From January 2016 to December 2019, there were 11,516,331 presentations to EDs in public hospitals across New South Wales. After excluding ineligible and non-LBP presentations, 176,729 were retained in the analysis. Characteristics of the included presentations are presented in Table 1. The mean (SD) age of the entire sample was 51.8, 28.4% were aged 65 and older, and 52% were female. Most presentations were classified as semi-urgent or urgent, and 92.5% of them were for LBP without neurological signs and symptoms.
Presentations occurred at 177 public hospital EDs across New South Wales. Of these, 142 were in rural or regional areas of New South Wales, and 35 in the metropolitan Sydney area. Thirteen EDs were in principal referral hospitals, 21 in major hospitals, 40 in district hospitals, 34 in community hospitals, 60 in multi-purpose hospitals, 2 in sub-acute hospitals, and 7 in ungrouped hospitals. Principal referral, major, and district hospitals were responsible for most presentations (n = 156,285, 88.4%). Across all seven peer-groups, presentations were similar in terms of age, percentage of presentations aged 65 and older. Principal referral hospitals had a higher proportion of presentations arriving by ambulance compared to major, district, community, multi-purpose, and sub-acute hospitals. They also had a lower proportion of presentations classified as non-urgent and higher proportion of presentations classified as urgent compared to all other peer-groups (Supplementary file 4).
There were 44,459 hospital admissions due to LBP from 2016-2019, representing an unadjusted hospital admission rate of 25.2%. Principal referral hospitals had the highest absolute number and proportion of patients admitted (n = 20,078, 35.5%), followed by sub-acute (33.9%), major (26.2%), ungrouped (21.6%), multi-purpose (17.4%), community (17%), and district (13.5%) hospitals. Within hospitals, the proportion of presentations that were admitted ranged from 0% (11 hospitals) to 46.8% (1 hospital).
Model fit and discrimination
The adjusted models (c-statistic = 0.824 for both) were better at correctly discriminating presentations that resulted in hospital admission compared to the fixed effects-only logistic regression model (0.781). Likewise, the adjusted models fitted the data well (Supplementary file 3). Both models are described in Table 2.
The ICC (95% CI) adjusted only for case-mix was 0.14 (0.12 to 0.17) and 0.10 (0.08 to 0.13) when also adjusted for hospital peer-group. This means that 14% of the variation observed in the model adjusted for case-mix only was attributed to differences across hospitals. This variation reduced to 10% when the model was adjusted for hospital peer-group. The MOR for the case-mix adjusted model was 2.03. After adjusting for hospital peer-group, the MOR reduced to 1.8.
Hospital-adjusted admission rates
There was marked variation in the HAAR across hospitals in both models. In model 1, HAAR ranged from 6.9% to 65.9%. This variation was observed in hospitals with both small and large numbers of LBP presentations (Figure 1). Upon stratifying results in model 1 (Figure 2, top row), larger hospitals had a higher proportion of hospitals classified as having a high admission rate. For example, all principal referral hospitals and 76.2% of major hospitals fell within that category compared to 7.5%, 35.3%, 18.8%, and 28.5% of district, community, multipurpose, and other hospitals, respectively. There were also hospitals that had a lower HAAR in relation to the state average, with 23.8% of major, 55% of district, 14.7% of community, 6.6% of multi-purpose, and 14.3% of other hospitals being classified as such.
Model 2, which adjusted the multilevel model for peer-group characteristics, provided distinct results. Less principal referral hospitals were classified as having high HAARs; 38.5% were classified as having admission rates comparable to the state average, and 23.1% as having admission rates lower than the stage average. Similar findings were noted for major hospitals, whereas more district hospitals had a HAAR within the 95% confidence limits (Figure 2, bottom row).