-19 deaths, (2) high GDP loss and high COVID-19 deaths, (3) high GDP loss and low COVID-19 deaths, and (4) low GDP loss and high COVID-19 deaths. They found that most of the countries and regions fall into group (1) or group (2). The reason is that due to government-mandated policies and self-protecting behaviour, countries started the pandemic with either a high or low mortality rate. Countries with a high mortality rate at the beginning of the pandemic applied stricter Non-Pharmaceutical Interventions (NPI) to reduce the number of deaths. Therefore, they suffered from both high GDP loss and high COVID-19 deaths. On the other hand, luckier countries with a low mortality rate at the beginning of the pandemic applied milder restriction policies and thus suffered less from economic loss. Therefore, they have had both low GDP loss and low COVID-19 deaths. Using epidemiological models, Phurichai [7] argued that there is a trade-off between health and mobility during an outbreak. To increase well-being, mobility must be reduced. They concluded that early implementation of NPI can effectively reduce the damages caused by a pandemic.

None of the studies mentioned above analyzed macroeconomics during the COVID-19 pandemic and across different countries incomes. Macroeconomic impacts of the global crisis have been studied before the COVID-19 pandemic based on country incomes [14–17]. However, to the best of our knowledge, this work is the first to compare the macroeconomic response of different countries to COVID-19 based on the country income groups defined by the WB.

Some papers have used social media to track economic factors such as the unemployment rate during COVID-19. In [45] three different datasets including Twitter dataset have been used to track the unemployment rate of the USA during COVID-19. The results show that the unemployment rate can be tracked using social media such as Twitter. In [26] the unemployment rate of South Africa was traced using Twitter data and Principal Component Regression (PCR) was applied to nowcast it. In this paper, we improve this method by combining four different machine learning algorithms and use the result to analyse macroeconomics during the COVID-19 pandemic compared to pre-COVID-19 years across country-income levels.

2.1. Unemployment and inflation rates for Canada

We obtained monthly data for the unemployment and inflation rates for Canada from [49, 50].

2.2. Unemployment and inflation rates for Nigeria and South Africa

Monthly data for the inflation rates for Nigeria and South Africa were obtained from [51] and [52], respectively. Since the unemployment rate of Nigeria and South Africa are not available monthly, we use social media to find the missing information. Social media is increasingly becoming popular. People share their thoughts, experiences, and concerns on social media. Therefore, it is a great source of information that can help retrieve data and derive different indicators such as the unemployment rate. We use a method similar to the one used in [26] to find the unemployment rates of Nigeria and South Africa. According to the International Labor Organization (ILO) unemployment rate can be estimated using Google Mobility Index (GMI) for countries with limited data [53]. GMI is an indicator that shows how many people have moved around six different locations, namely, retail and recreation, groceries and pharmacies, parks, transit stations, workplaces, and residential [54]. It has been temporarily available during the COVID-19 pandemic to help different sectors, especially health workers in making decisions. According to ILO and other references [6, 7], the unemployment rate of countries with limited data can be estimated using GMI [53]. However, in this work, we use GMI to only gather our dataset, not to estimate the unemployment rate. We use certain keywords related to the unemployment rate to search for geotagged tweets from Nigeria and South Africa. We used the same set of keywords in [26], i.e. unemployed, employed, and different forms of retrenching to find a dataset for South Africa. For Nigeria, we used the keywords employed, unemployed, and lost * job(s), where * indicates a wildcard, i.e. lost his job, lost her job, lost my job, and lost their jobs to gather a dataset.

The number of tweets with the mentioned keywords and the sum of sentiment scores of the tweets were respectively, positively and negatively correlated with GMI during the COVID-19 pandemic and quarterly unemployment rate before COVID-19 for Nigeria and South Africa. We used the volume of the tweets and their sentiment scores to predict the unemployment rate of Nigeria and South Africa with different machine learning algorithms.

Sentiment analysis is a text classification technique that classifies the sentiment of a text into negative, neutral, or positive sentiment. A sentiment score is a number between -1 and 1. Positive, neutral, and negative sentiments have a number close to 1, 0, and -1, respectively. Unlike in [26] where the authors used a pre-trained model of Bidirectional encoder Representation of Transformers (BERT) to perform sentiment analysis, we trained our model for sentiment analysis using BERT and used it to find unemployment rates of Nigeria and South Africa to compare the macroeconomic response of countries from different income levels during COVID-19 pandemic. There are two BERT models available for fine-tuning, BERT-large and BERT-base. We used BERT-base to fine-tune the model to our dataset. While the pretrained model used in [26] had 69% accuracy, our model had 76% accuracy in our dataset. Detailed information on our sentiment analysis model using BERT is available in S1 Appendix.

We applied different machine learning algorithms to the volume of the tweets and their sentiments to estimate the unemployment rate of Nigeria and South Africa, and depict their Phillips curves to analyze the macroeconomic changes. We used the average results of the four machine learning algorithms, i.e. RFR, SVR, XGBoost, and ARIMAX to predict the unemployment rate of Nigeria and South Africa. From the beginning of 2017 till now, the census unemployment rate is available and known for 9 and 19 months for Nigeria and South Africa, respectively. Starting from the second know month, we used the previous known month(s) to train the machine learning model and predict the unemployment rate for the coming months. Our algorithm for predicting the unemployment rate of South Africa and Nigeria using census measures is available in S1 Appendix (Fig A-5 in S1 Appendix).

We use the data on the unemployment and inflation rates to depict the Phillips curve for Canada, South Africa, and Nigeria, and compare and analyze the unemployment and inflation rate responses of these countries before and during the COVID-19 pandemic.

3.1. Data gathering and sentiment analysis

The number of tweets with all the keywords had a relatively strong correlation with the unemployment rate and GMI, respectively before and during COVID-19 for Nigeria and South Africa. Moreover, the total dataset had a strong correlation with the unemployment rate. Therefore, we used the volume of the tweets to predict the unemployment rates of Nigeria and South Africa. Next, we fine-tuned a sentiment analysis model using BERT and found 76% accuracy on our dataset. Fig 2 shows the confusion matrix of our model and Table 1 shows the different metrics of it.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Confusion matrix.

The confusion matrix of the sentiment analysis model.

https://doi.org/10.1371/journal.pone.0272208.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Machine learning metrics.

Different metrics of the sentiment analysis model.

https://doi.org/10.1371/journal.pone.0272208.t001

From the confusion matrix in Fig 2, we can see that the probability of our predictions not being accurate is pretty low. In addition, according to Table 1, our model has a high recall and precision, meaning that it is able to find each class (precision) and every one (recall) pretty well. Moreover, a high F1-score indicates that both recall and precision are high for all of the classes.

The sum of sentiment scores of the tweets had a negative correlation with the unemployment rate before and during COVID-19 pandemic for both South Africa and Nigeria. This was expected because as the unemployment rate increases, people become dissatisfied and sentiments decrease, and vice versa [26]. Table 2 shows the correlation between the sentiments and unemployment rate and our dataset for the period before and during COVID-19 pandemic for Nigeria and South Africa. In addition, Table 2 compares our trained BERT model with the pretrained BERT model used in [26]. The results show that the sentiment scores of our model have a higher correlation with the unemployment rate of South Africa before and during the COVID-19 pandemic. Detailed information can be found in S1 Appendix.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Correlation checking.

Correlation of the sentiment scores with unemployment rate and the dataset for Nigeria and South Africa.

https://doi.org/10.1371/journal.pone.0272208.t002

Unfortunately, GMI is not available for pre-COVID-19 times. Therefore, the unemployment rate cannot be estimated using that for the period before COVID-19. Moreover, it takes several days for GMI to get updated. However, Twitter data can be accessed in real-time. Therefore, social media is alternatively used to find the missing data on unemployment rate for South Africa and Nigeria.

3.2. Estimating unemployment rate

We applied four different machine learning algorithms, namely, RFR, SVR, XGBoost, and ARIMAX on the datasets for Nigeria and South Africa to predict the unemployment rates of the countries. With RFR, we searched for around ten different decision trees to find the best model, since increasing the number of trees to more than ten did not improved the accuracy. With SVR, we used Radius-Basis Function (RBF) as the kernel. Moreover, the best order that we found for ARIMAX was (1, 0, 1). We found that combining the four algorithms by averaging their results has the highest performance. Table 3 shows the performance of the different algorithms. We have compared the models in terms of Mean Average Percentage Error (MAPE), Root Mean Square Error (RMSE), and R²-score [55].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Method evaluation.

Accuracy of different machine learning algorithms used for predicting the unemployment rate.

https://doi.org/10.1371/journal.pone.0272208.t003

As can be seen in Table 3, averaging over all the machine learning methods reduces MAPE and RMSE by up to 76, 77 percent for Nigeria and 71, 77 percent for South Africa, and significantly increases R²-score. We used the estimated values to draw the Phillips curve for Nigeria and South Africa.

3.3. Analysis of macroeconomic responses to COVID-19 pandemic

Fig 3 shows the Phillips curve for Canada and South Africa and Fig 4 shows that for Canada, South Africa, and Nigeria.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Phillips curve.

Phillips curve of Canada and South Africa.

https://doi.org/10.1371/journal.pone.0272208.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Phillips curve.

Phillips curve of Canada, South Africa, and Nigeria.

https://doi.org/10.1371/journal.pone.0272208.g004

In Figs 3 and 4, The red, blue, and green lines are the Phillips curves for Canada, South Africa, and Nigeria during the COVID-19 pandemic, respectively. The black and yellow lines are the Phillips curves for years 2019 and 2017–2018 for the given country. Since the diagram of Canada is closer to the zero point of the coordinates, it is clear that Canada has a better economy compared to South Africa and Nigeria. Moreover, before COVID-19, the unemployment rate of Nigeria was lower than South Africa and the inflation rate of South Africa was lower than Nigeria. However, during COVID-19 pandemic, unemployment and inflation rates of Nigeria has increased more than Canada and South Africa.

As shown in Figs 3 and 4, in all the three countries, the unemployment rate has increased during 2020 compared to the previous years. For Canada and South Africa from the beginning of the pandemic until May 2020, with the unemployment rate rising, inflation rate decreased. This is in line with the Phillips curve short run. Since May 2020, Canada has been able to control and reduce the unemployment rate, but South Africa and Nigeria are still struggling with the high unemployment rate caused by the pandemic. With employment rate reducing for Canada, the inflation rate has increased. This again is in agreement with the Phillips curve short run. However, the inflation rate of Canada, especially after March 2021 has increased to its highest in more than fifteen years. Other than reduction of supplements during COVID-19 pandemic, the policies that Canada informed during COVID-19 pandemic by providing Canada implemented such as providing the Canada Emergency Response Benefit (CERB) and Employment Insurance (EI) to the employees who lost their jobs may have caused this high inflation rate.

Since May 2020, unemployment rate in South Africa has not decreased to pre-COVID-19 level, but the inflation rate has increased. Although inflation rate for South Africa is still comparable to pre-COVID-19 years, according to Phillips short run curve, it will increase as the unemployment rate decreases, and it may get higher than the years before COVID-19 pandemic.

Unemployment and inflation rates in Nigeria have increased fiercely compared to Canada and South Africa during the pandemic, and the country is experiencing a horrible stagflation. This shows how vulnerable lower-middle income countries could be to lockdowns, restrictions, and economic limitations.

As stated in previous works [14–17], in 2008–2009 global economic crisis, higher-income countries suffered more from the global recession. However, the economic burden of COVID-19 crisis is heavier on lower income countries. The reason is that in the 2008–2009 crisis, the recession was caused by the impairment of the communication net between countries, especially with the USA. However, the macroeconomic complexities of the COVID-19 pandemic are caused by lockdowns and economic inactivity. As a result, countries with less economic asset bear heavier setbacks.

Overall, in the near future, unemployment may not increase that much for different countries anymore. However, all the countries, no matter which income level they belong to, will most probably experience a rise in inflation rate due to supplement shortages caused by the pandemic.