key: cord-0016128-q90v5kn0
authors: Masquelier, Bruno; Hug, Lucia; Sharrow, David; You, Danzhen; Mathers, Colin; Gerland, Patrick; Alkema, Leontine
title: Global, regional, and national mortality trends in youth aged 15–24 years between 1990 and 2019: a systematic analysis
date: 2021-03-01
journal: Lancet Glob Health
DOI: 10.1016/s2214-109x(21)00023-1
sha: 0cf63517c3a7ece967588a66b443e112fc72d5da
doc_id: 16128
cord_uid: q90v5kn0

BACKGROUND: The global health community is devoting considerable attention to adolescents and young people, but risk of death in this population is poorly measured. We aimed to reconstruct global, regional, and national mortality trends for youths aged 15–24 years between 1990 and 2019. METHODS: In this systematic analysis, we used all publicly available data on mortality in the age group 15–24 years for 195 countries, as compiled by the UN Inter-agency Group for Child Mortality Estimation. We used nationally representative vital registration data, estimated the completeness of death registration, and extracted mortality rates from surveys with sibling histories, household deaths reported in censuses, and sample registration systems. We used a Bayesian B-spline bias-reduction model to generate trends in (10)q(15), the probability that an adolescent aged 15 years would die before reaching age 25 years. This model treats observations of the (10)q(15) probability as the product of the actual risk of death and an error multiplier that varies depending on the data source. The main outcome that we assessed was the levels of and trends in youth mortality and the global and regional mortality rates from 1990 to 2019. FINDINGS: Globally, the probability of an individual dying between age 15 years and 24 years was 11·2 deaths (90% uncertainty interval [UI] 10·7–12·5) per 1000 youths aged 15 in 2019, which is about 2·5 times less than infant mortality (28·2 deaths [27·2–30·0] by age 1 year per 1000 live births) but is higher than the risk of dying from age 1 to 5 (9·7 deaths [9·1–11·1] per 1000 children aged 1 year). The probability of dying between age 15 years and 24 years declined by 1·4% per year (90% UI 1·1–1·8) between 1990 and 2019, from 17·1 deaths (16·5–18·9) per 1000 in 1990; by contrast with this total decrease of 34% (27–41), under-5 mortality declined by 59% (56–61) in this period. The annual number of deaths declined from 1·7 million (90% UI 1·7–1·9) in 1990 to 1·4 million (1·3–1·5) in 2019. In sub-Saharan Africa, the number of deaths increased by 20·8% from 1990 to 2019. Although 18·3% of the population aged 15–24 years were living in sub-Saharan Africa in 2019, the region accounted for 37·9% (90% UI 34·8–41·9) of all worldwide deaths in youth. INTERPRETATION: It is urgent to accelerate progress in reducing youth mortality. Efforts are particularly needed in sub-Saharan Africa, where the burden of mortality is increasingly concentrated. In the future, a growing number of countries will see youth mortality exceeding under-5 mortality if current trends continue. FUNDING: UN Children's Fund, Bill & Melinda Gates Foundation, United States Agency for International Development.

Appendix Global, regional, and national mortality trends in youth (15) (16) (17) (18) (19) (20) (21) (22) (23) (24) Data sources for the mortality rates were nationwide vital registration and sample registration systems, nationally representative sibling histories from surveys, and recent household deaths reported in censuses. Mortality rates were recalculated from the survey microdata when available or based on published tabulations.

1.1 Vital registration data 1.1.1 Data availability Vital registration data were extracted from the WHO mortality database 1 . For 138 countries, death registration data were available for at least one year since 1990 2 . Mortality rates for the period before 1989 were not included in the statistical model as our focus is on trends in the period 1990-2019.

We assumed that death registration was complete in 38 countries included in the Human Mortality Database (HMD). For the other countries, we evaluated the completeness based on "death distribution methods" (DDMs). These methods compare the age distribution of deaths in the VR with the age distribution of the population enumerated in censuses to estimate the fraction of deaths that are reported. Each method has a standard version that applies to one census only but assumes a stable population. The methods have subsequently been revised for use in non-stable populations [1] . Three methods are applicable when two censuses are available, and deaths have been recorded in the intercensal period: the generalized growth balance method (GGB) [2] , the synthetic extinct generation method (SEG) [3] , and a combination of the two, called the hybrid method (or GGB>SEG). We employed here the GGB>SEG approach, which has been shown to perform better than the GGB and SEG methods used separately [4] .

All the DDMs make strong assumptions, that under-reporting of deaths is constant by age (over a certain age limit), that age reporting is accurate, and that the population is closed to migration (unless net migration rates are available by age and sex). To reduce the sensitivity of estimates to violations of these assumptions, demographers usually select an age interval to calculate completeness, eliminating the youngest and oldest age groups to avoid introducing errors due to small numbers of deaths, age misreporting or net migration. Based on simulations, recommend using the age group 40-70 for the GGB, 55-80 for the SEG, 50-70 for the GGB>SEG [5] . Measuring completeness from adults over the age of 40 (or 50 or 55) reduces potential biases introduced by migration, which is often more frequent in young adults. However, estimates based on older ages are more likely to be affected by age reporting errors. In another study, Hill and colleagues (2009) suggested using the age groups 5+ to 65+ for fitting purposes [4] . The selection of age groups remains an area requiring further research. Depending on the frequency of age errors and the extent of migration, various age groups should probably be selected for different methods and/or applications. In this study, the age range was selected based on visual inspection of the diagnostic plots. For example, in the case of the GGB method, it is usually recommended to plot b(x+), the partial birth rate, against d(x+), the partial death rate, for each age x. A regression line is fitted on these points, and the analyst examines the fit and residuals to select the points that provide the best fit [1] . The DDM package of R [6] replicates this manual approach by retaining the points that minimize the square of the average squared residual. We inspected all diagnostic plots, and in most cases, retained the age range that had been selected by the DDM package. In a few countries with substantial migration, or when completeness estimates were highly sensitive to the selection of age groups, we preferred to opt for older age groups, often selecting the age range 50-70 as recommended by .

Completeness was estimated for each country for periods between pairs of recent censuses for which an age distribution of the population was available in the Demographic Yearbook 3 . When the estimated completeness was less than 80%, mortality rates derived from vital registration data were excluded from the model fit. When completeness was greater than or equal to 95%, the registration was considered complete, and mortality rates were not adjusted upwards. If completeness was between 80 and 95%, we multiplied the inverse of the completeness rate by the number of deaths to obtain adjusted estimates.

We did not introduce any smoothing of completeness levels, and changed the level from one intercensal period to the next. Mortality trends are smoothed in a second step using the B3 model (see below). After the last census, it is not possible to estimate completeness with death distribution methods, so completeness levels were kept constant until 2019.

Table S1 below presents the intercensal periods considered, the completeness estimates obtained with the GGB>SEG method and the age groups considered for the fit. In 103 countries, vital registration data could be used for at least one year after 2010. This number is slightly larger than that for under-five mortality (94 countries). The completeness of death registration could be lower among children, due to common under-reporting of neonatal deaths [7] . Countries with at least one usable vital registration data point after 2010 represented 29% of the global population of adolescents and young people, but only 20% of deaths in this age group.

Our estimates of completeness are compared in Table S1 with those of two other sources:

• First, as part of the Global Burden of Disease Study, the Institute for Health Metrics and Evaluation (IHME) regularly estimates the completeness of death registration also based on DDMs. The GBD 2017 study combined estimates derived from the GGB, SEG and GGB>SEG methods in a statistical model predicting completeness in adults from that of the registration of under-five deaths 4 . Estimates were smoothed in time, as well as in space, and some information on completeness was also borrowed from neighbouring countries [8] . Completeness of the registration of under-five deaths was calculated from the ratio of the observed mortality rate from the vital registration to the final GBD estimate of child mortality. This method introduces a dependence on child mortality estimates and completeness of registration in children that we wish to avoid here. To our knowledge, the interrelationships between the completeness of child and adult registration have not been well studied. In some countries, under-registration of child deaths could occur, for example, because of difficulties in registering neonatal deaths, whereas registration of adults could be more systematic.

For example, formerly Soviet Republics used a definition of live birth that excluded a substantial proportion of early neonatal deaths [9] , and this will affect the completeness of child registration, but not that of deaths among older children and adults.

We also wish to avoid borrowing information from neighbouring countries, as situations can vary widely within regions. For instance, according to GBD 2017 estimates, the completeness of the registration of adult deaths in Honduras is only 28%, while registration is considered complete in Guatemala and Nicaragua 5 .

• The second set of completeness estimates comes from the United Nations Statistics Division (UNSD). The UNSD provides a recent estimate for most countries 6 , but no trend over time and a variety of sources are involved (including workshops with national representatives, reports from National Statistical Offices and estimates from the World Health Organization).

In most countries, our estimates based solely on the GGB>SEG method are consistent the GBD estimates (derived from more involved modelling) ( Figure S1 ). The mean absolute percentage error (MAPE) between our estimates and the GBD 2017 estimates is 5.4% and the concordance correlation coefficient is 0.79 (0.71-0.84) [10] . Comparison with UNSD estimates is more difficult because the UNSD values refer only to the most recent estimate available for each country, and for many countries, coverage is estimated in a range rather than a precise value. Several cases deserve further attention as estimates based on death distribution methods are quite higher than those provided by UNSD (e.g. Colombia, Guyana, Jordan, Trinidad and Tobago, Tunisia). Table S1 -Completeness of death registration from various sources for intercensal periods Note: GGB>SEG refers to our estimates for the population aged 5 and above, GBD to the completeness among adults in the GBD 2017, UNSD to the completeness for deaths of all ages. The GGB>SEG estimates refer to the intercensal period, GBD to the mid-point of the period, and reference years for UNSD are provided in the Table. All estimates of completeness derived from death distribution methods remain ques-tionable and should be compared with other studies, e.g. from capture-recapture studies [11] . demonstrate that uncertainty intervals around completeness estimates from DDM are quite large, roughly about 25% [5] . This is in part because death distribution methods are based on assumptions that are often poorly respected in practice. In some cases, we have preferred not to adjust upwards the deaths reported to vital statistics, either because of large migrations flows that could invalidate the method, or because published studies had investigated the completeness of death registration. These exceptions are as follows:

• Brazil: our estimate of completeness based on the GGB>SEG method was 0.96 for the period 1991-2000 and 0.94 for the period 2000-2010. Both the SEG and GGB methods taken separately are higher than 95% for this second period. Li and Gerland (2019) [12] use various death distribution methods and also obtain completeness estimates above 95% after 2000. Hence we did not adjust the vital registration estimates upwards for Brazil.

• Republic of Korea: estimates of completeness obtained from DDMs applied to the WHO Mortality Database are irregular and inconsistent with GBD estimates that place the completeness above 95% after 1990. Based on the study by Hill and colleagues (2005) [13] , we assumed that death registration was complete after 1990.

• Syria: It was not possible to use DDMs to estimate completeness as the VR data currently available only cover the years 1983, 1984 and 1998-2010, and the censuses were conducted in 1994 and 2004. VR estimates are currently used without adjustment after 1998.

• Turkey: We cannot estimate completeness after 2010 with DDMs as the last census was conducted in 2011.Özdemir and colleagues (2015) [14] provide estimates for the period 2009-2013 based on population data from TURKSTAT Address Based Population Registration System (ABPRS). They estimate that in the period 2009-2013, death registration was higher than 95% for both sexes. Prior to that, death registration was lower than 80%. In this study, we used the VR data only from 2010.

• Singapore, Bahrain, Sri Lanka: we assumed that the vital registration was complete in these three countries, even though death distribution methods suggest otherwise, most likely due to the perturbing role played by international migrations. In Sri Lanka, the civil war could have introduced disruptions in both the numerators and denominators, such as changes in coverage of the population enumerated in the censuses.

The calculation of the probability 10 q 15 was derived from a standard period abridged life • The same calculation was applied for 5 q 20 .

• The probability 10 q 15 was computed as 10 q 15 = 1000

The numbers of deaths and mid-year populations refer to both sexes. Mortality rates were recalculated by pooling deaths and population from successive years together such that the coefficient of variation (the ratio of the standard deviation to the mean) is less than 20%. The stochastic standard errors of the mortality rates were calculated using a Poisson approximation, based on the number of children turning 15 in each year, estimated from the 2019 World Population Prospects [15] . In this study, SSH were used to estimate the probability of dying in youth ( 10 q 15 ) for a period of 0-12 years prior to each survey. This period was divided in intervals of various length (6, 4, 3, 2, 1 years) depending on the coefficient of the variation of the estimates. The original data sets, which have one row for each woman aged 15-49, were reshaped into files containing one row for each sibling. Based on the dates of birth and death (imputed by DHS), each life course was split into different spells; a new spell was created at every birthday and every time the number of completed years preceding the survey changed. These spells were then aggregated to form a new dataset in which each row corresponds to a person-period (a unique combination of a given age group, sex, and time preceding the survey). For each person-period, the exposure time and the number of deaths were summed and weighted by the sample weights. Standard errors were obtained using a Jackknife variance estimation method [16] . Age-specific mortality rates were converted into probabilities 5 q 15 and 5 q 20 and these probabilities were chained together to form the probability 10 q 15 .

There is a debate in the literature about the presence of selection bias in sibling survival histories because larger sibships are over-represented in the sample, some sibships with no surviving members are not apparent in the data and the respondent is alive and reports on a group which she belongs to [17] . It has been demonstrated, however, that these selection biases cancel each other out when adult mortality is not associated with the number of adult siblings [18, 19] . We assumed here that this was the case and used the standard DHS approach to obtain mortality rates (excluding the respondent when computing person-years of exposure).

Sibling survival data have been evaluated in validation studies [20, 21] and compared to census or model-based estimates [22, 23] . Other studies have examined the consistency of mortality rates across surveys [24, 25] and the plausibility of reported sibship sizes against past fertility estimates [26, 27] . This literature suggests that sibling histories could lead to underestimates of adult mortality due to under-reporting of deaths, especially for reference periods located more than 10 to 15 years before the survey. Under-reporting of deaths could be amplified or attenuated by systematic misstatement of ages and the timing of deaths. The magnitude of these biases likely varies by age but few studies have assessed biases in the 15-24 age group specifically. In a validation study in a Health and Demographic Surveillance System (HDSS) in Senegal, Helleringer and colleagues (2014) observed that sibling histories led to an underestimation of mortality only from the age of 50 onwards [20] . In younger age groups, sibling survival estimates were consistent with mortality levels in the HDSS. Biases affecting sibling data may, therefore, be limited when estimating youth mortality but additional validation studies are needed to reach firmer conclusions. To account for the possibility of bias, we proceeded in two steps. First, we excluded surveys that presented implausible age patterns of mortality and second, we modelled reporting bias in the surveys that were retained.

First, we examined age patterns of mortality contained in surveys against a standard based on high-quality mortality data. This standard was built from the following sources:

• The 5x5 life tables from the Human Mortality Database [28] , representing the historical experience of 41 countries or areas with long statistical tradition of accurate vital registration;

• Life tables recalculated for periods of 2 or 3 years from publicly-available datasets from 48 Health and Demographic Surveillance Systems (HDSS) 10 ;

• Life tables based on vital registration data from Mauritius and sample vital registration data from India (obtained from the UN Population Division).

The relationship between 5 q 0 and 10 q 15 in these various life tables is presented in Figure  S2 below, using a log-log scale, alongside the relationship captured in historical model life tables (here the North and South models of Coale and Demeny life tables 11 ). From this database containing about 1300 data points, we used a log-quadratic regression model to predict, for a given level of under-five mortality, what would be the expected level of mortality between the ages of 15 and 25:

where y i is value of log( 10 q 15 ) for observation i, x i1 is the value of log( 5 q 0 ) for observation i and x i2 equals log( 5 q 0 ) 2 for observation i. The independent errors terms i are normally distributed. Prediction intervals around log( 10 q 15 ) for a new estimate of log( 5 q 0 ) were be obtained asŷ

whereŷ h is the fitted value when the predictor values are

is the t-multiplier and M SE + (se(ŷ h )) 2 is the standard error of the prediction.

Turning to sample surveys, we extracted the probability 5 q 0 from birth histories for the period 0-12 years prior to data collection and the probability 10 q 15 from sibling histories for the same period. We used Eq.2 to obtain the 95% intervals around the predicted value of log( 10 q 15 ) for the observed value of log( 5 q 0 ). We checked if these prediction intervals contained the observed value of log( 10 q 15 ). When they did not, the mortality pattern encompassed in survey estimates was deemed implausible, and all survey data points were excluded so that they did not inform the model fitting. This procedure assumes that birth histories are unbiased and that our standard adequately reflect the 10 q 15 -to-5 q 0 relationship in all countries for which we have sample surveys.

This quality requirement led to the exclusion of 9 DHS surveys (out of 148) 12 , 13 WHS surveys (out of 42) 13 and 3 MICS (out of 9) 14 . Judging from the proportions of surveys excluded from the model fit by type of series, sibling histories collected in the WHS and MICS surveys seem of lower quality than in DHS. Hence, biases in levels and trends were estimated from the B3 specifically for sibling histories from DHS on the one hand, and those from WHS and MICS on the other.

Finally, we modelled the bias in the remaining surveys, both in terms of the level and the trend, as a function of the retrospective period. This bias was estimated by contrasting the sibling-based estimates with vital registration data for overlapping time periods (see section 2.3 below). 12 The following DHS were excluded: Bangladesh 2001, Indonesia 1997 , Mali 1995 -1996 , Nigeria 1999 , Morocco 1992 and 2003 , Peru 2009 and 2012 . 13 We excluded the following WHS: Bosnia & Herzegovina, Chad, China, Comoros, Congo, Estonia, Malaysia, Mauritania, Pakistan, Senegal, Slovakia, Ukraine, Uruguay. The implausible trends in mortality in estimates extracted from the WHS surveys conducted in Philippines and Mexico, as well as the very wide uncertainty intervals around estimates in Georgia and Russian Federation led us to exclude these four surveys.

14 The following MICS were excluded: Bhutan 2010, Lao PDR 2012, Mauritania 2011.

Censuses often include questions on household deaths in the last 12 months, which can be used to calculate mortality estimates for the population aged 15-24. The calculation of the probability 10 q 15 was derived from a standard period abridged life table. As for sibling histories, we checked the 10 q 15 -to-5 q 0 relationship observed in the data against the predictions of the log-quadratic model above, and discarded estimates that fell outside of the 95% predictions intervals. This assessment led to the exclusion of about 29% of observations from censuses.

In most cases, we did not adjust for incompleteness of death reporting in surveys and censuses. However, in China, we used published estimates from census data from 1982 to 2000, adjusted for incompleteness of death reporting with the General Growth Balance method by Banister and Hill (2004) [29] . To be consistent with this study, we also evaluated the completeness of death reporting in the 2010 census with the General Growth Balance method [30] .

To obtain estimates of the probability 10 q 15 for all country-years based on all available data, we applied to the 15-24 age group the model used by the UN IGME to monitor trends in under-five mortality. This is a Bayesian penalized B-splines bias-reduction (B3) model. All details on B3 are available in Alkema and New (2014) [31] . We repeat below the main features of this model to highlight a few changes implemented for the 15-24 age group.

We let u i denote the observed probability 10 q 15 for the observation i in country c[i] and year t[i]:

where U c (t) denotes the true probability 10 q 15 and ε i > 0 is the error multiplier. On the natural log-scale, this corresponds to

where y i = log(u i ), f c (t) = log(U c (t)), and δ i = log(ε i ).

The regression spline model for f c (t) from Eq.(3) is given by:

where α c,k refers to splines coefficient k in country c and B c,k (t) the k-th spline, evaluated in year t, given by a third order B-spline [32, 33] . Equally spaced knots were used such that the resulting splines are non-zero for a total of 4 · I years, where I refers to the in-between knots interval length. We chose I = 2.5 years such that each spline is non-zero for 10 years.

In each country, the knots were placed such that the largest two splines B c,Kc−2 (t) and B c,Kc−1 (t) in the most recent observation year t = t nc have equal height while B c,Kc (t) is close to zero. When fitting the splines model from Eq.(4) to the observations, second-order differences in adjacent spline coefficients (∆ 2 α k = α k − 2α k−1 + α k−2 ) are penalized to guarantee smoothness of the resulting 10 q 15 trajectory and to project forward the recently observed rate of change past the most recent data. Let t c = (t c,1 , . . . , t c,nc ) refer to the vector of country-specific observation years with spline model estimate f c (t c ) = (f (t c,1 ), . . . , f (t c,nc )) .

The

. . , B c,Kc (t c )), α c = (α c,1 , . . . , α c,Kc ) , can be written as follows [34, 33] :

The first part in Eq. 

where variance σ 2 c determines the extent of smoothing; a smaller variance corresponds to smoother trajectories. In the limit when σ c decreases to zero, a linear fit for log( 10 q 15 ) is obtained.

The model is fit in the Bayesian framework. No information on levels or trends during the observation period is exchanged across countries when estimating the spline coefficients. Information exchanged across countries only concerns the variability of the (second order) difference in the spline coefficients through a multilevel model. The variance of e c,q = α c,q+2 -2α c,q+1 + α c,q is estimated hierarchically:

Vague prior distributions are used for the b c 's and the hyper parameters for the hierarchical model for the σ c 's [31] .

The error distribution for observations from complete VR or SVR is given by

where τ i /u i is the stochastic standard error. These errors are calculated using a Poisson approximation (using the numbers of adolescents turning 15 from the World Population Prospects 2019) and set to a minimum of 2.5%. If the Poisson sampling standard error cannot be calculated (e.g. for Sample Registration Systems), it is set to 10%.

For non-VR data, δ i , the error term on the log-scale, is specified as follows:

where E i is the mean bias, S i the scale parameter and X i determines the distribution for observation i.

For all non-VR data series with repeated observations, mean biases were modeled as a linear function of the retrospective period of the observation in the survey:

where β 0,s[i] +β 1,s[i] ·π i represents the bias in level and trend as a function of the retrospective period π i for observation i in data series s[i]. The retrospective period π i was centered at 5 years for this study. The bias in the level of the series, β 0,s is estimated with a multilevel model:

where d[s] refers to the source type of series s, based on data source (the source types with multiple observations per series are given by DHS Direct and Others Direct (MICS and WHS surveys)), and µ 0,d and γ 2 0,d represent source type-specific mean bias and betweenseries variance respectively. A similar approach is used to estimate the slope β 1,s :

where µ 1,d and γ 2 1,d represent the mean slope and the between-series variance for source type d. For single observations constructed from reported household deaths, and single observations obtained from reported life tables, we assume that

Scale parameter S i is modeled as a combination of sampling variance τ 2 i /u 2 i (based on sampling variance τ 2 i for 10 q 15 ) and non-sampling variance ω 2 d [s[i]] :

where source type d [s] for series s refers to a further breakdown of source types to distinguish between DHS, Other (MICS and WHS) and recent household deaths. Where the sampling variance for non-VR data is not reported, we assume a sampling standard error of 10%. Finally, the distribution for δ i is given by: 1) for Standard and Other DHS direct, t ν otherwise, with ν ∼ U (2, 30).

A t-distribution with ν degrees of freedom is used for observations that are not obtained from Standard or Other DHS, in accordance with the model used by UN IGME for underfive mortality. All model parameters in Eq.(8)- (12) were assigned vague prior distributions. In the UN IGME model for under-five mortality, an informative prior distribution was used for the mean bias µ 0,d for the DHS Direct series, but here we used vague prior distributions for all parameters. Figure S3 presents mean biases and 90% prediction intervals for "new" data points by type of data source. This visualization shows that for a "true" level of the probability 10 q 15 of 15 deaths before age 25 per 1000 adolescents at age 15, the mean bias will vary across data sources and with the length of the retrospective period (in the survey estimates). The prediction intervals based on uncertainty in the bias parameters only (dark colors) are large, indicating that there is substantial variability in biases across data series of the same source type. Mean biases tend to be negative in sample surveys and slightly positive for data on recent household deaths (represented in pink). The median of the predicted 10 q 15 for a retrospective period of 2 years (our first estimate from DHS) is 14.9 per thousand (12.4-17.3) . For a retrospective period of 10 years, it declines to 12.4 (10.3-14.6), corresponding to under-estimating the true level by about 17%. This is consistent with previous studies on sibling histories indicating that the quality of death reporting declines as the reference periods stretch back in the past [24, 25, 22] .

For extrapolations, we implemented a logarithmic pooling procedure to combine countryspecific posterior predictive distributions for changes in spline coefficients with a global posterior predictive distribution. This procedure was applied to modify the posterior predictive distributions for α c,k for k = K c , K c + 1, . . . , P c , where P c refers to the last spline in the projection period of interest. While α c,Kc was among the spline coefficients that were included in the observation period up to year t nc , it was included in the set of "projected" coefficients to be pooled because its estimate is based mainly on an extrapolation of past changes. The logarithmic pooling weight κ, which determines the extent of pooling, was set at 0.8 for this study, after examination of country plots for various weights and also based on validation outputs (see below). Further details on the logarithmic pooling procedure are provided in [31] .

A Markov Chain Monte Carlo (MCMC) algorithm was employed to sample from the posterior distribution of the parameters in the global and country-specific models with the use of the software JAGS. For the global run, 10 parallel chains with different starting points were run with a total of 75,000 iterations in each chain. Of these, the first 25,000 iterations in each chain were discarded as burn-in and every 20th iteration after was retained. The resulting chains contained 2,500 samples each. For the country-specific runs, we ran 10 chains with a total of 62,500 iterations in each chain. Of these, the first 25,000 iterations in each chain were discarded as burn-in and every 30th iteration after was retained. The resulting chains contained 1,250 samples each. A computationally cheaper model was implemented to allow for updates of countryspecific estimates with additional data without the need to re-run the global model. Noncountry-specific parameters were fixed at the posterior medians from the global model run. Both models resulted in very similar estimates.

In some countries, there was insufficient smoothing, resulting in unrealistic short-term changes for a subset of country-years without VR data. For these countries, instead of using country-specific smoothing determined by variance parameter σ 2 c from Eq.(7), we set σ c = exp(χ), whereχ refers to the posterior median of µ, which is referred to as the global smoothing level. This subset includes 1) countries with both VR and non-VR data, or both adjusted and non-adjusted VR data, 2) countries that have VR data with gaps more than 5 years in the data, as well as 3) small countries with less than 15,000 adolescents aged 15 in 2019 (based on the World Population Prospects 2019).

Adjustments were made to account for abrupt increases in mortality due to conflicts and/or disasters, which would otherwise not be present in the smoothed mortality curves obtained from the statistical model. Vital registration systems, survey programs and censuses might also miss some crisis-related deaths, due to disruptions in the notification of deaths, interruptions in data collection or because some households have been dissolved as a result of mass displacement.

WHO has been estimating conflict deaths for their Global Health Estimates (GHE) and has relied primarily on databases maintained by the Uppsala Conflict Data Program (UCDP) together with information from a number of country-specific data sources [35, 36] . Since 2014, the Armed Conflict Location and Event Data Project (ACLED) has been collecting and making available detailed disaggregated information on conflicts and political violence in over 100 countries in real-time [37, 38] . As of early 2020, ACLED included conflict deaths for 92 countries. ACLED includes all African countries from 1997 onwards. However, ACLED data has only become reasonably complete for several other regions from 2016 onwards and does not yet include countries in the Americas.

For updated estimates of conflict deaths for the period 1990-2019, ACLED data for conflict deaths to the end of 2019 were assessed for consistency with conflict deaths estimated using previous GHE methods. The latter were updated using the latest available data from UDCP for Battle-Related Deaths (version 19.1-1989-2018), Non-State Conflict Dataset (version 19. 1-1989-2018) , and One-sided Violence Dataset (version 19.1-1989-2018) [35] . Estimates for countries using other specific data sources [36] were also updated.

ACLED-based estimates are generally very consistent with the updated GHE estimates and were used to update years beyond 2006 for most regions. Additionally, for some highconflict countries such as Afghanistan, Syria, Iraq and Yemen where GHE estimates have in the past been based on country-specific data sources with somewhat ad-hoc methods, ACLED offers a single reasonably consistent alternative and was used to revise and update the conflict death time series.

Age-sex distributions for conflict deaths were applied to estimates of total conflict deaths by country-year based on available distributions of conflict deaths by age and sex for specific conflicts [36] .

Estimated deaths for major natural disasters up to and including 2019 were obtained from the EM-DAT/CRED International Disaster Database [39] . GHE age-sex distributions were used. These are based on a number of studies of earthquake deaths and tsunami deaths [36] .

Direct and indirect deaths caused by the Ebola epidemic in 2014-2016 in Guinea, Liberia and Sierra Leone were added based on WHO Global Health Estimates.

For inclusion of crisis deaths, the splines regression model is fitted to "extreme eventfree" observations (obtained as the observed values of the probability 10 q 15 minus the ex-treme events). The estimated additional mortality for the crisis event is then added. Uncertainty intervals for 10 q 15 for the crises-years are based on the uncertainty in crisis-free 10 q 15 . The following criteria were used to identify crises:

1. The crisis was isolated to a few years, 2. Crisis deaths among adolescents and young adults aged 15-24 were >10 % of non-crisis deaths in this age group, 3 . The crisis 10 q 15 was greater than 0.2 per 1,000, 4. The number of crisis deaths among youth was greater than 10.

In total, adjustments for crises were made in 61 countries, listed below.

Year ( When applied to under-five mortality, the B3 model also includes adjustments for biases introduced by HIV/AIDS, because survey data come from mothers. Among these mothers, some are HIV-positive and without treatment, they face excess mortality and risk of transmitting the virus to their children. Birth histories, therefore, underestimate under-five mortality in settings where treatment is not widely available, or for earlier periods before treatment became widespread. No adjustment is made here for mortality between 15 and 25 years of age, as the data are derived from information provided on sibling survival by adult respondents.

To assess the performance of the B3 model for 10 q 15 , we used out-of-sample validation. A training set was constructed by removing all data collected in or after 2013 (i.e. about 20% of observations were left out). All retrospective observations from a survey carried out in or after 2013 were left out of the training set, even if they referred to periods before 2013. We present below two sets of validation results: based on left-out observations (Tables S3 and  S4) , and based on the comparison between estimates obtained from the training and full data set (Table S5) . Results are presented for estimates without crises in high mortality countries, defined as countries with a probability 10 q 15 above 12 per 1000 youths aged 15 in 1990. We compare validation measures for 10 q 15 with those obtained by Alkema and New (2014) for U5MR 15 and those obtained when applying B3 to mortality in children aged 5-14 16 . Results are presented for two values of the pooling weight κ; 0.5 to compare across studies and 0.8 as it is the pooling weight retained for 10 q 15 . Table S3 presents the percentage of observations falling below and above the 90% uncertainty intervals based on the training set 17 . If the model is well-calibrated, around 5% of observations should fall below and above the 90% uncertainty intervals based on the training set. This is what we observed for the most recent period. There seems to be some asymmetry in coverage in the period before 2013, but this is based only on 26 highmortality countries only. 11% of observations fall below the lower bound obtained from the training set, corresponding to about 3 countries out of 26. If the model is well calibrated, the chance of having at least 3 countries outside the interval is 13.8% (if we take a random sample of size 26 with a probability of 0.05 of being outside). Hence this percentage should not be interpreted as a sign of poor performance.

Results reported in Table S4 refer to the mean and median relative error and absolute relative error. Errors are defined as i = u i −ũ i , whereũ i is the posterior median of the predictive distribution for a left-out observation u i based on the training set. The median or mean relative error (MRE) and the median or mean absolute relative error (MARE) are larger than those of under-five mortality for the period before 2012, but relatively similar for the most recent period. Table S5 presents the results based on a comparison between estimates based on the training and full data set. The error in the estimate based on the training sample is defined

where U c (t) refers to the posterior median estimate based on the training sample, and U c (t) refers to the 10 q 15 estimate obtained from the full data set for country c in year t. Relative error is defined as c,t / U c (t) · 100. Again, the validation measures are the mean and median relative error and absolute relative error, and the Table S3 -Validation results based on left-out observations: Median and SD of percentage of observations below and above 90% predictions intervals based on the training set, from 100 sets of left-out observations, in countries with data in both the training and test set and left-out observations in the period of interest. Table S4 -Validation results based on left-out observations: median or mean error (ME), median or mean relative error (MRE), median or mean absolute relative error (MARE) based on the training set. percentage of estimates obtained from the full dataset that are falling below or above the 90% uncertainty intervals obtained from the training set only. Overall, these results indicate that the model performance is not substantially different for 10 q 15 than when applied to the other age groups. 7 23.0 1.0 -6.6 14.4 0.9 5.6 3.7

In 37 countries with insufficient data sources, we estimated youth mortality based on the relationship observed between U5MR and the probability 10 q 15 in countries for which the B3 model was used. Our approach is similar to that used to exclude certain surveys on the basis of the plausibility of mortality age patterns (section 1.2.2), except that we can use the results of the B3 model rather than the standard constructed from vital registration and HDSS data. The B3 estimates offer more insights into the diversity of regional patterns than the standard mortality age pattern described above. We removed the quadratic term and fit a varying-slope and varying-intercept model to predict log( 10 q 15 ) from log( 5 q 0 ), allowing the slope of the log( 5 q 0 ) variable to vary by region [40] :

The fixed effect parameter estimates are presented in Table S6 and Figure S4 displays the model predictions for different regions (using the UNICEF classification), for the range of U5MR estimates observed in each region in the period 1990-2019. We combined the median value of U5MR estimates from UN IGME for each country-year between 1990 and 2019 with the fixed and random effects of this model to obtain a point estimate log( 10 q 15 ). We did not account for uncertainty quantified by σ 2 y or the uncertainty around the U5MR estimates to make these predictions.

For a given level of child mortality, there are significant regional differences in the expected levels of the 10 q 15 probability. For example, for a level of 50 per thousand, the predicted 10 q 15 probability will be 11 per thousand in Eastern Europe and Central Asia, compared to 16 in East Asia and the Pacific, and up to 24 in Eastern and Southern Africa. The estimates for data-poor countries are to be considered with caution. They will be revised in future iterations as the database develops, by making greater use of non-standard surveys for these countries (upon data availability). This modelling covers a very small share of the world population aged 15-24 since these 37 countries represent only 4.9% of the population in this age group 18 . 

The annual rate of reduction (in %) between two years, y 1 and y 2 (y 1 < y 2 ), is computed as:

3 List of data series Table S7 lists all countries included in the analysis, their region, the different data series included in the mortality database and indicates the type of data collection. Some data series were included in the database but excluded from the statistical model. The 'Inclusion' column identifies these series with a 0. Excluded series refer to country-years with vital registration data until 1989 (as these vital statistics were not used for modelling trends from 1990), incomplete vital registration data or surveys and censuses with implausible age patterns. This column is also set to '0' for countries for which trends in 10 q 15 were modelled based on trends in U5MR (because of insufficient data series). q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q This section briefly compares the 2020 IGME estimates with those that could be obtained by indexing model life tables solely on U5MR, and those of the World Population Prospects 2019 and the Global Burden of Disease 2019. The country-specific plots presented in the previous section also display the estimates from the WPP and GBD.

In countries where the B3 model was used, the median 10 q 15 -to-5 q 0 ratio was 0.56 in 2019, increasing from 0.35 in 1990. In 25 countries, this ratio had already exceeded 1 in 2019 (7 in 1990) . Figure S5 depicts the 10 q 15 -to-5 q 0 ratio estimated from the log-quadratic model based on high-quality vital registration and HDSS data and the ratio encompassed in two standard model life tables; the North and South models of Coale-Demeny life tables. While the B3 estimates conform well to the predictions of the log-quadratic model, they deviate substantially from the standard life tables. This suggests that youth mortality should be inferred from child mortality only as a last resort, and that caution should be exercised when using the traditional model life tables, especially in low-mortality countries.

As indicated in the main text, our estimates contrast with those of the World Population Prospects (WPP) 2019 or the Global Burden of Disease (GBD) 2019 Study. The difference between the latest GBD estimate of the probability 10 q 15 and our estimate was more than 20% in 95 countries in 2019. The difference was more than 20% in 74 countries when comparing our study with the WPP, and more than 20% in 100 countries when comparing the GBD with the WPP.

To summarize these deviations by region, Table S12 presents the numbers of deaths for the year 2019 in this study against the GBD and WPP estimates. We use here the GBD regional classification to present the uncertainty intervals around GBD values (these are 95% uncertainty intervals). Globally, our estimates fall between the two other series: the number of deaths among 15-24-year-olds in the GBD survey is 12% lower than the estimate from this study, and the WPP estimate is 13% higher than ours. In 6 regions, the absolute differences are greater than 10,000 annual deaths between at least two series: North Africa and Middle East, Southeast Asia, South Asia, Western Sub-Saharan Africa, Eastern Sub-Saharan Africa and Central Sub-Saharan Africa. In North Africa and the Middle East, the total number of deaths in the GBD and WPP are respectively 31% and 18% lower than ours. In South-East Asia, the IGME and WPP estimates are consistent and lower than the WPP, while in South Asia, our estimates are lower than the other two series. But the most striking discrepancies are in sub-Saharan Africa. In West Africa, the WPP estimates of deaths among 15-24-year-olds are almost three times higher than those of the GBD survey, and in Central Africa, almost twice as high. Differences between estimates of adult mortality between the ages of 15 and 59 and maternal mortality in Africa have already been much debated [41] , and this debate resonates here. Our estimates are consistently higher than the GBD values, while the consistency with WPP estimates varies across sub-regions. It is difficult to make sense of these variations, which are due both to adjustments introduced to account for biases in sibling histories and to differences in the methods for reconstructing mortality trends. Deviations in the population at risk of dying do not play a large role here (the GBD estimate of the population aged 15-24 in Sub-Saharan Africa in 2019 was 3% higher than the WPP estimate). Given the increasing concentration of deaths in Sub-Saharan Africa, more research is warranted to disentangle the main drivers of deviations between mortality estimates.

Countries included in this study are classified in geographical regions according to the UNICEF classification ( Figure S6 ). Aggregated estimates based on other regional classifications (SDG, WHO, etc.) are available at www.childmortality.org. 

It was formed in 2004 to share data on child mortality, harmonize estimates within the UN system, improve methods for child mortality estimation, report on progress towards child survival goals and enhance country capacity to produce timely and properly assessed estimates of child mortality. The organizations and individuals involved in generating the 2020 estimates of mortality in children, adolescents and youth are the following: • United Nations Children's Fund : Lucia Hug, Sinae Lee, Yang Liu, Anupam Mishra, David Sharrow, Danzhen You • World Health Organization: Bochen Cao

Simon Cousens (London School of Hygiene and Tropical Medicine), Trevor Croft (The Demographic and Health Surveys Program, ICF), Michel Guillot (University of Pennsylvania and French Institute for Demographic Studies

Tools for demographic estimation. Paris: International Union for the Scientific Study of Population

Estimating Census and Death Registration Completeness

Mortality estimation from registered deaths in less developed countries

Death distribution methods for estimating adult mortality: Sensitivity analysis with simulated data error

What Can We Conclude from Death Registration? Improved Methods for Evaluating Completeness

DDM: Death Registration Coverage Estimation

World Health Organization, Neonatal and Perinatal Mortality: Country, Regional and Global Estimates. World Health Organization

Global, regional, and national age-sex-specific mortality and life expectancy

Analytical methods to evaluate the completeness and quality of death registration: Current state of knowledge

A concordance correlation coefficient to evaluate reproducibility

Overview of the principles and international experiences in implementing record linkage mechanisms to assess completeness of death registration

Evaluating the completeness of death registration at old ages: A new method and its application to developed and developing countries

Unconventional approaches to mortality estimation

Reliable mortality statistics for turkey: Are we there yet?

World Population Prospects: The 2019 Revision

Child mortality estimation: appropriate time periods for child mortality estimates from full birth histories

Death by survey: estimating adult mortality without selection bias from sibling survival data

Adult Mortality from Sibling Survival Data: A Reappraisal of Selection Biases

Maximum likelihood estimation of the parameters of Coale's model nuptiality schedule from survey data

Reporting errors in survey data on adult mortality: results from a record linkage study in Senegal

Improving the quality of adult mortality data collected in demographic surveys: Validation study of a new siblings' survival questionnaire in niakhar, senegal

Adult Mortality in Africa

Age patterns and sex ratios of adult mortality in countries with high hiv prevalence

Measuring Adult Mortality Using Sibling Survival: A New Analytical Method and New Results for 44 Countries

Adult mortality in Sub-Saharan Africa: evidence from demographic and health survey

DHS Maternal Mortality Indicators: an assessment of data quality and implications for Data Use

Sibship Sizes and Family Sizes in Survey Data Used to Estimate Mortality

Data resource profile: The human mortality database (HMD)

Mortality in China

The generalized growth balance method

Global estimation of child mortality using a bayesian bspline bias-reduction model

Flexible smoothing with b -splines and penalties

Splines, knots, and penalties

Flexible smoothing with p-splines: a unified approach

Uppsala Conflict Data Program

WHO methods and data sources for life tables 1990-2016

Introducing ACLED: An armed conflict location and event dataset

Armed conflict location and event data project

EM-DAT: The International Disaster Database

Data analysis using regression and multilevel/Hierarchical models

A comparison of maternal mortality estimates from GBD 2013 and WHO

Young people aged 15 Note: the uncertainty intervals around the IGME estimates are 90% UIs, while the uncertainty intervals around the GBD estimates are 95% UIs.