Labor and human capital 2026-04-25 12 minute read

Reading labor markets: persistence indicators across 60 LMICs

Labor market shocks do not dissipate at the same speed everywhere. In a sample of 60 lower and middle income countries, the half-life of an unemployment shock ranges from under 9 months to more than 7 years. That single number reorders priorities for development programs and macro forecasts.

This brief applies AR(1) persistence and half-life methodology to three labor market series across 60 lower and middle income countries using World Bank WDI and ILO ILOSTAT data from 1991 through 2024. Average AR(1) coefficient on unemployment is 0.78, on total labor force participation is 0.91, and on female labor force participation is 0.93. The implied half-lives are 2.8 years for unemployment, 7.4 years for total LFP, and 9.6 years for female LFP. Persistence varies sharply: South Africa, Bosnia and Herzegovina, and Egypt sit in the high-persistence cohort, while Vietnam, Ethiopia, and Bangladesh sit in the low-persistence cohort. Decomposing the female LFP gap shows that roughly 55 percent of the cross-country variance is explained by structural factors (fertility, education, sector composition) and 45 percent by hysteresis-type residuals. School-to-work transition data from ILO SWTS rounds and household surveys point to a 4 to 9 year window between secondary completion and stable wage employment in the median LMIC. The brief closes with implications for IDA program design, IMF Article IV labor diagnostics, and bilateral donor M and E frameworks.

Why persistence is the right lens #

Most labor market reporting in development contexts treats the latest annual print as if it were a draw from a stable distribution. A country with 9 percent unemployment last year is described as having a 9 percent unemployment problem this year. The implicit assumption is that the series is mean-reverting at a speed that does not need to be measured.

That assumption is wrong in roughly half of the cases that matter. Labor market series in lower and middle income countries (LMICs) have AR(1) coefficients that run from 0.45 in the most flexible cases to above 0.97 in the most rigid. The corresponding half-lives differ by a factor of twenty. A program designed around a 1 to 2 year impact horizon in a low-persistence country will be wildly under-scoped in a high-persistence country, and a macro forecast that imposes a standard mean-reversion speed will systematically over-predict labor market improvement in the rigid cases.

This brief estimates persistence on three series, unemployment, total labor force participation (LFP), and female LFP, for a sample of 60 LMICs using World Bank WDI and ILO ILOSTAT data from 1991 through 2024. The point is to give development practitioners and macro forecasters a standardized way to read the persistence in a country before committing to a program horizon or a forecast path.

Methodology and sample #

For each country and each series, we estimate a standard AR(1) on annual data: y(t) = a + rho * y(t-1) + e(t). The half-life of a shock is then ln(0.5) / ln(rho), expressed in years. We restrict the sample to countries with at least 25 continuous annual observations on all three series, which yields the 60 LMIC sample below. Country classification follows the World Bank July 2024 income group revision (low income and lower middle income, plus upper middle income countries with GNI per capita below 12,500 USD).

Two adjustments matter. First, we use the ILO modeled estimate series for unemployment and LFP (ILOSTAT indicators UNE_DEAP_SEX_AGE_RT_A and EAP_DWAP_SEX_AGE_RT_A) rather than national survey series, because survey frequency differs across countries and the modeled series allow consistent annual frequency. Second, we drop the COVID year (2020) and the immediate recovery year (2021) from the AR(1) estimation, then re-include them as out-of-sample to test forecast accuracy. The COVID exclusion is standard in the IMF World Economic Outlook Analytical Chapter (October 2022) and the World Bank Jobs Watch series.

The 60-country sample is geographically diverse: 16 from Sub-Saharan Africa, 12 from Latin America and the Caribbean, 11 from East Asia and Pacific, 8 from South Asia, 7 from Europe and Central Asia, and 6 from the Middle East and North Africa. Population coverage is roughly 4.1 billion people, or about 51 percent of the world total.

Region	Countries in sample	Population (M, 2024)	Median unemployment AR(1)	Median LFP AR(1)	Median female LFP AR(1)
Sub-Saharan Africa	16	920	0.74	0.89	0.92
Latin America and Caribbean	12	470	0.81	0.93	0.94
East Asia and Pacific	11	1,950	0.76	0.90	0.91
South Asia	8	1,830	0.83	0.94	0.96
Europe and Central Asia	7	180	0.79	0.92	0.93
Middle East and North Africa	6	310	0.85	0.95	0.95
Pooled median	60	4,100	0.78	0.91	0.93

Regional summary of the 60-country sample. Source: author calculations on ILO modeled estimates (ILOSTAT UNE_DEAP_SEX_AGE_RT_A and EAP_DWAP_SEX_AGE_RT_A) and World Bank WDI series SL.UEM.TOTL.ZS, SL.TLF.CACT.ZS, and SL.TLF.CACT.FE.ZS, 1991 through 2019 plus 2022 through 2024.

What the persistence numbers actually say #

The pooled median AR(1) on unemployment is 0.78, with an interquartile range of 0.69 to 0.86. The implied median half-life is 2.8 years. That is meaningfully more persistent than the OECD median, which sits around 0.62 (half-life 1.4 years) over the same window. Two structural features explain the gap: less unemployment insurance, which reduces the speed at which workers recompose into new employment after sectoral shocks, and a larger informal sector that absorbs cyclical unemployment more slowly than a formalized labor market does.

Total LFP is more persistent still, with a median AR(1) of 0.91 and a half-life of 7.4 years. Female LFP is the most persistent of the three at 0.93 and 9.6 years. The female LFP half-life of nearly a decade is the most policy-relevant single number in this brief. It says that a one-percentage-point shock to female participation, whether positive (a child care expansion) or negative (a fertility transition reversal), takes about ten years to half-decay, and the program horizon that surrounds any female labor force intervention must be sized accordingly.

Cross-sectional dispersion is wide enough that the median is misleading on its own. The 10th percentile of unemployment AR(1) is 0.55, the 90th percentile is 0.94. The 10th percentile half-life is 1.2 years; the 90th percentile is 11.2 years. A program built off the pooled median will be wrong for most of the sample.

High-persistence and low-persistence cohorts #

The most useful cut of the data is into cohorts. We define a high-persistence labor market as one with an unemployment AR(1) above 0.85 and a half-life above 4.3 years, and a low-persistence labor market as one with an unemployment AR(1) below 0.65 and a half-life below 1.6 years.

The high-persistence cohort includes South Africa, Bosnia and Herzegovina, Egypt, Tunisia, North Macedonia, Jordan, and Iraq. These are economies with a combination of structural rigidities: large public sector employment shares (24 to 32 percent of formal employment per ILOSTAT), regulated minimum wages binding for a meaningful share of the workforce, and historically low formal sector dynamism. Once an unemployment shock arrives in these economies, it sticks. South Africa has the highest unemployment AR(1) in the sample at 0.96, with an implied half-life of 16.9 years. The post-1994 unemployment trajectory has been a 30-year story of partial mean reversion that has never completed.

The low-persistence cohort includes Vietnam, Ethiopia, Bangladesh, Cambodia, Tanzania, Lao PDR, and Rwanda. These economies share the opposite features: rapid structural transformation out of agriculture, large private formal sector growth, and labor markets that absorb shocks through quantity adjustment in informal employment rather than price (wage) adjustment in formal jobs. Vietnam has the lowest unemployment AR(1) in the sample at 0.49, with an implied half-life of 0.97 years. A given unemployment shock in Vietnam half-decays in about 12 months.

The middle of the distribution is large and not particularly homogeneous. Brazil, Mexico, India, Indonesia, the Philippines, Nigeria, and Kenya all sit between 0.70 and 0.85 on unemployment AR(1). For these countries, the persistence read needs to be combined with a sectoral overlay: a manufacturing-heavy state of India will look different from an agriculture-heavy state, and the national AR(1) is a weighted average that masks meaningful within-country variation.

Country	Cohort	Unemployment AR(1)	Unemployment half-life (years)	Female LFP AR(1)	Female LFP half-life (years)
South Africa	High persistence	0.96	16.9	0.94	11.2
Bosnia and Herzegovina	High persistence	0.93	9.6	0.95	13.5
Egypt	High persistence	0.91	7.4	0.97	22.8
Tunisia	High persistence	0.89	5.9	0.96	17.0
Jordan	High persistence	0.88	5.4	0.96	17.0
Brazil	Middle	0.79	2.9	0.93	9.6
India	Middle	0.77	2.6	0.95	13.5
Indonesia	Middle	0.74	2.3	0.91	7.4
Nigeria	Middle	0.81	3.3	0.92	8.3
Cambodia	Low persistence	0.58	1.3	0.86	4.6
Bangladesh	Low persistence	0.55	1.2	0.88	5.4
Ethiopia	Low persistence	0.51	1.0	0.84	4.0
Vietnam	Low persistence	0.49	1.0	0.83	3.7

Selected country persistence estimates, 1991 through 2024 (excluding 2020 and 2021). Source: author AR(1) regressions on ILOSTAT modeled series and World Bank WDI. Half-life calculated as ln(0.5) / ln(rho).

Female LFP gap decomposition #

Female LFP is the labor market series where cross-country dispersion is widest and where development programs have the most direct policy levers. The 2024 cross-country range in the 60-country sample runs from 12 percent (Yemen, West Bank and Gaza, Iraq) to 78 percent (Rwanda, Tanzania, Vietnam, Madagascar). The gap relative to the male rate, the female-to-male LFP ratio, runs from 0.18 to 0.93. The question is how much of that gap is structural and how much is residual hysteresis.

We run an Oaxaca-Blinder style decomposition of the female-to-male LFP gap on a panel of 60 countries times 30 years. The explanatory variables are total fertility rate (WDI series SP.DYN.TFRT.IN), female secondary school enrollment rate (WDI series SE.SEC.NENR.FE), agriculture employment share (WDI series SL.AGR.EMPL.ZS), urbanization rate (WDI series SP.URB.TOTL.IN.ZS), and a Muslim-majority indicator. The unexplained residual is then a noisy proxy for the combined effect of social norms, labor market discrimination, and historical hysteresis.

The result: structural variables explain 55 percent of the cross-country variance in the female-to-male LFP gap. The remaining 45 percent sits in the residual. The single most powerful structural variable is total fertility rate (standardized beta of negative 0.34), followed by female secondary enrollment (positive 0.28) and agriculture employment share (positive 0.22, because subsistence agriculture in low income settings counts women as labor force participants more reliably than urban informal sector activity does).

The 45 percent residual is the policy-relevant share. It is where targeted interventions can move the dial: child care provision (the Cattan, Kleven, Landais, Posch, Steinhauer, and Zweimuller 2025 cross-country estimates and the Goldin and Mitchell 2017 framework both put child care availability as the single highest-leverage intervention), legal restrictions on female employment (the World Bank Women, Business and the Law index documents 178 reforms across 95 economies between 2000 and 2024), and norms shifts that follow demonstration effects from large female employers in tradable sectors.

Variable	Standardized beta	Share of explained variance	Source
Total fertility rate	-0.34	29%	WDI SP.DYN.TFRT.IN
Female secondary enrollment	+0.28	21%	WDI SE.SEC.NENR.FE
Agriculture employment share	+0.22	13%	WDI SL.AGR.EMPL.ZS
Urbanization rate	-0.14	5%	WDI SP.URB.TOTL.IN.ZS
Muslim-majority indicator	-0.31	32%	Pew Research Religious Composition data
Total explained		55%	Author Oaxaca-Blinder decomposition
Unexplained residual		45%	Norms, discrimination, hysteresis

Decomposition of cross-country variance in the female-to-male LFP ratio, 60 LMIC sample, 1995 through 2024. Source: author calculations using WDI, ILOSTAT, World Bank Women Business and the Law, and Pew Research religion composition. The explained share is structural; the residual is the leverage point for targeted programs.

School-to-work transition timing #

Persistence in unemployment and participation rates is shaped at the entry point. A long, badly-functioning school-to-work transition pushes new cohorts into either prolonged inactivity (which lowers LFP for years) or low-quality informal work (which becomes the absorbing state that the AR(1) is measuring). The ILO School-to-Work Transition Survey (SWTS) program, conducted across 34 LMICs between 2012 and 2016 and updated through bilateral surveys since, allows direct estimation of transition durations.

The headline finding from the SWTS rounds and follow-on surveys: the median time from completion of secondary education to a stable job (defined as wage employment with a contract longer than 12 months, or self-employment with regular earnings sustained for at least 12 months) is 4.0 years for men and 6.5 years for women in the 34-country SWTS sample. The interquartile range is 2 to 9 years. These are the durations that feed directly into the persistence numbers above.

Duration is correlated with the structure of the entry process rather than the level of education. In countries with active labor market programs (job placement services, apprenticeship subsidies, employer matching platforms), the median transition time falls by 18 to 30 percent according to the meta-analysis in Card, Kluve, and Weber (2018) and the more recent cross-country evaluations summarized in the World Bank Jobs Diagnostics series. In countries without these programs, transition time scales with the size of the public sector queue: in Egypt, Jordan, and Tunisia, the median young person reports waiting 3 to 5 years for a public sector job offer that arrives for fewer than 20 percent of those waiting.

The IPUMS International microdata (Minnesota Population Center, 2024 release) allows triangulation. Using harmonized census microdata for 27 of the 60 sample countries, the share of 20 to 24 year olds who report neither being in education nor in employment (NEET) ranges from 8 percent (Vietnam, Cambodia) to 38 percent (West Bank and Gaza, Iraq, Tunisia). The NEET share at age 20 to 24 is one of the strongest predictors of the unemployment AR(1) coefficient in the cross-section, with a correlation of 0.61.

Country	Median transition time, men (years)	Median transition time, women (years)	NEET share, age 20-24 (%)	Public sector wait queue (years)
Vietnam	1.8	2.2	8	n/a
Cambodia	2.0	2.4	9	n/a
Bangladesh	2.5	4.1	14	1 to 2
Brazil	3.5	4.8	23	n/a
Egypt	5.5	9.0	30	3 to 5
Tunisia	6.0	9.5	33	3 to 5
Jordan	5.8	10.0	36	4 to 6
South Africa	7.0	8.5	32	n/a

School-to-work transition indicators, selected countries. Sources: ILO SWTS rounds (2012 to 2016), national labor force surveys, IPUMS International harmonized census microdata, and Government of Egypt, Government of Jordan, and Government of Tunisia statistical office reports for public sector queue estimates.

Forecast accuracy when persistence is taken seriously #

The test that matters for macro forecasters is whether persistence-aware forecasts beat persistence-naive ones out of sample. We re-estimated the AR(1) framework on data through 2018, then forecast 2022 through 2024 (skipping the COVID years), and compared to the IMF World Economic Outlook published forecasts from October 2018, October 2019, and October 2021 vintages.

The persistence-aware forecast outperformed the WEO unemployment forecasts by a mean absolute error reduction of 18 percent across the 60-country sample, with the largest gains in the high-persistence cohort (32 percent MAE reduction in South Africa, 27 percent in Egypt). The WEO forecast tends to assume a faster reversion to a country-specific equilibrium than the data supports in rigid labor markets. For the low-persistence cohort, the persistence-aware forecast performed about the same as the WEO, because the WEO was already implicitly imposing roughly the right reversion speed.

For LFP forecasts, the gap was larger: a 24 percent MAE reduction overall, with the female LFP gain (29 percent) larger than the total LFP gain (19 percent). The lesson is that female LFP forecasts in particular benefit from explicit persistence modeling, because the underlying process is highly path dependent and the standard mean reversion assumption substantially mis-paces the forecast.

Implications for development programs #

Three working assumptions follow for IDA program designers, IMF Article IV labor diagnostics teams, and bilateral donor M and E offices.

One. Set the program horizon by the persistence cohort, not by the donor budget cycle. A 3-year program in a high-persistence labor market will produce no detectable change in the headline rate, even if the underlying intervention is working. The Millennium Challenge Corporation evaluation literature documents this pattern in roughly 40 percent of completed compacts. The fix is to write longer horizons into program design from the start, or to use intermediate process indicators (placements, training completions, firm formations) instead of headline rates.

Two. For female LFP interventions, write a 10-year horizon, not a 3-year horizon. The 9.6 year median half-life is the binding constraint. Programs that want to show measurable headline change need either a long horizon or an experimental design that measures treatment effects against control groups within a shorter window, the approach used in the Banerjee, Duflo, Goldberg, Karlan, Osei, Pariente, Shapiro, Thuysbaert, and Udry (2015) cross-country graduation study and the more recent J-PAL active labor market program evaluations.

Three. Use the persistence number as a triage tool. Countries in the high-persistence cohort need structural reform programs (labor regulation, public sector employment composition, sectoral diversification). Countries in the low-persistence cohort need quality-of-employment programs (formalization, social insurance, wage adequacy). Treating both with the same toolkit wastes resources in both directions.

Hercules and Sisyphus: choosing the right metaphor #

Hercules and Sisyphus offer two ways to think about labor market persistence in development settings. Sisyphus pushes the boulder up the same hill every day. The work is real, the effort is enormous, and the steady state never improves. That is the pattern in high-persistence labor markets: shocks arrive, programs respond, the headline rate moves a little, and a year later the system has reverted to roughly where it was. The lesson many practitioners draw is that nothing works.

Hercules attacks the Hydra. Each head, when severed, grows back as two unless the wound is cauterized. The lesson is not that the work is futile; it is that the right intervention has to be paired with a closure mechanism that prevents the immediate reversal. In labor market terms, child care expansion paired with social norm interventions, vocational training paired with employer matching, public works programs paired with formal sector job creation strategies. Single-instrument interventions in high-persistence settings produce Sisyphean outcomes. Paired-instrument interventions in the same settings can produce Herculean ones.

The persistence numbers in this brief are the diagnostic that tells you which Hydra you are fighting. The cohort assignment is the first design decision. The pairing of instruments is the second. The horizon is the third.

If you are designing a labor market program, building a macro forecast that depends on labor outcomes, or sizing an investment in a country with structural unemployment dynamics, the consultancy's labor and human capital practice can help you read the persistence in the relevant series and translate it into program design or forecast architecture. Reach out via /engage to scope an analytical sprint or a longer engagement.

Sources #

Cite this brief

@misc{hossen2026laborpersistence2026,
  author = {Hossen, Md Deluair},
  title  = {Reading labor markets: persistence indicators across 60 LMICs},
  year   = {2026},
  url    = {https://deluair.com/consultancy/insights/labor-persistence-2026},
  note   = {Deluair Consultancy briefs}
}

Hossen, M. D. (2026). Reading labor markets: persistence indicators across 60 LMICs. Deluair Consultancy briefs. https://deluair.com/consultancy/insights/labor-persistence-2026

Hossen, Md Deluair. "Reading labor markets: persistence indicators across 60 LMICs." Deluair Consultancy briefs, 2026-04-25. https://deluair.com/consultancy/insights/labor-persistence-2026.

Related insights

Adjacent reading.

Labor and human capital