Scope and Purpose

This current review is an update of the 2010 review⁵³ that supported the 2012 USPSTF recommendation to prevent falls among older adults.⁵² The USPSTF will use this report to update its recommendation. Our update includes all studies from the previous review that met our updated inclusion criteria as well as studies published since the previous review. There is a concurrent systematic review⁵⁴ supporting a separate USPSTF recommendation⁵⁵ addressing vitamin D supplementation for the purpose of preventing fractures.

Key Questions and Analytic Framework

We developed an Analytic Framework (Figure 1) and two key questions (KQs) to guide the literature search, data abstraction, and data synthesis.

Figure 1

Analytic Framework.

Data Sources and Searches

In addition to re-evaluating all studies included in the 2010 review,^53,56 we searched the following databases for relevant English-language literature published between January 1, 2010, and August 30, 2016: MEDLINE, PubMed publisher-supplied records, Cumulative Index for Nursing and Allied Health Literature, and Cochrane Central Register of Controlled Trials. We worked with a research librarian to develop our search strategy (Appendix B) which was peerreviewed by a second research librarian. We also examined the reference list of a previously published systematic review⁵⁷ to identify additional studies for inclusion. We supplemented our searches with suggestions from experts and articles identified through news and table-of-contents alerts. We also searched ClinicalTrials.gov and the World Health Organization International Clinical Trials Registry Platform (ICTRP) (www.who.int/ictrp) for ongoing trials. We imported the literature from these sources directly into EndNote^® X7 (Thomson Reuters, New York, NY). Since August 30, 2016, we have continued to conduct ongoing surveillance through article alerts and _targeted searches of high-impact journals to identify major studies published in the interim that may affect conclusions. The last surveillance was conducted on November 22, 2017 and identified a network meta-analysis on fall prevention interventions.⁵⁸ After reviewing the included studies from that meta-analysis, two relevant studies were identified that met our inclusion criteria but did not change the conclusions: one multifactorial intervention⁵⁹ and one environment modification intervention.⁶⁰

Study Selection

We developed criteria for including or excluding studies based on the previous review⁵³ and our understanding of the literature (Appendix B Table 1). We included randomized placebocontrolled trials (RCTs) and cluster RCTs for intervention studies. All harms were restricted to studies included for KQ1, with the exception of medications and supplements. For the harms of vitamin D, we expanded our criteria to include systematic reviews. The population of interest was community-dwelling older adults (aged ≥65 years), including those residing in independent living facilities. We excluded trials that specifically recruited participants with specific diagnoses (e.g., neurologic diagnoses like dementia, Parkinson’s disease, stroke) because those populations may require specialized approaches to preventing falls. We also included any older adults the study investigators determined were at high risk of falling. Because age and a host of other individual and environmental factors determine fall risk, control group fall rates were calculated and presented in results sections as another indicator of fall-risk status. Studies were required to have a primary or secondary aim of preventing falls or an aim related to it (e.g., fear of falling) and falls measured as a primary or secondary outcome. If a study did not have an aim of fall prevention or a related aim (e.g., pneumonia prevention and walking capacity, increasing physical activity levels) or did have a fall-related aim without measuring falls as a primary or secondary outcome, we excluded the study. Interventions that were feasible or referable from primary care were included; while many fall-prevention interventions are implemented in the community (e.g., exercise, medication management, environmental hazard reduction), primary care clinicians may have a role in referring their patients to these programs. The intervention descriptors and how they were combined were based on taxonomy developed by researchers from the Prevention of Falls Network Europe (ProFaNE) group³⁵:

Multifactorial: Interventions in which two or more intervention components were given to participants but the interventions were linked to each individual’s risk profile. Each participant received a unique combination of intervention components.

Single: Only one major intervention component was provided to participants. Included intervention components: Exercise, Medication (including Medication Management and Vitamin D), Psychological, Environment/Assistive Technology, Knowledge

Multiple: Interventions in which two or more intervention components were offered to every participant in the intervention group of the fall-prevention program. Included intervention components: Exercise, Medication (including Medication Management and Vitamin D), Psychological, Environment/Assistive Technology, Knowledge

Using the inclusion and exclusion criteria as a guide, two independent reviewers independently screened in abstrackr⁶¹ all records in the updated searches on the basis of the titles and abstracts. Subsequently, at least two reviewers assessed the full text of potentially relevant studies in DistillerSR (Evidence Partners, Ottawa, Canada), including all of the previously included studies, using a standard form that outlined the eligibility criteria. Disagreements were resolved through discussion and consensus. We kept detailed records of all included and excluded studies, including the reason for exclusion.

Quality Assessment and Data Abstraction

Two reviewers independently used USPSTF criteria⁶³ to assess the methodological quality of all eligible studies by using DistillerSR (Evidence Partners, Ottawa, Canada), including the studies from the 2010 review.^53,63 We assigned each study a quality rating of “good,” “fair,” or “poor” according to study design-specific criteria (Appendix B Table 2). Good-quality RCTs had adequate randomization procedures and allocation concealment, similar groups at baseline, welldefined interventions, reliable outcome measures, blinded outcome assessment, and low attrition (≥90% of participants had followup data, with a less than 10 percentage-point difference in loss to followup between groups), and they used conservative data substitution methods for missing data. Trials were given a quality rating of fair if they were unable to meet the majority of the good-quality criteria but were not of poor quality. Trials were rated as poor quality if attrition was greater than 40 percent or differed between groups by 20 percentage points, the falls outcome was self-reported solely by the participant with recall more than 6 months and no other outcome of interest was reported, or there was any other flaw that seriously affected internal validity, as agreed upon by the two independent reviewers.

We abstracted descriptive and outcome data from each included study (both the original and updated studies) into detailed abstraction forms using DistillerSR. One reviewer completed primary data abstraction and a secondary reviewer checked all data for accuracy and completeness. Data collection included general characteristics of the study (e.g., author, year, study design), characteristics of the sample (e.g., age and clinical characteristics of a population, setting, country), description of the intervention (e.g., type, provider, frequency, duration), methods to collect information on falls, and results. A study in which participants prospectively collected information (e.g., onto calendars, postcards, or diaries) about their falls and sent the information to the research team was referred to as “diary” collection. When multiple intervention and/or control groups were available, we abstracted the most intense intervention group and the control group most similar to no intervention or usual care. If at any point followup in a study fell below 60 percent, we did not abstract and analyze outcomes at or past that point. We attempted to contact authors when data reporting was incomplete or particular data points required clarification.

Data Synthesis and Analysis

We synthesized data separately for each KQ. Many outcomes did not allow for quantitative pooling due to the limited number of contributing studies, so those data are summarized narratively. For outcomes with enough contributing studies (at least 50% of the included studies for that intervention component with very low heterogeneity or 5 or more studies in the presence of nontrivial statistical heterogeneity), we ran random-effects meta-analyses using the method of DerSimonian and Laird⁶⁴ to calculate the pooled relative risks. We did not pool study data for studies with interventions categorized as “multiple” because the interventions were clinically heterogeneous. When available, we favored the author-reported relative risks over those we calculated. When authors did not report relative risks, we calculated a crude effect estimate. If a CI for a relative risk was not reported, we calculated it from the reported p-value.⁶⁵ Within each study, we selected the longest followup available for pooled analyses and figures. Data from other followup times are presented in tables. As noted above, only one intervention and one control arm for each intervention category were abstracted and included in the analysis.

We grouped our outcomes as follows: falls, injurious falls, fractures, people experiencing a fall, people experiencing an injurious fall, people experiencing a fracture, and mortality, people transitioning to institutionalized care, people hospitalized, quality of life, ADL, and IADL. All fall and fall-related injury outcomes were reported either as an incident event (where a person could contribute more than one event to the analysis, e.g., falls, fractures) or the number of persons experiencing the event (where a person could contribute only once to an analysis, regardless of the number of times the event occurred, e.g., people experiencing a fall, people experiencing an injurious fall). For injurious fall outcomes, we included minor or severe injuries resulting from a fall, falls resulting in medical care, or any fall-related outcome the author categorized as injurious. The most inclusive outcome was used in meta-analysis if multiple outcomes in that injury category were reported (e.g., fall-related injuries instead of fall-related hospital admissions). For studies that did not report a composite injury outcome, we used the most prevalent outcome (e.g., falls leading to an emergency department visit was selected over falls leading to hospital admission). The number of injurious falls analyzed in the forest plots included both the number of fall-related injuries and the number of falls resulting in injury as reported in the trials. For fracture outcomes, we first selected fall-related fractures, but if that outcome was not available, we included data on hip fractures and overall fractures, even if the study may not have reported if the fracture was associated with a fall.

In cases where a cluster RCT was used but the authors did not account for the nested nature of the data, we adjusted for the clustering effect by applying a design effect, which was based on an estimated average cluster size (i.e., the total number of randomized participants divided by the total number of clusters) and multiplied by an estimated intraclass correlation. We estimated the intraclass correlation to be 0.05.

We examined statistical heterogeneity among the pooled studies by applying standard χ² tests and estimated the proportion of total variability in point estimates by using the I² statistic.⁶⁶ We applied the Cochrane Collaboration’s rules of thumb for interpreting heterogeneity⁶⁷: less than 40 percent likely represents unimportant heterogeneity; 30 to 65 percent, moderate heterogeneity; 50 to 90 percent, substantial heterogeneity; and more than 75 percent, considerable heterogeneity. In addition, we generated funnel plots to evaluate small-study effects (a possible indication of publication bias) and ran the Egger test to assess the statistical significance of imbalance in study size and findings that suggest a pattern.⁶⁸

We investigated whether the heterogeneity among the main results (the outcome of falls and the outcome of people who fall) was associated with any prespecified population or intervention characteristics of the studies. First, we used visual displays and tables grouped or sorted by these potentially important characteristics. Specifically, we examined the recruitment setting (emergency department, clinic, or a combination), mean age, percentage female, risk of falls (high or average risk, as defined by the authors), fall rate of the control group or the percent falling, country (United States vs. others), and study quality (fair vs. good) as they related to the effect estimates. For exercise interventions, we also examined the duration and intensity, exercise components (e.g., balance, flexibility, strength), number of components, and format (group, individual, or both). On the basis of visual examination of forest plots, we used metaregression to test for potentially significant sorting variables or groups, namely the recruitment setting for the falls outcome for multifactorial interventions. Due to the general lack of statistically significant meta-regression results, we ordered forest plots alphabetically. We used Stata version 13.1 (Stata Corp LP, College Station, TX) for all quantitative analyses. All significance testing was two-sided. Results were considered statistically significant if the p-value was 0.05 or less.

Grading the Strength of the Body of Evidence

We graded the strength of the overall body of evidence for each key question. We adapted the Evidence-based Practice Center approach,⁶⁹ which is based on a system developed by the Grading of Recommendations Assessment, Development and Evaluation (GRADE) Working Group.⁷⁰ Our method explicitly addresses four of the five Evidence-based Practice Centerrequired domains: consistency (similarity of effect direction and size), precision (degree of certainty around an estimate), reporting bias (potential for bias related to publication, selective outcome reporting, or selective analysis reporting), and study quality (i.e., study limitations). We did not address the fifth required domain—directness—as it is implied in the structure of the key questions (i.e., pertains to whether the evidence links the interventions directly to a health outcome).

Consistency was rated as reasonably consistent, inconsistent, or not applicable (e.g., single study). Precision was rated as reasonably precise, imprecise, or not applicable (e.g., no evidence). Reporting bias was rated as suspected, undetected, or not applicable (e.g., when there is insufficient evidence for a particular outcome). Study quality reflects the quality ratings of the individual trials and indicates the degree to which the included studies for a given outcome have a high likelihood of adequate protection against bias. The body of evidence limitations field highlights important restrictions in answering the overall key question (e.g., lack of replication of interventions, nonreporting of outcomes important to patients).

We graded the overall strength of evidence as high, moderate, or low. “High” indicates high confidence that the evidence reflects the true effect and that further research is very unlikely to change our confidence in the estimate of effects. “Moderate” indicates moderate confidence that the evidence reflects the true effect and that further research may change our confidence in the estimate of effect and may change the estimate. “Low” indicates low confidence that the evidence reflects the true effect and that further research is likely to change our confidence in the estimate of effect and is likely to change the estimate. A grade of “insufficient” indicates that evidence is either unavailable or does not permit estimate of an effect. Two independent reviewers rated each key question according to consistency, precision, reporting bias, and overall strength of evidence grade. We resolved discrepancies through consensus discussion involving more reviewers.

Expert Review and Public Comment

A draft of the Analytic Framework, KQs, and inclusion and exclusion criteria was posted on the USPSTF Web site for public comment from August 6, 2015, through September 2, 2015. Minor changes were made to the inclusion and exclusion criteria to clarify the included populations, interventions, and settings. No major changes were made to the research plan that altered the scope of the review or our approach to synthesizing the evidence.

Invited content experts and federal partners reviewed a draft of this report. Their comments were presented to the USPSTF during its deliberation of the evidence and were considered in preparing the final evidence review. Additionally, a draft of this report was posted for public comment on the USPSTF Web site from September 26, 2017 through October 24, 2017. A few comments were received during this public comment period; no significant changes were made to the report based on these comments.

USPSTF Involvement

We worked with USPSTF liaisons at key points throughout this review to develop and refine the Analytic Framework and KQs and to resolve issues regarding the scope for the final evidence synthesis.

This research was funded by the Agency for Healthcare Research and Quality (AHRQ) under a contract to support the work of the USPSTF. AHRQ staff provided oversight for the project, coordinated systematic review work with other related topics in the portfolio, reviewed the draft report, and assisted in the external review of the draft evidence synthesis.