Introduction
Similarities to and differences with common SRMA tools
Purpose | Tool | Intended effect | Description |
---|---|---|---|
Specification of search strategy | PICOS statement | Specify and clarify search criteria and eligibility criteria to guide the search based on the research question | Authors specify the: |
Patient population of interest | |||
Intervention/exposure | |||
Comparator/control | |||
Outcomes of interest | |||
Study types eligible for inclusion | |||
Pre-registration study protocol | For example: PROSPERO | Reduce bias caused by changes authors make based on findings Increase transparency | Registration in public database of study protocol. Contains items including: |
Search strategy, | |||
PICOS statement, | |||
Context/setting, | |||
Data extraction items, | |||
Planned risk of bias assessment, | |||
Strategy for evidence synthesis | |||
PRISMA-P | Guide reporting by providing a checklist for preferred reporting items for systematic reviews and meta-analysis protocols | Authors specify the | |
Funding and support sources and roles | |||
Rationale and objectives (PICO) | |||
Eligibility criteria | |||
Information sources | |||
Search strategy | |||
Study records (data management, selection process, data collection process | |||
Data items | |||
Outcomes and prioritization | |||
Risk of bias in individual studies | |||
Data synthesis | |||
Meta-bias(es) | |||
Confidence in cumulative evidence (such as GRADE) | |||
Risk of Bias assessment | ROB2 | Identify potential sources and level of risk of bias in RCTs | Identify risk of bias in the following domains and categorize these as low, some concerns, or high risk of bias |
1: bias from randomization process, | |||
2: bias due to deviations from intended interventions, | |||
3: bias due to missing outcome data, | |||
4: bias in measurement of the outcome, | |||
5: bias in the selection of the reported result | |||
ROBINS-I | Identify potential sources and level of risk of bias in non-randomized studies of interventions | Identify risk of bias in the following domains and categorize these as low, some concerns, or high risk of bias | |
1: pre-/at-intervention: participant selection, classification of interventions, | |||
2: post-intervention: missing data, bias in measurement of outcomes, bias in selection of reported result | |||
QUADAS-2 | Identify potential sources and level of risk of bias in diagnostic accuracy studies | Specify the relevant: | |
Population (presentation and prior test), | |||
Index test, | |||
Target condition, | |||
Reference standard | |||
Identify risk of bias in the following domains and categorize these as low, some concerns, or high risk of bias | |||
Patient selection (consecutive or randomized, case control, exclusions), | |||
Index test (blinding to reference standard result, pre specified cutoff), | |||
Reference standard (likely to correctly classify disease, blinded to index test result), | |||
Flow and timing (appropriate time interval, all participants receive same reference standard) |
How DAGs can improve the effectiveness of SRMAs
How DAGs can improve SRMA efficiency
Two examples
Research question | What is the effect of mindfulness-based interventions on medical students’ perceived stress? | What is the effect of maternal smoking on birthweight in newborns? |
---|---|---|
Population, intervention or exposure, comparator, outcome, study type | P: Medical students I: Mindfulness-based Interventions C: No mindfulness-based intervention O: Perceived Stress S: Randomized Controlled Trials | P: Newborns E: Maternal smoking C: No Maternal Smoking O: Birthweight S: Observational studies |
Potential DAG |
* | |
How our DAG may have aided our study design | We see that due to randomization the backdoor path from the treatment assignment to factors L to the outcome is normally closed, meaning we do not require adjustment. However, when the study has loss to follow up, we are conditioning on only the students for whom we observe follow-up, which opens the backdoor path by conditioning on C. If for example students lost to follow up have different stress levels than those who remain, this could introduce bias. To increase effectiveness, we could choose to only include studies that have near complete follow up or that accurately deal with loss to follow-up As it is difficult to blind students to whether they are in the control or the mindfulness-based intervention, this opens additional potential sources of bias. We could decide to perform a sensitivity analysis that explores the impact of only including blinded studies in our meta-analysis In non-randomized trials, there would be an arrow between factors L, such as age/sex/motivation/etc., and treatment. That is because we expect these factors to influence a persons’ choice of doing mindfulness-based interventions spontaneously, but we also think these factors influence their perceived stress. We assume we won’t have access to all potential influencing factors we deem relevant. We therefore decide to exclude studies that are non-randomized. We also show that we assume students’ gained attention control and cognitive reappraisal act as a mediator, which lies on the causal pathway between the exposure and outcome that fully explains the relationship. We therefore decide not to collect data on this parameter | Given our research question we will, due to ethical concerns, likely need to search for observational cohort studies When preparing for our SRMA we notice several studies adjusted for Gestational Age. We draw a DAG to visualize this relationship. Babies that are born earlier have a lower birthweight (arrow G–Y). If smoking causes babies to be born earlier (E–G), you are taking away part of the effect of smoking, ie. the part that is mediated through Gestational Age. We therefore choose to exclude papers that use Gestational Age as a confounder for this study (note that adjustments for G may be appropriate in other research questions) By adding an arrow between smoking and reported smoking (E*), we illustrate to our audience that we expect that there will be measurement error in our study, as individuals tend to provide incomplete information about behaviors that are typically perceived as detrimental to health We have additionally identified several potential confounders including SES, co-morbidity and age that could affect the relationship between smoking and birthweight, we therefore decide to include these items in our data extraction sheet and/or take note of how individual studies have addressed them |