Prior to the 2017–18 SHS collection, not all SHS agencies provided data to the AIHW (for example, in 2016–17, 97% of agencies provided all required data). Annual reporting of SHS data used weighting to compensate for missing agency data. For more information, see SHS Annual Report: Technical information.
Weighting has not been used for longitudinal analyses, so the interpretation of results should acknowledge that a small number of clients may be missing from the data (either completely or in part) before 2017–18. From 2017–18 onwards the SHS collection has nearly full coverage.
Relative Risk of service use
For study cohorts that include a comparison cohort, the longitudinal analyses include relative risk of service use between the study cohort and its comparison cohort. These relative risks are provided for all 3 study periods (retrospective, defining, prospective). For service use in the retrospective and prospective periods, clients (from both cohorts) that did not receive any services in that period are excluded from the calculation of relative risk. Therefore, relative risks for those two time periods are comparing the need for each service between clients that were in the SHS data in that period and in the study cohort (and therefore received a service, but not necessarily the one being examined) with the need (or lack of) for each service for clients in the comparison cohort that received at least one service in that period.
Differences in the number of SHS clients between longitudinal and annual data
The number of clients in longitudinal cohorts for a given reporting period will differ to the number of clients receiving services in the same reporting period, as reported in the AIHW SHS annual reports, for a number of reasons:
- The longitudinal cohorts are limited to clients that commenced a support period in a given defining study period (for example, 1 July 2015 to 30 June 2016). Conversely, the SHS annual reports focus on clients that receive support at any time during the reporting period, which will include clients that commenced supported in previous reporting periods (typically around 12% of clients commenced their support in previous reporting periods).
- The longitudinal cohorts are limited to clients that have closed support periods. That is, to be considered for a longitudinal cohort a client needs to commence their support during the defining study period and that support has to finish (at any time). This is so that outcomes can be measured for longitudinal clients. Clients in the annual data can have multiple support periods during the reporting period, some of which may be closed, but the last support period that starts during the reporting period can be open. Therefore, if a client only has open support periods in the reporting period, they will be excluded from the longitudinal data but may be included in annual reporting.
- Both the longitudinal analyses and the annual reporting exclude client with missing identifiers (statistical linkage keys) – this is to allow client data for multiple support periods to be grouped together. The longitudinal analyses apply stricter criteria to the quality of the identifier, however, and therefore excludes more clients.
Typically, a longitudinal cohort will have less clients than the corresponding annual cohort.
The SHSC longitudinal analyses make frequent use of binary outcomes as model targets. For example, analyses for the 2015–16 FDV cohort model the likelihood that clients will use SHS support in the future (binary outcomes: yes or no) based on client characteristics and client reasons for seeking assistance during the defining study period).
Modified Poisson Regression modelling with robust error variance was implement using the method of Zhao (2013) because this yields relative risk rather than the odds ratios of logistic regression (Zou 2003).
Importantly, the regression modelling undertaken with the SHSC longitudinal data is neither causal/explanatory nor predictive. That is, it neither aims to test causal hypothesis about what client factors (constructs) are relevant to outcomes (as per causal or explanatory modelling), nor is it predictive (that is, it is not intended to model future outcomes; Shmueli, 2010).
Instead, the modelling of the SHSC longitudinal data is descriptive. Their purpose is to measure the association between the dependent and independent variables Shmueli, 2010). The models are not refined nor are they internally validated; they therefore have no predictive value. Furthermore, there is no a priori development of causal theories that would allow the models to explain the reasons for given outcomes. Instead, the results serve as a starting point for further questions that could be subsequently addressed in exploratory models that explore individual causal factors.
Shmueli G (2010) To Explain or to Predict? Statistical Science, vol. 25, no. 3, pp. 289-310
Zhou K (2013) Proper Estimation of Relative Risk Using PROC GENMOD in Population Studies. SAS Conference Proceedings: Western Users of SAS Software 2013, November 13-15, 2013, Las Vegas, Nevada
Zou G (2003) A Modified Poisson Regression Approach to Prospective Studies with Binary Data. American Journal of Epidemiology, vol. 159, no. 7.