Modelled against measured

The gap the model is filling

A household survey costs money, requires trained enumerators, and takes time to process. Many low-income countries conduct them once every five to seven years; some go longer between rounds. When an international series needs a figure for a year in which no survey was fielded, someone has to decide what to publish. The decision is rarely noted in the headline number.

The World Bank's PovcalNet database — now succeeded by the Poverty and Inequality Platform — assembles welfare distributions from survey microdata where those data exist. Where they do not, the series interpolates between two survey years or extrapolates forward from the most recent one, typically using national accounts growth rates as a proxy for how household consumption has moved. The methodology note for the platform describes the procedure, but that note and the estimate appear in separate places; a user downloading the time series encounters no flag distinguishing a directly-surveyed year from an imputed one.

The practical consequence is that a country's poverty headcount for a given year can be, in effect, a national accounts estimate dressed in survey clothing. National accounts measure aggregate output; household surveys measure the distribution of that output at the household level. The two instruments diverge in ways that matter: consumption in national accounts includes imputed rent on owner-occupied housing and expenditures by non-profit institutions, items that do not appear in a household's own account of what it spent. Survey-based welfare aggregates and national accounts consumption aggregates for the same economy and the same year routinely differ by double-digit percentage points.

How models propagate into published series

When the World Bank or other producers use a national accounts growth rate to move a survey estimate forward, they are assuming that household consumption grew at the same rate as aggregate consumption in the accounts, and that the distribution of consumption — the share going to the poorest decile, say — held constant. Both assumptions can fail. During periods of commodity-price movement, distributional shifts, or large informal-sector fluctuations, the household-level reality may move quite differently from the national aggregate. The imputed estimate will nonetheless be published and will carry the same decimal precision as a directly measured one.

This is not concealment so much as a practical necessity acknowledged in technical literature but rarely transmitted to the reader of a policy document. The Millennium Development Goal framework, which tracked income poverty against a 1990 baseline calculated years after the fact, relied heavily on this kind of modelled fill-in — the baseline problem shaped the reported trajectory of progress from the outset. The fraction of country-years in the global poverty time series that rest on interpolation or extrapolation rather than a survey round has been estimated in academic literature at roughly half, though the exact proportion depends on which database and which period is examined.

Survey non-comparability adds a further layer. Household surveys are not standardised instruments. A consumption survey conducted in one decade may use a seven-day recall period for food; a later round in the same country may use a thirty-day recall period. The switch typically shifts measured consumption downward, not because households are worse off but because respondents recall food expenditure differently over different windows. The same welfare aggregate looks larger under a shorter recall period. When a survey with a different recall period replaces the one it was extrapolated from, the series can jump in ways that mix real change with methodological artefact.

What the published table does not show

International poverty statistics are published with confidence intervals rarely and with methodological provenance almost never. A table might show a country's poverty rate falling from 42 to 31 percent over a decade without indicating that the earlier figure came from a survey, the middle years were imputed from national accounts, and the later figure came from a survey conducted under a different methodology. A reader comparing the two endpoints would infer a fall of eleven percentage points; the inferential chain connecting them is considerably more complex.

National statistical offices, not international agencies, produce the underlying data that feed into global series. When a national office releases a new survey or revises an old one, the international series is updated — sometimes with revisions that run backward through the published history. Rebasing national accounts, a periodic exercise that updates the reference year and the basket of goods, can move the level of measured consumption in a way that changes the poverty headcount even if nothing in the underlying household survey has changed. The revision is absorbed into the series with little visible trace.

UNICEF and other agencies producing child-welfare indicators face the same structure. A country's child mortality rate for a year between surveys is typically a model-based estimate, often using a demographic model fitted to multiple data points of varying quality — civil registration data where systems are strong, survey-based estimates where they are not, and sometimes indirect estimates from questions about child deaths asked of mothers in reproductive-health surveys. The resulting time series is smooth in a way that real mortality rates never are; the smoothness is a property of the model, not of the world. The UN Inter-agency Group for Child Mortality Estimation documents its estimation methods and publishes uncertainty intervals, a practice more transparent than most poverty series, though the intervals themselves are rarely reproduced in the policy documents that cite the headline figures.

The discipline the uncertainty demands

None of this means the figures are wrong in a disqualifying sense. Modelled estimates, carefully constructed and honestly labelled, carry genuine information. A well-fitted interpolation between two survey rounds, in a country with stable consumption growth and reliable national accounts, may be a reasonable approximation of what a survey would have found. The problem is the absence of labelling and the absence of uncertainty ranges that would let a user calibrate their confidence.

The statistical discipline the situation demands is one the architecture of the MDGs and of aid effectiveness monitoring did not consistently apply. Targets were set against baselines, progress was assessed against those targets, and decisions about debt relief — the HIPC completion point, for example, turned partly on poverty reduction strategy documentation — rested on numbers whose epistemic status was not always clear to those using them. The Paris Declaration monitoring surveys tracked disbursement and process indicators rather than welfare outcomes, which in retrospect looks like an implicit acknowledgement that the welfare figures were too uncertain to manage against directly.

What remains is a structure in which the headline poverty figure for a given country and year may be a survey result, an interpolation, an extrapolation, or a model output from a neighbouring economy's parameters, and the published number does not say which. Improving that transparency — publishing the source classification alongside the estimate, and publishing uncertainty bounds that reflect the source — is a data infrastructure question as much as a statistical one. It has been discussed for decades and implemented unevenly.