Temporal generalized linear mixed models for monitoring national health behavior in the United States using subnational observational data
Article information
Abstract
OBJECTIVES
Since 2025, administrative changes have resulted in the dismissal of federal workers responsible for implementing United States national health surveys. While national data are crucial to understanding trends at the national level, it may be possible to complement such data sources using state-level observational data by using the latter to predict the former in statistical models. This study presents a proof-of-concept approach for estimating the national prevalence of health behaviors in the United States based on partial state-level prevalence estimates.
METHODS
Focusing on United States adolescent e-cigarette use, the prevalence of which informs regulatory efforts by the U.S. Food and Drug Administration’s Center for Tobacco Products, we implemented temporal generalized linear mixed models (GLMMs; with binomial or beta response distributions, a logit link function, and a spatial power covariance structure to account for autocorrelation). Models were fitted to national e-cigarette use prevalence from the National Youth Tobacco Survey and state-level e-cigarette use prevalence from 38 state-specific surveys administered by state health and education departments and other organizations from 2011 to 2025.
RESULTS
Beta models performed similarly to binomial models in year-level predictive performance, with modest mean absolute error (MAE) for in-sample predictions (beta: 1.5–1.8 percentage points [pp]; binomial: 1.4–2.6 pp) and out-of-sample predictions (beta: 2.3 pp; binomial: 2.2 pp). MAEs were generally higher for simpler comparator models: time-trend extrapolation models, 4.7–5.5 pp; 1-year-lag persistence models, 4.1–4.7 pp; and naïve mean models, 1.8–2.0 pp.
CONCLUSIONS
Temporal GLMMs may be useful tools for monitoring national trends in substance use behaviors when subnational data, but not national data, are available.
Key Message
In 2025, U.S. federal workers responsible for national health surveys were dismissed and federal health departments were restructured, creating uncertainty around future national health survey data collection, and an unmet need to understand national health behavior trends. This study explores the possibility of using state-level survey data to complement national survey data for years in which national data are not collected, or otherwise compromised, using statistical models. To this end, generalized linear mixed models for time series data are shown to predict national substance use prevalence from subnational data with some accuracy, which may help prevent epidemiological ‘blind spots.’
INTRODUCTION
Accurate and timely estimates of substance use prevalence are essential for understanding trends, informing prevention efforts, and evaluating existing interventions [1]. In the United States, the Department of Health and Human Services has historically funded multiple nationally representative surveys of health behaviors, including tobacco and other substance use, such as the National Survey on Drug Use and Health (NSDUH) and the National Youth Tobacco Survey (NYTS). Since 2025, administrative changes have resulted in the dismissal of federal workers responsible for implementing the NSDUH and NYTS [2,3]. Although these surveys are likely to continue [4,5], these administration changes have created uncertainty about the future of federal health surveys.
The NYTS is particularly important because researchers and the U.S. Food and Drug Administration (FDA) have used it to assess national trends in nicotine product use among adolescents [6,7], especially e-cigarette use, which was the most commonly reported individual nicotine product use behavior among United States adolescents in 2024 [8]. Ongoing monitoring of national adolescent e-cigarette use is therefore important for public health surveillance and regulatory strategy.
Given uncertainty around future national data collections, complementary methods for monitoring the same or similar survey endpoints may be informative. In the United States, many states conduct surveys of substance use and health behaviors among adolescents, such as the Florida Youth Tobacco Survey [9]. If all states collected data with relevant endpoints every year, state-level prevalence estimates could be aggregated straightforwardly to estimate national use prevalence. When complete annual state-level data are unavailable, however, a statistical model that maps partial state-level prevalence estimates for a substance use behavior to national prevalence estimates for the same behavior could estimate national prevalence when national data are unavailable or before national estimates are released.
Developing such a model presents several challenges. First, not all states collect data with relevant endpoints. Thus, it will not always be possible to simply aggregate data across all states to produce a nationally-representative estimate. Instead, the model must account for how closely states with available data track national-level trends when producing national-level predictions from those states. Second, state-level surveys are not all administered in the same years or at the same intervals; some are annual, some are biennial, and others are administered irregularly. This precludes model designs that require complete data for all included states in all years. Finally, within-state temporal correlation, or autocorrelation, must be accounted for to avoid underestimating variance.
This study aimed to develop a model that addresses these challenges and predict national prevalence from partial state-level data. We describe approaches implemented in SAS (SAS Institute Inc., Cary, NC, USA) and compare their year-level predictive performance both in sample, within the range of years with available national data, and out of sample, beyond the range of years with available national data. Such predictions may help fill years for which national data are unavailable or provide an early indication of national trends when state-level data become available first.
MATERIALS AND METHODS
Data source
National United States adolescent e-cigarette use prevalence estimates for each year were obtained from the NYTS, an annual cross-sectional survey of tobacco use among adolescent respondents in the United States. The NYTS was administered by the U.S. Centers for Disease Control and Prevention’s (CDC) Office on Smoking and Health from 1999 to 2024 and was transferred to the FDA’s Center for Tobacco Products in 2025. Annual e-cigarette use prevalence was available from 2011 to 2025. Annual NYTS sample sizes ranged from n=14,531 in 2020 to n= 29,861 in 2024, with a total sample size of n=318,197 across 2011–2025. NYTS data, codebooks, and methodological reports are publicly available from the CDC and FDA [10,11]. Briefly, the NYTS uses a complex sample design involving stratified cluster sampling and includes respondent-level sampling weights to produce nationally representative prevalence estimates.
State-level adolescent e-cigarette use prevalence estimates for each year in each state were obtained from state-specific surveys administered by state health and education departments, often in collaboration with local universities and other organizations. The most recent publicly available state-specific surveys with e-cigarette use prevalence estimates known to the authors at the time of analysis are described in Supplementary Material 1. The dataset for this work consisted of n=221 state-level prevalence observations from 38 state-specific surveys from 2011 to 2025. States with only 1 estimate (e.g., a single estimate in 2019 and no estimates for other years) were excluded to avoid model convergence and estimation issues. This list is not exhaustive; in this study, only 1 survey per state per year was included, namely that which provided the most recent estimate.
For maximum comparability, NYTS and state-specific surveys were restricted to high school students where possible. When results for a given state were reported only by individual grade (e.g., prevalence estimates reported separately for grades 10 and 12), estimates were averaged across grades and weighted by the sample size in each grade to produce a single high-school estimate for each state in each year. Details and exceptions are provided in Supplementary Material 1.
Measurement
Current e-cigarette use was defined as having had at least 1 puff of an e-cigarette on at least 1 day of the 30 days before the survey, consistent with the NYTS definition and the state-level estimates.
Model specification
To model the relationship over time between national e-cigarette use prevalence, as measured in the NYTS, and state-level e-cigarette use prevalence, as measured in state-specific surveys, generalized linear mixed models (GLMMs) were implemented in SAS version 9.4 (SAS Institute Inc.) using the PROC GLIMMIX procedure. The linear predictor included state-level prevalence, entered as a continuous numerical proportion ranging from 0 to 1; state identity, entered as a categorical variable; their interaction; and an intercept term. State-level prevalence and state identity were treated as fixed effects. Because the dependent variable was a proportion bounded between 0 and 1, a logit link function was used to relate expected national prevalence to the linear predictor. Because these were time-series data, state-level prevalence observations within the same state were expected to be temporally correlated (autocorrelation), with correlations decreasing as the time between observations increased. This correlation was modeled using a spatial power (SP[POW]) R-side covariance structure to avoid underestimating variance. SP(POW) was selected over other structures, such as AR(1), because the spacing of observations over time differed across states; some state-specific surveys were annual, others were biennial, and others were administered irregularly. SP(POW) allows for unequally spaced time points, whereas AR(1) assumes equal spacing. SP(POW) was considered a reasonable compromise between highly parameterized covariance structures, such as unstructured covariance, and simpler structures, such as compound symmetry covariance. Mathematical details are provided in Supplementary Material 2. Two approaches to the choice of response distribution were used, as described below.
Binomial models
The first approach to modeling the dependent variable used the “events/trials” syntax described in the SAS documentation [12]. In this syntax, “events” represented the weighted number of adolescents in the NYTS in a given year who reported e-cigarette use, and “trials” represented the weighted sample size in that same year; for example, “1215050/15644664” represented a prevalence of 7.8%. For the events/trials syntax, a binomial response distribution with the canonical logit link function was used; these are hereafter referred to as binomial models. Because NYTS sampling weights are required to produce nationally representative prevalence estimates in each year, the event and trial counts were survey-weighted counts. A down-weighting procedure was applied because the weighted event and trial counts were much larger than the corresponding unweighted counts (the NYTS sample size), which can cause issues in SAS; details are provided in Supplementary Material 3.
Beta models
For the binomial models described above, evidence of potential overdispersion was observed (see Results section). To address this issue, the second approach used the “response” syntax described in the SAS documentation [12], with the dependent variable specified as the weighted NYTS prevalence in a given year, expressed as a numerical proportion. For the response syntax, a beta response distribution with a logit link function was used; these are hereafter referred to as beta models. Because the beta models regress directly on weighted proportions rather than counts, this approach avoids the scaling issue described above for the binomial models.
Model diagnostics and evaluation
Goodness-of-fit
Models were fitted to national and state-level prevalence for the years 2011–T, where T>2011. For each model, overdispersion was measured using the “generalized chi-square/degrees of freedom (DF)” fit statistic in SAS. Residual analysis followed the approach of Gelman & Hill [13] for generalized linear models with logit link functions and responses bounded on the interval [0, 1], using plots of binned residuals. In this context, residuals were naturally binned by year. For well-calibrated models without obvious misspecification, residuals should be centered on zero, remain within theoretical 95% error bounds, and show no obvious patterns [13].
In-sample predictive performance
Adjusted means were derived from the model using SAS ESTIMATE statements to predict national prevalence each year based on state-level prevalence across states in that same year. These adjusted means were calculated for each year from 2011 to T. Because the model was trained on national data from 2011 to T, predictions of national prevalence from 2011 to T were in-sample predictions, within the range of years on which the model was trained. For each year in 2011–T, observed national prevalence, denoted , was compared with the model-based prediction ŷt (defined as the adjusted mean) by calculating the absolute error |ŷt-tt| in percentage points (pp). This approach quantified the accuracy of in-sample predictions. The mean absolute error (MAE) was calculated as the average absolute error across predictions for years 2011–T. This metric describes in-sample predictive accuracy on the same scale as the response, in pp [14]. MAE was used to compare models because there is no basis for comparing fit statistics with marginal models [15].
Out-of-sample predictive performance
Adjusted means were also calculated using SAS ESTIMATE statements and state-level prevalence for the year T+1 to predict national prevalence in the year T+1. Because the model was trained on national data from 2011 to T but not T+1, predictions of national prevalence in T+1 were out-of-sample predictions, beyond the range of years on which the model was trained; predictions for T+1 were made without using observed national prevalence in T+1. This process was repeated for values of T from T=2021, predicting T+1=2022, to T=2024, predicting T+1=2025. As of this writing, only 1 state-level prevalence estimate for 2026 was publicly available and known to the authors with which to make national predictions for 2026; since these data are so few, no predictions were made for 2026. The observed national prevalence in T+1, yT+1, was compared with the model-based prediction ŷT+1 by calculating the absolute error |ŷT+1-yT+1| in pp, thus quantifying the accuracy of out-of-sample predictions. The MAE was calculated as the average absolute error for predictions of T+1=2022, T+1=2023, T+1=2024, and T+1=2025. This metric describes out-of-sample predictive accuracy on the same scale as the response, in pp. For both in-sample and out-of-sample predictions, 95% confidence intervals were generated with SAS ESTIMATE statements and plotted.
States were included in a given model only if they had publicly available prevalence estimates for year T+1, so that they could be used to predict national prevalence in year T+1, and publicly available prevalence estimates for at least 2 years other than T+1, so that sufficient time-varying information was available to estimate model parameters.
Modeling was performed in SAS version 9.4, and SAS code is provided in an online GitHub repository [16].
Handling of missing data
State-level e-cigarette use prevalence was not available for every state in every year because state-level surveys were not administered annually in all states. No imputation was performed for years in which state-level prevalence data were unavailable; observations from missing years were simply not included. The binomial and beta models used in this study do not require observations for every state in every year or in the same years. SAS automatically sets up separate SP(POW) structures for each subject, here defined as state, in the model [17,18]. These structures need not have the same size, the same number of observations per state, the same years, or the same spacing between years for each state.
Technical notes on R-side vs. G-side modeling
In their text on GLMMs in SAS, Stroup et al. [15] compare R-side, or marginal/GEE-type, models with G-side models; these are 2 approaches to modeling repeated measures, such as state-level prevalence over time. A key difference is that G-side models are likelihood-based and can be estimated using either pseudo-likelihood or integral approximation, whereas R-side marginal models are quasi-likelihood models by definition and are necessarily estimated by pseudo-likelihood [15].
The models in the present study were used only for prediction; fitted model parameters were simply coefficients in prediction equations and were not interpreted [19]. Although Stroup et al. [15] recommend G-side models over marginal models when formal covariance testing is a priority, covariance testing was not a priority in this study. Model selection was not performed with likelihood-based covariance model selection tools but with the predictive performance metrics described above in Model Diagnostics and Evaluation. Alternative G-side model specifications with subject-specific random effects were also considered; however, for these data, marginal models generally outperformed G-side models in predictive performance. This finding is consistent with Stroup et al. [15]’s observation that “in many cases [R-side structure] adequately models the distribution... and the quasi-likelihood is perfectly well-defined for generalized linear model estimation.” For these data, marginal models with pseudo-likelihood techniques generally performed at least as well as G-side models with subject-specific random effects and integral approximation techniques. The final models presented in this work are marginal models with an R-side temporal covariance structure and no G-side random effects.
For beta and other 2-parameter distributions, marginal models are overparameterized because scale parameters are included for both the distribution and the R-side structure, which makes them confounded and harder to interpret. Stroup et al. [15] therefore state that “Both ‘G-side’ and ‘R-side’ covariance structures are problematic for two-parameter non-Gaussian distributions, e.g. negative binomial, beta, and gamma. This is an active area of GLMM research.” The models in the present study were nevertheless used only for prediction, and for these data, beta models performed similarly to binomial models in predictive performance (see Results section). Beta models are included here as a practical illustration because event and trial counts may not be available to analysts in other contexts, precluding use of the events/trials syntax in SAS.
Comparator models
The beta and binomial models described above were benchmarked against 3 simple comparator models using the same MAE metric: (1) the simple time-trend extrapolation model E[ŷT|T]=logit–-1(β0+β1T); (2) the 1-year-lag persistence model E[ŷT+1|yT]=logit–-1(β0+β1yT); and (3) a naïve mean model that computes the average state-level prevalence across states in a given year, weighted by state population, as the prediction for national prevalence in that year. These comparator models have clear disadvantages relative to the beta and binomial models. The time-trend model cannot easily handle changes in the direction of change of prevalence, such as an increase in one year followed by a decrease in the next; the 1-year-lag model cannot predict prevalence in years without NYTS data from the previous year (such as 2011, because e-cigarette use was not assessed in the NYTS in 2010); and the naïve mean model provides less principled uncertainty estimation because the standard errors of the means will be large when the number of observations is low. Nevertheless, these models are useful as rudimentary benchmarks.
Ethics statement
This study used only publicly available, de-identified secondary data and is not considered United States National Institutes of Health (NIH) human subjects research.
RESULTS
Raw prevalence (no model)
Figure 1 shows the prevalence of current e-cigarette use among United States adolescents in the NYTS, representing national prevalence, and in the state-specific surveys described in Supplementary Material 1. In the NYTS, current high-school e-cigarette use prevalence increased by 966.7% (+14.5 pp), from 1.5% in 2011 to 16.0% in 2015; decreased by 26.9% (−4.3 pp), to 11.7% in 2017; increased again by 135.0% (+15.8 pp), to a peak of 27.5% in 2019; and most recently decreased by 74.2% (−20.4 pp), to 7.1% in 2025, the lowest prevalence in more than a decade.
National prevalence from the National Youth Tobacco Survey and raw state-level prevalence of current (past 30-day) e-cigarette use among U.S. high school students over time.
Across individual states, prevalence generally increased from 2011 to peaks in 2019, or in 2018 for states that did not collect data in 2019, and then generally decreased through 2025. Prevalence differed widely across states, partly because of differences in the grades surveyed. The lowest prevalence was observed in California, where use peaked at 10.9% in 2018 and decreased to 5.0% in 2024. The highest prevalence was observed in Wyoming, where use peaked at 36.1% in 2018 and decreased to 17.0% in 2024. Variation in prevalence across states was greatest in 2018 and 2019, near peak prevalence, and was lower in 2025. Reasons for this variation are beyond the scope of the present study but may include differences in regulatory approaches to e-cigarettes across states and over time [20,21].
Binomial models
Figure 2 shows separate binomial models fitted to national and state-level prevalence data from 2011 to T and blinded to national data in T+1, for T=2021 (predicting T+1=2022) through T=2024 (predicting T+1=2025). The 95% confidence intervals, produced by ESTIMATE statements in SAS, are also plotted in Figure 2 for each adjusted mean, that is, each predicted value of national prevalence across states in a year.
National prevalence from the National Youth Tobacco Survey and predictions from binomial models based on state-level prevalence of current (past 30-day) e-cigarette use among U.S. high school students over time. (A) Model fitted with data from 2011–2021 to predict 2022. (B) Model fitted with data from 2011–2022 to predict 2023. (C) Model fitted with data from 2011–2023 to predict 2024. (D) Model fitted with data from 2011–2024 to predict 2025. Models are blinded to national data in the final year. CI, confidence interval.
Goodness-of-fit
All binomial models converged rapidly (≤10 iterations). All binomial models showed evidence of overdispersion (generalized chi-square/DF≫1). Residual plots showed values centered on zero, within theoretical 95% error bounds, and with no obvious patterns, suggesting fair model calibration across the range of predicted probabilities, reasonable fit, and no obvious misspecification (Supplementary Material 4).
In-sample predictive performance
Qualitatively, in-sample predictions from the binomial models generally followed the national prevalence trend with directional consistency (Figure 2). Quantitatively, the binomial model MAEs ranged from 1.4 pp to 2.6 pp. These errors were low relative to the range of national prevalence values, 1.5% to 27.5%, but comparable to some year-over-year changes in national prevalence, such as the −2.2 pp change from 10.0% in 2023 to 7.8% in 2024.
Out-of-sample predictive performance
The MAE for out-of-sample predictions was 2.2 pp for the binomial models. This error was low relative to the range of national prevalence values but comparable to some year-over-year changes.
Beta models
Figure 3 shows the beta models fitted to the same data used for the binomial models. The 95% confidence intervals for each adjusted mean are also plotted in Figure 3.
National prevalence from the National Youth Tobacco Survey and predictions from beta models based on state-level prevalence of current (past 30-day) e-cigarette use among U.S. high school students over time. (A) Model fitted with data from 2011–2021 to predict 2022. (B) Model fitted with data from 2011–2022 to predict 2023. (C) Model fitted with data from 2011–2023 to predict 2024. (D) Model fitted with data from 2011–2024 to predict 2025. Models are blinded to national data in the final year. CI, confidence interval.
Goodness-of-fit
Beta models also converged rapidly (≤19 iterations). No beta models showed evidence of overdispersion (generalized chi-square/DF=1). Residual plots showed residuals centered on zero, within theoretical 95% error bounds, and with no obvious patterns, suggesting fair model calibration across the range of predicted probabilities, reasonable fit, and no obvious misspecification (Supplementary Material 5).
In-sample predictive performance
Qualitatively, in-sample predictions from the beta models, expressed as adjusted means, followed the national prevalence trend relatively closely (Figure 3). Quantitatively, their MAEs ranged from 1.5 pp to 1.8 pp.
Out-of-sample predictive performance
For the beta models, the MAE for out-of-sample predictions was 2.3 pp.
Comparator models
By way of comparison, MAEs ranged from 4.7 pp to 5.5 pp for the simple time-trend extrapolation models, 4.1 pp to 4.7 pp for the 1-year-lag persistence models, and 1.8 pp to 2.0 pp for the naïve mean models. These values were generally higher than those for the beta and binomial models.
DISCUSSION
This study provides a proof of concept for modeling national-level substance use prevalence using partial state-level prevalence estimates for the same behavior. The GLMMs showed potential for this purpose, at least for 1-year-ahead prediction of national prevalence.
Importantly, these models should not be considered alternatives to, or replacements for, large federal surveys. However, they may provide useful complementary estimates for single years in which national survey data collection is suspended, delayed, or otherwise compromised. Although these models captured year-to-year changes with directional consistency in most cases (Figures 2 and 3), the MAEs were approximately 2 pp, indicating that out-of-sample predictions differed from observed national prevalence by approximately 2 pp on average. This error is comparable to some observed year-over-year changes. Therefore, although these models may be informative, their predictions should be interpreted with caution.
These model designs are not specific to nicotine use and could, in theory, be applied to other substance use behaviors or to other health behaviors, such as consumption of sugar-sweetened beverages. NSDUH, which was also reportedly affected by administrative restructuring at NIH [2], collects prevalence data on a wide range of substances, including tobacco, alcohol, and prescription drugs [22]. Provided that subnational data exist for a given substance or other health behavior, similar models could be developed to complement NSDUH data collection for affected years. Applying GLMMs to other behaviors is beyond the scope of this proof-of-concept study but represents a potentially informative direction for future observational research, because these methods may be applicable to surveys not focused on drug use and to other countries with subnational data. Importantly, the marginal models used in this work are not universally preferred, and alternative model specifications, including G-side models, may perform better with other data.
This study has several limitations. Substance use prevalence estimates derived from self-report surveys are subject to issues related to question understanding, recall bias, and social pressures that may affect response accuracy [23]. This limitation applies to both national and state-level prevalence estimates, and caution is therefore warranted when interpreting the prevalence estimates in this study. For the NYTS, changes in survey design likely affected the comparability of prevalence estimates over time. These changes included a change in the e-cigarette questions from select-all to forced-choice response options in 2014, a switch from a paper-and-pencil instrument to an electronic, tablet-based instrument in 2019, and a switch from in-school-only administration to mixed in-school and at-home administration for the 2021 survey, conducted during the coronavirus disease 2019 pandemic [24]. Because NYTS 2021 estimates based on at-home respondents have been shown to be systematically lower than those based on in-school respondents [25], this study used in-school respondents only for the NYTS 2021 estimate to maximize comparability with other NYTS years. Similar methodological changes may also have occurred in state-level surveys over time, but these changes are not well documented in publicly available materials. Questionnaire wording for e-cigarette use also varied across surveys and over time within the same surveys, including wording such as “vape” versus “e-cigarette” (Supplementary Material 1). The latter typically refers only to electronic nicotine delivery systems (ENDS), whereas the former can be used more broadly to describe cannabis or cannabinoid vaping. Although such assessment variation and imprecision limited the model’s ability to track ENDS use comparably across state and national data, the model estimates were relatively close to NYTS estimates despite these variations. The dataset was limited to n=221 state-level prevalence observations from 2011 to 2025; this relatively small number of observations, combined with many model parameters, may have caused overfitting. Larger datasets could improve the robustness of such models. Finally, these models’ predictions were limited in accuracy and precision and should be interpreted cautiously.
In conclusion, temporal GLMMs may complement national-level monitoring of health behaviors by using partial state-level data when national-level data are unavailable.
Data availability
NYTS and state-level data are available in publicly available online repositories; sources are provided in Supplementary Material 1. SAS code is also provided in an online GitHub repository (https://github.com/FloeFoxon/Temporal-GLMMs-to-Monitor-National-Health-Behaviour-Using-Subnational-Observational-Data).
Supplementary materials
Supplementary materials are available at https://doi.org/10.4178/epih.e2026035.
Supplementary Material 1.
Sources of data
Supplementary Material 2.
Mathematical description of models
Supplementary Material 3.
Weighting procedure
Supplementary Material 4.
Binned residual plots for the binomial models. Solid lines represent theoretical 95% error bounds.
Supplementary Material 5.
Binned residual plots for the beta models. Solid lines represent theoretical 95% error bounds.
Supplementary References
Notes
Conflict of interest
Since July 2020, the authors provide consulting services through Pinney Associates on tobacco harm reduction to Juul Labs. In the last 36 months, Pinney Associates also consulted to Philip Morris International solely on US regulatory pathways for non-combustible, non-tobacco nicotine products, which had no involvement in this work. Pinney Associates does not consult on combustible tobacco products. Author SS also owns interest in a novel nicotine gum that has neither been developed nor commercialized.
Funding
The development of the models and preparation of this manuscript were funded by Juul Labs.
Acknowledgements
The authors thank Ryan A. Black (formerly of Juul Labs).
The content and the decision to publish are the responsibility of the authors.
Author contributions
Conceptualization: Foxon F, Shiffman S. Data curation: Foxon F. Formal analysis: Foxon F. Funding acquisition: All authors. Methodology: All authors. Project administration: Shiffman S. Visualization: Foxon F. Writing – original draft: Foxon F. Writing – review & editing: Sembower MA, Shiffman S.
