Giters

mattlee821 / systematic_review_MR_adiposity

Data cam be browsed at https://mattlee.shinyapps.io/SR_meta_analysis_browser/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Systematic review: What has the application of Mendelian randomization informed us about the causal relevance of adiposity and health outcomes?

We performed a systematic review of adiposity on all health outcomes. Adiposity was defined broadly and included body mass index, birth weight, waist hip ratio and more terms.

Start here

Detailed info

EMBASE and MEDLINE were searched from inception (EMBASE = 1974; MEDLINE = 1946) until February 18th 2019 using detailed search strategies including free text and controlled vocabulary terms. The pre-print service bioRxiv was searched from inception (November 2013) until February 18th 2019. However, due to the limited search functionality and inability to include Boolean operators (‘AND’, ‘OR’, ‘NOT’) in bioRxiv searches, a restricted search strategy using four free text terms in four independent searches was used: ‘Mendelian randomization’, ‘Mendelian randomisation’, ‘causal inference’, and ‘causal analysis’.

The search strategy included synonyms for both adiposity and MR terms. For adiposity measures, this was to ensure searches returned all possible instances in which a measure of adiposity was used. For MR, synonyms were used as the term ‘Mendelian randomization’ has only been formalised recently and many early studies would have either been unaware they were performing an instrumental variable analysis or would have called the method something else.

After de-duplication, the titles and abstracts of all remaining articles from EMBASE, MEDLINE, and bioRxiv had their titles and abstracts screened by two independent reviewers (MAL and LJM) using Rayyan. Each reviewer screened all articles and discrepancies at this stage were resolved through discussion between the two reviewers. Studies from EMBASE, MEDLINE, and bioRxiv meeting the pre-defined inclusion criteria (see below) were combined and, in instances where the bioRxiv study had been published and this was identified in either the EMBASE or MEDLINE search, the bioRxiv version of the study was excluded. The full texts of the combined study dataset had their full text screened by the two reviewers.

For title and abstract screening and for full text screening, articles must have met the following pre-defined inclusion criteria:

  1. Be written in English
  2. Be available in full text (or in the case of conference abstracts, the authors must be contactable to obtain the relevant data)
  3. Be published in a peer-reviewed journal or bioRxiv
  4. Use MR methodology to investigate the causal effect of adiposity on any outcome
    1. Adiposity was considered to be any measure which aimed to assess the amount of adipose tissue an individual possessed
    2. If a study focused on adiposity alongside other exposures, the effect of each adiposity measure will be reported separately, if available, and report the joint effect with these other exposures, if not available.
    3. articles in which an MR approach is used but not explicitly called ‘Mendelian randomization’ will be included. More specifically, any study in which genetic variants are used as instrumental variables or the direct association between a genetic variant and outcome is employed will be eligible, provided it meets the other inclusion criteria.

In the first instance, data extraction was performed by nine reviewers (see Contributions), with articles split evenly between them, using a data extraction form and data extraction manual. One reviewer dropped out and their articles were split evenly between MAL and CAH. Once all articles had been reviewed, two reviewers (see contributions) extracted data on all articles they did not review in the first instance. The same two reviewers then checked all extracted data for discrepancies which were resolved through a third review of individual articles.

Articles included in data extraction may contain more than one relevant MR analysis. As such, study/studies refers to the MR analysis/analyses within an article. The following data were extracted from each articles studies: exposure(s), outcome(s), study design and sample characteristics, genetic variant and instrumental variable selection, MR methodology, sensitivity analysis, and causal estimates. Where relevant data was not reported by the article, “Not discussed” was entered into the data extraction form.

Once data extraction was completed, three additional columns were added to summarise the type of outcome being studied: column 1 (“outcome”) was used as a general categorisation of all outcomes across articles (e.g., the outcome “ER- breast cancer” would have the value “breast cancer”); column 2 (“outcome info”) reported the outcome-specific information that distinguished outcomes within categories defined in column 1 (e.g., column 2 would contain the value “ER-” for the same breast cancer example); and column 3 (“outcome group”) categorised outcomes more generally than values defined in column 1 (e.g., the breast cancer example would be categorised as “cancer”). Outcome categories were assigned based on prior biological knowledge and aimed to collapse the large number of outcomes. This could be achieved differently for some outcomes, for example smoking could go in a respiratory category or a behavioural category. Where there were few outcomes to make a category, they were grouped into an other category. This will include outcomes such as mortality, disease counts, epigenetic marker etc.

Quality assessment

There is currently no risk of bias tool to assess the quality of MR articles. We adapted the quality assessment tool used by Mamluk et al (2020) and assessed studies on a 3-point scale (low = 3, medium = 2, high = 1) across 12 questions. These 12 questions included the five used by Mamluk et al., one of which related to bias due to selection of participants, which we split into two questions for exposures and outcomes to accommodate two-sample MR analyses. In addition, questions for instrumental variable association, sample overlap, whether the study performed sensitivity analyses and whether these were biased, descriptive data, data availability (data missingness), and statistical parameters were included. Given no formal risk of bias tool exists, quality assessment here was not used as a prerequisite for inclusion/exclusion in the meta-analyses. Rather, it was used to supplement the meta-analyses and aid interpretation.

Meta-analysis

To identify studies which could be meta-analysed, a set of rules were used. These rules ensured that the exposure and outcome were consistent across studies, but also that there was no population overlap between the outcomes of different studies or between the outcomes and exposure of different studies. Sample overlap can induce bias in MR studies. Where there was overlap between the outcome of one study and the outcome of another study, or where there was overlap between the exposure of one study and the outcome of another study, the study with the larger sample size was retained. Excluding studies with overlapping outcomes or overlapping exposures and outcomes would involve including non-independent data and result in overly precise estimates. Finally, studies were excluded based on whether the MR method was comparable and then on whether the units where compatible with one another (e.g., where both studies reported a standard deviation increase in BMI). Studies which had overlapping exposure populations were included as the risk of bias is low. For completeness, studies were not excluded based on the quality assessment score, but are discussed later in this chapter when interpreting the meta-analysis findings. Inclusion and exclusion information is available as a text file.

Meta-analysis was performed using the meta package in R and the function metagen() using an inverse variance weighted random-effects model using estimates and standard errors. For binary outcomes, the relevant summary method was used for odds ratios, risk ratios, and hazard ratios etc. For continuous outcomes, the mean difference was used for the underlying summary method. For both binary and continuous outcomes, the Hartung and Knapp method to adjust confidence intervals to reflect uncertainty in the estimation of between-study heterogeneity was used. Between study variance was estimated for all meta-analyses using the Paule-Mandel estimator.

Narrative synthesis

A narrative synthesis was performed for all articles that were not included in the meta-analyses. The outcome categories used to categorise outcomes during data extraction were used to guide the synthesis. The direction of effect estimates across outcome categories were summarised across the exposures used for these analyses.

Author contributions

Matthew A Lee - conceptualization, data curation, formal analysis, investigation, methodology, project administration, software, supervision, validation, visualization, writing (original draft preparation), writing (review & editing)

Charlie Hatcher - data curation, investigation, validation, writing (review & editing)

Luke J Mcguiness - data curation, methodology, writing (review & editing)

Nancy McBride - data curation, writing (review & editing)

Thomas Battram - data curation, writing (review & editing)

Wenxin Wang - data curation, writing (review & editing)

Si Fiang - data curation, writing (review & editing)

Kaitlin H Wade - resources, data curation, supervision, writing (review & editing)

laura J Corbin - resources, supervision, writing (review & editing)

Nicholas J Timpson - resources, supervision, writing (review & editing)

Screening, data extraction, and analyses

Title and abstract screening - MAL and LJM

Full text screening - MAL and LJM

Data extraction was performed independently by 8 reviewers - MAL, CH, LJM, NM, TB, WW, SF, KHW

Data extraction was performed a second time, independently by MAL and CH

Analyses - MAL

About

Data cam be browsed at https://mattlee.shinyapps.io/SR_meta_analysis_browser/

Languages

Language:R 99.9%Language:CSS 0.1%