PurposeDirected acyclic graphs (DAGs) are critical in epidemiology and public health research for guiding study design and minimizing bias. Yet, developing DAGs for causal inference requires substantial domain knowledge. Given the vast amounts of training data for large language models (LLMs), this study assesses the effectiveness of prompt engineering for LLMs to generate DAGs that depict causal relationships in population health using OpenAI’s GPT-4o and GPT-o1.MethodsWe consider a hypothetical study on statins vs no treatment for prevention of cardiovascular disease in a general adult population. We assessed four types of prompt engineering strategies: zero-shot, one-shot, instruction based, and chain of thought (CoT) prompts. Generated DAGs were assessed based on consistency, acyclicity, accuracy of sources, completeness (based on ASCVD risk score criteria), and adherence to the prompt.ResultsWe found that all generated DAGs were acyclic, except for one run using the instruction-based prompt. Additionally, more than half of the DAGs included 6/7 of the ASCVD criteria, though race was absent from all. Overall, CoT resulted in the most complete DAGs and one-shot provided the most consistency across runs and adherence to the task in the prompt. The zero-shot prompt performed notably better on GPT-o1 compared to GPT-4o, consistently providing justifications and sources for variable inclusion.ConclusionWhile the findings suggest that LLMs have a baseline capacity to generate DAGs that adhere to basic epidemiological conventions, we also found several limitations including lack of justification, systematic omission of race, and frequent source hallucination, highlighting the need for human oversight and expertise. We conclude that contemporary LLMs cannot replace a domain expert’s judgment but may serve as a brainstorming or pre-analysis tool for DAG development when guided by well-engineered prompts.

Quasi-experimental approaches are used routinely in clinical epidemiological research to enable causal treatment/intervention effect estimation from observational data. One such approach is the regression discontinuity design (RDD): a method to estimate a causal effect in situations when a treatment or intervention is assigned to individuals according to an externally defined decision rule, based on a continuous, individual-level “assignment variable”. RDDs were developed originally for use in econometrics but their use in clinical epidemiology is increasing, particularly with the widening availability of electronic health records and the use of rule-based treatment/intervention decisions. In particular, an RDD can be a useful method to assess the effectiveness of clinical decision making. In this paper, we provide an overview of the RDD, describing the method and key assumptions that permit its use in observational clinical data. We outline the common continuity-based and local randomisation RDDs and demonstrate how these can be fitted in both R and Stata. A worked example is presented of an RDD to estimate the treatment effect of statins on low density lipoprotein (LDL) cholesterol level, when statins are prescribed according to a rule based on a cardiovascular disease risk score.

IntroductionComorbidity between disorders is pervasive, and its relationship to the main conditions under investigation needs to be addressed for robust causal inference. However, many clinical etiologic studies still fail to capitalize on the theoretical advancements and improved recommendations regarding covariate adjustment in this context. Specifically, studies often lack explicit causal assumptions about the role of comorbidity in exposure–outcome relationships, potentially leading to inappropriate accounting for comorbid conditions and resulting in biased effect estimates. This study aims to explore common causal structures involving comorbidity and provide guidance for handling it in etiologic research.MethodsWe use Directed Acyclic Graphs (DAGs) to depict six causal scenarios involving comorbidity as a confounder, mediator, collider, or consequence of the exposure or outcome, illustrated with real-world clinical examples. Simulations were conducted across 5,000 iterations for each scenario, assessing the impact of conditioning on comorbidity under four effect measures (risk difference, odds ratio, risk ratio, and mean difference). Bias was evaluated by comparing adjusted and unadjusted effect estimates to the true values.ResultsThe impact of conditioning on comorbidity varied by its causal role. Adjusting for comorbidity mitigated bias when it acted as a confounder but introduced bias when it was a mediator or collider. In instances where comorbidity was a consequence of either the exposure or outcome, the decision to adjust depended on the research objectives and could vary across effect measures.DiscussionExplicit causal assumptions are essential for selecting appropriate analytical strategies in etiologic research. This study provides practical guidance on analytical handling of the measures of comorbidity, highlighting the need for study design and analysis to align with research objectives. Future work should address more complex causal structures and other methodological challenges.

This manuscript describes the design and protocol of the Biobank for Metabolic Syndrome Consequences (BMSC), a prospective cohort study in Northwest China. Metabolic syndrome (MetS) is characterized by a group of interrelated disorders, including abdominal obesity, hyperglycemia, hypertension, and dyslipidemia. The presence of three or more of these conditions markedly increases the risk of multiple chronic diseases and mortality. The pathophysiology and natural course of MetS and its consequences are insufficiently understood. To improve our understanding, longitudinal research that combines biomarkers with longitudinal data measured over multiple time points is imperative. The BMSC, launched in August 2021 and still ongoing, is a prospective observational study of 2000 Chinese participants aged 18 to 75 years with MetS or relevant disorders living in the Northwest of China. At baseline survey, data on sociodemography, disease history, behavior and lifestyle, and mental health are collected by a structured questionnaire. The anthropometry is conducted by trained researchers. Fasting peripheral venous blood, urine, stool, and hair samples are collected according to standardized protocols. Extensive physical examinations are conducted in specific subgroups. Participants will be followed up every 3 months for at least 5 years for the incidence of MetS-related outcomes, such as cardiovascular disease, with clinical data and biological samples being collected at intervals similar to the baseline. These findings may contribute to improved prevention, early diagnosis, and personalized treatment of MetS-related conditions.

PurposeTo investigate the impact of light-moderate (up to 20 g/day) alcohol consumption on incidence of postmenopausal breast, kidney, lung, pancreatic, colorectal, postmenopausal ovarian and postmenopausal endometrial cancer among women.MethodsParticipants were 70,932 women aged 41–70 years, randomly recruited in the Norwegian Women and Health (NOWAC) cohort study from 1996 to 2004. We included women who reported that they consumed alcohol. Only postmenopausal women (N = 32,735) were included in the analyses for female cancers. Multivariable Cox proportional hazard models were used to estimate hazard ratios (HR) and 95% confidence intervals (CI).ResultsThe mean follow-up was 19 years. The estimated hazard ratio (HR) from each additional 12g/day of alcohol consumption for postmenopausal breast cancer was 1.20 (95% confidence intervals CI: 1.03 to 1.41), and for kidney cancer 0.42 (95% CI: 0.24 to 0.75). The corresponding estimates for postmenopausal breast cancer among women who used menopausal hormone therapy (MHT) were HR = 1.27, 95% CI: 1.05 to 1.54, and among women who never used MHT were HR = 1.12, 95% CI: 0.86 to 1.47. Compared to alcohol consumption of <3.5 g/day, consumption of 3.5–10 g/day revealed for lung cancer inverse association with risk of lung cancer among women who consumed primarily wine (HR = 0.65, 95% CI; 0.43 to 0.88), but not among other drinkers (HR = 1.10, 95% CI; 0.88 to 1.31). No associations were confined for pancreatic, colorectal, ovarian and endometrial cancers.ConclusionWomen drinking light-moderate alcohol level had a higher risk of postmenopausal breast cancer and a lower risk of kidney cancer incidence. Our results do not support the threshold of up to 1 drink/day as a safe limit for breast cancer, especially for postmenopausal women who use MHT. The inverse relationship found for lung cancer could be explained by the healthier lifestyle correlated with this light-moderate drinking.

PurposeThis study aimed to investigate the association between the use of sodium–glucose cotransporter 2 inhibitors (SGLT2i) and the risk of developing depression in patients with type 2 diabetes mellitus.Patients and MethodsThis study used Taiwan’s National Health Insurance Database and an active comparator new-user design to evaluate depression risk among 551,917 patients initiating SGLT2i or DPP4i between 2016 and 2018. The primary outcome was depression incidence, assessed over a three-year follow-up. Stratified Cox regression models were used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) between groups.ResultsAmong new SGLT2i users, 3255 cases of depression occurred (7.18 per 1000 person-years) versus 7190 cases among DPP4i users (10.12 per 1000 person-years). After adjustment for demographic and clinical covariates, SGLT2i use was consistently associated with a lower risk of depression in both the full cohort (adjusted HR = 0.77; 95% CI: 0.73–0.80) and the propensity score–matched cohort (adjusted HR = 0.77; 95% CI: 0.74–0.81). The association remained robust in multiple sensitivity analyses and across clinical subgroups.ConclusionSGLT2i use was associated with a reduced risk of depression among individuals with type 2 diabetes mellitus. These findings suggest potential neuropsychiatric benefits of SGLT2i and support further investigation into their broader therapeutic implications.