Dataset Information

Thirteen Questions About Using Machine Learning in Causal Research (You Won't Believe the Answer to Number 10!).

ABSTRACT: Machine learning is gaining prominence in the health sciences, where much of its use has focused on data-driven prediction. However, machine learning can also be embedded within causal analyses, potentially reducing biases arising from model misspecification. Using a question-and-answer format, we provide an introduction and orientation for epidemiologists interested in using machine learning but concerned about potential bias or loss of rigor due to use of "black box" models. We conclude with sample software code that may lower the barrier to entry to using these techniques.

SUBMITTER: Mooney SJ

PROVIDER: S-EPMC8555423 | biostudies-literature | 2021 Aug

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Thirteen Questions About Using Machine Learning in Causal Research (You Won't Believe the Answer to Number 10!).

Mooney Stephen J SJ Keil Alexander P AP Westreich Daniel J DJ

American journal of epidemiology 20210801 8

Machine learning is gaining prominence in the health sciences, where much of its use has focused on data-driven prediction. However, machine learning can also be embedded within causal analyses, potentially reducing biases arising from model misspecification. Using a question-and-answer format, we provide an introduction and orientation for epidemiologists interested in using machine learning but concerned about potential bias or loss of rigor due to use of "black box" models. We conclude with s ...[more]

PMID: 33751024

Similar Datasets

Project description:BackgroundThe huge amount of biomedical-molecular data increasingly produced is providing scientists with potentially valuable information. Yet, such data quantity makes difficult to find and extract those data that are most reliable and most related to the biomedical questions to be answered, which are increasingly complex and often involve many different biomedical-molecular aspects. Such questions can be addressed only by comprehensively searching and exploring different types of data, which frequently are ordered and provided by different data sources. Search Computing has been proposed for the management and integration of ranked results from heterogeneous search services. Here, we present its novel application to the explorative search of distributed biomedical-molecular data and the integration of the search results to answer complex biomedical questions.ResultsA set of available bioinformatics search services has been modelled and registered in the Search Computing framework, and a Bioinformatics Search Computing application (Bio-SeCo) using such services has been created and made publicly available at http://www.bioinformatics.deib.polimi.it/bio-seco/seco/. It offers an integrated environment which eases search, exploration and ranking-aware combination of heterogeneous data provided by the available registered services, and supplies global results that can support answering complex multi-topic biomedical questions.ConclusionsBy using Bio-SeCo, scientists can explore the very large and very heterogeneous biomedical-molecular data available. They can easily make different explorative search attempts, inspect obtained results, select the most appropriate, expand or refine them and move forward and backward in the construction of a global complex biomedical query on multiple distributed sources that could eventually find the most relevant results. Thus, it provides an extremely useful automated support for exploratory integrated bio search, which is fundamental for Life Science data driven knowledge discovery.

Project description:AimsAt the basis of many important research questions is causality - does X causally impact Y? For behavioural and psychiatric traits, answering such questions can be particularly challenging, as they are highly complex and multifactorial. 'Triangulation' refers to prospectively choosing, conducting and integrating several methods to investigate a specific causal question. If different methods, with different sources of bias, all indicate a causal effect, the finding is much less likely to be spurious. While triangulation can be a powerful approach, its interpretation differs across (sub)fields and there are no formal guidelines. Here, we aim to provide clarity and guidance around the process of triangulation for behavioural and psychiatric epidemiology, so that results of existing triangulation studies can be better interpreted, and new triangulation studies better designed.MethodsWe first introduce the concept of triangulation and how it is applied in epidemiological investigations of behavioural and psychiatric traits. Next, we put forth a systematic step-by-step guide, that can be used to design a triangulation study (accompanied by a worked example). Finally, we provide important general recommendations for future studies.ResultsWhile the literature contains varying interpretations, triangulation generally refers to an investigation that assesses the robustness of a potential causal finding by explicitly combining different approaches. This may include multiple types of statistical methods, the same method applied in multiple samples, or multiple different measurements of the variable(s) of interest. In behavioural and psychiatric epidemiology, triangulation commonly includes prospective cohort studies, natural experiments and/or genetically informative designs (including the increasingly popular method of Mendelian randomization). The guide that we propose aids the planning and interpreting of triangulation by prompting crucial considerations. Broadly, its steps are as follows: determine your causal question, draw a directed acyclic graph, identify available resources and samples, identify suitable methodological approaches, further specify the causal question for each method, explicate the effects of potential biases and, pre-specify expected results. We illustrated the guide's use by considering the question: 'Does maternal tobacco smoking during pregnancy cause offspring depression?'.ConclusionsIn the current era of big data, and with increasing (public) availability of large-scale datasets, triangulation will become increasingly relevant in identifying robust risk factors for adverse mental health outcomes. Our hope is that this review and guide will provide clarity and direction, as well as stimulate more researchers to apply triangulation to causal questions around behavioural and psychiatric traits.

Dataset Information

Thirteen Questions About Using Machine Learning in Causal Research (You Won't Believe the Answer to Number 10!).

Publications

Thirteen Questions About Using Machine Learning in Causal Research (You Won't Believe the Answer to Number 10!).

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets