There is a dearth of evidence regarding even the common problem of falls amongst the very old, an information gap that urgently needs filling. Few studies have included more than a small minority of octogenarian participants, even fewer nonagenarians, and falls rates in these age-bands have rarely been presented. Those studies which do report for these age-groups have had very small numbers, for example one reported rate of 152 falls/100-person-years amongst over-90-year-olds 2 had only 2% of their sample at this age on which to base this incidence. The picture of health conditions affecting older old people is different from that for younger old age 3 so simple extrapolation cannot provide adequate estimates on which to plan for future demographic change.

How falls data are gathered is important for validity and reliability. There are questions of defining or classifying falls and methodological issues including information sources, ethics and data collection tools. Taking a "fall history" is important in clinical practice and retrospective recall is often used in research data collection to save time and costs. However, the length of recall period used can have considerable effects on the data gathered 4. Under-reporting of falls is a largely un-quantified problem but has long been an acknowledged limitation with recall methods 5. A prospective approach has often been regarded as the method of choice in falls research, however fewer than half the studies in a recent systematic review of fall measurement used prospective data collection 6. The impact on estimates of using more economical methods is largely unknown. This study set out to address these gaps using data from two consecutive years gathered retrospectively and prospectively.

The Cambridge City over-75s Cohort (CC75C) is a longitudinal study of ageing which began in 1985 78. Participants were initially recruited from socially and geographically representative primary care practices, with a 95% baseline response rate equivalent to 40% of the total over-75-year-old population in Cambridge, UK. The cohort has been followed-up with successive interviews every few years, each approved by Cambridge Research Ethics Committee, with death the principal cause of attrition. As well as the standard CC75C nurse-administered interview, the Year 17 survey added a special focus on falls in advanced old age, including retrospective measures and prospective follow-up (2002–4).

All analyses were conducted in Stata (Intercooled version 9).

Ninety women and 20 men (84% of cohort survivors) were included in the survey: 90% personally took part (n = 99) while proxy informants provided all information for those unable to be interviewed and supplementary information for others (n = 34). Reasons for lack of interview with the participant themselves included failure to trace (n = 3), access denied by a relative or GP (n = 4), illness (n = 4), profound deafness (n = 2) and refusal (19), but proxy informants gave interviews for 11 of these.

Ages ranged from 91 to 105 years old, mean (SD) 94.4 (2.4), with no marked age differences in the sample by gender although the larger numbers of women gave a wider distribution. They were a predominantly frail population – 11% housebound, more than a third severely cognitively impaired, 2/3 with difficulty in at least two basic activities of daily living, 4/5 affected by at least five health conditions, most unable to climb stairs and only a fifth still able to walk around their local neighbourhood. Over half the participants were living in the community, a further sixth lived in sheltered accommodation and a quarter in institutional care, mainly residential homes.

The following findings describe the study's falls measures in terms of how common falling was amongst these over-90-year-olds (the "prevalence" of falling and repeated falling) and how often they fell (incidence of falls) using both retrospective and prospective data plus additionally, from the latter, how long before any fall occurred.

Well over half the participants (58%) reportedly fell in the year before interview and 59% of them, or 34% of the total, reported falling more than once in the past year. A fifth recalled falling at least three times in the last year, and three people a dozen or more times.

Even more of the study participants were reported to have fallen at least once in the past five years than in the past year (88% vs 58%). Figures 1a and 1b illustrate how the prevalence of falling, based on reports of remembered falls, increases with the length of time about which respondents are questioned regarding their fall history. However, by contrast, incidence rates based on of the number of falls reported as having happened within different time periods before interview increase with shorter recall periods from 1.6 falls/person-year taking the full previous year to 2.2, 2.4 and 3.1/person-year in the past 6, 3 and 1 months respectively.

Comparison of the time since the participants were said to have last fallen and the time since they said they last hurt themselves falling suggest that falls resulting in injury appear to be better recalled, with a longer median time to the last remembered injurious fall (9 months, IQR 2–30 months) than to any last fall (5.5 months, IQR 1–23 months).

In univariate analyses of association with socio-demographic factors (see Table 1) increasing age, higher educational level and living in any sheltered scheme or care home were all associated with having fallen in the past year, as recalled by participants and proxy informants. Only age was associated with recalled repeated falls – two or more falls in the year before interview. Regressions using the number of falls recalled as having occurred in the previous year also showed older age, even across such a limited age range, and living in any supported setting increased falls risk ratios. Almost identical proportions of both men and women remembered having fallen in the past year (55% and 58% respectively); differences widened in repeated falling measures with falls rates 50% higher in women, but these differences did not reach significance. Multiple variable regressions in these retrospective data modelled the effects on each socio-demographic variable of the others. Age was the only factor that remained significantly associated with increased risk of falls (Each year older than 91: IRR 1.1, 95% C.I. 1.0–1.3; Ageband ≥ 95 vs ≤ 94: IRR 1.8, 95% C.I. 1.0–2.9) and with repeated falling (Each year older than 91: RR 1.3, 95% C.I. 1.1–1.6; Ageband ≥ 95 vs ≤ 94: RR 2.8, 95% C.I. 1.7–3.6) when adjusted for other demographic covariates. However, place of residence (RR 1.4, 95% C.I. 1.0–1.8) and educational level (RR 1.5, 95% C.I. 1.0–1.7) as well as age (Each year older than 91: RR 1.1, 95% C.I. 1.0–1.2; Ageband ≥ 95 vs ≤ 94: RR 1.5, 95% C.I. 1.1–1.8) continued to be associated with having suffered any fall in the past year (adjusted risk estimates not tabulated).

This unique cohort remains a highly representative population-based sample, using systematic tracing and careful re-recruitment to ensure a low drop-out rate: for only 16% of living survivors was it not possible to obtain even proxy informant data. No other study to date has gathered prospective falls data specifically from older people of such advanced age, representative of their population base. Intensive 12 months' follow-up covered the full sample, with methods that proved to be feasible and well accepted. Indeed there was great willingness from participants, as well as both formal and informal carers, to help with fall reporting – an indication of the relevance and high importance attached to the problem of falling amongst older people. Building on a long-standing study with the methodology already in place clearly imposes some limitations: the sample size in this survey was pre-determined by the survival of the cohort, though participation still exceeded numbers in the "90 plus" age range included in many larger population studies. Interpretation of the significance or non-significance of the relationships examined between group characteristics and measures of falling requires caution given the small sample size.

Falls follow-up methods were intensive with phone-calls to cover those who missed returning one of the weekly calendar pages or who preferred telephone reports. Nevertheless, the possibility that there was under-reporting of some falls cannot be ruled out. Over-reporting is far less likely as details recorded ensured that it was soon detected if the same fall was mentioned twice. If there was marked under-reporting, the prevalence of falling could be even more widespread than found and falls incidence rates even higher.

Recall bias can affect any study; it is an anticipated problem in falls studies5. Moreover, in this study's age-group cognitive impairment is a likely reason for forgetting falls 4. The most cognitively impaired in this study all had proxy informants reporting on their behalf, but these proxies may not always have been aware of every fall. There is more scope for under-reporting from those with milder impairment, not all of whom had a proxy source.

It has long been known that the proportion of people who report past falls varies with the length of recall period questioned, with shorter intervals not necessarily providing the most accurate recall 5. Our survey gathered data in a format that allowed measurement of different recall time periods and also recorded time lapsed since the last recalled fall. This longer time frame revealed that a fall within the previous five years was remembered for 87% of the participants. As expected, the percentage remembering falls rose with longer time intervals but, as found in the EVOS study 1415, prevalence does not increase in proportion to length of recall period. This can be interpreted in different ways. It is plausible that falls which happened longer ago are less likely to be remembered, particularly non-injurious falls. Certainly in the current study the time lapsed since a "near fall" was generally short, suggesting these were often dismissed as unimportant and soon forgotten. Moreover falls that had resulted in injury were remembered as having happened longer ago than falls in general. On the other hand, it could be that people remember falls as having happened more recently than they actually did. The rising prevalence of falling and decreasing incidence of falls over longer recall periods shown in Figure 1 may help explain some of the differences found for some factors between the relative risks for falling or repeated falling and the incidence rate ratios for falls (see Tables 1 and 2).

Recall interval is also relevant when comparing falls reported from previous prospective studies that used different systems to record falls. A Japanese study illustrated how differential follow-up methods affected reporting: the prevalence of falling during a year in three comparable groups of men asked about their falls every month, every three months and just once at year end was 21%, 16% and 6% respectively, though no such pattern was found amongst women (26%, 18% and 21% respectively) 16. However, this gender distinction was not found in CC75C.

A recent systematic review of falls monitoring 4 took prospective data collection using on-going weekly or monthly calendars as their recommended criterion standard, though the authors found insufficient evidence to advise what time interval was optimal. This review concluded that recall methods could be highly specific (91–95%) but less sensitive (80–89%) than prospective. The broad concordance between retrospective and prospective findings shown in the few studies mentioned below, and close agreement found in the current CC75C survey, would fit these conclusions. However, even where similar proportions of people falling are reported for pre- and post-interview periods, these will not necessarily represent the same individuals. The Gloucestershire Longitudinal Study of Disability 17 reported that, across three years of general practitioner checks on over-75-year-olds' health and disability, falling was the most inconsistent measure annually: about 70% of people who had reported a fall in the previous three months no longer reported recent falls at the following year's interview, while 11% became new reporters of falling. Such discrepancies are particularly large for shorter recall times but the same point also applies to longer periods.

All risk ratios presented for falls adjust for clustering of some falls, based on the assumption that each participant's falls are not necessarily independent of each other. Although it could be argued that episodes involving a series of falls (perhaps attributable to a common factor) might unduly affect interpretation of the data, 32% of those who fell reported at least two falls in close proximity. Thus such reports cannot be discounted without grossly underestimating both prevalence and incidence rates. However, it is also possible that outlying fall frequency counts may inflate findings, therefore sensitivity analyses examined the effects of excluding multiple falls by different cut-points illustrated in Figure 3. For example three people fell more than four times within a week, each imminently preceding either death or hospital admission. Excluding either all three or just the multiple falls preceding death slightly reduced incidence rates (2.3 and 2.6 falls/person-year respectively). However, it is also arguable that such acute illness or end-of-life falls are not untypical in this advanced age-group and so should be included.

Analyses included four people who died within 8 weeks of interview, all with no reported falls, contributing a total of only 0.3 person-years of prospective follow-up. Although incidence rates are clearly unaffected, this approach may under-estimate prevalence because the true denominator was lower for most of the follow-up.

Few other studies have reported fall prevalence specific to the oldest old. The two "old old" studies that have reported falls, with recalled data only, each found very similar proportions had fallen at least once in the year before interview: 45% of the men and women in the Umeå 85+ study18, and 44–49% of women in two Leiden 85-plus Study interviews 19. Thus the CC75C findings are between a fifth and a third higher than recorded in these slightly younger cohorts. Cross-sectional and prospective studies of broader age-ranges of older people that have presented age-specific results report the proportion who fall each year as between 35% and 51% of people aged 85 or more 19202122, and between 29% and 41% of people aged 80 or older 2123242526272829 except in one small volunteer study that reported annual fall prevalence as 58% based on only 12 individuals aged over 803031.

Estimates of repeated falling – more than one fall in a year – range from 14% to 29% in the two studies that have reported these proportions for age-bands over 80-years-old 2425, even the higher figure only two-thirds of the proportion of recurrent fallers in CC75C.

Few studies have reported both retrospective and prospective falls data from the same sample, but one study in a younger old sample that also report recalled and followed-up falls from two consecutive years showed a similar trend to ours in reporting more repeated falls in follow-up, but also more single falls. This was the Australian Randwick Falls and Fractures study which found higher fall prevalence amongst women aged ≥ 65 years in follow-up than recall: 38% vs 20% for any falls, 21% vs. 14% for repeated falls 3233. A Dutch study of over-70-year-old general practice patients reported falls from different length periods before and after interview: 33% fell at least once during just 36-week's follow-up, a marginally higher proportion than could be expected from the 41% with any fall history in the year before interview, but the repeated falls were in similar proportions (16% in follow-up vs 26% recalled) 34.

Annual fall incidence rates have also rarely been reported for very old people. The Montreal study reported identical rates for men and women aged 80 or older (65.9 falls/100 person-years) 24, notably lower than in New Zealand's Dunedin study235 which did not report sex-specific rates but broke down their over-80-year-olds into three age-bands: incidence rose from 94 to 152 falls/100 person-years between the 80–84 years and over-90s age-bands. Although methodological differences may mean studies are not directly comparable, this puts the CC75C incidence at over 80% higher than the previously reported rate for nonagenarians.

Having fallen before has been identified repeatedly as a risk factor for falling again, both in major community-based studies 3134353637383940 and in institutional settings 414243. It has also been suggested that characteristics and risk profiles of people who fall repeatedly are different from those who report "just a one-off" fall 4445 and diverse studies have presented their findings in terms of which factors identify this higher risk group 232829334647. Although not all studies confirm this supposition 3148, it has been suggested that "once only fallers" have more in common with "non-fallers" than "twice or more fallers" 49. Amongst the over-90-year-olds in the CC75C study, retrospectively gathered fall reports appeared highly predictive of falls in prospective follow-up. Moreover, in this study sample, recalling more that one fall in the past year strengthened the risk estimates for subsequent falling and recurrent falling associated with any recalled fall in the past year. Adjustment for the potentially confounding effects of socio-demographic variables had minimal effects. Fuller examination of other potential risk factors with this study data 50 is beyond the scope of this report's focus on basic demographic descriptors of falls epidemiology in advanced old age. Falling is a multi-factorial problem and the analyses presented here have not attempted to further un-ravel factors that predict and perhaps pre-date any recalled falls. When asking about previous falls, it is important to also consider factors that contribute to this history of falling and thus also to future risk.