How advances in machine learning and statistical modelling have changed how we should learn about and support people with Down syndrome
- Frank Buckley (Down Syndrome Education International, Cumbria, UK, Down Syndrome Education USA, CA, USA)
Correspondence: frank.buckley@dseinternational.org
Parents and practitioners want clear and reliable answers to practical questions about how to best help the children they support. Ideally, these answers are based on valid inferences accurately derived from the available evidence and accompanied by an honest appraisal of uncertainty.
Over the past few decades, a variety of highly effective techniques for recognising patterns in data and for making predictions from data have been developed (Bishop, 2006; Bishop & Bishop, 2024). Over a similar period, old (Bayesian) statistical approaches have been revitalised by the rapid increase in the speed and capacity of affordable computers, and the development of statistical sampling algorithms (Gelman et al., 2013; McElreath, 2020). Machine learning algorithms and statistical modelling methods overlap substantially in methods, theory, and practice. They continue to evolve at accelerating rates, providing a rapidly expanding toolbox of techniques for anyone who needs to accurately describe things we have observed or predict what we expect to observe in the future.
These advances offer researchers and practitioners new opportunities (and, arguably, responsibilities) to develop transparent theoretical models, to explore and explain data more thoroughly, to draw inferences (including causal inferences) more clearly, and to offer more precise and intuitive estimates of uncertainty in their conclusions than has often been the case in the past. In doing so, these advances may be a part of the solution to the replication crisis facing science today (McElreath, 2020, Chapter 17; McElreath & Smaldino, 2015; Rouder et al., 2016; Van Lissa, 2023).
This series of presentations reports on our experiences of applying selected machine learning algorithms and statistical modelling methods to small and large data sets of relevance to families, professionals and organisations supporting people with Down syndrome. We explore how these new and evolving techniques can improve data analysis, inference and the communication of findings.
Bishop, C. M. (2006). Pattern recognition and machine learning. Springer.
Bishop, C. M., & Bishop, H. (2024). Deep Learning: Foundations and Concepts. Springer.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2013). Bayesian Data Analysis (3rd edn). Chapman and Hall.
McElreath, R. (2020). Statistical Rethinking: A Bayesian Course with Examples in R and STAN (2nd edn). Chapman and Hall.
McElreath, R., & Smaldino, P. E. (2015). Replication, Communication, and the Population Dynamics of Scientific Discovery. PLOS ONE, 10(8), e0136088.
Rouder, J. N., Morey, R. D., Verhagen, J., Province, J. M., & Wagenmakers, E.-J. (2016). Is There a Free Lunch in Inference? Topics in Cognitive Science, 8(3), 520–547.
Van Lissa, C. J. (2023). Developmental data science: How machine learning can advance theory formation in Developmental Psychology. Infant and Child Development, 32(6), e2370.
Introducing modern statistical and machine learning methods for developmental and educational research in Down syndrome
- Frank Buckley (Down Syndrome Education International, Cumbria, UK, Down Syndrome Education USA, CA, USA)
Correspondence: frank.buckley@dseinternational.org
Modern machine learning algorithms and statistical methods are transforming how research is conducted in many of the sciences today, although they are not yet widely used in developmental and educational research in Down syndrome.
These techniques are well suited to exploratory and experimental research aiming to describe models of development and learning that often involve high-dimensional predictors, measurement error, clustered data, longitudinal change, small samples and significant heterogeneity. Tree-based methods – especially ensemble approaches such as random forests and gradient-boosted trees – provide flexible tools for prediction and classification, and can be useful for exploratory identification of candidate predictors (Fife & D’Onofrio, 2023). Bayesian data analysis and multilevel modelling provide a coherent framework for estimating effects in the presence of group structure and individual differences, modelling linear and non-linear relationships (including interactions), quantifying uncertainty, and incorporating prior information (Gelman et al., 2013; McElreath, 2020). When combined with explicit causal assumptions and appropriate study designs, they can also support causal inference in both randomised and observational settings (Li et al., 2023; McElreath, 2020; Oganisian & Roy, 2021).
We introduce ongoing work to explore and evaluate the use of modern machine learning algorithms and statistical methods in developmental and educational research in Down syndrome. We will highlight a few of the techniques we have examined, their uses, and some of the practicalities involved in running machine learning algorithms and fitting complex statistical models with small and large data sets.
Fife, D. A., & D’Onofrio, J. (2023). Common, uncommon, and novel applications of random forest in psychological research. Behavior Research Methods, 55(5), 2447–2466.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2013). Bayesian Data Analysis (3rd edn). Chapman and Hall.
Li, F., Ding, P., & Mealli, F. (2023). Bayesian causal inference: A critical review. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 381(2247), 20220153.
McElreath, R. (2020). Statistical Rethinking: A Bayesian Course with Examples in R and STAN (2nd edn). Chapman and Hall.
Oganisian, A., & Roy, J. A. (2021). A Practical Introduction to Bayesian Estimation of Causal Effects: Parametric and Nonparametric Approaches. Statistics in Medicine, 40(2), 518–551.
Exploring predictors of recorded births of babies with Down syndrome using gradient boosting decision tree learning algorithms and Bayesian statistical models
- Frank Buckley (Down Syndrome Education International, Cumbria, UK, Down Syndrome Education USA, CA, USA)
- Gert de Graaf (Dutch Down Syndrome Foundation, Meppel, The Netherlands)
- Brian Skotko (Down Syndrome Program, Division of Medical Genetics and Metabolism, Department of Pediatrics, Massachusetts General Hospital, Boston, MA, Department of Pediatrics, Harvard Medical School, Boston, MA)
Correspondence: frank.buckley@dseinternational.org
Understanding the current and future characteristics of babies born is important for planning healthcare, education and social support services.
U.S. birth certificate data includes detailed information about babies born in the U.S. (Vital Statistics Natality Birth Data, 2025). Since 1989, this information has included the presence of Down syndrome. Unfortunately, congenital conditions are significantly underreported, and this underreporting may be biased such that characteristics of recorded cases are not an accurate reflection of the true population.
Here, we describe an ongoing study using gradient boosting decision tree learning algorithms to train a model to identify recorded cases of Down syndrome using records of 33.6 million births from 2016 to 2024. Given an estimated ~60% of true cases are not recorded, we treat recorded negatives as unlabelled data and therefore view the problem as a case of learning from positive and unlabelled data (Bekker & Davis, 2020). Starting with 48 candidate features, we identify important predictors using a combination of SHAP (Lundberg & Lee, 2017) and permutation importance (Breiman, 2001) values, and by examining distance correlations (Székely et al., 2007) between predictors.
We use the resulting trained model to identify the most likely missed cases in each month to estimate all Down syndrome births in each of these years. We quantify the characteristics of the total population of recorded and likely missed babies with Down syndrome, including maternal and paternal race and ethnicity and education, family income indicators, complications arising during pregnancy and delivery, and recorded congenital abnormalities. We compare these to other available estimates.
Using this augmented data set, we fit a Bayesian statistical model (Gelman et al., 2013; McElreath, 2020) to estimate the influences of predictors on the births of babies with Down syndrome in the U.S. from 2016 to 2024.
We describe some of the practical challenges we have encountered working with a large data set and how these techniques might be applied to other areas of research.
Bekker, J., & Davis, J. (2020). Learning from positive and unlabeled data: A survey. Machine Learning, 109(4), 719–760.
Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5–32.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2013). Bayesian Data Analysis (3rd edn). Chapman and Hall.
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Proceedings of the 31st International Conference on Neural Information Processing Systems, 4768–4777.
McElreath, R. (2020). Statistical Rethinking: A Bayesian Course with Examples in R and STAN (2nd edn). Chapman and Hall.
Székely, G. J., Rizzo, M. L., & Bakirov, N. K. (2007). Measuring and testing dependence by correlation of distances. The Annals of Statistics, 35(6), 2769–2794.
Vital Statistics Natality Birth Data. (2025, September 2). NBER.
Exploring factors influencing language and reading skills in children with Down syndrome using random forest regression algorithms and Bayesian statistical models
- Frank Buckley (Down Syndrome Education International, Cumbria, UK, Down Syndrome Education USA, CA, USA)
- Sue Buckley (Down Syndrome Education International, Cumbria, UK, Down Syndrome Education USA, CA, USA)
- Angela Byrne (Down Syndrome Education International, Cumbria, UK, Down Syndrome Education USA, CA, USA)
Correspondence: frank.buckley@dseinternational.org
Learning language and learning to read are complex, multifaceted and interacting processes (Share, 2025; Snowling & Hulme, 2025). Many factors influence the development of both language and reading skills over time, many of these factors interact and many of these interactions may be compounded over time. Disentangling this complexity requires techniques that can model these interactions and estimate their effects. It also requires detailed and longitudinal data sets.
Here, we describe an ongoing study to explore predictors of language and reading skills in children with Down syndrome utilising existing longitudinal data sets. We use random forest regression algorithms to identify important predictors of language and reading outcomes using a combination of SHAP (Lundberg & Lee, 2017) and permutation importance (Breiman, 2001) values, and by examining distance correlations (Székely et al., 2007) between predictors. We then use multilevel Bayesian statistical models (Gelman et al., 2013; McElreath, 2020) to explore and estimate the effects of these predictors on language and reading outcomes.
We describe some of the practical challenges we have encountered working with small data sets, how these techniques might be applied to other areas of research, and what we are learning about the implications of these approaches for future study designs.
Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5–32.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2013). Bayesian Data Analysis (3rd edn). Chapman and Hall.
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Proceedings of the 31st International Conference on Neural Information Processing Systems, 4768–4777.
McElreath, R. (2020). Statistical Rethinking: A Bayesian Course with Examples in R and STAN (2nd edn). Chapman and Hall.
Share, D. L. (2025). Blueprint for a Universal Theory of Learning to Read: The Combinatorial Model. Reading Research Quarterly, 60(2), e603.
Snowling, M. J., & Hulme, C. (2025). The Reading Is Language Model: A Theoretical Framework for Language and Reading Development and Intervention.
Székely, G. J., Rizzo, M. L., & Bakirov, N. K. (2007). Measuring and testing dependence by correlation of distances. The Annals of Statistics, 35(6), 2769–2794.
Utilising new and emerging statistical and machine learning methods for developmental and educational research in Down syndrome
- Frank Buckley (Down Syndrome Education International, Cumbria, UK, Down Syndrome Education USA, CA, USA)
Correspondence: frank.buckley@dseinternational.org
Developmental and educational research in Down syndrome faces persistent challenges. Funding is limited, and many studies involve small samples, which reduces precision and increases the risk of false negative findings, especially when true effects are modest. Small samples also make it harder to estimate heterogeneous effects reliably, whether using frequentist or Bayesian approaches. Many studies recruit participants who may not be representative of the wider population, limiting generalisability, and replication remains limited.
Longitudinal evidence spanning multiple developmental periods is scarce. Much of the available literature is cross-sectional, providing snapshots that may be insufficient for understanding the developmental processes that generate observed outcomes over time (D’Souza & Karmiloff‐Smith, 2017).
Null hypothesis significance testing (NHST) remains the dominant analytic approach in developmental research and in much of the published Down syndrome literature. Yet, NHST is widely criticised as being prone to a number of problems, including multiple misinterpretations, arbitrary binary decisions, deflecting from estimates of practical importance, providing limited tools for quantifying evidence in favour of the null, and sensitivity to model assumptions (Christensen, 2005; Gigerenzer et al., 2004; Greenland et al., 2016; Kline, 2023; Meehl, 1967; Wasserstein & Lazar, 2016).
Here, based on our ongoing exploratory studies, we discuss the potential benefits of newer statistical and machine learning approaches for developmental and educational research in Down syndrome. We also discuss practicalities and obstacles to adoption, with an emphasis on transparent reporting and reproducible workflows.
Christensen, R. (2005). Testing Fisher, Neyman, Pearson, and Bayes. The American Statistician, 59(2), 121–126.
D’Souza, H., & Karmiloff‐Smith, A. (2017). Neurodevelopmental disorders. Wiley Interdisciplinary Reviews: Cognitive Science, 8(1–2), e1398.
Gigerenzer, G., Krauss, S., & Vitouch, O. (2004). The Null Ritual: What You Always Wanted to Know About Significance Testing but Were Afraid to Ask. In D. Kaplan, The SAGE Handbook of Quantitative Methodology for the SocialSciences (pp. 392–409). SAGE Publications.
Greenland, S., Senn, S. J., Rothman, K. J., Carlin, J. B., Poole, C., Goodman, S. N., & Altman, D. G. (2016). Statistical tests, P values, confidence intervals, and power: A guide to misinterpretations. European Journal of Epidemiology, 31(4), 337–350.
Kline, R. B. (2023). Principles and Practice of Structural Equation Modeling: Fifth Edition (5th Edition). The Guildford Press.
Meehl, P. E. (1967). Theory-Testing in Psychology and Physics: A Methodological Paradox. Philosophy of Science, 34(2), 103–115.
Wasserstein, R. L., & Lazar, N. A. (2016). The ASA Statement on p-Values: Context, Process, and Purpose. The American Statistician, 70(2), 129–133.
