Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | In this example, each observation was assumed independent of the others, but this is not always valid, particularly when using multiple sources of data. For example, cancer incidence recorded in a population registry and population cancer mortality rates via death certificates are clearly not independent of each other. In infectious disease modelling, dependence can arise when some individuals are observed in multiple datasets; Alahmadi et al. [8] provided an example in influenza outbreak modelling where the same people could be captured in a ‘first few hundred’ study as well as in case-notification systems. When the dependence structure of the observation processes is non-trivial, we can use a directed acyclic graph [7] (DAG) to both define and visualise the structure of the random variables and the data in calibration [2]. In these graphs, nodes represent quantities in the calibration, and directed edges represent parent-child relationships, such that any node is independent of all other nodes conditional on its parent and descendants [7]. The model-parameter nodes have no parents and are (marginally) distributed as per the prior; the observed-data nodes have no children, and are distributed as per the observation process from which we have the likelihood; and quantities ‘within’ the model (which make up the map M ) sit in-between the two former types of node. These graphs can be generated using the BUGS language [7] and at their most generic, have the structure shown in Figure 1. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Figure 1: Directed acyclic graph outlining the conditional dependency of generic quantities in a Bayesian calibration. The direction of the arrow indicates parent-child relationships. A node is independent of all nodes conditioned on its parent and descendants. Model parameters, denoted Θ, have marginal distribution given by the prior, fΘ, and the model provides a map, M , from the model parameters to those of the likelihood of the observed data i.e. LX|Φ(X|Φ = M (θ)) ≡ fX|Θ. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | part of the calibration procedure. Jackson et al. [2] calibrated a human papillomavirus prevalence model using data from both a trial and a population screening registry, which did not agree. They estimated the discrepancy between the two as part of the calibration and this had a marked effect on estimates of disease progression rates when compared to calibrating under an assumption of zero-discrepancy. Lastly, data may be missing ‘not-at-random’, and a model of missing responses may also be included and estimated in Bayesian calibration. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | A suitable region of the parameter space to sample from is one where practical identifiability holds, as there is less risk of bias or non-convergence in the next step. Practical identifiability is the property that the distribution of the parameter-values can be uniquely determined by the observed calibration data. It has been further categorised as either ‘global’ or ‘local’, with the respective meanings that the identified value is unique among all possible values, or it is unique only within some neighbourhood of itself. We focus on methods to investigate local identifiability as they can be readily applied. Practical identifiability should be investigated with either the observed data, or using simulated data following the examples of Rutter et al. [22] or Alarid-Escudero et al [9]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | One approach to analyse local identifiability of parameter estimates is to use the profile likelihood [25] or the profile posterior [10]. Raue et al. [see 25, §2.2-2.3] defined (local) practical identifiability as; a parameter estimate is identifiable when its likelihood (or posterior)-based confidence interval (CI) is bounded given a pre-specified confidence level. Estimates such as the maximum likelihood estimate (MLE) or the maximum a posteriori (MAP) estimate are most useful [3] particularly if the likelihood or posterior are unimodal. Estimating the profile is computationally feasible (see §Numerical evaluation of profile posterior in Supplementary Material 1) for a small number of parameters but will suffer from the ‘curse of dimensionality’ in models with many parameters [9]. For large numbers of parameters, it may be more practical to consider sensitivity and collinearity indices used in environmental science and ecology [26, 27], particularly if computational effort of MCMC is comparable to estimation of profiles. If the collinearity index is above a heuristic threshold for a given set of parameters then they may be considered practically non-identifiable. Another approach to assess identifiability is to evaluate an overlap statistic of the posterior and the prior, where a high degree of overlap indicates that a parameter was not practically identifiable [22, 28]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We outline how to compute this statistic after obtaining the posterior sample in §Overlap statistic (Supplementary Material 1). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | In a calibration we must at least examine, if not formally establish, practical identifiability for parameters with improper priors otherwise the posterior will be improper [29] and posterior sampling methods such as Markov Chain Monte Carlo (MCMC) will not converge [10]. If the model predictions are insensitive to non-identifiable parameters, particularly those with non-informative priors, then perhaps the model could be simplified [26] and the previous steps are revisited with the simpler model. Otherwise, additional data either in the form of an informative prior or as part of the observation process (likelihood) is necessary to proceed. A value-of-information analysis may identify the observations that can provide the most additional information for the parameter [30]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Our next step is to obtain a sample from the posterior of the model’s parameters. MCMC methods are widely used for sampling from the posterior, in particular the Metropolis/Metropolis-Hastings, Gibbs, and Hamiltonian Monte Carlo samplers (references for specific algorithms provided below). These algorithms use a Markov chain whose steady-state distribution is the posterior. The sample is obtained by simulating a chain for long enough that convergence can be reasonably established, and the ensuing elements of the simulated chain form the sample. We outline the Metropolis algorithm in §Metropolis sampler in Supplementary Material 1 as an example. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Running multiple chains can assist in monitoring convergence by considering the between and within-chain variance [see 31, §6]. The ratio of these, after discarding a fixed portion of the start of the chain, should be close to one if the chain has converged. The converse statement does not hold, the chain may not have converged even if the ratio is close to one. The tail of the multiple chains, where they have converged, can then be pooled to form the posterior sample. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The starting position of each chain should be drawn from a distribution that is over-dispersed with respect to the posterior. This increases the likelihood that the eventual sample is representative of the domain of the posterior, and not only some neighbourhood of the starting positions. For problems where search algorithms can find such estimates; a suitable initial state for the chain could be drawn from the asymptotic distribution of the MLE (or similar for MAP estimate), with a scaled covariance to increase dispersion. Informative prior distributions may also be useful for drawing the initial state for each chain. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The effective sample size [see 32, §11.5 p286], denoted neff, can be used to determine at what point to end the simulation of (each) chain. The desired neff will depend on the precision required for predictions and tests that will be performed with the model. For example, the order of the error in the posterior mean of an approximately normally distributed parameter is ∼ 1/ neff, and the variance of the posterior sample is dominated by the variance of the posterior itself, as opposed to the uncertainty introduced via a finite sample thereof, with neff as low as 100 [see 32, §10.5 p267]. The algorithms can be tuned for better computational performance, although they are all restricted by how efficiently the likelihood can be calculated. Efficient implementations of the sampling algorithms are readily available, such as stan [33], NIMBLE [34] (itself an extension of BUGS [7]), and PyMC3 [35]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | using these measures. One or more of the best performing models would be selected from the available models to be used for making predictions. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Alahmadi et al. Alahmadi et al. [8] stated, “The gold standard for selecting between models while accounting for prior information is to either calculate Bayes factors or the model evidence.” However, Bayes factors can be computationally intractable and non-informative priors can lead to ambiguity [37]. As an alternative, information criteria such as the Deviance Information Criterion (DIC) and the Watanabe-Akaike Information Criterion [2] (and [see 32, §7.2]) can be used. They are easily evaluated with a sample from the parameter posterior and both criteria penalise models that may have worse predictive accuracy than a simpler model. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Neither the Bayes factor nor the information criteria are intuitive measures of whether a model is sufficient for a particular application. The sum of square error is easy to interpret and is a reasonable measure of performance when the error is normally distributed with zero mean [see 32, §7.1 p167]. However, most models are inadequate, giving rise to ‘model discrepancy’ (or ‘structural uncertainty’) which is the difference between the mean model and mean true output, given (true) input values [38]. The discrepancy function can be modelled and estimated, and summary measures thereof can describe performance, guide model refinement and reduce uncertainty, or the estimate may be included as a correction to the model [1]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Model calibration could be considered complete after the model selection step, however the parameter- posterior sample can be used in ways that a single estimate (or some other form of sample) from other procedures can not. The salient usage is to calculate summary statistics from the sample to describe uncertain knowledge of the parameters or predictions of the model. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We can describe the values of model-parameters compatible with the observed data using the mean, standard deviation, or quantiles of the parameter-posterior sample. Intervals such as the equal-tailed 95% posterior credible interval, which is estimated by the 2.5th and 97.5th percentiles of the sample, describe the range of values of the parameter that would be unsurprising to observe if we were able to measure the true value for the model. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | main drivers of prevalence decline have been price control (tobacco taxes), hard-hitting mass media antismoking campaigns and smoke-free public places [40]. The National Tobacco Strategy in 2012 set a policy target of 10% adult daily smoking prevalence by 2018 and an update on the strategy is yet to occur. A flexible and dynamic model of tobacco smoking prevalence for the Australian population can inform policy targets by quantifying the impact of ongoing and new tobacco control strategies on future trends in smoking, burden of disease, and cost-effectiveness evaluations. To this end, we built a model of the life-course smoking behaviour of the Australian population. The model was similar to the U.S. Smoking History Generator (SHG) developed by the Cancer Intervention and Surveillance Modeling Network [41] that has been used to estimate the impact of tobacco policy [42] and the benefits, harms and cost-effectiveness of lung cancer screening in the US [43, 44]. During the model building process we sought to, within practical limits, place calibration on firm statistical foundations to fulfil a decision maker’s need for dependable and meaningful measures of uncertainty [45]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We used a calibration methodology that incorporated identifiability analyses, Bayesian evidence synthesis of multiple data sources, and uncertainty quantification, as outlined in the first part of the manuscript. Through these analyses we gained the following advantages over earlier smoking behaviour model calibration efforts: 1) we could show that model predictions were measurably improved by allowing those who had quit smoking long-term to switch to reporting as having never smoked; 2) we quantified uncertainty in trends of smoking initiation and cessation due to the random sampling in the surveys and trials used to calibrate the model; and 3) we accounted for competing events exactly as all transitions were calibrated together. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The model structure was an extension of that by Gartner, Barendregt and Hall [11] who predicted smoking prevalence to the year 2056, using: Australian survey data on smoking from 1980 to 2007, population counts and mortality data, and smoking status-specific hazard ratios of death estimated in the American Cancer Society’s Cancer Prevention Study (CPS) II cohort. In brief, Gartner et al. calculated smoking prevalence by combining smoking-related excess mortality with: the estimated sex-specific trend in the proportion that ever-smoked in each cohort, and the age, sex, and year-specific annual proportional change in smoking prevalence. They calibrated the model via a weighted least-squares fit to the observed prevalence, and quantified uncertainty in the predictions using a bootstrapping procedure. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | In our model calibration we included additional surveys to extend the observation period of smoking behaviour to the years 1962-2016 [46]. We estimated smoking status-specific hazard ratios of death from the Sax Institute’s 45 and Up Study [47] using methods published previously [48]. Although clear trends in smoking status-specific hazard ratios of death have been observed over the last 50 years, primarily due to the relative improvement in non-smoker mortality [49], estimates derived from the 45 and Up Study were more relevant to the Australian population and reflected the smoking-related mortality for cohorts that made up much of the smoking survey observations used to calibrate the model. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The annual proportional change in smoking prevalence estimated by Gartner et al.[11] was a measurement of the effect of initiation, quitting and relapse events together, conditional upon one-year survival. In our formulation, we estimated the rate of quitting events alone, without needing to estimate unsuccessful attempts and subsequent relapse, with the help of a ‘recent quitter’ state. We also estimated the rate ex-smokers switched to reporting as a never smoker [50, 51] by the inclusion of a state within the model called ‘reporting as never’. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | determined that independent data on the hazard ratio of death for either ex-smokers or smokers relative to never smokers was necessary for successful calibration. The hazard ratio data were then incorporated into calibration as part of a prior estimate of the parameter values. We derived a mathematical expression to describe the survey sampling process and identified a suitable parameter- value neighbourhood for a Markov Chain Monte Carlo (MCMC) algorithm to sample from. We obtained 200 samples of parameter values using the MCMC algorithm for each model under consideration. We selected the best model out of eight and summarised trends in the life-course of smoking behaviour using a sample of its; proportion that initiate, quit rate, and rate that ex-smokers switch to reporting as never smokers. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | National Drug Strategy Household Survey (NDSHS) 1998, 2001, 2004, 2007, 2010, 2013 and 2016 [see 52, references therein]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | National Campaign Against Drug Abuse and Social Issues Survey (NCADASIS) 1991 and 1993 (including the Victorian Drug Household Survey 1993 sample), NDSHS 1995 (including the Victorian Drug Strategy Household Survey 1995 sample), Social Issues Australia (SIA) survey 1985 [see 53, references therein]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Risk Factor Prevalence Study (RFPS) 1980, 1983 and 1989 [see 54, references therein]. • CCV Australian adult smoking surveys in 1974, 1976, 1980, 1983, 1986, 1989, 1992 and 1995 [see 55, references therein]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Australian Gallup Polls (AGP) no. 158, 160, 168 and 193 (1962-1967) [56, 57, 58, 59]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We defined a current smoker as those who smoked on a daily/regular basis, with the remainder of non-missing cases classified as non-smokers. In all surveys except the Gallup Polls, non-smokers were further categorised as either ex-smoker, recent quitter or never smoker. We defined ex-smokers as non-smokers who had smoked on a daily/regular basis and had quit at least two years prior to the survey, while those who had quit less than two years prior were recent quitters. For surveys without a question on past daily smoking, or missing responses, we used an affirmative response as to whether a participant had smoked 100 cigarettes in their life to identify ex-smokers and recent quitters. We defined all other non-smokers as never smokers. For participants with sufficient information on a quit event, we further divided ex-smokers and recent quitters into age-at-quit categories. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The surveys contained 254,231 observations in total. We singly-imputed age in years for those whose age-at-survey was grouped using a Penalised Composite Link Model fitted to each survey (i.e. 60,131 imputed ages) [60]. We included participants of age 20-99 years at the time of the survey, born from 1910 onwards (229,582 participants). We excluded 3,394 observations for missing smoking status, leaving a total of 226,188 observations. The rate of missing information on whether a non-smoker, who had smoked in the past, had quit within two years prior to the survey was less than 8.6% in all but the 1985-95 NDSHS and the 1989 CCV survey. The 1985-95 NDSHS did not include the relevant questions, while only a subset of ex-smokers in the 1989 CCV survey were asked how long ago they had stopped. We defined all these missing cases as ex-smokers, as the clear majority (86%) of complete cases had quit at least two years prior. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Australian population data by age, sex and in some cases, other demographic variables. We applied an iterative proportional fitting procedure [61, 62] to all surveys to estimate a correction to the supplied weights such that the re-weighted response counts matched Australian Bureau of Statistics population counts by age, sex, state, and capital city versus non-capital city residence. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Mortality We obtained historical mortality data for the Australian population for the period 1930-2016 from the Human Mortality Database: mortality.org [63]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Hazard ratios We estimated smoking-status specific hazard ratios of death from the 45 and Up Study, a prospective cohort study of 267,153 residents of NSW aged ≥ 45 years at baseline (2006-2009), randomly sampled from the Services Australia (formerly the Department of Human Services and Medicare Australia) enrolment database [47]. People 80+ years of age and residents of rural and remote areas were oversampled by design, and the overall response rate was about 18% of invitees. The study oversampled higher income groups than the target population [64], however hazard ratios derived from the study may be generalisable to the population [65]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Fact of death to August 2017 was ascertained from the NSW Registry of Births Deaths and Marriages, with data linkage performed by the NSW Ministry of Health’s Centre for Health Record Linkage (CHeReL; www.cherel.org.au). We applied Cox proportional hazards regressions separately by sex for each 5-year age stratum from 45 to 99 years to estimate the hazard ratios of death for current smokers, recent quitters, and ex-smokers compared to never smokers using similar methods to those reported by Banks et al. [48] Briefly, we excluded participants for the following reasons: aged ≥ 100 years at baseline, data linkage errors, self-reported history of chronic disease at baseline (heart disease, stroke, blood clot and cancer except melanoma and non-melanoma skin cancer), and missing information on smoking status. Hazard ratios were adjusted for age. After exclusions, 191,031 participants were included for analysis. We assumed that current smokers and those who had quit within two years (recent quitters) had the same hazard ratio. The analysis was conducted within the Secure Unified Research Environment. Ethics approval for the 45 and Up Study was provided by the University of NSW Human Research Ethics Committee and approval for linkage by CHeReL was provided by the NSW Population and Health Services Research Ethics Committee. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Figure 2: Structure of the compartmental model of lifetime smoking behaviour. Line with filled circle represents transition to death. Transition rates between boxes depend upon age and sex, except for the transition from recent quitter to ex-smoker which occurs at two years in the former state. µN is the never-smoker mortality rate. µC = HRCµN and µF = HRF µN , where HRC and HRF are the hazard ratios of death for current smokers and ex-smokers, respectively, compared to never smokers. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The estimate of the quit rate was consistent if; all smokers initiate before the selected age; migration is independent of smoking status; and the model of differential mortality was consistent with the population. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We assigned an age-at-quit category so that those who smoked for a shorter duration may transition to reporting-as-never at a different rate from those who smoked for longer. This was implemented as separate states for each age-at-quit group within each smoking status that followed a quit event, as shown in Figure 2. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Never smokers and reporting-as-never smokers died at the same rate, denoted µN . • Current smokers and recent quitters died at the rate µC = HRCµN , with age-specific hazard ratio HRC. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Ex-smokers died at the rate µF = HRF µN with age-specific hazard ratio HRF . |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | amongst current smokers was the same for each cohort. The initial number of recent quitters was determined using assumptions given in §Single cohort equations (Supplementary Material 2). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | where ρN was the total proportion in the never smoker and reporting-as-never smoker categories, ρC was the total proportion in the current smoker and recent-quitter categories, and ρF was the total proportion in the ex-smoker categories. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Parameterisation The two key quantities for the initial population at age 20 years were the proportion of ex-smokers (amongst current and ex-smokers), denoted PF , and the proportion that initiated, denoted PI (one minus the proportion of never smokers). We modelled the logit-transformed proportion that initiated as a natural cubic spline-function of the birth year. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We denote the quit rate as λQ, and we modelled the log-transformed value as the sum of natural cubic spline-functions of age and calendar year (period). We assumed that the rate that ex-smokers transition to the reporting-as-never state, denoted λR, for each age-at-quit group (<30 years, 30-39 years, ≥40 years) was constant. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We pre-specified the knots for each spline term. Mathematical notation and details of the knot locations considered are given in the §Spline terms in Supplementary Material 2. Along with different numbers of knots, terms were included or excluded to form eight different models, labelled ‘null’ for the model with only intercept terms (and none reporting as never) and then labelled ‘A’ through ‘G’, listed in Table S2-1 in Supplementary Material 2. We assumed that the total number of degrees of freedom, including the age-group specific hazard ratios, will be small compared to the number of observations in the surveys so that the parameters might be practically identifiable in each model. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Parameters to calibrate To select suitable parameters to calibrate we informally examined structural identifiability. We considered a simplified scenario where the proportions of never, current and former smokers were known for a cohort. We argued, with details provided in §Structural identifiability of the smoking behaviour model (Supplementary Material 2), that the quantities PF and PI could be identified from the observed proportions at the starting age in the model. The remaining four quantities, λQ, λR, µC and µF , were not identifiable, but if any one of these were known, the remaining three would be identifiable. Therefore, if all of these parameters were included in the calibration, we required at least one to have an informative prior, rather than a non-informative prior. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Prior distribution We assumed that the joint prior of the parameters could be decomposed into independent priors for; the hazard ratios, the proportion PF , the coefficients of the spline describing PI (including an intercept), the coefficients of the sum of the splines in λQ (including an intercept), and each rate of switching to reporting as never. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We used an informative prior for the hazard ratios, given by the asymptotic distribution of the estimate from the Cox regressions. In doing so, we addressed the non-identifiability described earlier by suppling information on both µC and µF . Although our examination of structural identifiability suggested we only needed prior information on one of these two, there was no reason to leave additional readily available information out of the calibration process. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Model of smoking survey responses The dependence structure of the priors, the model, the smoking survey data, and the population mortality data is shown in Figure S2-1 (Supplementary Material 2). In order to sample parameter values within this structure we required an expression for likelihood of the smoking survey data. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We assumed each survey response was drawn from a categorical distribution, with sex-, age-, and birth year-specific probabilities. The three categories were named ‘current’, ‘ex-’ (by age-at-quit group), and ‘never’ smokers, with definitions as per above in §Data sources, except for the recent- quitter category being merged into the current-smoker category. We assumed the probabilities for each category were the same as the sum of one or more expected proportions in the model. The current-smoker probabilities were given by the sum of the model’s current smokers and recent-quitter proportions, ex-smoker probabilities were given by the model’s ex-smoker proportions, and never- smoker probabilities would be the remainder. We found the expected proportions by numerically solving the governing equations of the model, detailed in §Single cohort equations (Supplementary Material 2). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We accounted for uncertainty introduced by non-random sampling using the effective sample size of each cell in the cross-tabulation of each survey by sex, age and birth year. The likelihood for each independent cell was given by a Dirichlet distribution with parameters determined by observed weighted proportions and the effective sample size, details provided in §Survey data likelihood (Supplementary Material 2). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Suitable sampling region We found a suitable region of parameter values to sample from by investigating local practical identifiability of the maximum a priori (MAP) estimate of the parameters. For this step, only, we fixed the hazard ratios at their prior mode. We defined local practical identifiability as the posterior-based 95% ‘confidence’ interval (CI) of the estimate being finite in extent [10]. We used the optim() command in R [66] to find the MAP estimate. We calculated the profile posterior of each component using the predictor-corrector approach defined in the §Numerical methods for practical identifiability (Supplementary Material 1), in the neighbourhood given by ± 7.1 standard deviations of the asymptotic distribution of the MAP estimate. We then estimated the highest level of confidence for which the CI was contained within the neighbourhood; if the level was less than 95% then practical identifiability of the parameter near the MAP estimate was questionable. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The degree to which the hazard ratios were identified by the survey data was investigated with the overlap statistic [28, 22] which was calculated using the approach outlined in §Numerical methods for practical identifiability (Supplementary Material 1) and the posterior sample obtained in the next step; the threshold 0.35 or less was used as evidence of weak identifiability. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Sampler in Supplementary Material 2. We used this procedure for each model listed in Table S2-1 (Supplementary Material 2). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Model selection We estimated the Deviance Information Criterion (DIC) for each model to assess improvement in predictive accuracy between models. The procedure to calculate the DIC using the samples from the joint posterior is detailed in §Deviance information criterion (Supplementary Material 2). We applied some discretion in selecting a model if improvement in the DIC over a nested model was small, to reduce risk of over-fitting. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We summarised the model discrepancy in the proportion of smokers, the proportion of never smokers amongst non-smokers, and the proportion of ex-smokers that quit before age 30 years. The discrepancy was assumed to be a Gaussian process [67] and was estimated by regression of the residual proportions (the difference between the model-expected and survey values) using the mgcv package in R for Generalised Additive Models [68, 69]. The discrepancy was summarised using the expected mean and standard deviation on a log-odds scale of the discrepancy for the observed survey data; with the expectation taken over the parameter-posterior. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Cross-sectional smoking prevalence We generated 200 samples of the predicted proportions of never, current, and ex-smokers for each survey by sex and calendar year. Each sample corresponded to one prediction from the model using one of the samples of the parameter values. We compared the observed proportions in the NDSHS 2016 to model predictions via the probabilities that each model prediction was more extreme than the observed value. The probabilities, which are also known as Bayesian posterior predictive p-values, were estimated using the corresponding proportions of more extreme values in the sample. We modelled the predicted counts, used to determine the predicted proportions, as Dirichlet-Multinomial random variables as described in §Survey data likelihood in Supplementary Material 2. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | the expected proportion that initiated before age 20 years by birth year and sex; • the expected rate of quitting smoking by age, calendar year, and sex; and • the expected rate that ex-smokers would switch to reporting as never in a survey by sex and age-at-quit group. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Each sample corresponded to the expected value of the model using one of the samples of the parameter values. We examined evidence for differences between men and women using the expected probability that the value for men was more extreme than the value for women. We estimated the probabilities using the mean proportion of more extreme values in the men’s sample compared to each value from the women’s sample. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The quit rate cannot be validated directly using the survey data as the quit event in our model is conditional upon no future relapse and we do not know which non-smokers in the survey will relapse. However, in the ITC Four Country Survey, a longitudinal study, it was observed that 95% of those with at least two years abstinence went on to maintain abstinence over the next year [70], therefore we compared the model to an estimate of the quit rate from two years prior to a survey given sustained abstinence of at least two years. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | In the survey data, only about 70-85% of smokers within any birth cohort had started smoking by age 20 years. A more realistic age at completed initiation would be 25 years by which age 93-97% of smokers had started [46]. To assess the impact of assuming all smokers start by age 20 years, we calibrated the selected model using a starting age of 25 years. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | For simplicity (and identifiability) we chose that the quit rate was a sum of age and calendar year terms, which imply at most a linear relationship with cohort. We considered an alternative in sensitivity analysis by using a sum of age and cohort terms, implying at most a linear calendar year effect. The DIC can be used to determine which choice is more compatible with the survey data, i.e. which non-linear effect out of calendar year and cohort is more useful. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | To assess the sensitivity of the model to the assumption that those who smoked and quit after the age of 40 years would always report as an ex-smoker, and not switch to reporting as a never smoker, we calibrated the selected model with the assumption of a non-zero rate of switching for the later age-at-quit group. Similarly, we tested no reporting as a never smoker for all age-at-quit groups if the selected model did allow switching. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | In the main analysis we used the approximate effective sample size of the smoking survey data and individual weights to account for survey design effects. To determine how sensitive our results were to the adjustment for non-random sampling, we also calibrated the selected model ignoring the weights and effective sample size in the likelihood. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | ratios are summarised by their median and 5th and 95th percentiles in Table S2-3 (Supplementary Material 2), along with the overlap statistics. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The parameter-value samples from the MCMC algorithm satisfied our criteria regarding (non- )convergence. All samples reached an effective sample size of at least 50.0, and the estimated potential variance reduction was at most 1.11 for any component after 2472 iterations per chain. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The DIC estimate for model ‘F’ improved upon all simpler models, and model ‘G’ only demonstrated a marginal improvement for men, as shown in Table S2-4 (Supplementary Material 2). To minimise risk of over-fitting, we selected model ‘F’ for the main analysis. Model ‘F’ had: three degrees of freedom (d.f.) in the effect of birth year in the proportion of a cohort that initiated; two d.f. each for the effects of age and calendar-year in the quit rate; and non-zero rates of switching to reporting as never by age-at-quit group for those who quit before age 40 years. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The measures of discrepancy of the selected model were smaller than those for the ‘null’ model. The magnitude of the mean discrepancy - a measure of bias - in either the proportions of smokers in the population or the proportions of never-smokers amongst non-smokers, both on a log-odds scale, was at most 0.023 in the selected model, compared to at most 0.170 for the ‘null’ model. Likewise, the standard deviation - a measure of the unexplained variation - was no greater than 0.153 in the selected model, compared to at most 0.498 in the ‘null’ model. The magnitudes of the mean and the standard deviations of the discrepancy in the proportion of ex-smokers that had quit prior to age 30 years were greater than for the other proportions (for a given model), with values for the selected model at most 0.119 and 0.405 respectively. The mean and standard deviation in the discrepancy for each model and each proportion evaluated is shown in Table S2-5 (Supplementary Material 2). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Cross-sectional smoking prevalence The predicted cross-sectional proportion of Australian men and women age ≥20 years in each smoking status category is shown, by survey, in Figure 3, comparing the survey point estimate with the interval given by the 5th and 95th percentiles of predictions sampled from the model (the 90% equal-tailed intervals or “90% ETI” hereafter). 52.0% of the survey proportions were within the intervals, considerably less than the ideal value of 90%. The Pearson correlations between the survey data and the model-sampled predictions are displayed on each panel by sex and smoking status in Figure 3. The correlations were lower for the quantities with more non-linear calendar year trends, e.g. 90% of the correlations for the proportion of never smokers amongst women were in the interval [0.810,0.900], and were higher for simpler trends, e.g. 90% of the correlations for the proportion of current smokers amongst men were in the interval [0.969,0.982]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The proportions of never smokers, smokers and ex-smokers amongst men, women, and persons in the NDSHS in 2016 and the corresponding ETIs containing 90% of the sampled predictions from the model are shown in Table 1, along with the estimate of the p-value that the predictions were more extreme than the survey value. The p-values for the never-smoker proportions were all greater than 0.810, on the other hand, the p-values for the proportions of smokers and ex-smokers were smaller, near to 0.180 amongst the predictions for men, and no greater than 0.030 for women or persons. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Figure 3: Proportion of never smokers, smokers and ex-smokers amongst Australian men and women age ≥20 years, in four smoking survey series conducted between 1962-2016 (markers), and corresponding intervals given by the 5th and 95th percentiles of the predictions sampled from the selected model (error bars). Size of marker indicates number of respondents age ≥20 years in each survey. Median Pearson correlation between survey data and the model-samples (and interval given by the 5th and 95th percentiles) shown on each panel.Abbreviations; AGPS: Australian Gallup Poll series; CCV: Cancer Council Victoria adult smoking surveys; RFPS: Risk Factor Prevalence Study; NDSHS: National Drug Strategy Household Survey. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | the sample median of the proportions was 87.1% (90% ETI: [85.9%,88.4%]), to 1996, with median 22.7% (90% ETI: [21.1%,24.2%]). The mean proportion of sampled values for men that were more extreme than that of women was less than 5% up until birth year 1967, and again from 1981. The growth rate of the proportion that initiated is summarised in Table S2-6 9 (Supplementary Material 2), by sex, and by the same birth years as Table 2. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The sample obtained from the model of the quit rate per 100 person-years (PY) amongst Australian daily smokers is summarised in Table 3 by sex, age (at 30, 50, and 70 years), and ten-year increments of calendar year from 1940 to 2010 and the calendar year 2016. The sampled quit rates increased with calendar year throughout, while the effect of age varied by sex. In a given calendar year women quit at higher rates at earlier and later ages compared to age 50 years. In the last year of the calibration period, 2016, the median sampled quit rate per 100PY for women at age 30 and 70 years was 5.52 (90% ETI: [5.19,5.88]) and 6.05 (90% ETI: [5.36,6.72]), respectively, whereas at age 50 years, the sample median was 5.00 (90% ETI: [4.77,5.31]) events per 100PY. In a fixed calendar year, the quit rate for men increased with age. In the year 2016 the median sampled quit rate per 100PY for men at age 30 years was 3.67 (90% ETI: [3.44,3.92]), whereas at age 70 years the sample median was 4.70 (90% ETI: [4.13,5.29]). In 2016, the mean proportion of sampled values for men that were more extreme than that of women was less than 5% up until age 47 years, and again from age 65 years. The sample obtained for the growth (rate) of the quit rate is summarised in Table S2-7 (Supplementary Material 2), by sex, and calendar year for the same years as Table 3. The quit rate for women grew faster than for men through the years 1930 until 2016. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Female ex-smokers switched to reporting as a never smoker at a higher rate than male ex-smokers in both age-at-quit groups. The median sampled rate for men who quit before age 30 years was 2.07 (90% ETI: [1.94,2.23]) events per 100PY, and for those who quit between ages 30 and 39 years it was 0.31 (90% ETI: [0.15,0.45]) events per 100PY. For women, the median sampled rate for those who quit before age 30 years was 2.31 (90% ETI: [2.14,2.46]) events per 100PY, and for those who quit between ages 30 and 39 years it was 0.86 (90% ETI: [0.66,1.03]) events per 100PY. The mean proportion of more extreme values for men compared to women was at most 0.063. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The sample of the proportion of those who started smoking before age 20 years who had also quit smoking by age 20 years is shown, by sex, in Table S2-8 (Supplementary Material 2). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The comparison of the estimate of the quit rate from two years prior to a survey (given sustained abstinence of at least two years) to the model’s quit rate is shown in Figure S2-2 (Supplementary Material 2). The survey estimate is, on average, marginally higher than the model, which may be partly explained by relapse after two years amongst some non-smokers. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The estimate of the DIC for each sensitivity analysis, by sex, is shown in Table S2-9 (Supplementary Material 2) along with the estimate for the selected model. The models in each sensitivity analysis were less compatible with the data (greater value of DIC) compared to the selected model, except for the marginal improvement in DIC when female ex-smokers who had quit after the age of 40 years were allowed to switch to reporting as never. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Table 1: Proportions of never smokers, smokers and ex-smokers amongst Australians ≥20 years of age in 2016 observed in the National Drug Strategy Household Survey, and the sample of prediction thereof obtained from the selected model, for men, women and persons. Predictions are described by the median and the interval given by the 5th and 95th percentiles of their samples. The p-value is the proportion of the model-sampled predictions that were more extreme than the observed value in the survey. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Table 2: Sample of the proportion of Australians that initiated daily smoking obtained from the selected model, by sex (men, women, and persons) and ten-year increments of birth year from 1910 to 1990 and birth year 1996. Shown are the sample median and the interval given by the 5th and 95th percentiles of each sample. p-value: M ≡ W; the mean proportion of more extreme values for men compared to each value for women. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Table 3: Sample of the rate that Australian smokers quit daily smoking (per 100 person-years) obtained from the selected model by sex, selected ages (30, 50 and 70 years), and ten-year increments of calendar year from 1940 to 2010 and in 2016. Shown are the sample median and the interval given by the 5th and 95th percentiles of each sample. p-value: M ≡ W; the mean proportion of more extreme values for men compared to each value for women. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | the year 1991 that initiated is shown in Table S2-11 (Supplementary Material 2) along with the sample from the selected model. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The sample of the quit rate in 2016 given the varying the assumptions tested in sensitivity analysis is summarised in Table S2-12 in the (Supplementary Material 2) by sex and at age 30, 50 and 70 years. The greatest sensitivity was to the use of a non-linear cohort effect in the model of the quit rate compared to a non-linear calendar year effect as in the main analysis, resulting in a greater quit rate across all ages and for both men and women, however this was less compatible with the survey data than the main analysis according to the DIC. For all the other variations in the models tested, the effects were mostly seen at older ages, with one exception. The estimated quit rate at younger ages decreased marginally if no ex-smokers were allowed to switch to reporting as never. At older ages, the quit rate increased for both men and women when the prior hazard ratios of mortality for smokers and ex-smokers was taken from the CPS-I study. Amongst women, but not men, at older ages, the estimated quit rate increased marginally if assumptions about ex-smokers being able to report as never smokers were varied. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The sample of the rate of switching to reporting as a never smoker was steady across the sensitivity analyses. The samples are summarised in Table S2-13 (Supplementary Material 2) by sex and age-at-quit group. In the analysis allowing those who quit after age 40, only female ex-smokers in this age-at-quit category appeared to switch at a rate similar to age-at-quit 30-39 years. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | The posterior of the age-standardised hazard ratio of mortality for smokers and ex-smokers in each sensitivity analysis (except for the analysis using the CPS-I data), along with the selected model from the main analysis, is provided in Table S2-14 (Supplementary Material 2), along with the overlap statistic between the prior and the posterior. The posterior hazard ratio was consistent with the main analysis except for a marginal increase in the ex-smoker hazard ratio when ex-smokers were not allowed to switch to reporting as never. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We have presented an outline of the Bayesian calibration of simulation models common to several sciences. A Bayesian calibration is enabled by: identifiability analyses; quantifying prior information about parameters; statistical modelling of the observation process that was used to collect data; Markov Chain Monte Carlo methods; model selection guided by information criteria and other measures of predictive (or otherwise) performance; validation; and, lastly, effective communication of the predictions and the uncertainty that can be accounted for in the Bayesian framework. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Two of our steps focus on identifiability analyses and, within the Bayesian framework, these can inform data collection and evidence gathering, can inform the selection of the prior, and can signal a high risk of non-convergence of the procedures used in later steps. The purpose of identifiability analysis is no longer confined to demonstrating a desirable trait, it also guides decisions in calibration and can eliminate waste of computational effort, and often even a non-identifiable model can be calibrated when a suitable prior is available (what makes a prior suitable depends on the context and philosophical attributes listed by Gelman and Hennig, 2017 [72]). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Expertise across a number of disciplines will be required to succeed in formulating and parameterising both the simulation model and the model of the observation process used to collect data, and the vital step of constructing a map between the two. The expertise may be spread across mathematicians, computer scientists, and statisticians, as well as the need for experts or stakeholders in the process being modelled. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | validity. Menzies et al. [3] stated “. . . policy choice will not wait on all potential issues to be resolved. In this context, calibration should be considered an exercise in creating a reasonable model that produces valid evidence for policy”, which suggests that while not all issues need be resolved, some demonstration of validity belongs in calibration. Model validation is an evolving practise requiring ongoing research [36], and formal Bayesian frameworks have been developed [73, 5]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Estimates and predictions given by the sample of the posterior do not reflect all sources of uncertainty. The uncertainty due to the calibration data being a finite (random) sample is captured, while the uncertainty due to differences between the real process and the model is not. Conveying the meaning of summary statistics of the sample to the audience does not require that they are reported as “estimates of (a statistic) of (some) posterior distribution”. Using a cell division rate as an example, some audiences would better understand the statement “90% of our samples of the cell division rate were inside the interval [0.02,0.05] events per hour” than the statement “We estimated that the 90% posterior credible interval of the cell division rate was [0.02, 0.05]”. Discussion of which sources of uncertainty are accounted for in these statements will be required regardless of which statement is used, however the latter statement describes features that only some audiences may recognise or care about. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We calibrated a model of smoking behaviour in Australia using the steps outlined in §Bayesian calibration of simulation models: tutorial. The calibration process synthesised multiple data sources including 26 smoking surveys, population mortality data [63], and the 45 and Up cohort study [47], and we obtained samples from the model of: the predicted prevalence for each survey; the proportion of each birth cohort that initiated smoking; the rate of quitting; and the rate that those who had quit long-term switched to reporting as having never smoked. The median of these samples and the intervals given by their 5th and 95th percentiles describe the range of values that are most compatible with the data sources we used and they indicate the range that we would expect in the population. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Our model allowed those that had quit smoking to report as having never smoked, which is one of the mechanisms that may explain increases in the never-smoker proportion observed at early ages within a cohort [46]. This transition was more frequent for those who quit before age 30 years than later, and we speculated that it would also be more common for those who had smoked occasionally or at a lower intensity. The inclusion of this pathway decreased the discrepancy in the proportion of those who quit at younger ages, and was more compatible with the data than a model without this pathway. Removing this pathway decreased the estimated proportion that initiated and increased the quit rate at middle or older ages. Other models may obtain better estimates of the initiation rates and quit rates by including this effect. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We modelled the discrepancy in a similar, but not identical, fashion to that of Kennedy and O’Hagan, 2001 [38], and we estimated it with a widely available and efficient procedure [69]. We described bias and unexplained variance by age and cohort for a more diverse set of outcomes (expected proportions of; smoker vs non-smoker, ex-smoker amongst non-smokers, and those who quit before age 30 years amongst ex-smokers) than was presented for the United States CISNET Smoking History Generator [41] and the Australian model upon which this work was based [11]. This report may be a useful blueprint for assessing model performance in a manner that mutes the effects of stochastic (or aleatory) uncertainty. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | than revising the model and calibrating again, incorporate the estimate of the discrepancy into predictions [73]. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | We observed that only 52% of the observed cross-sectional proportions (of never smokers, smokers and ex-smokers) were contained in the 90% equal-tailed intervals. This deficiency could be caused by either an incorrect model of the survey data, for example our ambivalence to between-survey effects, or that the model is lacking some detail or effect. Changes of delivery and the sequence or questions included can affect the response rate, missing response rate and the classification of a respondent, and generate between-survey effects. Extensive changes were made to the questionnaire in the NDSHS in 1998, and estimates of the prevalence of current smokers and ex-smokers in the NDSHS and NCADASIS prior to 1998 seem like outliers relative to the 1998 (and later) NDSHS and the CCV surveys conducted contemporaneously, suggestive of between-survey effects which we did not account for. We briefly consider issues regarding face validity. We did not account for the effect of migration in and out of the population. Differences in smoking prevalence between immigrants and emigrants, combined with enough migration, would bias the estimates of the rates and the model predictions [11]. Earlier cohorts of women initiated smoking later [46] and this could cause under-estimation of the quit rate at younger ages in earlier cohorts; although this biased appeared small in sensitivity analysis. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | This study had some structural limitations. 1) We did not consider interactions between calendar year and age in cessation rates, meaning we could not examine differences in the long-term effects of social norms and tobacco policy between age groups. 2) Those who quit at younger ages have a lower risk of smoking-related mortality than those who quit at later ages, however our model did not vary the hazard ratio of mortality by age-at-quit, therefore our estimate of the quit rate at younger ages was likely biased. 3) Importantly, we have not considered how smoking intensity varies across age and between cohorts; a vital measure of exposure for use in analyses of health outcomes in the population. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Bayesian calibration enables the use of multiple data sources to inform model predictions, which is vital when a single data source either lacks critical information or is too imprecise on its own. The calibration process is supported by identifiability analysis to minimise waste of computational effort, and the robustness of the results can be strengthened by model selection and analysis of model discrepancy. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Our Australian smoking behaviour model was the first cohort-specific model of the life-course of smoking that was calibrated using Bayesian statistics and was the first to demonstrate the significant impact of recanting bias on the rates of initiation and quitting daily smoking. In recent years, fewer Australians within a birth cohort started daily smoking than at any time over the study period, and the rate that people quit daily-smoking is at its highest. There are signs of emerging differences between men and women, with fewer women starting smoking than men, and women who smoke are now quitting at a higher rate than men. This model can be used to forecast future smoking prevalence rates in the Australian population to assess the potential impact of new or ongoing interventions in tobacco control. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Ministry of Health; NSW Department of Communities and Justice; and Australian Red Cross Lifeblood. We thank the many thousands of people participating in the 45 and Up Study. |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Data linkage of 45 and Up Study and NSW Registryof Births Deaths and Marriages performed by the NSW Ministry of Health’s Centre for Health Record Linkage (CHeReL; www.cherel.org.au). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Victorian Drug Strategy Household Survey 1995. • National Campaign Against Drug Abuse and Social Issues Survey 1991, and 1993. • Victorian Drug Household Survey (VDHS) 1993. • Social Issues Australia Survey 1985. • Risk Factor Prevalence Study (RFPS) 1980, 1983, and 1989. • Cancer Council Victoria Australian adult smoking surveys 1974, 1980, and 1983. • Australian Gallup Polls (AGP) no. 158, 160, 168, and 193 (1962-1967). |
Processed_Bayesian_calibration_of_simulation_models:_A_tutor.txt | Some of data analysis for this paper was generated using SAS software, Version 9.4 Release M6 of the SAS System for Windows (x64). Copyright ©2018 SAS Institute Inc. SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc., Cary, NC, USA. |
Processed_Metallicity_distributions_in_and_around_galaxies.txt | Abstract: Metals are found in all baryonic phases and environments, and our knowledge of their distribution ‘in and around galaxies’ has significantly improved over the past few years. Theoretical work has shown that the fraction of metals in different baryonic components can vary significantly when different feedback schemes are adopted. Therefore, studies of element abundances provide important information about all gas-dynamical processes which determine the cosmic evolution of baryons. I give here a brief review of recent observational progress, describe the implications of recent theoretical studies, and discuss briefly future prospects. |
Processed_Metallicity_distributions_in_and_around_galaxies.txt | During the last decade, significant progress has been made in the study of element abundances. Accurate measurements of metallicities have been collected for a large number of individual stars for our own Galaxy, and for a number of the brightest members of the Local Group. Large scale surveys such as the 2dF galaxy red- shift survey and the Sloan Digital Sky Survey (SDSS) have allowed abundance determinations in hundreds of nearby galaxies. The new generation of large tele- scopes and sensitive detectors have allowed us to mea- sure a range of chemical elements in stars, HII regions, cold interstellar gas, and hot intergalactic medium, up to an epoch when the Universe was only about one tenth of its current age. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.