Case Definition

Numerous definitions have been used in fall studies and have been reviewed (Kellogg International Work Group on the Prevention of Falls by the Elderly, 1987; Hauer, Lamb, Jorstad, Todd, & Becker, Reference Hauer, Lamb, Jorstad, Todd and Becker2006; Zecevic, Salmoni, Speechley, & Vandervoort, Reference Zecevic, Salmoni, Speechley and Vandervoort2006). The multiplicity of definitions undoubtedly accounts for some of the heterogeneity seen in prevalence and incidence estimates, as well as in the strength of associations with risk factors noted across various studies (e.g., American Geriatrics Society, British Geriatrics Society, & American Academy of Orthopaedic Surgeons Panel on Falls Prevention, 2001). It is not surprising and is in fact appropriate that the earliest studies relied on a tacit understanding of the term “fall”, rather than attempt a formal operational definition. For example, in one of the earliest published studies, by Droller (Reference Droller1955), a sample of older people was simply interviewed on “the question of falling”. Only by using a broad definition could the early authors catalogue the range of events that older adults identify as falls, which in turn informed the first attempts at typologic classification (e.g., Droller, Reference Droller1955; Sheldon, Reference Sheldon1960). This early work formed the basis for operational definitions in later studies that allowed greater specificity of “geriatric syndrome” falls, (e.g., by excluding falls due to overwhelming external forces that would cause a fall in a young healthy person (Nevitt, Cummings, Kidd, & Black, Reference Nevitt, Cummings, Kidd and Black1989; Tinetti, Speechley, & Ginter, Reference Tinetti, Speechley and Ginter1988) and by focusing the analysis on those who fell repeatedly over the period of follow-up (Nevitt et al., Reference Nevitt, Cummings, Kidd and Black1989).

Occurrence

Fall occurrence can be expressed in numerous ways, and the resulting proliferation can make it difficult to compare estimates across studies. If the unit of analysis is the person, then the proportion experiencing at least one fall during a specified time interval indicates the fall risk of the observed group. In spite of differences in sampling and measurement (i.e., fall case definition), three seminal prospective cohort studies published between 1988 and 1993 produced similar annual fall risk estimates of 29 per cent (O’Loughlin, Robitaille, Boivin, & Suissa, Reference O’Loughlin, Robitaille, Boivin and Suissa1993), 32 per cent (Tinetti et al., Reference Tinetti, Speechley and Ginter1988), and 35 per cent (Campbell, Borrie, & Spears, Reference Campbell, Borrie and Spears1989). Interestingly, Droller’s non-prospective 1955 study produced a risk estimate of 35 per cent (Droller, Reference Droller1955).

Among those who fall, many are repeat fallers. The recurrent nature of falling means that either falls or fallers can be treated as the unit of analysis. Early case series such as Sheldon’s (Reference Sheldon1960) study sampled on the basis of 202 fallers but the analysis was based on 500 falls. The analysis of falls clustered with variable frequency within persons is problematic because they are not statistically independent events. As well, the highly right-skewed distribution of the number of falls suggests that means, and parametric statistics such as t-tests, are inappropriate.

Because the number of falls is a discrete variable, one valid approach to the recurrent nature of falling has been a simple dichotomy of one-time versus recurrent (two or more) fallers. Tinetti et al. (Reference Tinetti, Speechley and Ginter1988) reported that 14.8 per cent fell once and 17.3 per cent fell two or more times. In a sample on average about four years younger, O’Loughlin et al. (Reference O’Loughlin, Robitaille, Boivin and Suissa1993) reported that 17.6 per cent fell once and 11.5 per cent fell two or more times. Another use of this dichotomy has been to restrict the sample to those with a recent fall history, and the analysis to recurrent fallers (≥ 2 falls) ascertained prospectively, as done by Nevitt et al. (Reference Nevitt, Cummings, Kidd and Black1989).

Another main occurrence measure used for recurrent events is cumulative incidence rate, defined as the total number of events divided by the total person-time at risk. O’Loughlin reported an annual incidence rate of 41.4 falls per 1,000 person-months, compared to 56.8 falls per 1,000 person-months calculable from data reported by Campbell et al. (Reference Campbell, Borrie and Spears1989). Both fall risk and fall rate increase with age, and are higher in those in long-term care and other institutional settings (Rubenstein, Reference Rubenstein2006).

The challenges of analyzing fall data were well summarized 20 years ago by Cumming, Kelsey, and Nevitt (Reference Cumming, Kelsey and Nevitt1990). Recent statistical contributions to the literature include more-sophisticated methods to analyze both recurrent falls and person-time at risk using Poisson regression for discrete count variables, and variants of Cox proportional hazard survival analysis, either for the study of risk factors (Gill, Zou, Jones, & Speechley, Reference Gill, Zou, Jones and Speechley2009) or in fall prevention trials (Robertson, Campbell, & Herbison, Reference Robertson, Campbell and Herbison2005).

Causes of Falls I: Contributory Conditions and Typologies

Some early studies placed a heavy emphasis on understanding falls using a typology approach. For example, Droller (Reference Droller1955) distinguished between clinical conditions such as vertigo (by far the most common single factor he observed), and environmental conditions such as loose rugs. Droller also distinguished between “liability to fall” and “opportunity to fall” (now predisposing and precipitating factors) and clearly recognized that the causes of a particular fall were “frequently multiple and difficult to unravel”. Another early author (Sheldon, Reference Tinetti, McAvay and Claus1960) found it “surprisingly easy” to categorize 500 falls. For example, Sheldon reckoned that 171 (34%) were “accidental falls” that happened on stairs, due to slipping, and so on. Sheldon also recognized that his categories were not “self-contained”: vertigo was judged the main cause of 37 of the “non-accidental” falls as well as 12 of the 171 “accidental” falls.

By 1978, Isaacs had pointed out the limitations of using a mixture of descriptive terms (e.g., slips and trips, dizziness, drop attacks) and pathological terms (e.g., postural hypotension) and pointed out that attributing a fall to a “loss of balance” is a tautology, not an etiology (Isaacs, Reference Isaccs1978). Nonetheless, the tradition of classifying falls by type continues. Rubenstein’s summary (Reference Rubenstein2006) of several such studies is interesting for several reasons. For example, overall the most likely cause was “accident” or environment related, with a mean attributed percent of falls across studies of 31 per cent, but with a range among studies of 1 to 53 per cent. Second most frequent was gait or balance disorders or weakness, at 17 per cent (range: 4 to 39), followed by dizziness or vertigo (13%, 0 to 30%); drop attack (9%, 0 to 52%); confusion (5%, 0 to 14%); postural hypotension (3%, 0 to 24%); visual disorder (2%, 0 to 5%); syncope (0.3%, 0 to 3%); other specified causes (15%, 2 to 39%); and unknown (5%, 0 to 21%).

Although potentially useful for explaining some falls, the enormous ranges in the proportions across studies further underscores the difficulties of accurate retrospective causal attribution in a multifactorial condition, much less the valid differentiation (e.g., among syncope, orthostatic or postural hypotension, and dizziness or vertigo) obtained from self-report. As with previous authors, Rubenstein (Reference Rubenstein2006) has emphasized that many of the falls classified as “accidents” are actually the result of the interaction between environmental hazards and individual susceptibility.

Causes of Falls II: Individual Risk Factors

Although some researchers have noted that more than 400 fall risk factors have been reported (Masud & Morris, Reference Masud and Morris2001), a critical epidemiology would urge that a distinction be made among mere associated variables, risk markers, and truly causal risk factors that can be understood within a theoretical framework. Many fall researchers have come to support the multifactorial risk factor theory of falling, which posits that most falls in older adults are the result of the gradual accumulation of both normal age-related and specific disease-related declines in the key systems underlying postural stability (Tinetti, Reference Tinetti2003; Rubenstein & Josephson, Reference Rubenstein and Josephson2006; Rubenstein, Reference Rubenstein2006; American Geriatrics Society, British Geriatrics Society, and American Academy of Orthopaedic Surgeons Panel on Falls Prevention, 2001). First mentioned by the early authors (Droller, Reference Droller1955; Sheldon, Reference Sheldon1960), quantitative evidence for this theory came from prospective cohort studies that observed a gradient of increasing fall risk across strata of increasing numbers of baseline risk factors (Tinetti et al., Reference Tinetti, Speechley and Ginter1988).

This theory in turn gained support from advances in biomechanical and physiological understanding of human balance (Carter, Kannus, & Khan, Reference Carter, Kannus and Khan2001; Lord, Menz, & Tiedemann, Reference Lord, Menz and Tiedemann2003). Although frail (or intoxicated) people can fall from a bed or a chair, the first geriatric syndrome falls often occur from a standing height. Each fall from a standing height requires two factors: an initial perturbation that displaces the body’s center of mass, and the failure rapidly to regain balance using the trunk, hips, feet, hands, or arms. Walking, an important feature of independent mobility, involves deliberately transferring balance from one foot to the other. Given the centrality of balance and the challenges posed by walking, it is not surprising that balance and gait problems have been repeatedly identified as fall risk factors (e.g., American Geriatrics Society, British Geriatrics Society, and American Academy of Orthopaedic Surgeons Panel on Falls Prevention , 2001; Tinetti et al., Reference Tinetti, Speechley and Ginter1988). However, rather than be analyzed alongside age-, injury-, and disease-related factors, perhaps balance and gait should be considered the key proximal systems through which the distal “upstream” factors combine to produce many geriatric falls.

Human balance is extremely complex; however, a simple physiological model gives prominence to three physiological systems: (a) the visual, (b) the somatosensory, and (c) the vestibular (Lord et al., Reference Lord, Menz and Tiedemann2003). Each system automatically sends signals to the brain to help maintain balance. In young healthy people, redundancy among the systems compensates for temporary limitations in any one system (e.g., walking in darkness). With age, injury, and disease, however, redundancy is reduced such that balance is not as easily maintained when presented with a challenge.

This physiological model with balance at its core, and assessments that directly assess various aspects of balance and associated physiologic systems (Lord et al., Reference Lord, Menz and Tiedemann2003), provides a clear conceptual basis for understanding the causal role of intrinsic risk factors such as muscle weakness, foot problems, and visual deficits (American Geriatrics Society, British Geriatrics Society, and American Academy of Orthopaedic Surgeons Panel on Falls Prevention , 2001). The causal role of other identified risk factors such as arthritis can be understood with reference to effects on balance through decreased range of motion, and on safe mobility via effects such as antalgic gait. Similarly, the causal role of factors such as activities of daily living (ADL) and other physical functional limitations is probably via numerous pathways such as muscle weakness, sensory limitations, and specific diseases. Similar causal interpretations can be made for factors such as advanced age and female gender, in which sarcopenia might cause leg weakness and balance and gait problems. A balance-and-gait–centric model provides a mechanistic interpretation of many causal risk factors for most geriatric syndrome falls, although by definition it excludes those due to sudden loss of consciousness.

Given the observation of compromised balance and gait through myriad mechanisms, we can then understand how precipitating factors (e.g., dimly lit monochromatic stairs with no handrail) interact with predisposing factors (e.g., visual deficit, heavily callused feet, tranquilizers) to cause individual falls. This in turn helps us understand the absence of a simple main effect of the number of environmental hazards present. Specifically, Tinetti et al. (Reference Tinetti, McAvay and Claus1988) found that, on average, people who fell actually had fewer baseline environmental hazards (13.6) than those who did not (14.4).

Another supportive finding is that the risk of serious injury posed by situational factors is very low in those with no intrinsic risk factors, but mounts rapidly when both factors are present (Tinetti, Doucette, & Claus, Reference Tinetti, Baker, King, Gottschalk, Murphy and Acampora1995). A potential tripping hazard such as a raised doorsill between rooms will not cause a fall as long as the occupant has the memory to recall it, the eyesight to see it, and the ankle dorsiflexion and hip flexion to step over it. It is only when the internal balance mechanisms are compromised that previously benign environmental features become fall hazards. This interpretation also explains why environmental modifications have minimal effects in general populations of older adults, but dramatic effects in those with severe visual impairment (Gillespie et al., Reference Gillespie, Robertson, Gillespie, Lamb, Gates and Cumming2009). By considering how each age-, injury-, and disease-related limitation can mechanistically cause falls via balance-and-gait systems, we can move beyond simple “lists of risk factors” (e.g., American Geriatrics Society, British Geriatrics Society, and American Academy of Orthopaedic Surgeons Panel on Falls Prevention , 2001) that make no distinction about the temporal position of the various factors in the causal pathways of individual falls.

Another advance in the study of intrinsic risk factors is an awareness of the distinction between increasing fall risk resulting from normal physiologic aging (e.g., decreased visual acuity, muscle strength, reaction time, range of motion) and those due to specific conditions such as Parkinson’s disease. The early community-based risk factor studies typically had very few, mostly pragmatic, exclusion criteria (e.g., severe cognitive impairment) but otherwise attempted to enroll representative cohorts of older adults. But because both the frequency of falling and the role of many risk factors in Parkinson patients who fall are likely to be somewhat specific to that condition, it is now appropriate for studies to enroll cohorts of patients with the same fall risk-increasing condition. In addition to Parkinson’s disease (Latt, Lord, Morris, & Fung, Reference Latt, Lord, Morris and Fung2009), studies of people with stroke (Weerdesteyn, de Niet, van Duijnhoven, & Geurts, Reference Tinetti, Speechley and Ginter2008) and dementia (Härlein, Dassen, Halfens, & Heinze, Reference Härlein, Dassen, Halfens and Heinze2009) have been recent examples of the use of sample restriction, a very strong design stage (as opposed to analysis stage) method of enhancing signal detection by increasing sample homogeneity.

We can appreciate the importance of this advance by examining past practice. For example, one review of fall risk factors showed that cognitive impairment or dementia was detected in only four of the 11 studies that examined it, with the largest odds or risk ratio equal to 2.3 and the lowest one equal to the null (American Geriatrics Society, British Geriatrics Society, & American Academy of Orthopaedic Surgeons Panel on Falls Prevention, 2001). It is highly likely (and, with the null result, virtually certain) that these figures substantially underestimate the true causal effect of cognitive problems on falls at the population level. As well, these values highlight the fallacy of basing fall prevention policies and practices on the size of odds or risk ratios. Specifically, it is a fallacy to interpret the size of an odds or risk ratio as indicative of a risk factor’s biological causal potency in individuals. In fact, the size of an estimate represents the prevalence of the risk factor in the causal complements of the individual causal models present in the sample. In other words, the routine exclusion of people with severe cognitive impairment from community studies has introduced a selection bias that would in turn reduce the prevalence of this risk factor in the causal models of the individual fallers in these studies, resulting in a biased estimate towards the null of the true population effect.

Some psychological fall risk factors such as depression (American Geriatrics Society, British Geriatrics Society, and American Academy of Orthopaedic Surgeons Panel on Falls Prevention, 2001) can also be understood mechanistically through a balance-and-gait–focused model. For example, reviews of longitudinal studies have shown that baseline depression levels are associated with reduced physical activity (Roshanaei-Moghaddam, Katon, & Russo, Reference Roshanaei-Moghaddam, Katon and Russo2009), which in turn could increase fall risk partly through deconditioning and muscle weakness. Other psychological factors, such as fear of falling, might be similarly posited to increase fall risk through a deconditioning mechanism, although recent studies suggest that reduced gait velocity in those with fear of falling is adaptive to optimize balance, rather than a sign of balance impairment (Reelick, van Iersel, Kessels, & Rikkert, Reference Reelick, van Iersel, Kessels and Rikkert2009).

Of all of the characteristics that have been associated with falling, fear of falling is probably the best example of a factor that is both a cause and an effect of falls. Prospective studies suggest that fear of falling can arise due to factors other than recently experiencing a fall, and that falling and fear of falling share many risk factors in common (Murphy, Dublin, & Gill, Reference Murphy, Dubin and Gill2003; Friedman, Munoz, West, Rubin, & Fried, Reference Friedman, Munoz, West, Rubin and Fried2002). Understanding a factor with bidirectional effects that is also correlated with other risk factors is challenging and is an excellent example of the power of prospective designs over retrospective designs. As well, because fear of falling is observed in older adults who have fall risk factors but have not themselves had a recent fall, the possibility that fear of falling could be an adverse event of programs that seek to prevent falls purely by “increasing fall risk awareness” (e.g., pamphlets full of frightening statistics about sequelae) should receive more attention.

Some of the other previously identified fall risk factors are probably better understood as risk markers. For example, a recent fall history is one of the most consistently identified factors associated with fall risk, and is used appropriately as one of two entry points (the other being emergency room admission for a fall) to a fall prevention algorithm (American Geriatrics Society, British Geriatrics Society, and American Academy of Orthopaedic Surgeons Panel on Falls Prevention, 2001). To call it a risk factor is, however, misleading, because recent fall history is merely an occurrence of the outcome of interest a year earlier in a natural history. Put another way, people who fall in Year 2 and who reported having fallen in Year 1 are simply recurrent fallers observed over a two-year risk window.

Many of the remaining 400 “risk factors” identified in studies probably have little or no true causal role in most individual fall models, and arose in studies due to chance sampling variability and by confounding through their association with true causal factors. But pursuing this line of thinking raises another methodological issue: what is the logic of using multivariable statistical models to estimate the independent effects of single factors on a multifactorial recurrent outcome in which many factors are at least moderately intercorrelated and in which the outcome can cause its own risk factors? The concern goes beyond the statistical nuisance of multicollinearity and the limitations of stepwise regression models. Specifically, is it logical to try to estimate the independent effect of, say ADL limitations, adjusted for the effects of weakness in the legs (or vice versa)? If intrinsic risk factors interact both with each other and with situational factors to cause falls, perhaps our analytic emphasis should shift to the study of effect measure modification, rather than adjusted main effects of collinear variables. Bellinger (Reference Bellinger2000) argued that the study of neurotoxicants in children’s health had reached the point where the next advances would come from studying effect measure modification, and the fall literature may have now reached a similar situation.

As mentioned, the epidemiologic cycle is complete when the causal theory is demonstrated by experimental modification of putative risk factors that are correlated with changes in outcome probabilities. The multiple risk factor theory of falling reached this stage in 1994 when Tinetti et al. (Reference Tinetti, Baker, McAvay, Claus, Garrett and Gottschalk1994) were able to explain the positive result of their FICSIT trial by demonstrating relatively more risk factor changes in the treatment group (Tinetti, McAvay, & Claus, Reference Tinetti, McAvay and Claus1996). That the pooled estimate of efficacy observed in the 2003 Cochrane review (Gillespie et al., Reference Gillespie, Gillespie, Robertson, Lamb, Cumming and Rowe2003) was weaker in subsequent reviews (Gates, Fisher, Cooke, Carter, & Lamb, Reference Gates, Fisher, Cooke, Carter and Lamb2008) is more readily explained by interventions of insufficient intensity (Tinetti, Reference Tinetti2008) than by the sudden invalidity of the theory.

Injurious Falls and Other Outcomes

Because most falls in older adults do not result in injury (Tinetti, Reference Tinetti2003; Rubenstein & Josephson, Reference Rubenstein and Josephson2006; Rubenstein, Reference Rubenstein2006; American Geriatrics Society, British Geriatrics Society, and American Academy of Orthopaedic Surgeons Panel on Falls Prevention, 2001; O’Loughlin et al., Reference O’Loughlin, Robitaille, Boivin and Suissa1993), falling is a necessary but insufficient cause of a fall injury (just as serious injury is necessary but insufficient for other outcomes such as hospitalization and institutionalization). Most early studies placed greater emphasis on detecting falls, and less on injurious falls. Because injurious falls can also be recurrent events, all of the statistical challenges of analyzing falls are present with injurious falls. However, because only a minority of fallers experiences an injury, the precision of the statistical estimates from cohort studies of a set sample size will be lower because the outcome is rarer. Added definitional challenges are that injuries differ in terms of the nature of the injury and anatomic site and severity, and a single fall can produce multiple injuries.

Together, these challenges may partly explain heterogeneities in occurrence estimates and risk factors identified across studies. For example, Nevitt, Cummings, and Hudes (Reference Nevitt, Cummings and Hudes1991) reported serious injuries (fracture, dislocation, or laceration requiring suture) in 6 per cent of 539 falls among recurrent fallers whereas O’Loughlin et al. (Reference O’Loughlin, Robitaille, Boivin and Suissa1993) reported serious injury in 3 per cent of 197 falls, including one-time fallers, using a similar definition of serious injury. Tinetti et al. (Reference Tinetti, Speechley and Ginter1988) reported serious injuries (fractures as well as soft-tissue injuries that required medical attention) in 11 per cent of 272 falls. As noted with falls, different units of analysis (e.g., 23 to 24% serious injurious falls among fallers; Tinetti, Doucette, Claus & Marottoli, Reference Tinetti, Doucette, Claus and Marottoli1995; Tinetti et al., Reference Tinetti, Speechley and Ginter1988) cannot directly be compared to figures with falls in the denominator.

Using hospitalization or emergency room visit data to estimate occurrence will introduce further selection biases due to referral, admission, and access-to-care factors. Hip fracture is perhaps one exception where hospitalizations will approximate population incidence; however, hip fractures are seen in only 1 per cent (Rivara & Thompson, Reference Rivara and Thompson2000) to 1.5 per cent (Tinetti et al., Reference Tinetti, Speechley and Ginter1988; O’Loughlin et al., Reference O’Loughlin, Robitaille, Boivin and Suissa1993) of falls from prospective series.

Case Definition

The earliest published studies in the 1940s (Sheldon, Reference Sheldon1948) and 1950s (Droller, Reference Droller1955) used a tacit definition which could include “near falls”. For example, Sheldon’s data came from a health survey of a representative sample of older English adults. Amazingly, a question on falls was added only after the study began, when the researchers realized the importance of falling (Sheldon, Reference Sheldon1948). Interestingly, Sheldon reported a figure of 36.2 per cent of the subjects being “liable to falls”, including 10 “near fallers” but excluding all cases of a “genuine slip such as might happen to anyone” (Sheldon, Reference Sheldon1948, p. 96). Despite these limitations, if the 10 “near falls” are excluded, Sheldon’s estimate of 32.6 per cent is squarely in the range of the estimates from the later prospective cohort studies (29 to 35%) that also excluded near falls (Tinetti et al., Reference Tinetti, Speechley and Ginter1988; O’Loughlin et al., Reference O’Loughlin, Robitaille, Boivin and Suissa1993; Campbell et al., Reference Campbell, Borrie and Spears1989).

Sampling

The early authors were clearly aware of the differences between community-dwelling and institutionalized older adults, and sampled accordingly. For example, Droller’s study (1955) was set in the community whereas Fine’s study (1959) was set in hospital. The legacy of this distinction can be traced to the present day, and to the decision to produce separate Cochrane systematic reviews of fall prevention programs in the two settings (Gillespie et al., Reference Gillespie, Robertson, Gillespie, Lamb, Gates and Cumming2009; Cameron et al., Reference Cameron, Murray, Gillespie, Robertson, Hill and Cumming2010). Whereas Sheldon in Reference Sheldon1960 was aware of the “heavy pathological bias” of studying hospitalized people, his sample consisted of fallers seen in the emergency room, hip fracture patients, and fallers reported by family physicians, which probably introduced a selection bias due to injury severity (Sheldon, Reference Sheldon1960). The authors of an early longitudinal study of falling in a home for the aged (Gryfe et al., Reference Gryfe, Amies and Ashley1977) argued that residents with the highest levels of independence living in such a setting might be generalizable to older adults in the wider community. Indeed, their estimated rate of 668 falls per 1,000 person-years (55.7 per person-month), while higher than the 41.4 falls per 1,000 reported by O’Loughlin et al. (Reference O’Loughlin, Robitaille, Boivin and Suissa1993), is very close to the 56.8 per 1,000 figure calculable from data reported by Campbell et al. (Reference Campbell, Borrie and Spears1989).

Another early study sampled on the basis of injury and included both community and institutional subjects (Lucht, Reference Lucht1971). Interestingly, this 1971 study is the earliest publication on falling in older adults that appears in PubMed, and was also among the first papers to address the potential for falls to be prevented. In this comprehensive prospective case series of 472 people over 60 years of age who had been treated in hospital after falling in their homes or institutions, Lucht estimated the annual incidence of serious injury due to a fall at 14 per 1,000 population. This same study noted that three times more injured fallers were women than were men, and that fall frequency increased sharply at age 75. In the discussion, Lucht directly addressed the potential for fall risk reduction through a combination of medical treatment and environmental modification. In another early study that used a mixed sampling approach of community, residential care, and hospital residents, Campbell, Reinken, Allan, and Martinez (Reference Campbell, Reinken, Allan and Martinez1981) retrospectively estimated a fall risk of 34 per cent, again squarely in the range of the prospective cohort estimates in community populations, probably because their sample was overwhelmingly of community origin.

Research Designs

The descriptive case series was gradually replaced by the first controlled investigations, but surprisingly this did not really occur in the community setting until the early 1980s (Prudham & Evans, Reference Prudham and Evans1981). A common approach has been to identify recent fallers and non-fallers from a single sample, and then to compare these two groups with respect to potential risk factors. Although this resembles the case-control design, it is properly called a backward prevalence design (Kleinbaum, Kupper, & Morgenstern, Reference Kleinbaum, Kupper and Morgenstern1982). The first assumption is that people who have recently fallen and those who have not form two distinct groups, and that differences in prevalence of certain characteristics indicate potential risk factors. Underlying this assumption is that the risk factors identified existed prior to the fall. While immutable factors such as age and gender meet this assumption, other characteristics such as balance, gait, sensory deficits, and functional limitations clearly do not. This design is particularly problematic with falls because “risk factors” can in fact be outcomes of falls in some cases (e.g., gait abnormality due to fall injury), or both risk factors and outcomes in others (e.g., fear of falling).

The limitations of this design were noted in the early 1990s (Cumming et al., Reference Cumming, Kelsey and Nevitt1990) coincident with the seminal prospective cohort studies previously mentioned (Nevitt et al., Reference Nevitt, Cummings, Kidd and Black1989; Tinetti et al., Reference Tinetti, Speechley and Ginter1988; O’Loughlin et al., Reference O’Loughlin, Robitaille, Boivin and Suissa1993). These cohort studies ensured that factors present at baseline existed prior in time to falls measured over the follow-up period. The possibility that some baseline factors were actually caused by falls occurring before the study can be partially controlled by including fall history as a covariate in multivariable models.

With the temporality of previously identified risk factors confirmed in the cohort studies, it was appropriate to move on to experimental manipulations of modifiable risk factors in randomized trials. The recent Cochrane reviews summarized the results of 111 trials involving more than 55,000 community residents (Gillespie et al., Reference Gillespie, Robertson, Gillespie, Lamb, Gates and Cumming2009) and of 41 trials involving more than 25,000 residents in institutions (Cameron et al., Reference Cameron, Murray, Gillespie, Robertson, Hill and Cumming2010).

These reviews are invaluable in identifying which interventions seem to work, and how well, in which populations. Many interventions have stronger effects when rate ratios, as opposed to risk ratios, are analyzed, and in many cases the former is statistically significant and the latter is not. This pattern suggests that many interventions are more effective in reducing the average number of falls among fallers than they are in reducing the proportion of people who fall. In community dwellers (Gillespie et al., Reference Gillespie, Robertson, Gillespie, Lamb, Gates and Cumming2009), tai chi, individually prescribed multi-component home exercise, and multi-component group exercise fit this pattern, as do multidisciplinary multifactorial interventions among long-term care residents (Cameron et al., Reference Cameron, Murray, Gillespie, Robertson, Hill and Cumming2010).

These results also provide important clues about which interventions do not appear to be effective, and help address the issues of high-risk versus general-population approaches, and multifactorial versus single-factor interventions. For example, home safety interventions reduce fall rate by 41 per cent among those with severe visual impairment, and vitamin D supplementation reduces fall rate by 43 per cent among those with deficiency, yet neither intervention has a statistically (or clinically) significant effect in unselected community populations (Gillespie et al., Reference Gillespie, Robertson, Gillespie, Lamb, Gates and Cumming2009). Moreover, multifactorial risk factor assessment coupled with active intervention results in a 30 per cent reduction in pooled fall rate, compared to a 16 per cent reduction if the intervention stops at referral or recommendations for treatment. Multifactorial interventions in long-term care residents with cognitive impairment, and single exercise modalities, appear to be ineffective (Cameron et al., Reference Cameron, Murray, Gillespie, Robertson, Hill and Cumming2010).

It is hoped that future studies will be designed to build systematically on these results (e.g., by using fall rate as the primary outcome measure). One possible explanation for heterogeneity in effect sizes across intervention studies in institutions is that “usual care” differs across institutions and within them over time. Although the routine clinical care of complex patients can never be standardized as can be done in placebo-controlled drug trials, careful descriptions of the control condition and of steps taken to prevent conscious or subconscious enhancements of “usual care” during a study will improve the estimates of the efficacy of specific interventions and ease the interpretation of future meta-analyses.

Individual Fall Risk Prediction

Because falling is so profoundly multifactorial, falls among individuals, as well as different falls by an individual, can be the result of different sets of risk factors as they combine to produce minimal causal sufficiency. Researchers who have attempted to predict individual fall risk actually faced the same problem encountered by earlier researchers who attempted to categorize falls by “type”. Consider two people each with the same five baseline risk factors: (a) leg weakness, (b) weak grip strength, (c) severe foot calluses, (d) poor vision, and (e) the same nine prescription medications. The first person has four falls over the next year. The first fall is caused primarily by the first three risk factors; the second one by factors unknown or unmeasured; the third is caused by interaction among three of the medications and two of the other factors; and the fourth is caused in part by injuries sustained in the first three. The second person does not fall over the next year. A third person, with no known risk factors, does fall. Each of these contingencies is possible, and this complexity is not present in most health outcomes we wish to predict. It is probably for this reason, in addition to the use of backward prevalence designs and the effects of chance sample variability, that most scales to predict individual fall risk perform poorly (Gates, Smith, Fisher, & Lamb, Reference Gates, Smith, Fisher and Lamb2008).