Screening to prevent fragility fractures among adults 40 years and older in primary care: protocol for a systematic review

Since individuals without prior fracture but at risk for incident fragility fracture are asymptomatic, screening should be able to identify those who are at greater risk of fracture and potential candidates for preventive intervention. Information from screening may be used, along with patient values and preferences, to inform decisions about treatment that might decrease future risk of fracture and related morbidity [45]. Thus, the aim of screening is not to detect the existence of osteoporosis but rather to reduce fracture-related burden of morbidity, mortality, and costs.

Screening to prevent fragility fractures involves a sequence of activities, not simply one test. The activities include a systematic offering of screening in a specified population of asymptomatic people with the intent to identify those at increased risk for fractures in order to provide preventive treatment and improve health outcomes. The effectiveness is ideally measured over the entire population being offered the screening program, relying upon trials that directly evaluate long-term outcomes from screening compared with no screening, or between different screening programs, in primary care populations. Inferences about the effectiveness of screening programs to prevent fragility fractures, however, have mostly relied upon indirect data (linked evidence) from individual components of an end-to-end screening program. These indirect data include information about the accuracy and performance of risk assessment tools and the effectiveness of treatment among people at increased risk for fracture.

International guidelines (Additional file 1) vary in their current recommendations on screening approaches, based on the country-specific population burden of fragility fractures and mortality, competing societal priorities, and resource availability [9]. Several screening strategies exist in clinical practice, and in most cases, recommendations differ by population group based on sex, menopausal status, and age. For women 65 years or older (or postmenopausal), many North American organizations recommend either only using BMD assessment [46, 47] or assessing BMD in all women and integrating this with other clinical risk factors into an absolute fracture risk for treatment decision-making [1, 10, 12, 48]. More common in European guidelines for this population group (and oftentimes across all populations >50 years) is an assessment of absolute risk using clinical factors before deciding whether to further stratify risk by assessing BMD [49‐51]. For women who are not menopausal (or < 65 years) and for men, many recommendations are to first assess risk based on clinical factors and use BMD in those considered at-risk. In some approaches, BMD assessment is also recommended in all men of a certain age category (e.g., ≥ 50 [12], ≥ 65 [1], or ≥ 70 years [52, 53]). Shared decision-making is incorporated in few recommendations; the Institute for Clinical Systems Improvement recommends shared decision-making about BMD testing, but only in specific population subgroups: men 70 years and older; adults with a known condition associated with low bone mass/bone loss; and organ transplant patients [54]. The European Society of Endocrinology’s guidelines for postmenopausal women recommend that patient values and preferences be considered when deciding who to treat [55]. When BMD testing follows a clinical risk assessment, it is not always clear if this is used independently or integrated (as possible) into a total clinical risk score. Moreover, in some jurisdictions, the indication for BMD testing may be restricted to instances where the absolute fracture risk is predicted to be intermediate to moderate (i.e., close to the level where treatment would be considered), whereby further information from the test may better inform treatment decisions. In these guidelines (e.g., United Kingdom), BMD testing would not be indicated when absolute risk is either well below or far above treatment thresholds [56]. The definition of the intermediate risk category may be determined based on other considerations such as resource availability and funding, and the risk profile of the target population.

There are at least 12 published tools to predict fracture risk [16, 19]. These tools combine an individual’s known clinical risk factors for fragility fracture into a single total estimation of absolute fracture risk over a certain time period (commonly 5 or 10 years) [16]. The main difference between various tools is the number of factors assessed and how these factors are weighted in the models. Certain prediction tools (e.g., FRAX) require calibration to the population context in which they will be used to account for differences in fracture incidence and mortality across geographic regions [57]. Not all tools have been validated in populations outside of their derivation cohort, limiting transferability of these risk prediction models [58]. Some tools (e.g., FRAX, Garvan) allow for, but do not require, inclusion of BMD results; others (e.g., CAROC) require BMD. Tools generally incorporate easily obtained clinical risk measures, but may be enhanced by simple arithmetic procedures (e.g., falls history or level of exposure to glucocorticoids added to FRAX [56]).

Most guidelines recommend that when BMD is assessed it should be measured at the femoral neck via DXA [1, 19, 50, 59], because measurements at this site can be incorporated into many risk assessment tools [1, 19, 50, 59], and the use of multiple sites does not appear to improve the accuracy of fracture prediction [60, 61]. Lumbar spine BMD is also commonly reported and may be used by some practitioners in their decision-making on fracture risk assessment. For example, procedures have been developed and endorsed by the International Society for Clinical Densitometry and International Osteoporosis Foundation [62], to adjust FRAX probabilities when large discordance exists between lumbar spine and femoral neck BMD [63‐65]. Some DXA instruments also offer vertebral fracture assessment, which can be used as a complement to BMD assessment to identify existing vertebral fractures [24]. Though these fractures are generally asymptomatic, clinicians should be aware that emerging evidence suggests that they strongly and independently predict incident clinical fracture outcomes (including hip fracture), independent of FRAX score [24, 41]. Further evidence, controlled for important confounding variables (e.g., hip fracture), is needed to confirm these findings. Current Canadian guidelines recommend vertebral fracture assessment via DXA or spine radiography when other clinical evidence suggests that a vertebral fracture is likely to be present (e.g., height loss) and may be used among those in moderate risk categories to help inform treatment decisions [1]. Analysis of data from the Canadian Multicentre Osteoporosis Study [66] indicates that Jiang et al.’s algorithm-based qualitative approach [67], which focuses on depression of the vertebral endplate, is the preferred approach to defining vertebral fractures (compared to the widely used Genant semiquantitative method [68]). Other less common BMD assessment methods (e.g., quantitative ultrasound, peripheral DXA, quantitative computed tomography scan, bone turnover markers) are typically used outside the scope of a population-based primary screening program [19, 59, 69].

Many systematic reviews on fracture risk assessment tools have focused on discrimination (i.e., ability to distinguish between people who develop fractures versus those who do not; measured by area under the receiver operating characteristics curve and other accuracy measures [e.g., sensitivity, specificity] relying on particular thresholds) as their primary, or only, outcome. On the other hand, primary care providers and patients may find calibration (i.e., accuracy of absolute risk prediction within a population) to be a more clinically meaningful measure to inform shared decisions about management.

Treatment thresholds vary considerably across countries and may take into account variation in population-specific risk of fracture and mortality [57], competing health care priorities, patient willingness-to-pay for fracture-related health care, resource availability (e.g., access to BMD assessment tools), and pre-existing reimbursement criteria [9, 56]. The United States National Osteoporosis Foundation [70] recommends initiating pharmacological treatment in individuals with osteoporosis or with low BMD (T-score between − 1.0 and − 2.5, osteopenia) and either a 10-year hip fracture probability ≥ 3% or a 10-year major osteoporosis-related fracture probability ≥ 20% (using FRAX). This decision was supported by a cost-effectiveness analysis based on assumptions from one-step BMD screening followed by treatment with a generic bisphosphonate (assumed relative fracture reduction of 35%), and a willingness-to-pay threshold of $60,000 per quality-adjusted life-year gained [71, 72].

Canadian guidelines [1, 73], as well as those developed in several other countries (e.g., Austria [74], Greece [75], Hungary [76], Malaysia [77, 78], Mexico [79], the Philippines [80], Saudi Arabia [81], Poland [82], Slovakia [83], Slovenia [84], Spain [85‐87], Taiwan [88], Thailand [89]), that are based on country-specific FRAX models, use a fixed 20% 10-year probability of major osteoporotic fracture as a treatment threshold [56]. In many (but not all) cases, the choice of the 20% intervention threshold is without a specific rationale, but instead based on the threshold used in the United States. Some guidelines also use a fixed 3% 10-year hip fracture probability as an alternative intervention threshold [56]. Another less common approach is to use intervention thresholds that increase with age [56]. The threshold is based on the rationale that because individuals with a prior fracture can be considered for treatment without the need for further assessment, other individuals of the same age with a similar fracture risk but no prior fracture should also be eligible [51]. Recent strategies adopt a hybrid approach (i.e., incorporating both fixed and age-dependent intervention thresholds) [51, 90, 91]. For example, the National Osteoporosis Guideline Group for the United Kingdom recommends that the treatment threshold increase with age for individuals up to 70 years to align with the level of risk associated with a prior fracture (ranges from approximately 7 to 24% 10-year probability of fracture; equivalent to the risk probability of a woman of the same age with a prior fragility fracture) [51]. After age 70, a fixed threshold is used to account for the reduced sensitivity of the risk probability algorithm for those without a prior fracture, which becomes most apparent at advanced age [51].

Treatment decisions may best be based on patient preferences, including their competing priorities and assessment of the relative importance of benefits and harms, and shared decision-making between patients and their healthcare providers [92]. Although treatment efficacy appears to be an important variable when choosing between different treatments [92], a major factor impacting the effectiveness of any treatment, and therefore screening program, is medication adherence. A study in the United States showed that close to 30% of patients provided with a prescription for osteoporosis treatment do not fill their prescription [93]. Of those initiating treatment, only half are still taking their medication at 1 year [94]. Predominant factors affecting adherence include dosing frequency, side effects of medications, costs, and lack of knowledge about the implications of osteoporosis [94]. One study conducted in the United States showed that in 2009, half of women (mean age 69 years; 30–40% with osteoporosis or prior fracture; perceived risk for 10-year fracture about 40%) who were provided information regarding fracture risks and treatment risks and benefits reported that they would accept prescription osteoporosis treatment at the threshold currently recommended by national physician treatment guidelines; 18% of the women would not accept treatment even at 50% fracture risk levels [95]. Willingness to accept treatment increased at higher levels of fracture risk and was higher in those with greater acceptance of the risks of medications [95]. There is large variation between patients regarding their treatment preferences, which supports a shared decision-making approach in place of recommended treatment thresholds based on fracture risk [92].

According to the 2010 Osteoporosis Canada guideline, for postmenopausal women, the first-line therapy is either one of three bisphosphonates (i.e., alendronate, risedronate or zoledronic acid), denosumab, or raloxifene (a selective estrogen receptor modulator) [1]. Hormone therapy may be considered for women experiencing vasomotor symptoms [1], and etidronate (another bisphosphonate) may be considered for those who are intolerant of first-line therapies [96]. As of October 2013, calcitonin is no longer approved by Health Canada for the treatment of osteoporosis due to concern about the increased risk of malignancies associated with the drug [97]. Moreover, systematic reviews evaluating etidronate have failed to demonstrate an impact on fracture reduction [19, 98] and this medication is used infrequently in Canada. For men, Osteoporosis Canada recommends bisphosphonates (i.e., alendronate, risedronate, zoledronic acid) as first-line therapy [1]. More recent guidelines from the American College of Physicians (2017) [99] and American Association of Clinical Endocrinologists/American College of Endocrinology (2016) [100] recommend alendronate, risedronate, zoledronic acid, and denosumab as first-line treatments for preventing fractures. Furthermore, use of hormone therapy for the prevention of fractures in postmenopausal women is not recommended [101].

In 2018, the United States Preventive Services Task Force (USPSTF) reviewed the effects of pharmacological treatments on preventing fragility fractures, using data from studies where the majority of the participants had no prior fracture [19]. Compared with placebo, moderate-certainty evidence was found for bisphosphonates in reducing the primary outcomes of vertebral and nonvertebral fractures in women, although low-certainty evidence found no difference in reducing the secondary outcome of hip fracture alone [19]. To explain this, it has been reported that only one of the three trials with hip fracture as an outcome was adequately powered to detect a significant difference [102]. Moreover, only one of the trials reporting on bisphosphonates was conducted in men [103]. One trial (n = 7868) of denosumab compared with placebo showed a decrease in vertebral, nonvertebral, and hip fractures in women [19]; the certainty of evidence was assessed as low for these outcomes. Few trials reported data on all clinical fractures or clinical vertebral fractures, and the reviewers did not assess the certainty of evidence for these outcomes. Trials have based their inclusion criteria on BMD (levels ranging from osteopenic to osteoporotic) rather than absolute risk for fractures, such that findings may not be applicable to those with high risk for fractures but with normal BMD. Similarly, beneficial effects may be obscured by inclusion of patients with low BMD but without higher fracture risk.

Non-pharmacological interventions (e.g., vitamin D, calcium, exercise, falls prevention) are considered as adjuncts to pharmacological treatment in primary care [1] and are considered to be out of scope for the current review.

The screening tests may expose individuals to small amounts of radiation from DXA scans (with or without vertebral fracture assessment/spinal radiography) [104]. Costs for the patient and healthcare system include the time, effort, and expense related to attending appointments and the resources used to screen in clinical settings, to organize and perform tests, and to interpret results [19]. Patients may not always fully understand the meaning of risk assessment results, nor the consequences of an asymptomatic finding that cannot easily be conceptualized [105, 106]. Individuals undergoing screening, and those who perceive their predicted risk for fragility fracture to be high, may experience anxiety and feelings of uncertainty [105, 107]. These people may become overly cautious, limit their activities, and become less independent [107, 108]. They may feel stigmatized if they are labelled as “old” or “frail” [105]. However, quantitative data from a recent (n = 12,483) randomized controlled trial of screening in the United Kingdom examined the effect of the screening on anxiety and quality of life and suggested that the risk of these harms is small [4]. Individuals who were screened had levels of anxiety and quality of life that were very similar to those who were not screened [4]. One reason for this finding may be related to patient attitudes and beliefs. For example, a qualitative study of patients aged 50 and older in Canada showed that individuals perceived fractures and osteoporosis not to be serious health conditions and believed that they had negligible impact [109]. More research is needed to better understand the factors that influence a patient’s desire to have or avoid screening for osteoporosis-related fracture risk.

Individuals can experience physical and psychological harm if their risk of fracture is over- or under-estimated (e.g., due to inaccurate measurement or interpretation of BMD or risk assessment results). When a patient is identified as having a higher risk of fracture than they truly have, they may experience unnecessary anxiety, and these individuals may be subjected to unneeded treatments that can have adverse effects with little or no benefit. Alternatively, a patient may be identified as having a lower risk of fracture than they truly have, which may be especially likely when BMD alone is used to estimate risk [110]. Based on false reassurance, these individuals may not make useful lifestyle modifications. They may also not have access to available treatments that could ultimately decrease their risk of fracture when screening program eligibility criteria are based on fracture risk rather than shared decision-making.

Two systematic reviews have assessed adverse events for multiple bisphosphonates as well as for denosumab. Based on moderate-certainty evidence, the USPSTF’s 2018 systematic review did not find increased discontinuation rates due to the composite outcome “any adverse events,” upper gastrointestinal events, or serious adverse events for bisphosphonates over placebo. Insufficient evidence was found for cardiovascular events, osteonecrosis of the jaw, and atypical femoral fractures. For denosumab, in women, there was insufficient evidence for discontinuation due to adverse events, and low-certainty evidence found no significant increase in serious adverse events and serious infections [19]. The evidence used for this review was limited due to its focus on randomized controlled trials and studies of patients without previous fracture or secondary causes of osteoporosis, even though it may be argued that the harms of treatment are unlikely to differ substantially between somewhat different patient populations. Using a broader patient population, and thus a larger and more comprehensive evidence base, a 2012 systematic review by the Agency for Healthcare Research and Quality [94] reported different findings. For example, the review found high-certainty evidence for an increased risk of mild upper gastrointestinal events (e.g., acid reflux, nausea, vomiting) with alendronate, low-certainty evidence of an increased risk for bisphosphonate-related osteonecrosis of the jaw and atypical femoral fractures, and high-certainty evidence that denosumab increases infections [94]. Authors of both reviews considered the evidence insufficient for serious cardiovascular events (e.g., atrial fibrillation, acute coronary syndrome) and cancers (e.g., esophageal, gastrointestinal) [19, 94, 99]. For several outcomes (e.g., serious cardiovascular events), observational evidence was only considered when no trials existed. More recently, evidence has emerged to suggest the possibility of rapid bone loss or risk of multiple vertebral fractures due to rebound increased bone resorption after discontinuation of treatment with anti-RANKL antibodies (i.e., denosumab) [111]. However, supportive evidence of these effects from extensions of clinical trials is currently limited [112, 113].

Although the result of the screening test—a risk for future fracture—is not a diagnosis of a condition or a disease, it has similar consequences because certain risk levels lead to labeling of patients as “at high risk,” and at one point a certain threshold has to be chosen by care providers either to serve as a threshold for treatment or to start a conversation with a patient about treatment. Overdiagnosed patients may be considered to be those who are deemed to be at excess risk of fracture—either according to a set threshold or based on shared decision-making—but who would never have known they were at risk because, without screening, they would not have experienced a fracture. Using a shared decision-making perspective, overdiagnosis leading to overtreatment may be conceptualized as patients who had a risk assessment and following shared decision-making decided to start treatment but would never have sustained a fragility fracture regardless of screening efforts.

KQ1b: Does the effectiveness of screening to prevent fragility fractures vary by screening program type (i.e., 1 step vs 2 step) or risk assessment tool?

KQ2: How accurate are screening tests at predicting fractures among adults ≥ 40 years?

KQ3a: What are the benefits of pharmacologic treatments to prevent fragility fractures among adults ≥ 40 years?

KQ3b: What are the harms of pharmacologic treatments to prevent fragility fractures among adults ≥ 40 years?

KQ4: For patients ≥ 40 years, what is the acceptability* of screening and/or initiating treatment to prevent fragility fractures when considering the possible benefits and harms from screening and/or treatment?

*Acceptability indicators include positive attitudes, intentions, willingness, and uptake

Figure 1 shows the analytical framework that depicts the population, KQs and outcomes, as well as key screening characteristics that will be considered. A staged approach to the evidence will be undertaken.

At the first stage, we will focus on identifying direct evidence from screening for fragility fracture on benefits and harms that are patient-oriented and either critical or important to clinical decision-making (KQ1a). We will prioritize evidence from randomized controlled trials, as these studies generally provide the highest internal validity. We will also consider evidence from controlled clinical trials (i.e., that includes a comparison [control] group and contains all of the key components of a true experimental design other than randomization: assignment of groups is determined by study design, and the administration of screening and endpoint ascertainment follows a protocol) if certainty in the evidence from randomized controlled trials is limited and poses a barrier to the development of recommendations, and the Task Force believes that further evidence from controlled clinical trials may influence their recommendations. We expect that this could occur due to limited available evidence overall or lack of evidence for selected subgroups (e.g., by age, sex, or different risk assessment approaches). If evidence for KQ1a indicates that screening for fragility fracture reduces fracture risk, we will examine whether this effectiveness varies by screening approach (e.g., 1 step vs. 2 step) or by risk assessment tool (KQ1b). We will review evidence related to the acceptability of screening and/or treatment (KQ4), as well as indirect evidence on the accuracy of screening tests (KQ2), concurrently with KQ1. We will proceed with KQ3 (treatment benefits and harms) only if the Task Force believes that further indirect evidence would influence their recommendations.

We will inspect studies for methodological and clinical heterogeneity, and if appropriate, for KQ1 (benefits and harms of screening) and KQ3 (benefits and harms of treatment), we will pool data for each outcome via pairwise meta-analysis using the DerSimonian and Laird random effects model [136] in Review Manager (version 5.3, The Cochrane Collaboration, Copenhagen, Denmark). In the case of rare events (< 1% event rate, e.g., adverse events), we will instead consider using the Peto odds ratio [137] method in order to provide a less biased effect estimate [138]. We will pool the data from randomized controlled trials and controlled clinical trials separately from observational studies. We will report risk ratios (RRs) or rate ratios between groups and corresponding 95% CIs for dichotomous or count data, respectively. When zero events are reported for at least one of the intervention groups, we will report the risk difference (RD) and 95% CI. For continuous outcomes, we will report the mean difference (MD) and 95% CI when all data are collected using the same measurement tool, or the standardized mean difference (SMD) and 95% CI when a variety of tools are used to describe a similar construct. When data for multiple time-points are available, we will choose to include data from the longest length of follow-up within the following categories: 6 to 12 months, 13 months to 5 years, 6 to 10 years, > 10 years.

If appropriate, we may pool data from studies of different bisphosphonates together, then analyze each bisphosphonate separately (i.e., as a subgroup) and compare estimates of effect for individual drugs to the class of bisphosphonates. For the clinical fracture and serious adverse event outcomes, we will preferentially analyze dichotomous data using a RR (primary outcome). If this is not reported by the authors, we will also consider analyzing count data using a rate ratio (surrogate outcome). The only instance in which we may consider combining dichotomous and count data in one analysis (assuming RR and rate ratios are very similar) is after clinical and statistical consultation confirms that events are rare enough and would be highly likely to have occurred in distinct patients and only once during follow-up.

We will calculate absolute effects for each outcome-comparison by applying the risk ratio from the meta-analysis to the median control group event rates from the included studies. If statistically significant, we will also calculate numbers needed to screen or treat.

If appropriate, for KQ2 (accuracy of screening tests), we will pool model calibration data for each identified screening method separately using the restricted maximum likelihood estimation approach and the Hartun-Knapp-Sidnick-Jonkman correction to derive 95% CIs [139, 140]. We will rescale total observed versus expected fracture event ratios and their variance (standard error (SE)) on the natural log scale prior to entering these into meta-analysis to achieve approximate normality [141‐143]. We will report the observed versus expected fracture ratio and 95% CIs for calibration. When studies report calibration slope and/or calibration within categories (e.g., quintiles of risk), we will summarize the overall results narratively rather than extracting data for each category. We will consider model calibration to be “good” when the summary observed vs. expected fracture ratio is between 0.8 and 1.2 (i.e., there are 20% more or less events than are expected) [128].

Because discrimination outcomes (e.g., C-statistic/area under the receiver operating characteristics curve, sensitivity, specificity, positive and negative predictive values) were not rated as important by the Task Force, these will not be systematically reviewed by the Evidence Review Synthesis Centre. We will, however, present model discrimination information narratively and/or in tables as reported in the USPSTF review. We will consider model discrimination to be “good” when the summary C-statistic is > 0.75 (where 0.5 indicates no concordance and 1.0 indicates perfect concordance) [98].

We expect to perform a narrative synthesis given the likely heterogeneity in study designs, exposure characteristics (e.g., differences between studies in presentation of information on screening or treatment effects), populations, and outcomes reported across the studies. We will generally follow the guidance developed by Popay et al. [144] recognizing that our question of acceptability differs to some extent from questions about intervention effects or implementation factors. We will begin with a preliminary synthesis of the findings across studies and follow this with an exploration of the relationships between the studies, focusing on our population and exposure subgroups of interest (see Table 4) as well as other factors such as methodological quality. We will attempt to provide a best estimate of the acceptability of screening and/or treatment initiation (e.g., by people having information on the benefits and harms in absolute terms and with similar magnitude as thought to be applicable to the population of those at general risk for fracture), as well as factors that may impact the acceptability.

If data required for meta-analysis are not directly reported by individual studies, whenever possible, we will compute or estimate these using other statistics presented in the studies, based on available guidance [128, 145]. If necessary, we will substitute means with medians. If standard deviations (SDs) or SEs are not reported, we will compute these from CIs, z- or t-statistics, or p values [146]. When computing SDs for change from baseline values, we will assume a correlation of 0.5 unless data pertaining to the actual correlation are available. If none of these data are available, we will approximate the SD using the range or interquartile range [147]. If it is not possible to compute or estimate the SD from other available data and the number of missing SDs is small, we will impute the mean SD from other studies in the meta-analysis, as this approach has been shown to minimally impact average effect estimates and their 95% CIs [148]. For KQ2 (accuracy of screening tests), we will estimate the log of the observed versus expected fracture ratio and its variance using available data (e.g., observed vs. expected fracture ratio, observed and expected events, observed and expected outcome probabilities, calibration-in-the-large) and standard formulae [128, 149, 150].

Our approach to subgroup analysis for KQs 1–3 will be to first report on within-study subgroup data for our pre-specified subgroups of interest (see Tables 1, 2, 3). Within-study findings are usually not available across all studies and can be difficult to conceptualize across a body of evidence. Thus, we will further explore heterogeneity in effects (i.e., in direction or magnitude of effects) using an exploratory between-study approach whereby we will categorize studies into subgroups; for population subgroups, we will use a large majority (e.g., ≥ 80% of participants) for classifying groups. To assess differences across subgroups, we will use appropriate statistical techniques (e.g., meta-regression if more than 8–10 studies) or stratify the meta-analysis by subgroup. We will interpret the plausibility of subgroup differences cautiously using available guidance, without relying on statistical significance [151, 152]. To assist in our interpretation of plausibility for KQ2 (accuracy of screening tests), we will calculate the 95% prediction interval as an estimate of the range of potential model performance in a new validation study and present these values along with the results of meta-analyses [128, 153].

When appropriate, we will perform sensitivity analyses (e.g., variability in overall or domain-specific risk of bias across studies, study design [randomized versus nonrandomized trials], differences in outcome definitions or adherence rates between studies) by removing certain studies from the analysis to see whether findings are different. For KQ1 and KQ3, we will perform sensitivity analyses if we have uncertainty about combining count and binary data. If substantial heterogeneity is present and cannot be plausibly explained via subgroup or sensitivity analyses, we may decide to suppress the pooled estimate of effect and instead present the findings of the comparison narratively.

When meta-analyses include at least eight studies of varying size, we will test for small study bias by visually inspecting funnel plots for asymmetry and quantitatively using Egger’s regression test (KQ1 and KQ3) [154] or the funnel inverse variance test (KQ2) [155] (significant at P < 0.10).