Fifteen studies (n=704,134) were included in the review: seven cohort studies, one case-control study nested in a well-defined cohort and seven observational longitudinal retrospective studies of large administrative databases with no control group. All studies met at least 77% of the quality criteria and six met all quality criteria.
The pooled persistence mean was 184.1 days (95% CI 163.9 to 204.3; five studies) and the pooled MPR mean was 66.9% (95% CI 63.3 to 70.5; five studies) at one year follow-up.
Low compliance when compared with high compliance was significantly associated with increased overall fracture risk (RR 1.46, 95% CI 1.34 to 1.60; six studies) from one to 2.5 years after starting treatment. Compared to high compliance, low compliance was significantly associated with increased non-vertebral fracture risk (RR 1.16, 95% CI 1.07 to 1.26; three studies) from 1.9 to 2.2 years, increased hip fracture risk (RR 1.28, 95% CI 1.06 to 1.53; four studies) from 1.9 to 2.4 years and increased vertebral fracture risk (RR 1.43, 95% CI 1.26 to 1.63; two studies) from two to 2.2 years follow-up.
Sensitivity analyses found differences in the summary estimates according to drug used. The risk ratio for studies with participants only taking bisphosphonates was 1.59 (95% CI 1.38 to 1.83; four studies) and the risk ratio for participants taking bisphosphonates plus hormone replacement therapy was 1.18 (95% CI 1.16 to 1.20; two studies).
When studies were categorised according to different thresholds of compliance, the fracture risk estimate changed. The risk ratio for studies where compliance was defined as at least 80% versus less than 80% was 1.51 (95% CI 1.36 to 1.68; five studies) and the risk ratio for studies where compliance was defined as at least 90% versus less than 20% to 50% was 1.37 (95% CI 1.28 to 1.48; three studies). Estimates did not markedly change in the sensitivity analyses according to age and confirmed osteoporotic diagnostic.
Statistically significant heterogeneity was identified in all analyses. There was no evidence of publication bias with reference to Egger's test for any of the outcomes. Inspection of the funnel plot for all-sites fracture risk indicated asymmetry, so publication bias could not be ruled out.