Alcohol hospitalisations by LGA

Estimates of the numbers and rates of deaths and hospitalisations attributable to the use of tobacco, alcohol, to high body mass and other risk factors used age and sex-specific aetiologic fractions developed by the School of Population Health, University of Queensland and the Australian Institute of Health and Welfare and published in 2007 (Begg et al. 2007).

The contribution of 14 health risks to the total burden of disease was assessed by the School of Population Health, using methods developed by the WHO Comparative Risk Assessment project (Ezzati et al. 2004). Earlier work by English and colleagues (English et al. 1995) was also used with reference to risks from the use of drugs and alcohol by the researchers from the School of Population Health. The main elements of the methodology are the prevalence of exposure to a health risk in a population and information on the risk of disease, injury or death from this exposure, which is derived from meta-analysis of published scientific literature. Calculations result in estimates of the proportions of cases of specific diseases and injuries that could be attributed to each risk factor.

For this report, electronic files of the aetiologic fractions developed by Begg and colleagues were obtained directly from the School of Population Health, University of Queensland by the Centre for Epidemiology and Research. The disease and injury groupings used in these files were defined using coding developed for the Burden of disease study (BOD), but a mapping to ICD-10-AM codes was also provided.

There are two steps in applying the aetiologic fractions to a death or hospitalisation dataset:

(a) ill-defined categories (e.g. ICD-10-AM: heart failure, unspecified diabetes mellitus and injuries with unspecified intent) are redistributed into specific BOD categories based on other information in the record and/or on a pro rata basis;

(b) the aetiologic fractions are applied to categorised records.

References

Begg S, Vos T, Barker B. The burden of disease and injury in Australia, 2003. Cat. no. PHE 82 edition. Canberra: AIHW, 2007. Available at http://www.aihw.gov.au/publications/index.cfm/title/10317

English DR, Holman CDJ, Milne MG. The quantification of drug caused morbidity and mortality in Australia. Canberra: Commonwealth Department of Human Services and Health, 1995.

Ezzati M, Lopez AD, Rodgers A, Murray CJL (eds). Comparative quantification of health risks: Global and regional burden of disease attributable to selected major risk factors. Geneva: World Health Organization, 2004. Available at http://www.who.int/healthinfo/global_burden_disease/cra/en/

Local Government Areas (LGAs) are the spatial units which represent the geographical areas of incorporated local government councils. There were 153 LGAs in NSW in 2008, using the 2006 Australian Bureau of Statistics (ABS) Australian Standard Geographic Classification (ASGC) boundaries.

The ABS publishes preliminary estimates of the residential population of LGAs in an annual March report including estimates concerning the previous year (ABS 3218.0 2010).

Population estimates used in this report

Due to the spatial isolation of Lord Howe Island, both the health data and population from Lord Howe Island SLA (population 364 persons) were omitted from the analyses in this report and only Unincorporated Far West SLA (population 1,116 persons) was included as Unincorporated NSW LGA. The 153 LGAs ranged in population from 1,116 to over 280,000 (based on population estimates as at June 2006). The detailed distribution of the population across the local government areas is shown in the Population topic in the report.

There were 13 LGAs with total populations less than 3000, and of these four had populations less than 2000 (based on population estimates as at June 2006). The areas with the smallest populations are particularly vulnerable to variation in their numbers and statstics due to chance.

Smoothing of estimates for small areas using Bayesian smoothing methods in this report

The term ‘small area’ refers to a small geographical area and a small population. Data from a small area are characterised by considerable variability. Smoothing is a general term for methods of minimising variability in data. Examples include rounding, moving averages, extending the period of time in which cases are counted or increasing the size of areas considered. In addition, Bayesian statistical smoothing can be used to adjust raw estimates in small areas by taking into account information from adjacent areas (local or spatial variability) and from the whole state (global or non-spatial variability).

The local government areas (LGA) are the smallest level at which data are analysed in this report. 'Statistical smoothing' using Bayesian smoothing methods are used to control for random variability in the small area estimates and result in more conservative estimates for small areas. These methods are described in general in Smoothing of estimates for small areas using Bayesian smoothing methods and in detail in two sections: Smoothing of estimates for population-based indicators and Smoothing of estimates for binary-type indicators. These Methods sections are available with all LGA based indicators.

References

Australian Bureau of Statistics. Regional population growth, Australia, 2008-09. 3218.0. Canberra: ABS, 2010. Available at http://www.abs.gov.au/AUSSTATS/abs@.nsf/mf/3218.0

Definitions of risk and high risk alcohol drinking

In 2009, the NHMRC published new guidelines to reduce the health risks from drinking alcohol. These guidelines focus on the effects of alcohol during, and immediately after drinking, and introduce the concept of lifetime risk of alcohol related disease or injury. Guideline 1 states that the lifetime risk of harm from alcohol-related disease or injury is reduced by drinking no more than two standard drinks on any day when drinking alcohol. To reduce the risk of injury on a single occasion of drinking the guidelines state that healthy men and women should drink no more than four standard drinks on a single occasion (Guideline 2). Not drinking alcohol is the safest option for children and young people under 18 years of age (Guideline 3), and for women who are pregnant or planning a pregnancy, or who are breastfeeding (Guideline 4) (National Health and Medical Research Council 2009).

These definitions vary from the previous NHMRC guidelines which defined 'any risk-drinking behaviour' as one or more of the following: consuming alcohol every day; consuming on average more than four if male or two if female standard drinks per day; or consuming more than six if male, or four if female, standard drinks on any occasion in the past four weeks (National Health and Medical Research Council 2001). These earlier guidelines also included a level regarded as 'high risk alcohol drinking' (consuming 11 or more standard drinks in any one day if male, and 7 or more if female) (National Health and Medical Research Council 2001). When interpreting changes in the prevalence of risk drinking behaviour over time, changes in these definitions should be considered, along with possible changes in behaviour.

In this report 'risk drinking' has been defined as drinking more than 2 standard drinks on any day when drinking alcohol. Previously used concept of 'high risk drinking' has been discontinued.

References

National Health and Medical Research Council. Australian guidelines to reduce health risks from drinking alcohol. Canberra: NHMRC, 2009. Available at http://www.nhmrc.gov.au/_files_nhmrc/file/publications/synopses/ds10-alcohol.pdf

National Health and Medical Research Council. Australian Alcohol Guidelines: Health Risks and Benefits. Canberra: NHMRC, 2001. Available at http://www.nhmrc.gov.au/publications/synopses/ds9syn.htm

The smoothing used for the population-based indicators is obtained by applying the convolution or Besag, York and Mollie (BYM) model (Lawson et al, 2003). This model incorporates both spatially correlated and uncorrelated variation. It accounts for variability across the entire state (uncorrelated or global variation) as well as variability amongst the local government areas immediately adjacent to the area in question (spatially correlated or local variation). It is a fully Bayesian model which has been used substantially for disease mapping since it was introduced by Besag, York and Mollie in 1991.

Under the BYM model, the smoothed SIR/SMR (or relative risk) is implemented using Gibbs sampling within WinBUGS. The sample values for the parameter of interest ( i) obtained by running this model in WinBUGS range in value within each LGA. This range of values form what is known as the posterior distribution of i, which represents the expected distribution of the SIR/SMR for each LGA when adjusted (smoothed) for the two types of variability mentioned above. For each LGA the mean of this posterior distribution was used as the best estimate of the smoothed SIR/SMR, and the proportion of the probability distribution above or below unity (one) was used to estimate the statistical significance of the small area estimate relative to the state average. It should to be noted that the posterior distribution is dependent upon the expected number of cases: the higher the expected number of cases the smaller will be the standard error of the distribution, and hence the posterior distribution of i will be 'tighter' around its mean.

A Bayesian 95% credible interval, which is obtained by selecting the 2.5th and 97.5th percentiles of the posterior distribution, is analogous to the more common 95% confidence interval used in frequentist-based analyses. It can be interpreted as the range in which 95% of the SIR/SMR estimates are located.

As noted previously, the proportion of the posterior probability distribution that lie above or below unity (one) was used to estimate the statistical significance of the small area estimate relative to the state average. A two-sided test was used, so that if the proportion of the posterior distribution above or below unity was less than 0.025, then that area was considered to have significantly decreased or increased risk at the 5% level of significance respectively.

All tabular output is in alphabetical order by LGA. The columns in the output present the smoothed number of hospitalisations or deaths per year (suppressed if this was less than five); the smoothed SIR/SMR; the upper and lower 95% credible interval endpoints for the smoothed estimates of SSR/SMR; and the level of significance in relation to the state average indicated as follows:

Smoothing of estimates for small areas using Bayesian smoothing methods in this report

The Bayesian smoothing methods are described in general in Smoothing of estimates for small areas using Bayesian smoothing methods and in greater detail in two other Methods sections: Smoothing of estimates for population-based indicators (this document) and Smoothing of estimates for binary-type indicators. These Methods sections are available with all LGA based indicators, general description with every map indicator and one of the detailed descriptions depending on the type of analysis in an LGA map indicator.

References

Lawson AB, Browne WJ, Vidal Rodeiro CL. Disease mapping with WinBUGS and MLwiN (statistics in practice). Chichester: Wiley and Sons, Ltd, 2003.

Smoothing of estimates in this report

Smoothing is a general term for methods used to minimise random variability in data. Examples include rounding, moving averages, extending the period of time in which cases are counted or increasing the size of areas considered. In addition, Bayesian statistical smoothing can be used to adjust raw estimates in small areas by taking into account information from adjacent areas (local or spatial variability) and from the whole state (global or non-spatial variability).

In this report, Bayesian statistical smoothing was used for estimates in Local Government Areas.

These methods are described in general in Smoothing of estimates for small areas using Bayesian smoothing methods (this document) and in greater detail in two other Methods sections: Smoothing of estimates for population-based indicators and Smoothing of estimates for binary-type indicators. These Methods sections are available with all LGA based indicators, general description with every map indicator and one of the detailed descriptions depending on the type of analysis in an LGA map indicator.

Smoothing of estimates for small areas using Bayesian smoothing methods

Mapping cases or rates of events of interest, such as rates of deaths, cases of a disease, or rates of smoking, can be very informative in understanding the geographical distribution of the events. However, low numbers and rates can occur if the event is rare or if the areas studied have small populations (‘small areas’). If numbers or rates are low, then they will also be very variable, since chance events will have an undue effect on the total number. Consequently, estimates of numbers or rates may be too changeable to be reliable for most purposes. Occasionally, there may not be any cases of interest at all in an area, and the estimated rate for that area would be zero.

More reliable estimates of numbers and rates can be obtained by extending the length of time within which the cases are counted, or by increasing the size of the areas considered, but each of these methods may contradict the purpose of the study and undermine the usefulness of the data.

Another option is to apply statistical smoothing methods to calculate more reliable estimates with data collected in a shorter, more up-to-date period. Rates can be estimated even when there were no cases in an area in the relevant period of time.

Bayesian statistical smoothing methods are used to improve the estimates for individual areas by including information on events in adjacent areas which are expected to be similar, and overall variability between all areas. Smoothing has the greatest effect for areas where the number of cases is the lowest.

In this report, Bayesian smoothing was used to adjust raw estimates in Local Government Areas by taking into account information from adjacent areas (local or spatial variability) and from the whole state (global or non-spatial variability).

For population-based indicators such as rates of hospitalisation and rates of death, Bayesian smoothing was performed using the convolution or Besag, York and Mollie (BYM) model (Lawson et al, 2003). This model is widely used for disease mapping. The smoothed estimates calculated are the age-standardised relative risk for each area compared to NSW. That is, the standardised incidence ratio (SIR) for hospitalisations and the standardised mortality ratio (SMR) for deaths. All models were fit using WinBUGS 1.4.3 (WinBUGS, 2007). The details are further discussed in Methods section Smoothing of estimates for population-based indicators.

For indicators such as smoking in pregnancy and attendance at antenatal care, which are based on binary outcomes, smoothing was obtained by modelling the data using a binomial distribution with a logit link function. Smoothed proportions incorporate both local and global information, but are not age standardised. The details are further discussed in Methods section Smoothing of estimates for binary-type indicators.

The results of the Bayesian smoothing were used to determine whether the results obtained from individual areas are significantly different from the estimate of the average for all areas. Smoothed estimates are displayed on choropleth maps for all indicators. The intensity of the colour of a LGA increases as the ratio increases, and the same scale is used for all maps. The level of significance and the direction of difference from the state average is shown using plus and minus signs. One plus sign means that the smoothed estimate for a LGA is significantly greater than the state average at the 5% level of significance and two plus signs mean that the estimate is significantly greater at the 1% level of significance. Conversely, one minus sign means that the smoothed estimate for a LGA is significantly lower than the state average at the 5% level of significance and two minus signs refer to the 1% level of significance. If an area does not differ from the state, then no symbol is shown.

The success of the Bayesian smoothing method depends largely on the degree of similarity between areas that are used in the calculations. In the case of Local Government Areas in NSW, similarity is very high and the method works well.

References

Lawson AB, Browne WJ, Vidal Rodeiro CL. Disease mapping with WinBUGS and MLwiN (statistics in practice). Chichester: Wiley and Sons, Ltd, 2003.