The Outcomes That Change the Verdict
This chapter explains how to read evidence about fall-prevention interventions without allowing one attractive number to decide the whole question. For community-dwelling older adults, the difference between fewer falls, fewer people who fall, fewer injurious falls, and fewer fractures is not technical decoration; it changes what an intervention has actually been shown to do. The goal is to connect each estimate to its denominator, uncertainty, patient importance, and applicability.
A fall count and a faller count answer different questions
Suppose two groups each contain 100 older adults. In one group, a few people fall repeatedly. In the other, more people fall once but nobody falls repeatedly. The groups could have different total numbers of falls and different numbers of people who fall. An intervention that prevents repeat events may reduce the fall rate without changing the proportion of people who experience at least one fall. An intervention that prevents some people from falling at all may reduce the number of fallers even if the total count is less affected.
This is why the reviews separate outcomes that are often blended in everyday language. “Falls” usually counts events, so one person can contribute more than one event. “People with one or more falls,” or fallers, counts people once. “Injurious falls” adds a consequence to the event, while “people with an injurious fall” asks how many individuals experienced that consequence. Fractures are rarer still and may require larger studies or longer follow-up to estimate reliably.
The distinction is visible in the USPSTF evidence review. Across 29 exercise trials, exercise was associated with a lower rate of incident falls: incidence rate ratio (IRR) 0.85, with a 95% confidence interval from 0.75 to 0.96. It was also associated with a lower risk of having at least one fall: relative risk (RR) 0.92, 95% CI 0.87 to 0.98. The same review found a lower rate of injurious falls, IRR 0.84, 95% CI 0.74 to 0.95, but not a statistically significant reduction in the risk that an individual experienced an injurious fall, RR 0.90, 95% CI 0.79 to 1.02. These are related results, not interchangeable ones.
The multifactorial evidence shows the same interpretive problem in a different direction. The pooled incidence rate for falls was lower, IRR 0.84, 95% CI 0.74 to 0.95. But the risk of having at least one fall was not statistically significantly lower, RR 0.96, 95% CI 0.91 to 1.02. The estimates for injurious falls and fractures also did not establish a statistically significant reduction. It would be too strong to say that multifactorial programs do nothing. It is more accurate to say that the clearest pooled benefit was limited to the count of falls, and that the evidence was less consistent for the outcomes involving people, injuries, and fractures.
The diagram is a reminder to ask “what was counted?” before asking “did it work?” It is also a warning against treating a result that is not statistically significant as proof of no effect. A confidence interval that crosses the point of no difference says that the data are compatible with little or no benefit as well as with some benefit; the width of the interval indicates how precisely the study or pooled analysis distinguishes those possibilities.
Magnitude needs a denominator and a measure of uncertainty
An IRR compares the number of events accumulated over time between groups. It is useful when people can experience multiple falls. An RR compares the probability that a person experiences an outcome, usually once during the study period. Neither measure alone says how many events a program prevents in an actual population. A relative effect must be interpreted against a baseline risk.
A simple hypothetical example shows why. If an intervention produced an RR of 0.92, an untreated group with a 40% risk of at least one fall would move to an estimated 36.8% risk—a reduction of 3.2 percentage points, or about 32 fewer people per 1,000. With the same RR, a group starting at 10% would move to 9.2%, a reduction of only 0.8 percentage points, or about 8 fewer people per 1,000. The relative result is the same; the practical number of people affected is not. This is why a relative estimate without a meaningful baseline can make a modest effect look larger or smaller than it is for the population being considered.
The USPSTF review supplied one hypothetical absolute frame. Based on national fall rates, exercise was estimated to prevent about 107 falls, 22 people experiencing a fall, and 27 injurious falls per 1,000 people treated. Multifactorial intervention was estimated to prevent about 114 falls per 1,000, but the interval around that estimate ranged from 36 to 186. These projections are not individual guarantees and do not establish that multifactorial intervention prevents more falls than exercise. They illustrate two points instead: baseline risk changes the number of events that can be prevented, and an imprecise estimate can span a wide range of practical results.
A confidence interval has two jobs in this context. It shows the uncertainty around the estimated effect, and it helps reveal whether the data can rule out effects that would matter in practice. An interval clustered tightly around no difference provides a different kind of evidence from an interval that stretches from a meaningful benefit to a meaningful harm. Statistical significance is therefore not the only question. The direction, plausible range, baseline risk, outcome importance, and consequences of being wrong all matter.
The pooled fall estimates also had substantial heterogeneity. In the USPSTF review, the I² statistic was 82.7% for exercise and 85.0% for multifactorial interventions. I² describes how much variation among observed estimates is not explained by sampling error under the analysis used; it does not identify the cause of the variation. The studies differed in populations, settings, exercise content, supervision, duration, assessment, and usual care. The reviewers examined several possible explanations but did not find a single variable that accounted for the heterogeneity.
This matters for the title’s question. A similar average effect across two categories does not mean the categories are equally reliable in every setting. If a package contains many routes to benefit, it may also contain many routes to variation: different assessors, referrals, providers, home environments, and adherence patterns. A pooled average can describe the evidence base while still leaving the reader uncertain about which version of the package will work in a particular context.
The appropriate response is not to discard pooled evidence. It is to state the claim at the level the evidence supports. The USPSTF review supports saying that both exercise and multifactorial interventions were associated with fewer falls in multiple good-quality trials. It supports saying that exercise showed the most consistent statistically significant benefit across several fall-related outcomes in that review. It does not support saying that every exercise protocol is superior, that multifactorial care cannot prevent injuries, or that a relative effect will transfer unchanged to a lower-risk population.
What a network meta-analysis adds
The second review broadens the comparison by including 219 randomized trials and 167,864 participants for benefits and harms. It coded 59 intervention nodes and ran eight network meta-analyses. A conventional pairwise meta-analysis can directly compare exercise with usual care, for example. A network meta-analysis can also use a connected chain of comparisons: if exercise has been compared with usual care and tai chi has been compared with exercise, the network can use both direct and indirect evidence to estimate tai chi relative to usual care.
That extra reach is valuable when direct comparisons are scarce. It lets researchers examine intervention attributes such as supervision, duration, individual versus group delivery, and home-hazard assessment in a shared analytical framework. The review judged the overall assumptions of transitivity and coherence adequate. Transitivity means that the studies are similar enough in important effect modifiers—such as population, risk, setting, and follow-up—for indirect comparisons to be meaningful. Coherence means that direct and indirect evidence do not show major disagreement.
Neither assumption turns indirect evidence into direct evidence. A network can be well connected while some nodes have few studies, imprecise estimates, or limited outcome data. In this review, only about half of the nodes in each network had at least low-certainty evidence for benefit. For the outcome of people who fell, 18 of 57 nodes had moderate-certainty evidence. For fractures, function, quality of life, and long-term-care admission, many interventions had very-low-certainty evidence, often because the relevant data were missing or sparse.
The review therefore did not rely on a simple league table. It ranked interventions by the number of outcomes showing benefit and the certainty of those findings, rather than treating a surface ranking such as SUCRA as a final answer. This is an important methodological choice. An intervention with a large estimated effect but a very wide interval and very-low-certainty evidence should not automatically outrank a smaller, more consistently supported effect.
A network result also inherits the way interventions were coded. If one review treats a program that gives exercise to everyone as multifactorial and another treats it as mixed, the same trial can contribute to different categories. The network is not only a mathematical object; it is a set of decisions about which components belong together and which differences are important enough to separate. That is why the architecture learned in the previous chapter remains necessary when interpreting a sophisticated comparison.
The network findings reinforce rather than erase the central distinction. The strongest patterns were concentrated in supervised, long-duration balance and resistance exercise, group tai chi, and selected education, cognitive-behavioral, whole-body-vibration, and mixed interventions. No needs-based multifactorial intervention by itself showed moderate-certainty evidence for benefit across the outcomes considered. The review also concluded that adding other interventions to exercise did not appear to substantially increase benefits in general. That is a statement about the evidence and its certainty, not a law that add-ons can never help.
Patient importance can reorder the apparent winners
A reduction in falls is relevant, but it is not the final measure of value. The Canadian review assessed how much different outcomes reduced health-related quality of life. Admission to long-term care had the highest estimated disutility, about 1 in one low-certainty study, though that estimate should not be treated as a universal valuation. Hip-fracture disutility was about 0.53 in the first three months after injury and 0.16 at 12 months. For non-hip fractures, the estimates were about 0.57 early and 0.19 at 12 months. The estimated disutility of a fall over the preceding year was about 0.09, and functional impairment in at least one activity of daily living was about 0.12.
These values do not mean that a fall is unimportant or that every person would rank outcomes identically. They show why an intervention that reduces frequent falls but has not demonstrated a reduction in fractures cannot be judged solely by the number of events prevented. Fractures were less frequently measured and often had lower-certainty evidence, so the clinically most important endpoint is also one of the harder endpoints to study. No included study measured the disutility of an injurious fall, leaving another gap between what trials count and what patients may experience.
This creates a three-part reading discipline. First, identify the outcome and whether it counts events or people. Second, inspect the size and uncertainty of the estimate, including baseline risk and heterogeneity. Third, ask how much the outcome matters relative to fracture, function, and independence. A broader program may look attractive because it reduces several kinds of risk, but that promise is not the same as evidence that it changes the most consequential outcomes.
The two reviews can differ in detail without one invalidating the other. The USPSTF review focuses on 83 trials and emphasizes pairwise, primary-care-feasible evidence for exercise and multifactorial interventions. The Canadian review includes a broader intervention set, a different coding taxonomy, indirect network comparisons, outcome valuation, and patient preferences. The Canadian authors note that some reviews classify packages as multifactorial that they classify as mixed, and that conclusions can change with classification, follow-up, and the way certainty is assessed. Both reviews support a benefit signal for exercise and some reduction in fall rates with multifactorial interventions. Neither supports the blanket idea that a more complex package must be superior.
Once the verdict is separated into outcome, magnitude, certainty, and patient importance, exercise is no longer merely the “simple” option. It becomes a specific intervention whose content and delivery can be examined. The next question is what balance, resistance, supervision, and duration are doing—and why the evidence signal is stronger for some exercise designs than for the label alone.