Mortality, Fertility, and Economic Growth in Nigeria, 1990–2024: Time-Series Evidence and Growth Forecasts
Abstract:
Population health and demographic dynamics influence economic development, yet the distinct associations of mortality and fertility with economic growth in Nigeria remain unclear. This study examined the relationships between changes in crude death and birth rates and real gross domestic product (GDP) growth in Nigeria from 1990 to 2024 and evaluated alternative time-series forecasting models. Annual data from official statistical sources were analyzed using unit-root tests, regression models with Newey–West heteroskedasticity and autocorrelation-consistent standard errors, and autoregressive moving average (ARMA) models. Unit-root tests indicated that GDP growth was stationary, whereas crude death and birth rates were integrated of orders one and two, respectively. Accordingly, first-differenced mortality and second-differenced fertility indicators were incorporated into the regression models. Robustness was assessed using lagged demographic indicators and GDP growth, while forecasting accuracy was evaluated through hold-out and rolling-origin validation. A significant negative association was identified between changes in crude death rates (CDRs) and GDP growth (β = −17.46, p = 0.003), suggesting that greater mortality reductions were associated with higher contemporaneous economic growth. Mortality changes were jointly significant across contemporaneous and lagged specifications, although their association weakened after controlling for lagged GDP growth. Fertility changes exhibited only weak statistical evidence of an association (p = 0.096). Although the second-order MA model provided the best in-sample fit among models fitted to the undifferenced series, the first-order AR model was preferred for its parsimony, residual diagnostics, and forecasting accuracy. Annual GDP growth was projected at approximately 3.9–4.4% during 2025–2030, with the 2025 forecast (3.91%) closely matching reported growth (3.87%). These findings suggest that mortality reductions are associated with favorable short-term economic outcomes, whereas fertility-related benefits may emerge over longer periods. The results underscore the relevance of primary healthcare and family planning policies to Sustainable Development Goals 3 and 8, without implying causal relationships.
1. Introduction
Health is part of a nation’s human capital. Healthy workers are more productive and lose fewer working days, children who survive and stay healthy learn more, and households that do not have to pay for illness can save and invest (Bloom et al., 2004; Weil, 2007). Nigeria’s health system has struggled to secure these gains. Under the 2001 Abuja Declaration, African Union governments pledged to allocate at least 15% of their budgets to health, yet health received about 4.6% of Nigeria’s 2024 federal budget (Chukwunedu, 2024) and about 4.3% of the approved 2025 budget (Taiwo, 2025). Nigeria accounted for an estimated 28.7% of all maternal deaths worldwide in 2023, the highest share of any country (World Health Organization, 2025). The 2023–2024 Nigeria Demographic and Health Survey reported an under-five mortality rate of 110 per 1,000 live births and a total fertility rate of 4.8 births per woman (Anyanwu, 2025; Federal Ministry of Health and Social Welfare of Nigeria et al., 2025).
Two bodies of theory explain why population health outcomes may be related to economic growth, and they predict different channels and time horizons for mortality and fertility.
The first is health human capital theory. Grossman (1972) treats health as a stock of capital that yields “healthy time” for work and other activities. At the macroeconomic level, a healthier population supplies more effective labor per worker, attends school longer and learns more, and has a stronger incentive to invest in skills because working lives are longer (Bloom et al., 2004; Weil, 2007). A falling death rate is the most direct population-level sign of these improvements. Part of its effect on output should appear quickly, because lower adult mortality and morbidity raise the effective labor supply in the same or the following year. The theory does not predict an unambiguous effect, however. Acemoglu & Johnson (2007) argue that falling mortality also raises population growth, which can dilute capital per worker and weaken growth in income per capita, at least in the medium term. Because this study analyses aggregate real gross domestic product (GDP) growth rather than growth per capita, the labor-supply channel is expected to dominate. The first hypothesis is therefore that years of faster decline in the crude death rate (CDR) are associated with higher real GDP growth.
The second is demographic transition theory. As mortality falls, fertility falls with a lag, and the population’s age structure shifts towards working ages. This lowers the dependency ratio and can create a “demographic dividend” (Bloom & Williamson, 1998; Bloom et al., 2003; Lee, 2003). The fertility channel works slowly: a decline in births changes the size of the working-age population only after 15 to 20 years, and the dividend depends on whether the additional workers find productive employment (Canning et al., 2015; Gbehe et al., 2024). In the short run, a fall in births mainly reduces the number of young dependants. The second hypothesis is therefore that any short-run association between the crude birth rate (CBR) and growth is weak. If such an association exists, it is more likely to reflect the pace at which the fertility decline accelerates than the year-to-year level of births.
These two theories guide the choice of variables. The CDR (deaths per 1,000 population) and the CBR (births per 1,000 population) are the only population health and demographic indicators with a consistent annual series for Nigeria from 1990 to 2024, and they measure the two channels separately. Both are population outcomes, not direct measures of health system performance. The CDR is also affected by the population’s age structure, income, education, conflict and epidemics. The CBR depends on age structure as well as on fertility behavior and access to family planning. The CBR is also not the same as the total fertility rate: the CBR relates births to the whole population, whereas the total fertility rate measures the average number of births per woman over her reproductive life. Throughout this study, the two indicators are therefore interpreted as population health and demographic outcomes that partly reflect the reach of health services. They are not treated as measures of healthcare performance.
Table 1 compares recent studies directly relevant to the present research in terms of geographical context, indicators, and methods. Three patterns emerge. First, cross-country studies of sub-Saharan Africa that used life expectancy or mortality as the health indicator consistently found a positive association between better health and growth (Akinbode et al., 2021; Gebrihet et al., 2024). The size of the association was found to depend on institutional quality (Gebrihet et al., 2024) and differ across the growth distribution (Asongu & Odhiambo, 2023). However, these panel estimates are averages across countries and provide limited insight into individual countries’ experiences. Second, most Nigerian time-series studies have treated health as the outcome of health spending or macroeconomic conditions rather than as a determinant of growth. Their findings have also been mixed. Anagun et al. (2025) and Awoyemi et al. (2023) found that public health spending improved mortality and life expectancy, while Irikefe et al. (2025) found that the composition of spending mattered more than its volume and that some short-run effects were perverse. Adeosun et al. (2022) found a positive short-run association between income per capita and mortality but no significant long-run relationship. Third, demographic studies showed that fertility decline in sub-Saharan Africa began later and progressed more slowly than elsewhere, with the working-age population share yet to reach its peak in most countries (Dörflinger & Loichinger, 2024; Garenne, 2023). The growth benefit of falling fertility appears over the long run and depends on structural change (Gbehe et al., 2024).
The comparison identifies a specific gap. The evidence that health outcomes are associated with growth comes mainly from cross-country panels. The Nigerian literature has mostly examined the relationship in the opposite direction, from spending to health, and has not separated the fast mortality channel from the slow fertility channel. Nigerian studies have also rarely checked whether a regression is balanced in terms of the orders of integration of its variables, even though health indicators exhibit strong trends while economic growth does not. As a result, they risk spurious associations (Wooldridge, 2025). Finally, growth forecasts for Nigeria have seldom been compared across models on data that the models have not seen (Hyndman & Athanasopoulos, 2021).
Study | Setting and Period | Health or Demographic Indicator | Method | Main Finding |
|---|---|---|---|---|
Akinbode et al. (2021) | Sub-Saharan Africa, panel data | Health outcome indicators | System generalized method of moments | Better health outcomes raised growth. |
Asongu & Odhiambo (2023) | Sub-Saharan Africa, panel data | Health performance indicators | Quantile regression | Association differed across the growth distribution. |
Gebrihet et al. (2024) | Sub-Saharan Africa, panel data | Health outcomes | Panel models with institutional interactions | Institutional quality strengthened the health–growth link. |
Gbehe et al. (2024) | 26 sub-Saharan African countries, 1992–2019 | Age structure | Pooled mean group autoregressive (AR) distributed lag, fully modified ordinary least squares, dynamic ordinary least squares | Positive long-run effect of age structure via structural change |
Dörflinger & Loichinger (2024) | Global, including 49 sub-Saharan African countries | Fertility, working-age share | Demographic projections | Slowest fertility decline in sub-Saharan Africa; dividend not yet realized |
Adeosun et al. (2022) | Nigeria, 1990–2019 | Mortality rate | AR distributed lag | Short-run link with income; no significant long-run relationship |
Awoyemi et al. (2023) | Nigeria, time series data | Health outcome indicators | Time-series regression | Health spending improved health outcomes. |
Irikefe et al. (2025) | Nigeria, 1990–2023 | Life expectancy, infant mortality | AR distributed lag (Grossman framework) | Composition of spending mattered; some perverse short-run effects |
This study | Nigeria, 1990–2024 | Crude death rate, crude birth rate | Unit-root-consistent regression; Box–Jenkins forecasting | Findings are presented in Sections 3 and 4. |
This study addresses these gaps for Nigeria from 1990 to 2024. Its objectives are to (i) test the two hypotheses above in a regression that is consistent with the orders of integration of the series, and examine the stability of the results with lagged indicators and a dynamic control; and (ii) compare Box–Jenkins models with simple benchmarks in a hold-out sample and in a rolling-origin evaluation, and forecast real GDP growth for 2025–2030. The analysis relates directly to two Sustainable Development Goals (United Nations, 2015). Sustainable Development Goal 3 (good health and well-being) targets lower maternal and child mortality (Targets 3.1 and 3.2) and universal access to family planning (Target 3.7), which drive the CDR and CBR. Sustainable Development Goal 8 (decent work and economic growth) calls for sustained economic growth (Target 8.1). The study examines whether progress towards Sustainable Development Goal 3 is reflected in progress towards Sustainable Development Goal 8 in Nigeria.
2. Methodology
The study used annual national data for Nigeria for 1990–2024 (n = 35). The CDR (deaths per 1,000 people; indicator code: SP.DYN.CDRT.IN) and the CBR (births per 1,000 people; indicator code: SP.DYN.CBRT.IN) were taken from the World Bank’s World Development Indicators (World Bank, n.d.). These are annual modeled estimates from the United Nations World Population Prospects 2024 (United Nations DESA, 2024). Real GDP growth (annual percentage change; indicator code: NY.GDP.MKTP.KD.ZG) for 1990–2023 was taken from the World Development Indicators and was based on Nigeria’s national accounts at constant 2010 prices. For 2024, the growth rate of 3.40% published by the National Bureau of Statistics (NBS) on the same 2010 base was used (National Bureau of Statistics, 2025). The series was therefore constructed using data based on a consistent base year.
Nigeria rebased its national accounts from 2010 to 2019 prices in July 2025. The rebasing raised the level of GDP. Under the rebased series, the NBS reported real growth rates of 3.38% for 2024 and 3.87% for 2025 (National Bureau of Statistics, 2026). The 2024 figure was almost identical to the 2010-base figure used in this study. The 2025 value was not used to fit any model and served only as an external check of the 2025 forecasts, with the caveat that it was derived from the rebased series. The complete dataset, including GDP growth figures for 2020–2024 (−1.79%, 3.65%, 3.25%, 2.86%, and 3.40%, respectively), is provided in Table A1 in the Appendix and in the supplementary data file.
The models were compared based on the coefficient of determination (R²), the root mean squared error of the fitted values and the significance of the highest-order term. The exponential trend was not fitted to GDP growth, which was negative in several years.
Each series was tested in levels and, where necessary, in differences with the augmented Dickey–Fuller regression (Dickey & Fuller, 1979):
The trend term δt was included for series exhibiting a trend. The lag length k (maximum of 3) was chosen using the Akaike Information Criterion based on a common sample, and the regression was then re-estimated on the largest available sample. The null hypothesis of a unit root (H₀: γ = 0) was tested against the alternative hypothesis of stationarity (H₁: γ < 0). The Phillips–Perron test (Phillips & Perron, 1988) was used to correct the Dickey–Fuller statistic non-parametrically for serial correlation and heteroskedasticity, using the Newey–West long-run variance with a Bartlett kernel and a bandwidth of 3. Because the sample was small, p-values were obtained by simulating the distribution of each statistic under H₀ for the relevant sample size (20,000 replications).
A regression is balanced when all of its variables are integrated of the same order. Section 3.2 shows that GDP growth is integrated of order zero, I(0), the CDR is integrated of order one, I(1), and the CBR is integrated of order two, I(2). The main model therefore incorporates each health indicator into the difference that makes it stationary:
where, ΔCDRₜ = CDRₜ − CDRₜ₋₁ is the annual change in the CDR, and Δ²CBRₜ = ΔCBRₜ − ΔCBRₜ₋₁ is the change in the annual change in the CBR. The two transformations measure different things. The first difference of the CDR (ΔCDR) measures how fast mortality is falling in a given year. Therefore, a negative α₁ means that years of faster mortality decline are years of higher growth. The first difference of the CBR (ΔCBR) measures how fast the birth rate is falling, while the second difference of the CBR (Δ²CBR) measures whether that decline is accelerating (Δ²CBR < 0) or slowing (Δ²CBR > 0). A negative α₂ therefore means that years in which the fertility decline accelerates are years of higher growth. The specification using ΔCBR instead of Δ²CBR is reported for comparison only, because ΔCBR remains non-stationary.
Statistical inference was conducted using Newey–West standard errors robust to heteroskedasticity and autocorrelation (three lags) (Newey & West, 1987). Multicollinearity was assessed using the variance inflation factor, and the residuals using the Durbin–Watson, Ljung–Box and Jarque–Bera statistics. Because the regression contained only two health indicators, the CDR coefficient may partly reflect other economic changes. With 33 observations, a large model was considered inappropriate, and three modest checks were conducted. First, the model was re-estimated with lagged GDP growth as a single control variable to account for growth persistence and the smooth trends in the health indicators. Second, the model was re-estimated with one-year lags of ΔCDR and Δ²CBR to allow for the delayed effects predicted by the theory. Third, the model was re-estimated without the 2020 COVID-19 year. All estimates were interpreted as associations, not causal effects.
Real GDP growth was modeled with the autoregressive moving average (ARMA) family of models (Hyndman & Athanasopoulos, 2021; Shumway & Stoffer, 2025). The ARMA model of orders p and q, denoted as ARMA(p, q), is:
The model reduces to an AR model of order p, denoted as AR(p), when q = 0, and to a MA model of order q, denoted as MA(q), when p = 0. The autoregressive integrated moving average (ARIMA) model, denoted as ARIMA(p, d, q), applies the same structure to the d-th difference of the series:
where, L is the lag operator. Candidate model orders were identified using the autocorrelation and partial autocorrelation functions. All models with orders of up to 2 were estimated using exact Gaussian maximum likelihood (Kalman filter), with stationarity and invertibility imposed. In-sample fit was compared:
where, k is the number of estimated parameters, including σ². Residual adequacy was checked with the Ljung–Box statistic (Ljung & Box, 1978), with m = 10:
Residual normality was also assessed using the Jarque–Bera test (Jarque & Bera, 1987; Uyanto, 2022).
In-sample fit and out-of-sample accuracy address different aspects of model performance; therefore, both were examined. The Box–Jenkins models were compared with four benchmarks: the naive (no-change) forecast, the three-year MA, simple exponential smoothing, as shown in Eq. (8), and Holt’s linear trend method, as shown in Eq. (9):
The smoothing constants were chosen to minimize the sum of squared one-step-ahead errors (Hyndman & Athanasopoulos, 2021). Two evaluations were carried out. First, all methods were estimated using data from 1990 to 2019 and used to forecast 2020–2024 (a single five-year hold-out). Second, because a single origin of five years may produce a fragile ranking, a rolling-origin evaluation was also conducted (Tashman, 2000). Each method was re-estimated using an expanding window. For one-year-ahead forecasts, the estimation window ended in each year from 2004 to 2023, yielding 20 forecasts for 2005–2024. For three-year-ahead forecasts, the forecast origins ranged from 2004 to 2021, yielding 18 forecasts for 2007–2024. Accuracy was measured using:
Differences in one-step-ahead squared errors were tested with the Diebold–Mariano test with the small-sample correction proposed by Harvey et al. (1997). The forecasting model was chosen based on four criteria: parsimony, white-noise residuals, in-sample fit, and out-of-sample accuracy in the rolling-origin evaluation. The chosen model was re-estimated using the full sample and used to forecast 2025–2030. The 95% prediction intervals were calculated as the point forecast ± 1.96σ̂ₕ, where σ̂ₕ is the standard error of the h-step-ahead forecast.
3. Results
Real GDP growth averaged 4.22% (standard deviation = 3.85) and ranged from −2.04% in 1993 to 15.33% in 2002 (Table 2; Figure 1a). Growth was highest in 2000–2009 (average = 7.68%), slowed to 3.65% in 2010–2019 and averaged 2.27% in 2020–2024, which included the 2020 contraction (−1.79%). The CDR fell overall from 18.86 to 11.64 deaths per 1,000 people, a decline of 38.3%. The CBR changed little in the 1990s and then fell faster after 2010, reaching 32.49 births per 1,000 people in 2024 (Figure 1b and Figure1c). The two health indicators were highly correlated in levels (r = 0.87), reflecting their common downward trends, whereas their correlations with GDP growth were weak (r = 0.05 and 0.32, respectively).
Variable | Mean | Median | Standard Deviation | Minimum | Maximum | Skewness | Jarque–Bera Statistic (p-Value) |
|---|---|---|---|---|---|---|---|
Real gross domestic product growth (%) | 4.22 | 4.20 | 3.85 | −2.04 | 15.33 | 0.53 | 2.12 (0.346) |
Crude death rate (per 1,000 people) | 15.46 | 15.07 | 2.44 | 11.64 | 18.97 | 0.07 | 2.87 (0.238) |
Crude birth rate (per 1,000 people) | 40.80 | 42.90 | 4.04 | 32.49 | 44.51 | −0.99 | 6.23 (0.044) |
For the CDR, the linear trend fitted closely (19.715 − 0.2366t; R² = 0.985; root mean squared error = 0.291). The quadratic term was not significant (p = 0.408); therefore, the linear model was selected. The CDR fell by an average of 0.24 deaths per 1,000 people a year. For the CBR, the quadratic trend fitted much better than the linear and exponential forms (43.214 + 0.2879t − 0.01784t²; R² = 0.974; root mean squared error = 0.637). The negative quadratic term was statistically significant (p < 0.001), confirming an accelerating decline. Projected to 2030, the CDR would reach 10.0–10.7 deaths per 1,000 people and the CBR would be between 25.0 (quadratic model) and 32.7 (linear model) births per 1,000 people. However, the quadratic projection should be treated with caution. GDP growth showed no significant linear trend (slope = −0.037, p = 0.570). A quadratic trend was significant (p = 0.021) but explained only 16.4% of the variation and extrapolated to negative growth in 2025. Therefore, trend models were not used to forecast growth.
GDP growth was stationary in levels with a constant according to both the augmented Dickey–Fuller test (τ = −3.803, p = 0.012) and the Phillips–Perron test (Zτ = −3.984, p = 0.005), indicating that the series was integrated of order zero, I(0) (Table 3). With a constant and a trend included, the Phillips–Perron test also rejected a unit root (p = 0.023) and the augmented Dickey–Fuller test rejected it at the 10% level (p = 0.069). The CDR was non-stationary in levels but stationary after first differencing, indicating that it was integrated of order one, I(1). The CBR became stationary only after second differencing, indicating that it was integrated of order two, I(2). These findings were consistent with the smooth, accelerating decline observed in the modeled demographic series. However, the statistical power of these tests was limited due to 35 observations.
Series | Test Form | Augmented Dickey–Fuller τ (Lags) | Augmented Dickey–Fuller p-Value | Phillips–Perron Zτ | Phillips–Perron p-Value | Decision |
|---|---|---|---|---|---|---|
Gross domestic product (GDP) growth | Level, constant | −3.803 (0) | 0.012 | −3.984 | 0.005 | Stationary |
GDP growth | Level, constant + trend | −3.707 (0) | 0.069 | −3.901 | 0.023 | Stationary |
Crude death rate | Level, constant + trend | −2.945 (3) | 0.241 | −2.375 | 0.424 | Non-stationary |
ΔCDR | First difference, constant | −3.598 (0) | 0.020 | −3.604 | 0.011 | Stationary |
Crude birth rate | Level, constant + trend | −1.665 (1) | 0.781 | −0.771 | 0.960 | Non-stationary |
ΔCBR | First difference, constant | −1.811 (0) | 0.406 | −1.773 | 0.416 | Non-stationary |
ΔCBR | First difference, constant + trend | −2.360 (0) | 0.479 | −2.436 | 0.392 | Non-stationary |
Δ²CBR | Second difference, constant | −5.801 (0) | < 0.001 | −5.861 | < 0.001 | Stationary |
The main regression in Eq. (3), which is balanced, was significant overall (F(2, 30) = 5.62, p = 0.008; R² = 0.273) (Table 4). The coefficient of ΔCDR was negative and significant (−17.458; Newey–West standard error = 5.374; p = 0.003). A year in which the death rate fell by an additional 0.1 deaths per 1,000 people was associated with about 1.7 percentage points higher GDP growth. The coefficient of Δ²CBR was negative but significant only at the 10% level (−4.173; standard error = 2.425; p = 0.096), indicating that years in which the decline in the birth rate accelerated tended to have somewhat higher growth. The two regressors were not collinear (variance inflation factor = 1.04). The residuals were normal (Jarque–Bera p = 0.808) and showed no autocorrelation (Ljung–Box test, Q(10), p = 0.599; Durbin–Watson statistic = 1.52).
The birth-rate result depended on the transformation, as anticipated in Section 2.4. With the use of ΔCBR, as in the original specification, the birth rate coefficient was positive and not significant (1.736, p = 0.367). With the use of Δ²CBR, the coefficient was negative and weakly significant. The two results addressed different questions. The first examined whether growth was related to the speed of the fertility decline, and the second examined whether it was related to its acceleration. Because ΔCBR was non-stationary, only the second specification satisfied the stationarity requirements for regression with stationary real GDP growth. Nevertheless, the evidence for a short-run fertility association was weak, consistent with the second hypothesis.
The association between the CDR and economic growth was stable in direction and size, but was not fully independent of the persistence of growth (Table 4). The coefficient remained significant when ΔCDR was included alone (−15.58, p = 0.012) and when the 2020 COVID-19 year was excluded (−15.99, p = 0.012). With one-year lags of both indicators included, the current and lagged ΔCDR coefficients were individually imprecise because the two were correlated (r = 0.62). However, they were jointly highly significant (F = 12.40, p < 0.001), and their combined effect was −23.40 (p < 0.001). The lagged ΔCDR alone was also significant (−19.22, p = 0.001), consistent with a delayed effect. When lagged GDP growth was added as a control variable, the ΔCDR coefficient fell to −10.32 and was no longer significant (p = 0.118), while lagged growth was significant (0.410, p = 0.044). Part of the contemporaneous association therefore overlapped with the persistence of growth itself. The regression using the levels of the health indicators yielded statistically significant coefficients for both indicators (p < 0.01), but it was unbalanced and was shown only to illustrate the importance of establishing stationarity first.
Specification | ΔCDR Coefficient (p) | Birth-Rate Coefficient (p) | Other Terms (p) | R² | n |
|---|---|---|---|---|---|
Main model: ΔCDR, Δ²CBR | −17.458 (0.003) | −4.173 (0.096) | – | 0.273 | 33 |
ΔCDR only | −15.576 (0.012) | – | – | 0.201 | 33 |
ΔCDR, ΔCBR (original specification)ᵃ | −13.388 (0.013) | 1.736 (0.367) | – | 0.254 | 34 |
Main model excluding 2020 | −15.988 (0.012) | −3.947 (0.122) | – | 0.238 | 32 |
Main model + lagged GDP growth | −10.322 (0.118) | −1.639 (0.631) | GDPₜ₋₁: 0.410 (0.044) | 0.390 | 33 |
Main model + lagged indicators | −11.882 (0.224) | −3.799 (0.135) | ΔCDRₜ₋₁: −11.520 (0.190); Δ²CBRₜ₋₁: −3.750 (0.181) | 0.394 | 32 |
Lagged indicators only | – | – | ΔCDRₜ₋₁: −19.219 (0.001); Δ²CBRₜ₋₁: −3.555 (0.227) | 0.303 | 32 |
Levels: CDR, CBR (unbalanced)ᵇ | −1.506 (0.001)ᶜ | 1.097 (< 0.001)ᶜ | – | 0.323 | 35 |
The autocorrelation function of GDP growth was significant at lags 1 and 2 (0.441 and 0.413, respectively, against bounds of ±0.331) and then gradually declined. The partial autocorrelation function was significant only at lag 1 (0.441), suggesting a low-order AR process. Among models for the undifferenced series, the MA(2) model had the lowest Akaike information criterion (188.45) and Bayesian information criterion (BIC) (194.67), but the AR(1) model had almost the same BIC (194.93) and required one fewer parameter (Table 5). The additional terms of the AR(2) and the ARMA(1,1) model were not significant at the 5% level, and their residuals failed the normality test (p = 0.036 and p = 0.039). The AR(1) estimates were μ̂ = 4.393 (standard error = 1.073) and φ̂₁ = 0.483 (standard error = 0.155; p = 0.002). Therefore, about 48% of a deviation of growth from its long-run mean of 4.39% carried over to the following year. The residuals were white noise (Ljung–Box test, p = 0.217) and normal (Jarque–Bera test, p = 0.115). The ARIMA(1,1,1) model was fitted to the differenced series. Its Akaike information criterion value was not directly comparable with those of the other models, and differencing a stationary series was not warranted.
Model | Parameters | Log-Likelihood | Akaike Information Criterion | Bayesian Information Criterion | Ljung–Box p-Value | Jarque–Bera p-Value |
|---|---|---|---|---|---|---|
AR(1) | 2 | −92.134 | 190.27 | 194.93 | 0.217 | 0.115 |
AR(2) | 3 | −91.093 | 190.18 | 196.41 | 0.627 | 0.036 |
MA(1) | 2 | −93.956 | 193.91 | 198.58 | 0.069 | 0.154 |
MA(2) | 3 | −90.224 | 188.45 | 194.67 | 0.919 | 0.413 |
ARMA(1,1) | 3 | −91.632 | 191.26 | 197.48 | 0.374 | 0.039 |
ARMA(2,1) | 4 | −89.655 | 189.31 | 197.09 | 0.631 | 0.373 |
ARMA(1,2) | 4 | −90.177 | 190.35 | 198.13 | 0.819 | 0.323 |
ARIMA(1,1,1)ᵃ | 2 | −90.315 | 186.63 | 191.21 | 0.750 | 0.786 |
In the single 2020–2024 hold-out evaluation period (Table 6), every method over-predicted the 2020 contraction. The ARIMA(1,1,1) model and the naive forecasting method achieved the lowest root mean squared errors of 2.00 and 2.05, respectively, whereas the ARMA(1,1) model achieved the lowest mean absolute error of 1.68. Excluding 2020, the ARMA(1,1) model was most accurate (root mean squared error = 0.97), followed by the naive forecasting method (1.12) and the AR(1) model (1.22). With five observations obtained from a single origin, these rankings should be interpreted with caution.
Method | 2020 | 2021 | 2022 | 2023 | 2024 | Root Mean Squared Error | Mean Absolute Error | Root Mean Squared Error Excluding 2020 |
|---|---|---|---|---|---|---|---|---|
Actual | −1.79 | 3.65 | 3.25 | 2.86 | 3.40 | |||
Naive forecast | 2.21 | 2.21 | 2.21 | 2.21 | 2.21 | 2.049 | 1.664 | 1.117 |
Three-year MA | 1.65 | 1.65 | 1.65 | 1.65 | 1.65 | 2.142 | 2.002 | 1.668 |
Simple exponential smoothing | 2.00 | 2.00 | 2.00 | 2.00 | 2.00 | 2.066 | 1.789 | 1.320 |
Holt | 2.81 | 3.52 | 4.23 | 4.94 | 5.65 | 2.508 | 2.006 | 1.606 |
AR(1) | 3.49 | 4.11 | 4.41 | 4.56 | 4.63 | 2.601 | 1.965 | 1.218 |
MA(1) | 4.08 | 4.61 | 4.61 | 4.61 | 4.61 | 2.892 | 2.233 | 1.355 |
MA(2) | 3.40 | 2.81 | 4.59 | 4.59 | 4.59 | 2.602 | 2.056 | 1.311 |
ARMA(1,1) | 3.25 | 3.75 | 4.08 | 4.29 | 4.42 | 2.416 | 1.684 | 0.971 |
ARIMA(1,1,1) | 1.87 | 2.18 | 1.89 | 2.16 | 1.91 | 2.003 | 1.734 | 1.295 |
The rolling-origin evaluation provided a more reliable comparison (Table 7). For one-year-ahead forecasts over 2005–2024, the ARMA(1,1) model and the AR(1) model achieved the lowest root mean squared errors of 2.21 and 2.23, respectively, followed by simple exponential smoothing (2.28) and the naive forecasting method (2.34). The Diebold–Mariano test revealed no significant difference between the AR(1) model and the ARMA(1,1) model (p = 0.671), simple exponential smoothing (p = 0.839) or the naive forecasting method (p = 0.771). The MA(2) model, which had the best in-sample fit, was the least accurate Box–Jenkins model (root mean squared error = 3.10). Its forecasting accuracy was significantly lower than that of the AR(1) model (p = 0.007), a clear case of better in-sample fit not translating into better forecasts. For three-year-ahead forecasts, the ARMA(1,1) model, simple exponential smoothing, and the three-year MA method achieved the lowest prediction errors, with root mean squared errors of 2.86, 2.92, and 2.93, respectively, with the AR(1) model performing comparably well (root mean squared error = 3.00).
Method | One-Step-Ahead Root Mean Squared Error | One-Step-Ahead Mean Absolute Error | One-Step-Ahead Root Mean Squared Error Excluding 2020 | Three-Step-Ahead Root Mean Squared Error |
|---|---|---|---|---|
Naive forecast | 2.341 | 1.715 | 2.219 | 3.177 |
Three-year MA | 2.532 | 2.022 | 2.475 | 2.927 |
Simple exponential smoothing | 2.282 | 1.713 | 2.174 | 2.923 |
Holt | 2.909 | 2.342 | 2.792 | 5.519 |
AR(1) | 2.231 | 1.677 | 1.942 | 3.000 |
MA(1) | 2.482 | 1.975 | 2.162 | 3.116 |
MA(2) | 3.099 | 2.486 | 2.948 | 3.103 |
ARMA(1,1) | 2.209 | 1.587 | 1.949 | 2.859 |
ARIMA(1,1,1) | 2.614 | 2.175 | 2.547 | 3.032 |
Taking all four criteria together, the AR(1) model was retained as the main forecasting model. It was the most parsimonious model that fit essentially as well as the best in-sample model, with a difference in the BIC (ΔBIC) of 0.26 relative to the MA(2) model. Its residuals were white noise and normal, unlike those of the ARMA(1,1) model. Its one-year-ahead accuracy in the rolling-origin evaluation was statistically indistinguishable from the best method. The ARMA(1,1) model forecast slightly better in some evaluations, but the gain was small and not significant, while requiring an additional non-significant parameter. Therefore, its forecasts were reported alongside those of the AR(1) model as a check.
When re-estimated using the full sample, the AR(1) model forecast growth of 3.91% in 2025, rising slowly to 4.38% by 2030 as growth returned to its long-run mean (Table 8; Figure 2). The 2025 forecast differed by only 0.04 percentage points from the growth rate of 3.87% reported by the NBS using the rebased series. The ARMA(1,1) model forecast growth of 3.76% in 2025, whereas the naive forecasting method forecast 3.40%. The 95% prediction interval was wide (−3.12% to 11.88% by 2030), reflecting the volatility of economic growth in Nigeria.
Year | AR(1) | AR(1) 95% Prediction Interval | ARMA(1,1) | Naive Forecast | Holt |
|---|---|---|---|---|---|
2025 | 3.91 | −2.66 to 10.48 | 3.76 | 3.40 | 3.78 |
2026 | 4.16 | −3.14 to 11.46 | 3.96 | 3.40 | 4.09 |
2027 | 4.28 | −3.18 to 11.74 | 4.10 | 3.40 | 4.39 |
2028 | 4.34 | −3.15 to 11.83 | 4.19 | 3.40 | 4.70 |
2029 | 4.37 | −3.14 to 11.87 | 4.26 | 3.40 | 5.01 |
2030 | 4.38 | −3.12 to 11.88 | 4.30 | 3.40 | 5.32 |
4. Discussion
Nigeria’s death rate fell overall over 1990–2024, and its birth rate began to fall clearly only after 2010. This is the pattern of a country in the middle of the demographic transition, in which fertility decline follows mortality decline after a long lag (Lee, 2003; United Nations DESA, 2024). Both rates remain high, however. A CDR of 11.6 deaths per 1,000 people in a population as young as Nigeria’s, together with very high maternal and child mortality (Federal Ministry of Health and Social Welfare of Nigeria et al., 2025; World Health Organization, 2025), indicates that the population health gains still to be made are large.
The results support the first hypothesis. Years of faster mortality decline were associated with significantly higher GDP growth. The association remained statistically significant without the COVID-19 year, and the current and lagged mortality changes were jointly significant, with a combined coefficient of about −23. This agrees with the health-related human capital view (Bloom et al., 2004; Weil, 2007) and with evidence from panel studies in sub-Saharan Africa that better health outcomes are associated with higher growth (Akinbode et al., 2021; Gebrihet et al., 2024). The present study extends that evidence to a single-country time series while accounting for the orders of integration of the variables. The association weakened, however, when lagged growth was controlled for. Part of the observed relationship reflects periods in which faster mortality decline and stronger growth occurred together, such as the 2000s, rather than a separate effect of mortality. This finding, together with evidence from Nigeria that income and healthcare spending may also influence health outcomes (Adeosun et al., 2022; Anagun et al., 2025; Awoyemi et al., 2023), means that the estimates should be interpreted as associations. The health–growth relationship is likely to be bidirectional.
The results also support the second hypothesis. Changes in the CBR showed, at most, a weak short-run association with economic growth, primarily through an accelerated decline in the birth rate. Its sign changed between the ΔCBR and Δ²CBR specifications, which measure different aspects of the fertility transition. This finding is consistent with demographic transition theory. The economic benefit of lower fertility comes through a larger working-age share, which takes decades to build. In sub-Saharan Africa, where the fertility decline has been late and slow, this benefit has largely not yet been realized (Dörflinger & Loichinger, 2024; Garenne, 2023). Whether Nigeria gains a demographic dividend will depend on whether its growing labor force finds productive employment (Gbehe et al., 2024).
The forecasting results show why model fit and forecast accuracy must be distinguished. The MA(2) model provided the best in-sample fit but performed worst in the rolling-origin forecast evaluation. The simple AR(1) model was among the most accurate and is easy to interpret: economic growth reverts to a long-run mean of about 4.4% with moderate persistence. Its 2025 forecast (3.91%) was close to the reported value (3.87%), but the prediction intervals were wide, and no univariate model could have anticipated the 2020 contraction. These findings have specific implications for the Sustainable Development Goals. For Sustainable Development Goal 3, Nigeria’s under-five mortality of 110 deaths per 1,000 live births and its share of global maternal deaths show that it is far from Targets 3.1 and 3.2. The association between mortality decline and economic growth suggests that progress on these targets is compatible with, and may support, progress towards Sustainable Development Goal 8. For Sustainable Development Goal 8, the forecast aggregate growth of about 4% a year is modest for a population growing by more than 2% a year (United Nations DESA, 2024). It implies slow growth in income per person. Accelerating the fertility decline through family planning (Target 3.7) could reduce the number of young dependants over time, but the results suggest that its growth benefits should not be expected in the short run.
Four policy implications can be drawn from these findings and Nigeria’s circumstances. First, because the association between mortality decline and economic growth was the most robust finding, investment in mortality-reducing primary healthcare, immunization, safe delivery and malaria control is supported on economic as well as health grounds. That investment should move health financing towards the Abuja target. Second, because evidence from Nigeria shows that the composition of health spending matters (Irikefe et al., 2025), additional funds should be directed towards services with proven effects on mortality. Third, family planning and girls’ education should be treated as long-run investments in the demographic transition, whose growth returns depend on job creation. Fourth, growth projections used for planning should be accompanied by their wide uncertainty and checked against simple benchmarks.
The study has limitations. The CDR and CBR are population outcomes, not direct measures of health-system performance. They are influenced by age structure, income, education and conflict, and they are modeled estimates interpolated between censuses and surveys (United Nations DESA, 2024). They are therefore smoother than measured rates. With 35 annual observations, model complexity and the power of unit-root tests are limited. The regression identifies associations, not causal effects. GDP growth for 2020–2024 may be revised again when the rebased national accounts are incorporated into international databases. Future work should use direct measures of the health system, such as health expenditure per capita, health workers per 10,000 people and immunization coverage. It should also use outcome measures such as life expectancy and under-five mortality, and vector AR or AR distributed lag models with causality tests.
5. Conclusions
Over 1990–2024, faster declines in Nigeria’s CDR were associated with higher real GDP growth, both in the same year and with a one-year lag. This association was partly explained by the persistence of economic growth. Changes in the CBR showed, at most, a weak short-run association through the acceleration of the fertility decline. This is consistent with demographic transition theory, which places the growth benefits of falling fertility in the long run. Real GDP growth was found to be stationary with moderate persistence. An AR(1) model, chosen for its parsimony, white-noise residuals and rolling-origin accuracy rather than in-sample fit alone, forecast growth of about 3.9–4.4% a year for 2025–2030 in the absence of major shocks. The findings suggest that progress towards Sustainable Development Goal 3, especially through reductions in mortality, is associated with progress towards Sustainable Development Goal 8. These findings support continued investment in mortality-reducing primary healthcare and family planning, with realistic expectations regarding the timing of the economic returns.
Conceptualization, B.A.G.; methodology, B.A.G.; formal analysis, G.N.P. and B.A.G.; data curation, G.N.P.; validation, B.A.G. and M.U.B.; writing—original draft preparation, G.N.P.; writing—review and editing, B.A.G. and M.U.B.; supervision, M.U.B. All authors have read and agreed to the published version of the manuscript.
The data supporting the results are publicly available from the World Bank World Development Indicators (indicators NY.GDP.MKTP.KD.ZG, SP.DYN.CDRT.IN and SP.DYN.CBRT.IN) and the NBS GDP reports. The extract used in the analysis is listed in Table A1 and provided as supplementary material.
The authors declare no conflicts of interest.
Declaration on the Use of Generative AI and AI-assisted Technologies
During the preparation of this manuscript, the authors used Claude (Anthropic), an AI assistant, to support the re-analysis code, language editing and formatting. The authors checked all data, analyses and references and take full responsibility for the content of the work.
Table A1. Real GDP growth, crude death rate, and crude birth rate in Nigeria, 1990–2024
Year | GDP Growth (%) | Crude Death Rate (per 1,000) | Crude Birth Rate (per 1,000) |
|---|---|---|---|
1990 | 11.78 | 18.86 | 44.51 |
1991 | 0.36 | 18.97 | 44.32 |
1992 | 4.63 | 18.85 | 44.21 |
1993 | −2.04 | 18.68 | 44.11 |
1994 | −1.81 | 18.65 | 43.99 |
1995 | −0.07 | 18.54 | 44.08 |
1996 | 4.20 | 18.51 | 43.85 |
1997 | 2.94 | 18.36 | 43.51 |
1998 | 2.58 | 18.15 | 43.22 |
1999 | 0.58 | 17.80 | 43.45 |
2000 | 5.02 | 17.44 | 43.89 |
2001 | 5.92 | 17.15 | 43.93 |
2002 | 15.33 | 16.80 | 43.82 |
2003 | 7.35 | 16.37 | 43.71 |
2004 | 9.25 | 15.99 | 43.47 |
2005 | 6.44 | 15.69 | 43.37 |
2006 | 6.06 | 15.43 | 43.17 |
2007 | 6.59 | 15.07 | 42.90 |
2008 | 6.76 | 14.80 | 42.80 |
2009 | 8.04 | 14.55 | 42.50 |
2010 | 8.01 | 14.36 | 42.14 |
2011 | 5.31 | 14.20 | 41.79 |
2012 | 4.23 | 14.02 | 41.12 |
2013 | 6.67 | 13.86 | 40.44 |
2014 | 6.31 | 13.74 | 39.89 |
2015 | 2.65 | 13.66 | 38.95 |
2016 | −1.62 | 13.41 | 37.77 |
2017 | 0.81 | 13.19 | 36.69 |
2018 | 1.92 | 12.95 | 35.69 |
2019 | 2.21 | 12.68 | 34.69 |
2020 | −1.79 | 12.59 | 33.73 |
2021 | 3.65 | 12.31 | 33.55 |
2022 | 3.25 | 11.95 | 33.19 |
2023 | 2.86 | 11.74 | 32.95 |
2024 | 3.40 | 11.64 | 32.49 |
Note: GDP = gross domestic product. GDP growth data were obtained from the World Bank (n.d.) for 1990–2023 (at constant 2010 prices) and the National Bureau of Statistics (2025) for 2024. Crude death rate and crude birth rate data were obtained from the World Bank (n.d.), based on the United Nations DESA (2024).
Supplementary Information
The online version contains supplementary material available at: https://doi.org/10.56578/ocs050205.
