META-ANALISI BAYESIANA DI STUDI OSSERVAZIONALI SUI FATTORI DI RISCHIO PER LE CADUTE NELL'ANZIANO

Background. Despite the widespread application of Bayesian methods in meta-analysis, the use of clinical informative priors is still lacking. Methods. Using MEDLINE and previous reviews, we searched for prospective studies investigating risk factors for falls among community-dwelling older people. For 31 risk factors we computed pooled odds ratios (ORs) with random-effects frequentist models. For five risk factors (benzodiazepines use, female sex, history of falls, urinary incontinence, antiepileptic use) we computed pooled ORs both with frequentist and Bayesian methods using four different priors: non informative, clinical, skeptical and enthusiastic. Except for non informative one, the other priors were based on a clinical belief elicited from experts by mean of a specifically designed questionnaire. Clinical priors were included in the model as normal distributions computed with the moment method; the enthusiastic and skeptical priors were normal and skewed-normal distributions, which described the higher and lower estimates provided by the experts. The analyses were performed using the Markov Chain Monte Carlo approach. Results. A total of 74 studies met the inclusion criteria and 31 risk factors were included in the frequentist analysis. Fourteen experts (eight geriatricians and six general practitioners) provided ORs estimates for each selected risk factors for two categories of elderly people (75 years old, more than 80 years old). Geriatricians’ clinical priors gave better results in terms of estimate consistency between frequentist and Bayesian models and reduction of credibility interval. Enthusiastic and skeptical estimates appeared to be heavily driven by the prior distribution and they are not consistent with frequentist, non-informative and clinical ORs. Conclusions. In a meta-analysis a Bayesian model including an informative clinical prior can provide external information and could be useful in selected cases (i.e. insufficient number of studies available, confidence intervals too wide and including 1.00).