The unequal impact of the coronavirus pandemic: Evidence from seventeen developing countries

The current coronavirus pandemic is an unprecedented public health challenge that is having a devastating economic impact on households. Using a sample of 230,540 respondents to an online survey from 17 countries in Latin America and the Caribbean, the study shows that the economic impacts are large and unequal: 45 percent of respondents report that a household member has lost their job and, among households owning small businesses, 59 percent of respondents report that a household member has closed their business. Among households with the lowest income prior to the pandemic, 71 percent report that a household member lost their job and 61 percent report that a household member has closed their business. Declines in food security and health are among the disproportionate impacts. The findings provide evidence that the current public health crisis will exacerbate economic inequality and provides some of the first estimates of the impact of the pandemic on the labor market and well-being in developing countries.


Questionnaire
The questionnaire collects data on monthly total household income in January 2020 and expected monthly total household income in April 2020 in ranges constructed as multiples of the minimum wage in that country. The questionnaire asked if total household income was reduced during the past week. The questionnaire asked whether a household member lost their job or closed their business. The recall period for these questions was randomized between 1 week, 2 weeks, and 1 month. For respondents that remain in the labor market, the questionnaire asks whether the respondent worked outside the home or from home during the past week and we code a variable to represent teleworking if the respondent reported working from home. Using the past week as the recall period, the questionnaire asks if any member of the household went hungry due to lack of food and asks respondents to report the strength of their agreement with a statement "I am eating less healthy than normal" on a scale from 1 (complete disagreement) to 5 (complete agreement). The questionnaire asks if any member of the household received (gave) a transfer or a loan from (to) another household during the past week. The questionnaire also asks respondents to report the strength of their agreement with a statement that COVID-19 should be the top priority of the national government on a scale from 1 (complete disagreement) to 5 (complete agreement) and asked whether respondents think that non-essential businesses should close.

Data Construction
The questionnaire was implemented online using Qualtrics. Column (1) of Table S1 displays the launch date of the survey in each country. We construct the data set in several steps and the resulting number of observations that comprise the sample for each country is shown in column (2) of Table S1. First, we restrict the sample to completed surveys. Overall, approximately 59% of surveys that are started are completed. Second, we restrict the sample to surveys associated with IP addresses within the borders of the country for which the respondent is completing the survey. Across all countries in the sample, over 99% of completed surveys comply with this criteria. Third, Qualtrics flags surveys that are completed on the same device and likely to be repeat surveys completed by the same individual or household based on cookies. This is an imperfect filter. For example, it will not recognize repeated surveys by the same individual or household that are completed on different devices. Surveys flagged as repeats comprise less than 2.3% of completed surveys and we drop these surveys from the sample.

Estimation of weights
For each country, we modeled the probability of being in the nationally representative sample as a function of demographic, time-invariant characteristics such as gender, age and education of the respondent, indicators for the presence of children of 5 years old or younger in the household and elderly (60 years old or older) in the households, as well as the number of household members and the number of children enrolled in school. Finally, we also include region-specific indicators.
We conducted this process country by country. For each country, we used the most recent nationally representative survey available in the Inter-American Development Bank harmonized survey data repository. We estimated the model by fitting a logistic function (logit) and computed predicted probabilities of being in the nationally representative data set (p i,c ) for each respondent i in country c (see Table S2). We then used inverse probability weights (ipw i,c = 1/(1 −p i,c )) to, at least in terms of observable characteristics, resemble those from nationally representative surveys.
To prevent differences in response rates from driving the results, we re-scale the within-country weights ipw i,c by the inverse share of re-weighted number of responses per country, relative to the country's population size (P opulation c /( Nc i ipw i,c )).

Out-of-sample validation
We validate the country-specific weights by looking at their out-of-sample performance. For each country we randomly selected 60% of the observations in the online survey and the household (field) surveys. We used this sample to estimate logit models. We then use the resulting models to predict the probability of being observed in the household survey over the observations in the remaining 40% of the sample, which we call the testing sample. We then re-weight the observations in the testing sample and compare the adjusted means of demographic characteristics to those obtained from household (field) survey data. Table S3 shows this exercise in Columns 3 and 4. We were able to reduce the differences in respondent characteristics between the online and household surveys. This out-of-sample exercise suggests that the improvement in balance is not driven by overfitting. We also show that our online data does not differ substantially from the household survey data in non-targeted moments such as the proportion of respondents by income category.

Countries by type of mobility-restriction policies
The following list details the type of mobility restrictions implemented by each country. Data was collected from official government websites and press articles. The information was collected on April 23rd 2020.
No mandatory policies: Bahamas, Suriname, Trinidad and Tobago, Uruguay. Curfews: Dominican Republic, Guyana. Local Quarantines: Chile, Jamaica, Mexico. National Quarantines Barbados, Bolivia, Colombia, Costa Rica, Ecuador, El Salvador, Panama, Peru.  Job loss is a dummy variable that equals one if the respondent indicates that they lost their job within the last week, two weeks or month (these horizons were randomly assigned). Similarly, the variable for business closures equals one if subjects report that their business was forced to close by the government or by lack of demand within the same randomly assigned time horizon.

Statistical Analysis Details
To explore differences in job loss and business closure across the income distribution, responses are first aggregated at the country and income bracket level using within-country weights as described in the Supplementary Information Appendix. Country averaged outcomes are then regressed by OLS on indicators for income bin and control for country fixed effects, weighting by country population. Because there are only have 17 countries in the sample, using standard clustered standard errors would result in incorrect inference. Therefore, the analysis follow [1] by aggregating the data at the country level by income category. Fig 1 shows the point estimates and 95 percent confidence intervals for each income bin average. As robustness, results are shown for the raw pooled data, without applying weights in Figure S2.  For this, respondents that reported that at least one household member lost their job during or had to close their businesses during the survey reference period were identified, and calculated shares using weights to correct for sampling issues (See Supplementary Information section). Data corresponding to the share of self-employed workers was obtained from the World Bank's World Development Indicators.

Figure 3
The share of household income in each income bin for incomes reported by respondents corresponding to January 2020 and April 2020 are presented in Fig 3. The sample is restricted to responses collected between April 13th to May 1st 2020 to include all countries. Shares are calculated by re-weighting counts by within-and across-country weights as detailed in Supplementary Material section Estimation of weights. As robustness, results are shown for the raw pooled data, without applying weights in Figure S3.

Figure 4
Fig 4 shows the share of respondents supporting extending the lockdown policies by at least one month over time since the first COVID-19 case in the country. One challenge to explore how support changes over time is that the timing of responses varies between countries. Over time as new countries enter the sample, the change in sample composition could bias our results. We address this issue in two steps. First, we construct new weights and re-weight respondents after day 31 to match those responding at the beginning of the series (days 30 and 31) on observable demographic characteristics.
A model was estimated by fitting a logistic function (logit) and computed predicted probabilities of having responded to the survey on day 30 or 31 relative to first case. Explanatory variables included were: demographic characteristics such as household size, age indicators for: presence of children or elderly, having felt hunger, woman, education primary or less, college education or higher, household income categories (<1, 1-2,2-3, 8-11,>11), and for day of week. Then, inverse probability weights were estimated based on the propensity score obtained.
Second, OLS regressions were separately estimated for respondents whose household member lost job or closed business and for those who did not, where the dependent variable is an indicator variable that equals one if they support extending the lockdown by at least a month, and zero otherwise. The independent variables are indicators for days since first COVID-19 case, day of week indicators, and country fixed effects. The regression is weighted by the inverse probability weight described in step one and employs robust standard errors. Fig 4 presents the point estimates and 95 percent confidence intervals for these estimates. Similar estimates when using the pooled raw data without re-weighting are presented in Figure S4. Table 1 Table ?? reports comparisons of differences in the relevant outcomes between respondents whose households experienced a loss of livelihood during the pandemic period and those that did not. To control for time-varying characteristics, we perform comparisons focusing on respondents in the same locality and who completed the survey during the same day. We exploit granular data from over 3000 localities in 17 countries. Thus, we focus our analysis on the subset of 18,000 locality-date-of-response cells that include more than one observation. This approach allows us to isolate time-varying locality shocks, and thus purge regional confounding factors. In addition we control for industry sector fixed effects, to prevent differences in the sectors related to the household's main economic activities from driving the results, as exposure to the effects of the pandemic may vary across sectors.
We operationalize our approach by estimating the following specification, which is similar to that used in reference [2]: Here, Y i,l,c,t denotes the outcome of interest corresponding to respondent i in locality l from country c collected in date t. Lost livelihood i,l,c,t is an indicator of whether any member of the respondent's household lost her/his job or closed her business during the the past week, two-weeks, or month. X i,c,l,t is vector of demographic characteristics of the respondent (age, education level and gender) and of household characteristics (household size, presence of children younger than 5, presence of school-age children, and people 60 years old or older). δ l,c,t denotes locality-date fixed effects, and θ s denotes industry fixed effects based on the main pre-pandemic source of income of the respondent's household. i,c,l,t denotes unobserved shocks. To account for possible serial correlation of outcomes within localities, we cluster the standard errors at the locality-country level.
The parameter of interest, β, is reported in Table 1 and captures within-locality differences in outcomes between respondents whose households experienced a loss of livelihood during the pandemic and those who didn't. The reported models are estimated using weights to achieve country-level representativeness, to correct for differences in sample size across countries, and to provide higher based on country population size. All results are robust to excluding these weights (See Robustness Section in Supplementary Material).
Dependent variables: Went hungry is an indicator of whether any household member went hungry during the past week due to lack of food. Eats less healthy takes the value of one if the respondent somewhat or totally agrees with the statement "I eat more unhealthy foods than normal". Gifts/Loans is an indicator of whether the respondent's household received a gift or transfer from either friends or relatives during the preceding week. Gov. Priority is an indicator of whether the respondent somewhat agrees or totally agrees with the statement "The government's priority should be to stop the spread of the pandemic". Lockdown (>= month) is an indicator of whether the respondent reports agreeing with closing non-essential business for one month or longer. As this question was asked only to people that reported agreeing with policies that require non-essential businesses to close, Lockdown(>= month) takes the value of zero when the respondent reported not supporting measures of keeping non-essential businesses closed at all, regardless of the time.

Figure 2
To investigate whether the correlates of job loss or business closures and household outcomes are stronger in countries with high levels of informality, we also estimate the following specification: (2) +X i,c,l,t Σ + δ l,c,t + θ s + i,l,c,t Where Self-employment denotes the share of self-employed workers in country c and was obtained from the World Bank's World Development Indicators, using the most-recent observations for each country. We also report results using the share of informal workers (as a share of non-agricultural workers) in each country. This information was not available for Suriname, Jamaica, Trinidad and Tobago, Barbados, and Bahamas. See results in Table S4.   Date of launch is the date on which the social media posts began. The date that the survey was rolled out in each country was largely determined by bureaucratic processes and approvals. With the except of Costa Rica, data collection in each country continued until April 30, 2020. The number of observations for each country reflected the number in the sample after data cleaning. The percent of localities is the percent of localities of each country for which we have one or more observations in the sample. Household size -0.002 -0.575*** -0.206*** -0.123*** -0.073*** -0.061*** -0.084*** -0.047*** -0.195*** -0.326*** -0.048*** -0.085*** -0.097*** -0.197*** -0.067*** -0.125*** (0   -5.185*** -1.623*** -6.316*** -6.035*** -5.904*** -5.698*** -6.427*** -6.900*** -6.614*** -6.726*** 0.000 -6.797*** -6.041*** -8.058*** -10.927*** -5.377*** ( The table presents estimates of coefficients from a logit model of the probability of being the nationally representative household survey as a function of demographic characteristics. The models were estimated separately for each country. All models also included region-specific indicators, except in the case of Ecuador in which data regarding regions was not available. Empty cells in the table imply that the relevant variable was not available in the household surveys. Robust standard errors are presented in parenthesis. The table presents means of household and survey respondent demographic characteristics using data from the online survey and nationally representative surveys, and pooling observations from all study countries (weighting by country size). Column (1) reports raw means using all the observations from the online surveys. Column (2) reports means using all available observations in the household (field) surveys using sampling weights. Column (3) reports means from the online survey data using only data from the testing sample (i.e., the sample not used for the estimation of the inverse probability weights). Column (4) reports means using data from the household (field) surveys corresponding to the testing sample. The testing sample corresponds to a randomly selected subsample corresponding to 40% of all the observations in the online and household (field) surveys. MW stands for national minimum wage. The inverse probability weights are computed based on logit models of the probability of being observed in the household survey which are estimated country by country. The models include including age, gender, and education categories of the respondent as well as household-level demographic characteristics such as the presence of children younger than 5 years old, the presence of elderly children in the household, # of children enrolled in school, household size, as well as region fixed effects. Panel A reports unweighted regression coefficients capturing the relationship between livelihood losses during the pandemic and outcomes. Each column reports results of a regression of the dependent variable on an indicator of whether any household member either lost a job or closed a business and a vector of covariates. In addition, all regressions control for locality × day of survey completion fixed effects (18,764), as well as economic-sector fixed effects. Standard errors are clustered at the locality level (3,165). Panels B and C replicate the results using country-date fixed effects (300) and standard errors clustered at the country level (17). See the Empirical Methods section in the main text for more details. Panel A reports regression coefficients capturing the relationship between livelihood losses during the pandemic and outcomes, as a function of the share of self-employed workers in each country. Each column reports results of a regression of the dependent variable on an indicator of whether any household member either lost a job or closed a business, and an interaction term of job loss or business closure with the share of self-employed workers in each country, and a vector of covariates. In addition, all regressions control for locality × day of survey completion fixed effects (18,764), as well as economic-sector fixed effects. Standard errors are clustered at the locality level (3,165). Panels B replicates the results using the percentage of informal workers as in each country instead of the share of self-employed workers. See the Empirical Methods section in the main text for more details.