Sampling wesley lifeforce networks
In 2019, there were 92 networks in operation, with around a third established in 2017 or later . To be included, LifeForce networks had to be operational and established between 2001 and before 2017, leaving 60 Wesley LifeForce networks included. There were more Networks in regions (n = 30.50%) than in large cities (n = 18.30%) or remote regions (n = 12.20%), which fits the profile of all networks . The distribution of networks across Australian states or territories was also broadly representative of the profile of all networks. The networks served areas with a total population of 3,500,951 (average over the period), with a median population of 28,884. The first network was created in 2007 and the most recent in 2016, with 30 (50%) established in 2014 or later.
Suicide counts were obtained from the National Coronial Information System (NCIS) on all closed cases of intentional self-harm (with a final determination of suicide) that had been notified to a coroner between 2001 and 2017. NCIS is a repository of online data for all external cause deaths in Australia. The completeness of NCIS data (i.e. the cause of death was determined and the coroner made a finding) ranged from 95.8 to 99.0% over the study period . Data after 2017 has not been included as it can take up to 3 years for a file to be closed. Date of notification and place of residence were collected for each case. The geographic location and size of the catchment area of the LifeForce networks has been provided by Wesley Mission and confirmed with each network to the extent possible. General demographic data for the LifeForce network catchment areas and control communities were obtained from the Australian Bureau of Statistics (ABS) on population size, remoteness category (large city, inner region, outer region, remote and very remote, Australian Bureau of Statistics ), and relative socio-economic disadvantage, which is an index that ranks areas according to household income, qualification and occupation .
The study was approved by the University of Melbourne Human Research Ethics Board (Ethics ID 1954813.3). Ethical approval was also obtained from the NCIS Research Committee (MO446), the Victorian Department of Justice and Community Safety Human Research Ethics Committee (CF/20/6638), the Coroners Court of Victoria Research Committee (RC 344) and the Western Australian Coronary Ethics Committee (EC02/2019).
Data mapping and selection of control areas
Each LifeForce network provided us with its postcode and a Geographic Information System (GIS) and 2016 ABS suburban and postcode numerical boundaries were used to model the networks’ catchment areas. As some postcodes cover large areas and contain several suburbs, the catchment areas of each network were modeled by selecting all the suburbs that intersected the area of the postcode. Each LifeForce network was given a list of names of suburbs that tentatively represented their catchment area and were asked to review the data by confirming, deleting, or adding additional suburbs as necessary. The LifeForce Networks catchment areas that provided feedback have been edited in the GIS where appropriate and are referred to as the “confident boundary” in the following. Suicide data compiled from NCIS was matched to networks and control areas based on ABS Statistical Areas Level 2 (SA2)Footnote 1 person’s residence code.
Control areas without established LifeForce Networks but with similar demographic characteristics were identified and matched to LifeForce Networks in a 1:1 ratio, based on key criteria including remoteness, relative socio-economic disadvantage, and population size, using ABS population data from the 2016 Census . To maintain similar catchment area sizes in LifeForce networks and control areas, ABS Statistical Areas Level 3 (SA3)Footnote 2 were used to model control areas in metropolitan areas and ABS SA2 statistical areas were used to model control areas in regional and remote areas.
The characteristics of the 60 control areas were similar to those of the network areas. There was no significant difference in socioeconomic disadvantage scores between network and control areas, t(59)=-1.74, p=0.087. Networks and controls are also perfectly adapted to the remote area. Mean population was significantly lower in control areas (M = 29,724, SD = 39,359) than in network areas (M = 58,349, SD = 71,386), t(59) = 3.52, p
Longitudinal mixed-effects models for count data models were developed to examine the effect of network establishment on suicide rates. Due to the frequency of suicide-free months, the data was aggregated to form a quarterly count of suicides by site. The effect of individual sites was included as a random intercept. The numbers of suicides were modeled as a Poisson distribution, as a preliminary test of the likelihood ratio showed that the negative alternative binomial distribution was not significantly superior. The model contained fixed effects for type of site (network or control), intervention status, and time. Intervention status referred to how long a network has been in operation for a particular site and its implementation as one or more binary dummy variables to model change at different times after the establishment of a network on a site.
Given evidence of a nonlinear suicide rate over the period 2001-2017, the shape of the time trend was chosen using fractional polynomials implemented with the fp command in Stata. This allows for a wide variety of functional forms  and showed that a trend including linear and quadratic terms was the best fit. This indicates a decrease in the suicide rate followed by an increase, which is consistent with national data for the period . As suicide rates show a seasonal pattern, with higher rates in spring (September to November in Australia) or early summer [24, 25], the models included a variable for quarter of the year, to adjust for differential trends in suicide rates over the calendar year. Site population size was included as an exposure variable so that model parameters of suicide counts could be interpreted as rates. Population size was calculated from the Australian Bureau of Statistics, which provides annual population data disaggregated by SA2 or SA3 .
We first investigated a model testing whether the introduction of LifeForce Networks led to a dramatic change in suicide rates, with the effect of the intervention modeled as an indicator variable with a value of 0 until establishment and of 1 thereafter. We then explored the pattern of non-constant change in suicide rates attributable to the introduction of the program. While a gradually increasing effect of an intervention is a plausible mode of change, an unconstrained effect is implausible: this would imply that suicide rates continue to decline each quarter after the network is in place. Accordingly, the time after network establishment was fictitiously coded as quarters 1 through 4 (covering the first year after establishment) or quarters beyond the first year. The quarters preceding the establishment of the network constituted the reference category. For control areas, all quarters were coded as zero (reference). We examined whether a site was particularly influential on the model parameters by using a jackknife approach to estimate the model dropping one site at a time.
Effects are expressed as incidence rate ratios (IRR). IRRs below 1 indicate a decrease in suicide rates and IRRs above 1 indicate an increase. The significance level was set at p