Session: Methodological Innovations in Longitudinal Population Studies
A new perspective on the Pubertal Development Scale (PDS) for epidemiological studies (Luciana Maestri Karoleski, Federal University of Sao Paulo)
Importance: Assessing puberty is vital for detecting neuroendocrine disorders, particularly in populations with limited access to specialized care. Objective: To evaluate the Pubertal Development Scale (PDS) as a tool for identifying developmental subgroups in epidemiological research. Methodology: Using female data from the Millennium Cohort Study (age 11, n=5,367; age 14, n=5,716), we employed Confirmatory Factor Analysis (CFA) to verify construct validity and Latent Class Analysis (LCA) to identify developmental classes. Results: A four-class model best fit the data at both ages: pre-pubertal, early, mid, and late pubertal development. Conclusions: The PDS is a valid, reliable self-assessment instrument. It effectively captures pubertal changes, identifies risk subgroups, and serves as a robust screening tool for large-scale clinical and population studies.
Extending a data-driven approach to non-response in the Millennium Cohort Study: predictors of non-response at sweep 8 (Martina Narayanan, UCL Centre for Longitudinal Studies)
Non-response in longitudinal studies can reduce statistical power and introduce bias. Building on a data-driven framework applied to sweeps 1–7 of the UK Millennium Cohort Study (MCS), we examine predictors of non-response at sweep 8 (age 23). This is the first sweep to collect data from cohort members only, marking a key transition in study design and participation. Using data from earlier sweeps, we identify predictors of non-response at sweep 8. Preliminary findings show continuity with earlier sweeps, with predictors of sweep 8 non-response including previous sweep non-participation, lower parental social class, parental non-participation in general elections, and lower cognitive test scores. New predictors also emerge, including male sex and lower school readiness and reading ability. These findings highlight changing mechanisms of non-response for sweep 8 data.
Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks (Mack Nixon, UCL)
Frontier LLMs excel at coding but their adoption in quantitative research on longitudinal population studies is constrained by governance requirements prohibiting sensitive data transmission to external services. We investigate lightweight agents powered by locally deployable open-weight models for AI-assisted analysis satisfying these constraints. We introduce an open-source framework evaluating AI agents on a key bottleneck in longitudinal research: data preparation. It comprises a ground-truth dataset (human-curated preparation using the Next Steps dataset), task definitions targeting data cleaning and variable harmonisation, and automated routines evaluating LLM-produced R code and data outputs. Benchmarking models from frontier to lightweight, open-weight models on consumer-grade hardware shows promise towards AI-assisted data preparation in governance-restricted settings.
Unconditional and conditional incentives in a UK birth cohort: a randomised 3×2 factorial experiment embedded in the Early Life Cohort Feasibility Study (Konstantinos Tsigaridis, UCL Centre for Longitudinal Studies)
Declining response rates threaten longitudinal birth cohort studies, yet evidence on optimal incentives for families with young children is limited. We embedded a 3×2 factorial experiment in the Early Life Cohort Feasibility Study (ELC-FS), randomising 3,628 UK families with infants aged 9–13 months to one of three unconditional incentives (none, baby bib, £5 cash) crossed with one of two conditional vouchers (£10 or £20). Household response was 47.3%. Adjusted logistic regression showed £5 cash significantly outperformed both the bib (OR=1.29, p=0.006) and no incentive (OR=1.21, p=0.039); doubling the voucher gave weaker evidence of improvement (OR=1.13, p=0.062). No interaction or differential effects by deprivation, maternal age, ethnicity or UK nation emerged. Small monetary unconditional incentives outperform non-monetary gifts in birth cohort recruitment.
