Beyond its impact on health, the COVID-19 pandemic impacted poverty through job losses, shutdowns, and much more. Yet, obtaining timely and comparable estimates of poverty during the pandemic proved challenging. With social distancing measures in place, the primary method of data collection for estimating poverty, conducting household surveys, became impractical in numerous countries.
As a result, we and others quickly turned to producing poverty nowcasts—timely estimates of poverty based on modelling from available data sources. In April 2020, we released poverty nowcasts using the latest country-level data from the Poverty and Inequality Platform (PIP) extrapolated forward with GDP growth projections. Prior to COVID-19, this simple approach was shown to work relatively well compared to more complex methods, yet it is unclear how well it performed during a time of an unprecedented global crises like COVID-19.
Over the past couple of years, survey-based estimates of poverty for several countries have become available in PIP. We now have 64 countries with an estimate of poverty in 2020 and 70 in 2021. However, there is uneven regional representation—34 out of 62 countries in Europe and Latin America have survey estimates for 2020 while only 6 out of 70 countries in the Middle East & North Africa, Sub-Saharan Africa, and South Asia combined have survey estimates for 2020. Nonetheless, this data allows us to finally, partially assess the accuracy of our COVID-19 poverty nowcasts. Though some estimates for 2020 and 2021 are still being processed, a full assessment of the nowcasts will never be possible due to the paucity of survey data collection during the pandemic.
After the initial nowcasts, we generated seven other nowcasts at the poverty lines typical of low-, lower-middle, and upper-middle income countries: $2.15, $3.65, and $6.85 per day. The nowcasts coincide with the updates to World Bank GDP projections (through the Macro and Poverty Outlooks and Global Economic Prospects). From October 2022 onwards, we also relied on microsimulations and phone survey data where possible.
Our nowcasts were focused on predicting global poverty, so we start by evaluating how well we predicted changes in poverty from 2019-2020 at the global level. As these global estimates in PIP still partially are based on extrapolations and interpolations, we also evaluate how our nowcasts fared at the country-level where we have more reliable survey-based estimates. The country evaluations focus on countries with comparable survey-based poverty rates in 2019 and 2020 (or 2020 and 2021), which reduces our sample to 54 in 2020 and 50 in 2021.
We calculate the difference between the actual change in poverty rates from 2019 to 2020 (or from 2020 to 2021) and the nowcasted change in poverty rates. An estimate closer to zero indicates better accuracy and a positive difference implies an underestimation of poverty changes. If a country already had a survey-based estimate of poverty in 2020 or 2021 at the time a nowcast was conducted (which increasingly is the case the later the “nowcast” was conducted), we remove it from the evaluation, as the nowcast was no longer in need.
Why were the nowcasts in some cases so off target?
In Brazil, the GDP growth estimates for 2020 suggested a large contraction of the economy, which we predicted would increase poverty. Yet, the government implemented large cash transfers to 67 million individuals which offset the negative growth and in fact reduced poverty. In 2021, the reverse is the case: the economy rebounded, so we predicted lower poverty; however, the new social protection programs were largely stopped in 2021, thus raising poverty.
In Kenya, the 2021 nowcasts predicted a decline in poverty emanating from a large recovery in real GDP per capita, yet the survey estimates of poverty suggest that extreme poverty increased by about 1 percentage point. This is in part driven by the consumption aggregate not including housing costs nor accounting for price differences within Kenya, in part due to the different prices used to inflate the national poverty line and international poverty line from 2020 to 2021, and in part because GDP growth is driven by the service sector, which has little impact on the mostly rural poor.
Lessons learned
In a time of unprecedented uncertainty, the error of the GDP-based model to nowcast poverty was relatively small at the global level, in part because country-level errors offset each other. For most countries, the projections were not much off target, but for certain countries, there were large deviations. This applies especially to countries where sizeable social programs were implemented (and later removed) and where there is a notable misalignment between growth in GDP and growth in welfare from household surveys, often because of measurement error in one of the two. In such cases, when producing nowcasts for a particular country, one may consider using different approaches, for example incorporating the latest or expected policy decisions in a model. Yet it is challenging to produce timely, country-tailored models spanning the entire world, making the GDP-based approach a practical alternative. The GDP-based approach can be applied to nearly all countries, and is easy to communicate, replicate, and cross-validate.
We gratefully acknowledge financial support from the UK Government through the Data and Evidence for Tackling Extreme Poverty (DEEP) Research Program.
Join the Conversation