The review process is summarised in according to QUOROM guidelines.26
89 intervention trials meeting our inclusion criteria are summarised in webtable 1
with full details available from the authors. Trials assessed geriatric assessment in elderly people representing the general population (n=28)27–53
or those selected as frail (n=24),54–77
community-based care after hospital discharge (n=21),78–98
fall prevention (n=13),51,99–110
or group education and counselling (n=3).111–113
All trial interventions were complex and many individuals would have been eligible for any of the them.
Randomisation was by individual or household (n=80) or by clusters of general practices, community groups, or municipalities (n=9). The total number of people randomised was 97 984 with a median of 321 (range 54–43 219) in trials. One large study randomised 43 219 people.33
Mortality rates ranged from 0 to 60·8% per year, with a median of 6·8%. Trials assessing geriatric assessment in general and frail populations had median mortality rates of 5·4% (0–10·5%) and 6·1% (1·1–60·8%), respectively, suggesting that the frail category was often subject to selection, probably indicating eligibility issues. In trials of community-based care after hospital discharge, the median mortality rate was 16·2% (6·3–53·0%); for fall prevention and group education, it was 4·3% (0–11·6%) and 3·4% (2·7–4·4%), respectively.
Losses to follow-up were used as a marker of study quality. In trials with death as an outcome, 40 (48%) of 84 had losses to follow-up of 1% or less (range 0–27·6%). For physical function, few trials included people who had died or moved to nursing homes in their analyses; exceptions were Close100
and their colleagues' trials. 15 (35%) of 43 trials had losses of participants to follow-up for interview of 5% or less (0–33%).
The allocation process was described in 61 (69%) of 89 trials, but difficulties of assessing concealment and masking in complex intervention trials are unlikely to have been fully addressed.114
Intervention activity in control groups was evident in 40 (45%) of 89 trials.
Data for variability between clusters were insufficient and the effect of analysis errors arising from inclusion of cluster randomised trials was explored by sensitivity analysis. Inspection of funnel plots at all stages of the review gave no indication of selection bias in studies included in the analysis (data not shown).
Outcomes are summarised by type of intervention in the . The outcome of living at home at follow-up was available for 51 interventions; in a further nine trials death and nursing-home admission were used, with the consequent inclusion of the large Medical Research Council (MRC) trial.33
Overall, 60 (67%) of 89 trials reported living at home at follow-up or an estimate. However, this outcome was reported in only 4 [31%] of 13 trials.
Relative risk (95% CIs) of outcome by intervention context (standardised mean difference for physical function) and I2 heterogeneity statistic
In a meta-analysis of 60 trials with 79 578 individuals (), the overall risk of not living at home was lower in the intervention group (RR 0·95, 95% CI 0·93–0·97) than in the control group. Geriatric assessment of general elderly people and community-based care after hospital discharge were the only types of intervention that had a significant effect on the risk of not living at home (). Removal of trials with estimated values had little effect (0·95, 0·90–1·00). Heterogeneity was only manifest in trials of geriatric assessment in general populations and those selected as frail ().
Relative risk (RR) of not living at home
If typical rates of not living at home of about 7·6% (median in trials; range 0–12·1) per year for the general population are used, and the reduction in risk from intervention is 5%, a number needed to treat of 263 is obtained. For the increased rates of not living at home in people receiving community-based care after hospital discharge of about 25% per year with an RR reduction of 9%, the number needed to treat is 40.
Data for death were available for 84 (94%) of 89 interventions including 93 754 people (webfigure 1
). Interventions had no overall benefit (RR 1·00, 95% CI 0·97–1·02) and the only appreciable benefit by type of intervention was noted in 11 trials targeting fall prevention (0·79, 0·66–0·96; webfigure 1
). Slight heterogeneity (I2
=10·6%) was almost exclusively limited to trials of geriatric assessment in general elderly populations (I2
=39·7%; webfigure 1
Data for nursing-home admission (31 trials) or for residence at follow-up (23 trials) were available for 79 575 people (webfigure 2
) and were widely reported in trials of geriatric assessment in general (20 [71%] of 28) or populations selected as frail (16 [67%] of 24) and community-based care after hospital discharge (14 [67%] of 21), but not in trials of fall prevention (3 [23%] of 13).
For combined nursing-home outcomes, risk of admission was reduced in the intervention group (RR 0·87, 95% CI 0·83–0·90; webfigure 2
). Only a marginal effect was seen for residence at follow-up (0·93, 0·79–1·09). Geriatric assessment and community-based care after hospital discharge were the only types of intervention to have a significant effect on the combined outcome (webfigure 2
). Some heterogeneity was recorded in trials (I2
=29·0%), mainly in geriatric assessment in general populations (I2
=47·5%; webfigure 2
For nursing-home care, typical median rates for trial populations were 2·2% (range 0·1–5·4) per year for the general population and 11·1% (2·1–40·2) per year for people receiving community-based care after hospital discharge, generating numbers needed to treat of 354 and 39, respectively.
Hospital admissions were reported in between 5 (38%) of 13 (falls prevention) and 18 (86%) of 21 (community-based care after hospital discharge) trials. The most commonly reported outcome—number of people having an admission (41 trials with 20 047 people [webfigure 3
])—was used in the meta-analysis. Risk of hospital admission was reduced by interventions (RR 0·94, 95% CI 0·91–0·97; webfigure 3
). Geriatric assessment in elderly people selected as frail and community-based care after hospital discharge were the only types of interventions to show significant effect on this outcome. Heterogeneity (I2
=43·0%) was largely restricted to geriatric assessment in general elderly patients and community-based care after hospital discharge (61·4% and 57·0%, respectively; webfigure 3
). In random-effects meta-analysis, the overall RR was similar (0·94, 95% CI 0·89–0·99).
Trials with data that were incompatible with the meta-analysis were inconsistent, with hospital admissions reduced in seven,33,40,42,56,94,100,113
similar in four,68,70,80,87
and increased in five trials.46,53,54,58,61
The large MRC trial reported slightly reduced total admissions in the intervention group (RR 0·96, 99% CI 0·79–1·16).33
All 13 studies targeting fall prevention reported individuals who had fallen, whereas falls were less likely to be reported in trials in general elderly people (6 [21%] of 28 trials), frail elderly people (5 [21%] of 24), and those with community-based care after hospital discharge (5 [24%] of 21). No trials of group education reported falls. An overall benefit was noted in 25 trials including 15 607 people (RR 0·90, 95% CI 0·86–0·95; webfigure 4
). Interventions targeting fall prevention contributed 66% of the weight. Only trials of geriatric assessment in general elderly populations and those of fall prevention showed significantly reduced falls (webfigure 4
). Heterogeneity (I2
=52·8%; webfigure 4
) was restricted to trials of community-based care after hospital discharge (I2
=40·3%) and fall prevention(I2
=65·8%). Use of random-effects meta-analysis led to wide CIs including unity for the interventions targeting fall prevention (0·91, 0·82–1·00), but the overall effect including all trials was broadly similar (0·89, 0·83–0·96). The four trials that did not report individuals who had fallen showed benefit with reduced total falls in intervention groups.29,66,91,102
Physical function outcome was measured in 73 trials. The Barthel index of activities of daily living restrictions (n=14) and SF-36 physical function dimension (n=7) were frequently reported. Information on change and functional deterioration was available for only nine and 16 studies, respectively, and we used the widely available physical function at follow-up in our analyses.
Meta-analysis included 43 interventions with 21 651 individuals (webfigure 5
). Sources of variance data are available from the authors. In Reuben and colleagues' trial,68
substantial differences at baseline between randomised groups were reported, and analyses were done both with and without these data. In 36 trials with activities of daily living outcome, an overall benefit for interventions was noted (SMD −0·08, 95% CI −0·11 to −0·04) and in 14 trials with a generic physical function outcome that did not specifically focus on disability, the effect was similar (−0·09, −0·13 to −0·05). Exclusion of the trial with baseline differences had little effect on the SMD, and heterogeneity was evident for both outcomes (I2
=37·5% and I2
When SF-36 physical function means and variances from the 1992 Office for National Statistics survey were used,115
an SMD of 0·09 translated as an improvement in a representative elderly population of between 3·3% and 7·2% dependent on age. For the Barthel index, in the trials included in the review, the SMD of 0·08 equated to about half a point improvement in the 20-point score.
Combination of activities of daily living and generic outcomes (only used if no activities of daily living outcome reported) showed a similar benefit (SMD −0·08, 95% CI −0·11 to −0·06; webfigure 5
). Heterogeneity was little different in the combined analysis to that seen when activities of daily living and generic measures were analysed separately and was mainly restricted to trials of geriatric assessment in elderly people selected as frail. However, RRs were much the same in random-effects meta-analysis (−0·08, −0·13 to −0·04). Geriatric assessment in general populations and falls interventions showed benefit for physical function at follow-up when grouped by context.
Physical function was reported in a form unsuitable for meta-analysis in 30 trials. Two interventions showed improvement in activities of daily living76,97
and five showed weak evidence of benefit.33,50,73,81
However, no improvement was noted in 19 trials,30,31,38,42,44,47,48,59,71,77,83,86,93,95,98,102,108,111,113
and in four trials generic physical function was largely unaffected by interventions.51,59,86
Study quality in terms of losses to follow-up (webtable 2
) and randomisation process did not affect our findings—eg, RR of not living at home in 19 trials in which the randomisation process was not clear was 0·92 (95% CI 0·85–0·98), similar to that reported in trials with a clear description of randomisation (0·95, 0·93–0·98).
Analysis of results excluding trials with cluster randomisation had little effect on the overall RRs and variances. The contribution of the cluster randomised MRC trial to the meta-analyses was large, providing 71%, 58%, and 73% of not living at home, death, and nursing-home admission events, respectively. However, after exclusion of the trial, the results were reasonably consistent with RRs of 0·94 (95% CI 0·90–0·98) and 0·90 (0·83–0·97) for not living at home and nursing home admissions, respectively. When the MRC trial was excluded, the RR of death was reduced, although the 95% CI included unity (0·96, 0·92–1·00).
In trials with increased death rates, the RR of not living at home was reduced (second quartile of death rate 0·91, 95% CI 0·84–0·98, p=0·02; third quartile 0·96, 0·93–0·98, p=0·01; and fourth quartile 0·88, 0·79–0·96, p=0·05; webtable 3
). Similarly, nursing-home admissions were reduced after intervention in populations with high death rates, which differed significantly from one for the two highest quartiles (third quartile of death rate 0·86, 0·82–0·90, p<0·0001, and fourth quartile 0·75, 0·63–0·89, p=0·01; webtable 3
In trials with recruitment dates before the median of 1993, interventions showed benefit with a combined RR of not living at home of 0·89 (95% CI 0·84–0·93; webtable 4
), whereas in trials from 1993 onwards the RR was 0·97 (0·94–0·99). Removal of the MRC trial from the analysis made this difference even more pronounced (1·04, 0·96–1·12) in trials from 1993 onwards. In meta-regression, the outcomes of not living at home, death, and nursing home admission all showed increased risk reduction in studies before 1993 (webtable 4
). This increased risk reduction was also apparent for specific contexts—eg, for community-based care after hospital discharge, RR of not living at home in studies started before 1993 was 0·82 (0·73–0·93) compared with 1·01 (0·87–1·17) in later studies. Younger participants (≤74 years) tended to benefit more than elderly participants for all outcomes except hospital admission (webtable 5
) and nursing-home admission.
Evidence did not suggest that interventions with an increased intensity were more effective in improving any outcome than those that had less direct health-professional involvement, shorter duration, and number of visits (webtable 6
). Similarly, evidence did not exist for benefit of those interventions with multidimensional assessment compared with those with one discipline (one discipline: RR 0·95, 95% CI 0·93–0·97; at least three disciplines: 0·97, 0·89–1·07). No benefit for intense interventions was evident when interventions were grouped by type, including those after hospital discharge. Intervention activity in the control group did not affect outcomes (webtable 7