Data-Driven Insights into Fertility Trends: An Explainable AI Approach to Forecasting and Policy Implications

Abstract

Declining fertility rates in the United States present complex challenges for healthcare planning, economic stability, and reproductive justice. While prior demographic research has documented these shifts, relatively few studies have employed interpretable machine learning methods that combine forecasting accuracy with transparency. This study examines fertility dynamics in California and Texas from 1973 to 2020, applying artificial intelligence (AI) models to project future birth totals and evaluate reproductive health indicators. Using data obtained from the Open Science Framework, the Prophet time-series model was applied to forecast annual births through 2030 and benchmarked against a linear regression baseline. Prophet consistently outperformed regression, yielding substantially lower error in both states (California: RMSE = 6,231.41, MAPE = 0.83%; Texas: RMSE = 8,625.96, MAPE = 1.84%). To enhance interpretability, XGBoost regression combined with SHapley Additive exPlanations (SHAP) quantified the relative influence of predictors. Miscarriage totals, abortion access, and state-level variation emerged as the most influential drivers of fertility outcomes. Forecasts for both states suggest continued long-term declines in births, punctuated by short-term oscillations in the late 2020s, which may reflect the influence of policy or economic volatility. These results underscore the importance of combining forecasting accuracy with interpretability. By integrating Prophet with SHAP, this study provides transparent, data-driven insights into the demographic and policy factors shaping reproductive outcomes. The findings demonstrate the potential of explainable AI to inform healthcare planning, guide reproductive health policy, and support equitable responses to demographic change.

Keywords: fertility forecasting, reproductive health, time-series modeling, SHAP (SHapley Additive exPlanations), Prophet, XGBoost, Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Linear regression

Introduction

Fertility rates¹ in the United States have been steadily declining for several decades, raising urgent questions about demographic sustainability, economic growth, and reproductive justice. Lower fertility rates¹ influence multiple dimensions of society, including workforce participation, healthcare system planning, and population structure. As policymakers and healthcare providers grapple with these long-term shifts, there is growing recognition of the need for more precise forecasting tools² that can illuminate both the direction of fertility trends and the factors driving them.

Prior research has identified several contributors to fertility decline, including contraceptive use, abortion access, socioeconomic conditions, and cultural shifts such as delayed childbearing. However, most of this work has relied on descriptive statistics or traditional forecasting models² that often lack the transparency needed to interpret underlying drivers. Furthermore, many studies have not fully connected observed fertility patterns to the broader policy landscape, including watershed legal decisions like Roe v. Wade (1973)³ and Planned Parenthood v. Casey (1992)⁴, or more recent changes in reproductive healthcare access.

Artificial intelligence (AI) and machine learning provide new opportunities to address these gaps. Unlike traditional methods, interpretable AI models not only enhance predictive precision but also reveal the relative influence of social, medical, and policy factors on fertility outcomes. SHapley Additive exPlanations (SHAP)⁵, for example, allow researchers and policymakers to examine which predictors exert the strongest influence on model outputs, offering insights that are both statistically grounded and actionable.

This study applies time-series forecasting² and interpretable AI to fertility data from California and Texas spanning nearly five decades (1973–2020). These two states provide instructive contrasts: California with its comparatively broad reproductive healthcare access, and Texas with more restrictive policies and demographic volatility. The study has three main objectives:

1. Forecast annual birth totals in California and Texas from 2021–2030 using the Prophet⁶ model.
2. Identify and interpret the most influential reproductive health predictors using SHAP⁵ applied to XGBoost⁷ regression.
3. Evaluate model accuracy against a linear regression baseline using RMSE⁸ and MAPE⁹.

By integrating forecasting accuracy² with interpretability, this research advances the study of fertility decline beyond descriptive analysis. It contributes to demographic and public health scholarship by providing transparent, data-driven insights into how healthcare access, policy environments, and demographic change collectively shape reproductive outcomes.

Methodology

This study employed an observational, cross-sectional, and retrospective research design. Rather than intervening or manipulating variables, the study analyzed historical fertility-related data to uncover patterns and generate forecasts. This design was chosen to capture long-term demographic trends and evaluate predictive modeling performance without the constraints of experimental research.

No human participants were directly involved. The sample consisted of state-level reproductive health statistics for California and Texas spanning 1973–2020. Data included annual totals of births, abortions, pregnancies, and miscarriages. Because the dataset was aggregated, anonymized, and publicly available, there were no individual-level demographic identifiers such as age, race, or income.

Data were obtained from the Open Science Framework (OSF)¹⁰ U.S. fertility measures dataset (1973–2020). This repository compiles official state-level statistics from government sources such as the CDC¹¹, NCHS (National Center for Health Statistics)¹², and state public health departments¹³, ensuring historical consistency and reproducibility. The dataset is freely available for download in CSV format from the OSF¹⁰ platform. No new surveys, experiments, or recruitment were conducted; the study relied entirely on archival data.

Variables and Measurements:

• Forecast Variable: Annual total births (1973–2020).
• Predictor Variables: Abortion totals, miscarriage totals, pregnancy rates, abortion rates, and state identifiers.
• Measurement Tools: Prophet⁶ (time-series forecasting²), XGBoost⁷ (machine learning regression), and SHAP⁵ (interpretability and feature importance).
• Evaluation Metrics: Root Mean Squared Error (RMSE)⁸ and Mean Absolute Percentage Error (MAPE)⁹ were used to assess predictive accuracy.

Software and Environment:

Analyses were performed in Python (version 3.9). All scripts were run in Jupyter Notebook. Required open-source libraries included:

• pandas (for data cleaning and manipulation)
• numpy (for mathematical operations)
• prophet⁶ (time-series forecasting², installed via pip install prophet)
• scikit-learn (for model validation and baseline regression)
• xgboost⁷ (machine learning regression, installed via pip install xgboost)
• shap⁵ (interpretability, installed via pip install shap)
• matplotlib and seaborn (for data visualization).

Procedure

1. Data Download and Preparation

• a) Download fertility data CSV files from OSF¹⁰ .
• b) Import data into Python using pandas.
• c) Filter the dataset to include only California and Texas (1973–2020).
• d) Ensure consistency of column names and convert all date fields into datetime format.
• e) Handle missing values by forward-filling (ffill) or interpolation where appropriate.
• f) Create separate time-series objects for births, abortions, miscarriages, and pregnancies.

2. Forecasting with Prophet⁶

• a) Prophet requires two columns: ds (date) and y (value). Data were reformatted accordingly.
• b) Separate Prophet models were trained for each variable (births, abortions, miscarriages, pregnancies).
• c) Forecast horizons were set to extend through 2030.
• d) Prophet decomposed the series into trend and seasonal components, producing forward projections with confidence intervals.

4. Interpretability with SHAP⁵

• a) SHAP values were calculated for each predictor.
• b) Feature importance plots and dependence plots were generated to interpret how predictors influenced birth totals.
• c) Results quantified which reproductive health indicators most strongly contributed to the model’s predictions.

5. Validation Against Baseline Models

• a) A linear regression model was trained as a baseline.
• b) RMSE⁸ and MAPE⁹ were calculated for Prophet⁶ , XGBoost⁷ , and baseline models.
• c) Model performance comparisons ensured that improvements were attributable to advanced methods rather than chance.

Analysis combined time-series forecasting² and explainable AI. Prophet⁶ decomposed time-series into trend and seasonal components, enabling forward projections. XGBoost⁷ modeled non-linear relationships among predictors, while SHAP⁵ quantified each feature’s contribution to predictions. Visualization of forecasts and SHAP⁵ values ensured interpretability.

Because the dataset was public, anonymized, and aggregated at the state level, no institutional review board (IRB) approval was required. No private medical records or identifiable personal data were used. The study adhered to principles of responsible AI research, emphasizing transparency, fairness, and interpretability. Potential biases due to state-level policy differences and dataset limitations were addressed through model validation and explainability techniques.

Results

Figure 1 illustrates California’s abortion¹⁴ rate over time, which rose sharply after Roe v. Wade (1973)³, peaked in the late 1980s, and declined steadily from the 1990s onward. This trend reflects the expansion and subsequent regulation of abortion access, as well as the growing influence of contraception and reproductive health policies.

From a modeling perspective, these historical patterns are essential to both Prophet⁶ and XGBoost⁷. Prophet captures the long-term decline and short-term fluctuations, using past variability to forecast future abortion¹⁴ totals. Meanwhile, XGBoost incorporates abortion totals and rates as predictor variables, with SHAP⁵ analysis confirming their strong influence on birth outcomes. This connection demonstrates how shifts in abortion access—visible in Figure 1—directly inform the model’s ability to predict fertility dynamics.

Figure 2 shows pregnancy rates increasing through the late 1970s, stabilizing in the 1980s, and declining consistently from the mid-1990s onward. These changes align with improved access to contraception, delayed childbearing, and evolving cultural and socioeconomic pressures.

Prophet⁶ models these declines by decomposing the time series into trend and seasonal components, producing forward projections that reflect the continuation of these downward trajectories. In XGBoost⁷, pregnancy rates serve as a key predictor variable; SHAP ⁵ values demonstrate their consistent contribution to explaining birth totals. Thus, the trend captured in Figure 2 not only reveals historical shifts but also strengthens model interpretability by showing how reductions in pregnancies correlate with future declines in births.

Synthesis of graphs

Together, Figures 1 and 2 demonstrate a parallel decline in abortion¹⁴ and pregnancy rates after the 1990s, reflecting broader improvements in contraception, family planning, and reproductive autonomy. The simultaneous reduction in unintended pregnancies and abortion demand highlights how structural changes in healthcare and policy shape multiple aspects of reproductive behavior.

For Prophet⁶, the alignment of declining abortion¹⁴ and pregnancy rates reinforces the model’s forecasts of sustained long-term declines in births, with only minor rebounds. For XGBoost⁷, the synthesis underscores why abortion and pregnancy variables are among the most influential predictors: SHAP analysis⁵ shows that both features strongly shape model outputs, and their joint decline helps explain the projected decreases in future fertility. By linking these two indicators, the models capture not only individual variable effects but also their combined impact on long-term demographic change.

Utilizing SHAP Library

This is a SHAP (SHapley Additive exPlanations)⁵ summary plot visualizing feature importance in a machine learning model. The x-axis represents the mean absolute SHAP value, indicating the average contribution of each feature to the model’s predictions. The y-axis lists the features in descending order of their importance.

The most impactful features on the model’s output are:

1. state_US
2. miscarriagetotal
3. abortionstotal

The SHAP⁵ plot highlights the significant influence of factors like state US, miscarriage total, and abortion total on a machine learning model’s predictions. These variables can be tied to broader public health concerns, access to healthcare, and policy differences across states. For instance, the importance of abortion total and abortion rate total may reflect the ongoing impact of restrictive abortion¹⁴ laws and disparities in access to reproductive healthcare. Following the overturning of Roe v. Wade³, many states implemented strict abortion laws, creating significant disparities in access to abortion services. This ties directly to current debates surrounding reproductive rights, federal vs. state policies, and the challenges many individuals face in obtaining care.

Similarly, miscarriage total¹⁵ being a key factor in the model could indicate disparities in access to maternal healthcare, highlighting broader systemic issues such as rising maternal mortality¹⁶ rates and inconsistent quality of care. Current events, such as the ongoing efforts to expand Medicaid coverage for maternal care and postpartum support, align with this insight. Meanwhile, the prominence of state_US underscores how state-specific policies, healthcare systems, and cultural differences impact health outcomes. This ties into larger discussions about the need for consistent, equitable healthcare access across all states, regardless of state-level policies. To mitigate these issues, several strategies can be implemented. First, expanding access to reproductive healthcare is crucial. This can include increasing funding for clinics that offer affordable and comprehensive services, such as prenatal care, miscarriage management, and abortion services, especially in underserved areas. Supporting telemedicine for reproductive health can also help bridge access gaps in states with restrictive laws. Additionally, advocating for federal policies to ensure equitable access to healthcare across all states is vital, as state-level disparities often leave many without essential care.

Education and awareness campaigns also play a key role in addressing these issues. Public education initiatives can help reduce stigma around reproductive health while providing accurate, culturally relevant information to diverse communities. Investments in maternal and child health programs are another important step. Programs that offer prenatal care, miscarriage¹⁵ support, and postpartum assistance, particularly in underserved areas, can improve outcomes significantly. Collaborating with local organizations to provide holistic care can also help address the needs of vulnerable populations.

Finally, addressing the social determinants of health is essential for long-term change. Tackling root causes like poverty, lack of education, and inadequate healthcare infrastructure can reduce disparities in reproductive health outcomes. Partnerships between healthcare providers and social services can create integrated systems of care that address these underlying issues. By leveraging data-driven insights, advocacy groups and policymakers can create compelling narratives that drive systemic change, ensuring better health outcomes for all individuals.

This highlights the SHAP⁵ values for the year feature, with the color gradient representing the total number of miscarriages¹⁵ (miscarriage total). The increasing SHAP values over time indicate that the influence of year on the model’s predictions has grown. This trend could reflect improvements in data collection, changes in healthcare systems, or shifts in societal attitudes toward reproductive health. The concentration of high miscarriage totals (marked in red) in recent years might point to factors like delayed pregnancies, rising chronic health conditions, or greater awareness and reporting accuracy.

This graph underscores the importance of considering how historical and temporal trends impact reproductive health. For example, advancements in healthcare technologies or policies may have improved reporting and accessibility, but the increase in high-risk pregnancies due to lifestyle or medical factors also warrants attention.

This examines the SHAP⁵ values for birth rate total, showing its relationship to birth rates and miscarriage¹⁵ totals. The relatively flat distribution of SHAP values suggests that birthrate has a more limited direct impact on the model’s predictions compared to year. However, clusters of red points (indicating higher miscarriage totals) near average birthrate levels suggest a nuanced relationship. This could indicate disparities in healthcare access, social or economic conditions, or other confounding factors that influence both birth rates and miscarriages.

This graph highlights the need to further explore systemic inequities affecting birth and miscarriage¹⁵ rates. For example, communities with average birthrates may still face barriers to adequate prenatal care or education about reproductive health, contributing to higher miscarriage rates.

This graph plots the SHAP⁵ values for the feature abortion rate total against the abortion¹⁴ rate. The color gradient reflects the total number of miscarriages¹⁵ (miscarriagetotal). The negative correlation between SHAP values and abortion rate suggests that higher abortion rates may reduce the feature’s importance in the model’s predictions. However, clusters of red points (high miscarriage totals) around moderate abortion rates highlight a potential overlap of factors influencing both abortions and miscarriages, such as healthcare access or societal norms regarding family planning.

This graph points to the need for nuanced analysis of how abortion¹⁴ rates and miscarriage¹⁵ trends intersect, focusing on improving healthcare systems that address both outcomes comprehensively.

This graph illustrates the SHAP⁵ values for pregnancy rate total against pregnancy rates, again with miscarriage¹⁵ totals as the color gradient. The relatively uniform SHAP values indicate that pregnancy rates contribute consistently to the model’s predictions, regardless of fluctuations in the number of pregnancies. However, the concentration of red points around average pregnancy rates could suggest underlying stressors or risk factors that increase miscarriage totals even in regions or groups with typical pregnancy rates.

This insight highlights the importance of addressing underlying causes of miscarriages¹⁵, such as prenatal care quality or maternal health conditions, across diverse populations.

The final graph shows a strong linear correlation between the SHAP⁵ values and the miscarriage total feature, with a color gradient representing the total number of abortions (abortionstotal). The increasing SHAP values alongside miscarriage totals indicate that this feature is highly predictive in the model. High miscarriage totals (red points) are aligned with higher SHAP values, showing their dominant influence on the model’s predictions.

This graph emphasizes the central role of miscarriage data in reproductive health models. It calls for targeted interventions in areas with high miscarriage rates, ensuring that affected individuals receive adequate healthcare, counseling, and preventive measures.

Time Forecasting Model

The Prophet⁶ forecasts for California and Texas consistently project long-term declines across births, abortions¹⁴, miscarriages¹⁵, and pregnancies through 2030. Although both states follow a downward trajectory, the patterns differ in magnitude and volatility. California’s outcomes generally display smoother declines with modest late-decade rebounds, while Texas shows sharper oscillations, reflecting its demographic volatility, restrictive reproductive health policies, and uneven access to care.

In both states, birth totals are projected to decrease steadily, punctuated by fluctuations in the late 2020s. California shows relatively small oscillations, suggesting short-term rebounds that do not alter the overall downward trend. Texas, by contrast, exhibits sharper swings, with temporary recoveries around 2027–2028 followed by renewed declines. These dynamics suggest that external shocks—such as economic cycles, healthcare access, or immigration flows—may more strongly affect fertility outcomes in Texas.

Forecasts mirror the decline in births, with intermittent rebounds late in the decade. California’s trajectory reflects broader improvements in contraception, lower pregnancy rates, and demographic shifts such as delayed childbearing. Texas shows more volatility, where restrictive reproductive policies and disparities in healthcare access amplify fluctuations. These differences underscore how state-level policy environments directly shape reproductive outcomes.

Both states show a gradual decline with rebounds in the late 2020s, though California’s trajectory is smoother while Texas experiences greater variability. Miscarriage¹⁵ trends appear especially sensitive to maternal health factors, including delayed childbearing, chronic health conditions, and disparities in prenatal care. The widening confidence intervals highlight growing uncertainty tied to healthcare systems and demographic change.

Pregnancy totals are projected to decline steadily in both states, with modest rebounds in the late 2020s. California’s declines are gradual and shaped by delayed family formation, economic pressures such as housing and childcare costs, and cultural shifts toward higher education and career prioritization. Texas shows sharper oscillations, influenced by restrictive reproductive health laws and uneven healthcare infrastructure. Immigration may buffer declines in Texas by sustaining a younger childbearing population, but policy barriers may offset these demographic advantages.

Synthesis of graphs

Taken together, the forecasts reveal a consistent long-term decline across fertility-related indicators in both states, with intermittent rebounds that do not alter the overall trajectory. California’s trends are relatively stable, while Texas exhibits more volatility due to demographic growth, restrictive policies, and healthcare inequities. Across both states, widening uncertainty intervals underscore the influence of unpredictable external factors, including economic recessions¹⁷, policy shifts, and public health crises. These findings highlight that while structural demographic shifts drive the overall decline, state-specific policy contexts amplify or moderate fluctuations.

Prophet Evaluation

To evaluate the accuracy of the Prophet⁶ forecasting model, its performance was compared against a simple linear regression baseline using two widely accepted time-series error metrics: Root Mean Squared Error (RMSE)⁸ and Mean Absolute Percentage Error (MAPE)⁹. Prophet achieved an RMSE of 6,231.41 and a MAPE of 0.83%, while linear regression produced a much higher RMSE of 49,041.27 and MAPE of 10.13%. These results indicate that Prophet’s predictions were very close to the actual fertility counts, with deviations averaging less than 1% relative to the observed values. In contrast, linear regression struggled to capture the dynamics of the data, resulting in large errors that reflect its inability to model nonlinear trends and seasonal variation.

The substantial gap in performance highlights Prophet’s⁶ strength as a time-series forecasting tool, particularly for demographic datasets where fertility outcomes are shaped by both long-term trends and fluctuations in response to policy and economic conditions. Prophet’s capacity to incorporate seasonality and uncertainty intervals enables it to outperform simpler models that assume linear change over time. The validation results reinforce that Prophet provides a reliable forecast of California’s declining fertility trends, supporting its use as a robust model for anticipating demographic shifts and informing healthcare and policy planning.

To validate the accuracy of the Prophet⁶ model in Texas, forecasts were again compared against a linear regression baseline using RMSE⁸ and MAPE⁹. Prophet achieved an RMSE of 8,625.96 and a MAPE of 1.84%, whereas linear regression performed considerably worse, with an RMSE of 42,771.15 and a MAPE of 10.68%. These results show that Prophet was able to closely reproduce the observed fertility patterns in Texas, with average prediction errors under 2% of the actual values. In contrast, the linear regression model significantly misrepresented the data, reflecting its limitations in capturing nonlinear fertility dynamics.

The evaluation underscores Prophet’s⁶ superior ability to handle state-level fertility time series compared to simple baselines. Although Prophet’s error rates in Texas were slightly higher than those observed for California, they remain low enough to ensure that forecasts provide meaningful insights into demographic change. By capturing fluctuations and long-term decline more effectively than linear regression, Prophet offers a reliable approach to modeling future fertility trends in Texas, highlighting its robustness across different states with distinct demographic and policy contexts.

Discussion

This study combined machine learning and time-series forecasting to analyze fertility dynamics in California and Texas from 1973 to 2020 and to project future birth totals through 2030. The Prophet⁶ model revealed sustained long-term declines in fertility across both states, punctuated by fluctuations in the mid- to late 2020s that suggest temporary rebounds followed by continued decreases. California’s projections indicated relatively modest oscillations, while Texas displayed sharper variability, reflecting the state’s distinctive demographic growth, immigration dynamics, and restrictive reproductive policy environment.

Beyond forecasting, the integration of SHAP⁵ analysis with XGBoost⁷ regression identified miscarriage¹⁵ totals, abortion¹⁴ access, and state-level differences as the most influential predictors of fertility outcomes. These results emphasize how reproductive health trends are not only demographic in nature but also deeply intertwined with access to care and the sociopolitical landscape. The comparison between California and Texas underscores this point: while both states share a general downward fertility trajectory, the drivers and magnitude of decline vary according to policy context and healthcare availability.

The methodological contribution of this study lies in demonstrating how forecasting accuracy can be paired with interpretability. Prophet ⁶ outperformed linear regression in both states, with error rates under 2% of actual values, while SHAP⁵ provided transparent explanations of model predictions. Taken together, these approaches illustrate how interpretable AI can enrich demographic forecasting, supporting policymakers, healthcare providers, and researchers in anticipating fertility shifts and planning responsive interventions.

The study successfully met its objectives by generating accurate forecasts, identifying influential predictors, and demonstrating the utility of explainable AI for public health research. Nevertheless, limitations remain. The reliance on state-level aggregated data constrains generalizability, and long-term projections are inherently uncertain, subject to external shocks such as recessions¹⁷, legislative changes, or public health crises. Moreover, while SHAP⁵ illuminates associations, it does not establish causality, and results must therefore be interpreted with caution.

Future research should expand to multi-state or national datasets, incorporate socioeconomic and healthcare access variables, and explore hybrid models that combine demographic theory with advanced machine learning. Such approaches could further strengthen the balance between predictive power and interpretability.

Overall, this study demonstrates the potential of interpretable AI to deepen our understanding of fertility decline. As reproductive health in the United States continues to be shaped by evolving policies, economic conditions, and cultural shifts, transparent, data-driven forecasting tools² like Prophet⁶ —combined with explainable methods such as SHAP⁵ —offer critical guidance for equitable healthcare planning, resource allocation, and policy design.

References

1. Livingston, G. (2019). Why is the U.S. fertility rate falling? Pew Research Center. https://www.pewresearch.org/short-reads/2019/05/22/why-is-u-s-fertility-rate-falling [↩] [↩]
2. Taylor, S. J., & Letham, B. (2018). Forecasting at scale. The American Statistician, 72(1), 37–45. https://doi.org/10.1080/00031305.2017.1380080 [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩]
3. Roe v. Wade, 410 U.S. 113 (1973). Oyez. https://www.oyez.org/cases/1971/70-18 [↩] [↩] [↩] [↩]
4. Planned Parenthood v. Casey, 505 U.S. 833 (1992). Human Life International. https://www.hli.org/resources/planned-parenthood-vs-casey/ [↩]
5. Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems (NeurIPS). https://shap.readthedocs.io [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩]
6. Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and practice (3rd ed.). OTexts. https://otexts.com/fpp3/prophet.html [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩]
7. Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). ACM. https://doi.org/10.1145/2939672.2939785 [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩]
8. University of Illinois Urbana-Champaign. (n.d.). Root mean square error (RMSE) – Statistics with Python. Discovery Partners Institute. https://discovery.cs.illinois.edu/guides/Statistics-with-Python/rmse/ [↩] [↩] [↩] [↩] [↩]
9. Scikit-learn developers. (n.d.). sklearn.metrics.mean_absolute_percentage_error. https://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_absolute_percentage_error.html [↩] [↩] [↩] [↩] [↩]
10. OSF. (2020). Fertility rate measures by state and year (1973–2020). Open Science Framework. https://osf.io/duj6a [↩] [↩] [↩]
11. Centers for Disease Control and Prevention. (2023). National vital statistics reports. https://www.cdc.gov/nchs/nvss [↩]
12. National Center for Health Statistics. (2020). Births: Final data for 2020. https://www.cdc.gov/nchs/data/nvsr/nvsr70/nvsr70-17.pdf [↩]
13. Hamilton, B. E., Martin, J. A., Osterman, M. J. K., & Rossen, L. M. (2022). Births: Provisional data for 2021 (Vital Statistics Rapid Release; No. 20). National Vital Statistics System. https://www.cdc.gov/nchs/nvss/vsrr/births.htm [↩]
14. Guttmacher Institute. (2023). State abortion policy landscape. https://www.guttmacher.org/state-policy [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩]
15. Rossen, L. M., Ahrens, K. A., & Branum, A. M. (2018). Trends in risk of pregnancy loss among U.S. women, 1990–2011. Paediatric and Perinatal Epidemiology, 32(1), 19–29. https://doi.org/10.1111/ppe.12417 [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩] [↩]
16. Petersen, E. E., Davis, N. L., Goodman, D., Cox, S., Syverson, C., Seed, K., … Barfield, W. (2019). Racial/ethnic disparities in pregnancy-related deaths—United States, 2007–2016. MMWR Morbidity and Mortality Weekly Report, 68(35), 762–765. https://doi.org/10.15585/mmwr.mm6835a3 [↩]
17. Morgan, S. P., Cumberworth, E., & Wimer, C. (2011). The Great Recession’s influence on fertility, marriage, divorce, and cohabitation. In R. R. Rugh (Ed.), The Great Recession (pp. 220–245). Russell Sage Foundation. [↩] [↩]