Return on investment 2: evaluating the cost-effectiveness of needle and syringe programs in Australia 2009

Appendix B: Input data and assumptions for epidemic model

Page last updated: 2009

Mathematical models of epidemiology heavily rely on the behavioural, social, epidemiological, clinical, and biological data available. They are only as good as the quality of input data provided. For this report, sufficient data was accessible from various sources to inform model inputs. However, it should be noted that if the data are biased, particularly if trends in data over time are not truly reflective of the real world population, then this will directly bias the results and implications of the model. To compensate for potential bias in reported data, uncertainty with some parameters, and intrinsic heterogeneity associated with some factors of our model, we conducted rigorous uncertainty and sensitivity analyses. This involved defining a probability density function for each model parameter, instead of using point estimates, based on plausible ranges established from data published in the international peer-reviewed literature and Australian reports and databases. Exhaustive searching of the literature and other data sources was carried out. All inputs and assumptions of the epidemic model, including justifications and data sources, are provided in Table B.1, Table B.2 and Table B.3. Table B.1 lists demographic, epidemiological, and behavioural parameters; Table B.2 lists HIV disease parameters; and Table B.3 lists HCV disease parameters.

Latin Hypercube Sampling96-102, a type of stratified Monte Carlo sampling, was employed to stratify the defined probability density functions of each parameter (into N=10,000 equiprobable intervals) and the value of each input parameter was randomly chosen for a given simulation of the model. Each input value was used only once in the entire sampling analysis. Thus, we simulated the epidemics of HIV and HCV among Australian IDUs 10,000 times, producing 10,000 epidemic trajectories. Distributions of the outcome variables (e.g. incidence, prevalence, and the number of people in each health state over time) were then produced from the 10,000 model simulations. The appropriateness of each model simulation was determined by comparing (through a standard χ2 statistic) the model-based incidence and prevalence estimates of HIV and HCV with prevalence and notifications estimates from annual NSP survey data54 and Australia's surveillance system.6, 23 A form of Monte Carlo filtering was performed:103, 104 simulations that were ranked in the top 1,000 best fits over the period 1999-2008 for both HIV and HCV epidemics were then retained and other simulations were excluded from further analysis. Results of the model calibration are shown in the epidemiological impact chapters. Sensitivity analyses were then conducted to study how the uncertainty in the output of the model can be apportioned to sources of uncertainty in the model inputs.96, 105 The techniques employed were calculation of partial rank correlation coefficients (PRCCs),96-99, 106-117 calculation of standardised regression coefficients between model parameters and outcomes,118 and factor prioritisation by reduction of variance (calculating first order sensitivity indices).119-130 Sampling of parameter distributions and all sensitivity analyses were performed with the SaSAT software package.131

Table B.1 Demographic, epidemiological and behavioural parameters

Table B.2 HIV disease parameters

Table B.3 HCV disease parameters