In August 2023, over one million Paraguayans disappeared. Results from the 2022 census, released that month, put the number of Paraguayans at 6.1 million — 1.4 million fewer than the numbers projected by the 2012 census. The numbers left Paraguayan policymakers reeling. As the minister of finance, Carlos Fernández Valdovinos, told reporters, “We will basically have to plan for a new Paraguay.”

Paraguay’s experience is not unique. In 2025, Indian enumerators are planning to conduct the country's first population census since 2011. Retrospective surveys done by India's National Population Register suggest that the earlier census could have missed up to 28 million people — roughly the size of the state of Punjab at the time.

In Nigeria, the most populous country on the African continent, there hasn’t been an official census since 2006. That, too, was marred by controversy and accusations of miscounting. As Eze Festus Odimegwu, the head of Nigeria’s National Population Commission at the time, put it, “No census has been credible in Nigeria since 1816. Even the one conducted in 2006 is not credible.” 

Good population estimates are critical to core functions of the state, including taxation, political representation, and the planning of health and education systems. To see why they matter, and why they can be so hard to get right, it helps to understand where the census came from and how it evolved in tandem with the state itself.

At its most basic level, a census counts the number of people living within a geographic area and records simple demographic information like age, gender, and occupation. Since this information is key to mobilizing troops and raising taxes — two of the most primordial functions of the state — the origins of the census go almost hand in hand with the formation of centralized political societies.

In ancient Egypt, the first recorded census appeared some 3,700 years ago. According to Herodotus, Pharaoh Ahmose I, who reigned in the 16th century BC, required every Egyptian to declare annually to the provincial governor “whence he gained his living.” This was used to map out the scale of the Egyptian labor force, which in turn informed the planning of state infrastructure projects, such as canals and irrigation systems. 

Through the Middle Ages, the census continued to be used most often for coercive purposes, such as taxation and conscription. One of the most famous censuses was the Domesday Book, ordered by William the Conqueror after his 1066 invasion of England. The census intended to take stock of the spoils of William’s conquest, including the people and land newly under his command. Its name, given in the 12th century, was given because no one was to escape the finality of its judgment, much like the doomsday in the Bible.

By the 19th century, as British society moved toward greater public participation in governance — with parliamentary supremacy established in 1689 and voting rights expanding through Reform Acts in 1832 and 1867 — the nature of the census evolved too. In 1841, the Office of National Statistics in the UK conducted what is considered to be the first modern census. For the first time, rather than being counted by state administrators, households filled out the census forms themselves. This shift in responsibility — from the enumerator to the enumerated — reflected a broader change in the imagining of the role of the state in British society. In addition to maintaining order, the state became increasingly viewed as a vehicle through which public services could be delivered. 

In the 1851 census, forms asked about disabilities for the first time. This information was not particularly relevant for raising taxes or building an army, but it was highly relevant for the design of public health and education systems — programs that continued to gather momentum through the century. 

Nowadays, it is taken virtually for granted that the responsibilities of the state extend to public services like health and education. Accurate population estimates are vital to delivering these services effectively. Put yourself in the shoes of a state administrator in Bauchi, a state in northern Nigeria consisting of 20 local government areas. Without knowing how many people live in each LGA, it’s hard to know how many teachers ought to be sent to each district, where medicines should be distributed, and where school and road building should be prioritized.

These types of decisions don’t only affect state administrators. For nongovernmental and philanthropic organizations trying to deliver and evaluate health and education programs, knowing how many people live where is critical. At GiveWell, where I work, if you dig deep enough into a cost-effectiveness model, you will more often than not run into a population estimate. For example: 

How cost-effective would it be to accelerate the rollout of the RTS,S malaria vaccine in eastern regions of Ghana? This depends on how many children would be vaccinated, which depends on how many children are living in the region. If the number of children is overestimated by 20%, cost-effectiveness will also be overestimated by 20%, making it harder to rely on these estimates for effective resource allocation decisions.

Or take a question that seems deceptively simple: How much does it cost to treat one child with preventative malaria medicine during the high transmission season? This is more than just the cost of the drugs, as the cost of getting them into the country and paying community health workers to deliver them to households must also be accounted for. The “all in” cost per child can be reduced to a seemingly simple equation:

Cost to run a campaign / Number of children treated in campaign

But the denominator here hides a great deal of complexity. How do we know how many children got treated? One way is to ask community health workers to track every pill they’ve dispersed. This places a lot of pressure on each health worker reporting accurately — which is especially challenging when inventories are tracked using pen and paper. It can also lead to overestimates if, as is common, drugs are mistakenly given to older children who aren’t targeted in the intervention.

It is more common to estimate coverage via a coverage survey. After the campaign, enumerators visit a randomly selected group of households in the targeted community and record the fraction of targeted children who received the drugs. To convert this to an estimate of the number of children treated, the estimated coverage rate is multiplied by the number of children estimated to be living in the area, which — crucially — relies on a population estimate:

Number of children treated = Coverage rate (%) * Number of children

Wherever you see an estimate of cost per child reached, there is probably a population estimate lurking underneath.

Unfortunately, reliable, up-to-date estimates can be hard to come by. Over half of all Africans live in a country where there hasn’t been a census since 2009. Next year will mark 20 years since Nigeria conducted theirs. The Democratic Republic of the Congo, the continent’s second-most populous country, hasn’t held a census since 1984.

Given the stakes — economic, political, bureaucratic — of inaccurate estimates, it’s reasonable to ask why, in many countries, censuses occur so infrequently. The first and most obvious reason is that conducting them is hard. Running a census is expensive and logistically challenging, especially in countries with poor transport and communications infrastructure.

In the UK, the census is now mostly digital, with 89% of households completing the 2021 census via an online survey. The remaining 11% sent off their census form via the post — made possible by a functioning postal system, a good road network, and generally up-to-date electoral registers. 

Compare this with the DRC, where a population of some 100 million is scattered across a country the size of Western Europe, much of it blanketed by the world’s second-largest tropical rainforest. The GDP per capita is just over $740. Less than one-third of people use the internet. Less than 2% of the road network is paved, which means the majority of the rural population lives at least two kilometres away — often more — from an all-season road. Finally, the east of the country has been historically unstable. The current conflicts have displaced 5.6 million people. Against this backdrop, the idea of counting everyone living in the DRC seems a Herculean task.

When full enumeration is not possible, sample-based techniques can be used, though these have limitations. To know whether we have a nationally representative sample, we need to know what we’re sampling from. This leads to a somewhat paradoxical situation: If we have a representative sample, we can make inferences about the population, but to know whether we have a representative sample, we need to know about the size and basic characteristics of the population, which requires a full enumeration. 

A second reason conducting any census is hard stems from the inherently political nature of the exercise. Since its inception, the census has been a vital tool for determining how the state sees its citizens and how its citizens make themselves seen. Both parties have incentives to manipulate this lens. Those incentives are shaped, in turn, by the nature of the relationship between citizens and the state. 

In The Art of Not Being Governed, political scientist James C. Scott argues that for the majority of human history, this relationship has been an antagonistic one. The state attempts to tax, bully, and coerce its citizens on the one hand, and citizens try to deceive, manipulate, and hide from the state on the other. 

In 15th-century Florence, a census was botched after two-thirds of respondents refused to give surnames, having (correctly) inferred that doing so would make them easier to track and hence vulnerable to future tax and military demands. In the Qing dynasty of 18th-century China, households undercounted their members in an attempt to avoid tax. Retrospective estimates suggest that underreporting may have been as high as 50% in some provinces.

Perhaps most audaciously, in the early 20th century, Britain sought to enumerate all of the people living under its empire, including the 300 million subjects then estimated to be living in India. At the time, this was lauded as the most ambitious census attempt ever, but the people living in India — who knew the numbers would likely be used to levy demands on them — did not share the same enthusiasm. According to National Geographic:

The wave of decolonization that swept across Africa and Asia in the mid-20th century fundamentally reshaped this dynamic. Having spent decades trying to avoid the eye of colonial rulers, different groups now clamoured to be counted — and to do the counting — in order to increase their claims to political representation and state resources.

For example, in Nigeria the first postcolonial census was held in 1962 — two years after the country had gained independence, but two years before the first postcolonial elections. This census found that the population in the southern regions of the country had increased by 70% over the previous decade, compared with just 30% in the north. Most northern Nigerians, then and now, are Muslims belonging to the Hausa or Fulani peoples, while southerners are typically Igbo or Yoruba, and Christian. 

Had these figures been accepted, it would have upset the country’s delicate religious and ethnic balance of power. But the census was rejected by the Northern People’s Congress, which ordered a recount. This time, the north’s population was found to have increased by 84% over the previous decade — faster than the south and enough to ensure it continued to have more than half of Nigeria’s population, buttressing claims to state resources and legitimacy. 

Nigeria ran further censuses in both 1991 and 2006. During this period, the population of Nigeria was estimated to have increased by over 50%, from around 90 million to 140 million. Despite this large increase, each state's share of the total population was implausibly estimated to have remained exactly the same. ¹

State representatives of Lagos, concerned that the census was being manipulated to support the northern states’ claims to resources and representation, ordered their own state census soon after. This count put the total population of Lagos at 17 million — almost twice as high as what the national census had estimated. Disputes turned violent as seven enumerators were reportedly attacked with acid by members of the Movement for the Actualisation of the Sovereign State of Biafra, a secessionist movement demanding an independent southern state. 

Given the logistical difficulties and political controversies associated with censuses, it can be tempting to give up on on-the-ground data collection entirely and go all-in on remote sensing methodologies. These methodologies reinforce traditional censuses with big data, such as nightlights, satellite images, and mobile phone subscriptions. Two of the most well-known estimates come from WorldPop, an academic research group based at the University of Southampton, and the High Resolution Settlement Layer, part of the company Meta’s Data for Good initiative. WorldPop uses machine learning to redistribute census data into high-resolution grids by modeling population density based on satellite imagery and environmental variables. Meta’s approach, more bottom-up, detects individual buildings from satellite imagery and allocates population counts directly to those structures to create highly localized population maps.

These approaches can be cheaper, since they mostly use existing data, and less prone to manipulation, since it's harder to hide from a satellite than a census enumerator. However, they aren’t a panacea. For one thing, these estimates rarely line up, especially at the subnational level. Put yourself back in the shoes of the state administrator in Bauchi, northern Nigeria, and consider these competing estimates of the under-5 population in different LGAs produced by WorldPop and Meta. The differences are stark — whether there are 160,000 or 75,000 children under 5 living in Ganjuwa clearly matters for how many preschool teachers to hire, how many vaccines to procure, or where community health workers should be distributed. 

Noisiness in these estimates is one concern; bias is another. Since all of these estimates are based on similar inputs — most importantly the latest census — biases in the inputs may filter through to produce biases in the estimates themselves. A recent paper in Nature Communications claims that gridded population estimates, including those produced by WorldPop and Meta, could be systematically undercounting rural populations by between 53% and 84%. To arrive at these estimates, the team used resettlement numbers from 307 dam construction projects in 35 countries. Since dams often create a large floodplain, people in the affected areas are usually resettled. This prompts governments to run highly detailed surveys — effectively a mini-census — of the people living in the affected area.

The authors use these mini-censuses as “ground truth” data on the number of people living in each area and then compare these with the model’s predictions about the number of people living in the same area. Amazingly, they found that WorldPop’s models underestimated headcounts by around 50%, and other models underestimated them even more dramatically.

This paper has limitations. Biases in the resettlement surveys, or comparing geographic areas that don’t perfectly align, could undermine the apples-to-apples comparison. Nonetheless, if this validation exercise is even close to being accurate, it has stark implications for resource allocation decisions being made right now. If rural populations are being systematically undercounted in censuses and wider population estimates, then perhaps they’re also being systematically underinvested in. This could go some way in explaining the stark rural-urban divides in literacy, health, and poverty seen in many countries around the world.

Discrepancies between present-day population estimates are concerning, but they’re not the whole story. To effectively allocate resources, we’d need know how many people live where today, and how many people we can expect to live where tomorrow. Without knowing this, strategic planning for the number of hospitals to build, the size of social security programs, and the optimal path of government debt becomes extremely difficult.

Unsurprisingly, noisy present-day estimates lead to noisy forecasts. For example, in 1990 the UN predicted that there would be 281 million Nigerians by 2025; in 2000 this forecast fell to 203 million; in 2010 it rose again to 230 million. For the politician in Abuja, planning for the county’s future is difficult when forecasts fluctuate with an error bar in the tens of millions. 

Unfortunately, things could get worse before they get better. To produce reliable forecasts, you need a reliable base estimate (i.e., a reliable, recent census) and good data on “flow” variables like birth and mortality rates. In rich countries, these statistics are taken from vital registration systems, where births and deaths are reliably recorded in hospitals or local register offices. In poor countries, these systems are far less reliable. In Nigeria, for example, only 43% of births were estimated to have been registered in 2021, while in Lagos state, only 2% of deaths were reportedly registered.

In the absence of reliable administrative data, representative household surveys are used to estimate these crucial flow variables. The most important of these is the Demographic and Health Survey, a cross-country initiative responsible for collecting and disseminating accurate, nationally representative data on health and population in low- and middle-income countries. Since 1985, when the DHS began, between 25% and 40% of the inputs used by the UN to derive their total fertility rate estimates have come from a DHS question that asks women to list the number of children they have given birth to.

The DHS is funded by both the countries hosting the surveys and international donors — the most important of which was USAID, which in 2024 signed a $250 million contract to support the continuation of DHS surveys in over 50 countries between 2025 and 2030. In February 2025, this contract was terminated, leaving the project currently on pause. If these surveys aren’t restarted, critical inputs to population forecasts could be lost, making error bars even more inflated and decision-making even more challenging. 

The consequences for human welfare could be large. The UN estimates that there will be 700 million additional people on Earth by 2034, 637 million of whom will be living in low-income and lower-middle-income countries. It can be easy to dismiss counting these people as a mundane statistical task, less important than the more visibly urgent development goals. But ensuring adequate health care, infrastructure, and education depends fundamentally on the more ancient task of counting people — a task that is no different, in essence, from the one faced by enumerators in ancient Egypt four millennia ago. If we want these people to grow up in a world with well-stocked clinics and properly resourced schools, we need to make sure they are counted.