Aboriginal and Torres Strait Islander Peoples: Smoking Trends, Australia methodology

Latest release
Reference period
1994 to 2014-15
Released
19/10/2017
Next release Unknown
First release

Explanatory notes

Introduction

1 This publication describes trends in smoking habits among the Aboriginal and Torres Strait Islander population using data from a number of ABS surveys conducted throughout Australia, including remote areas, over the 20 year period from 1994 to 2014–15. Comparisons with trends for the non-Indigenous population during the period 1995 to 2014–15 are included where relevant.

2 The Menzies School of Health Research made a significant contribution to this publication, which is an outcome of a joint ABS and Menzies project to examine trends in smoking by Aboriginal and Torres Strait Islander peoples over a 20 year period by applying regression modelling to a subset of data from 12 health and social surveys by the ABS.

3 Further information on the smoking questions asked in each survey which are relevant to this piece of research is provided in Appendix 1: Smoking Questions by Survey. Appendix 2: Results of Modelling provides detailed information on the results of the modelling for average annual relative and absolute changes in tabular form, including significance levels and confidence intervals.

4 This release includes supporting time series tables which can be accessed via the Data downloads section. For the 2012–13 data in these tables, the 2012–13 NATSIHS has been used to provide consistency across the tables. This is because some of the smoking questions were not asked of every person in the 2012–13 AATSIHS.

5 Throughout this release, the term ‘Aboriginal and Torres Strait Islander’ refers to all persons who identified themselves as being of Aboriginal, Torres Strait Islander, or both Aboriginal and Torres Strait Islander origin.

6 Explanations of terms and concepts are provided in the Glossary.

Modelling estimates of average annual relative and absolute change

7 Regression modelling was used to estimate average annual relative and absolute change during the period 1994 to 2014–15. Further information about the modelling processes that were undertaken is provided in the Technical Note: Modelling Average Annual Change.

Dataset used for regression modelling

8 The dataset used for the regression modelling comprises Aboriginal and/or Torres Strait Islander persons and non-Indigenous persons who were selected to participate in any of the 12 national health and social surveys conducted by the ABS during the period from 1994 to 2014–15 and who were asked questions about smoking (Table 2.1). Predetermined survey reference points were used to measure the approximate length in time in years between 1 July 1994 and the relevant survey, and for determining whether a survey would be included in the analysis of trends before and after 1 January 2008. Surveys designated as occurring within a calendar year (e.g. 2008 NATSISS) were assigned a reference point of 1 July during that year. Surveys designated as occurring within a financial year (e.g. 2007–08 NHS) were assigned a reference point of 31 December during that year.

Table 2.1 ABS national health and social surveys used for regression modelling

SurveySurvey reference pointSample size (no.) for analysis
Smoking prevalence (18 years and over)Smoking prevalence (15- 17 years)Quit ratio(a)
ABORIGINAL AND TORRES STRAIT ISLANDER POPULATION
1994 National Aboriginal and Torres Strait Islander Survey1 Jul 19947 700893
2002 National Aboriginal and Torres Strait Islander Social Survey1 Jul 20028 4638265 974
2004–05 National Aboriginal and Torres Strait Islander Health Survey31 Dec 20045 7564 188
2008 National Aboriginal and Torres Strait Islander Social Survey1 Jul 20087 1636605 166
2012–13 Australian Aboriginal and Torres Strait Islander Health Survey31 Dec 20128 1577575 738
2014–15 National Aboriginal and Torres Strait Islander Social Survey31 Dec 20146 6044184 725
NON-INDIGENOUS POPULATION
1995 National Health Survey1 Jul 199538 55419 921
2001 National Health Survey1 Jul 200117 6939 192
2004–05 National Health Survey31 Dec 200419 27010 551
2007–08 National Health Survey31 Dec 200715 5719358 164
2011–12 Australian Health Survey31 Dec 201124 3961 42112 611
2014–15 National Health Survey31 Dec 201414 2538157 012

a. 18 years and over who have ever smoked
 

9 Persons for whom smoking status was unknown were excluded (5,403).

10 Four of the surveys did not ask smoking questions of people aged 15–17 years (1995 NHS, 2001 NHS, 2004–05 NHS, 2004–05 NATSIHS). Non-smokers in the 1994 NATSIS were not asked if they were ex-smokers or had never smoked, and so could not be included in the analyses of the quit ratio. The education variable was not available for the 1995 NHS. Further information about which questions were asked in which survey can be found in Appendix 1: Smoking Questions by Survey.

11 Starting with the 2004–05 NATSIHS and 2004–05 NHS, persons who did not currently smoke and who had never smoked daily were asked additional questions as to whether they had ever smoked 100 cigarettes or more in their life or smoked pipes, cigars or other tobacco products at least 20 times in their life. Those who answered yes to either of these additional questions were classified as ex-smokers. This had the effect of increasing the number of ex-smokers and decreasing those who had ‘never smoked’ relative to the methodology used in the earlier surveys.

12 The non-Indigenous surveys excluded Very Remote areas of Australia and discrete Aboriginal and Torres Strait Islander communities but these exclusions are unlikely to affect national estimates for the non-Indigenous population.

Variables used when estimating average annual relative change

13 The following table indicates which variables were used when estimating average annual relative change in smoking prevalence, smoking initiation and smoking cessation and the reference categories used for comparisons.

Table 2.2 ABS national health and social surveys, variables used for estimating average annual relative change - personal characteristics

VariableCategoriesModel  
Smoking prevalenceSmoking initiationSmoking cessation
PERSONAL CHARACTERISTICS
Sex1MaleXXX
2Female (reference)   
Age group218–24 years (reference)X X
325–34 years   
435–44 years   
545–54 years   
655 years and over   
Age115 years (reference) X 
216 years   
317 years   
Indigenous status1Non-Indigenous (reference)XXX
2Aboriginal and/or Torres Strait Islander   
Highest year of school completed1Year 12 or equivalent (reference)X X
2Year 11 or below or never attended school   
3Not stated or not applicable(a)   
Smoker status1SmokerXX 
2Non-smoker (reference)   
Quit status – persons who have ever smoked1Ex-smoker  X
0Smoker (reference)   
GEOGRAPHICAL CHARACTERISTICS
JurisdictionNSWNew South Wales (reference)XXX
VICVictoria   
QLDQueensland   
SASouth Australia   
WAWestern Australia   
TASTasmania   
NTNorthern Territory   
ACTAustralian Capital Territory   
Remoteness1Non-remote (reference)XXX
2Remote   
TIMES
Number of years between survey reference point and 1 July 1994Continuous variable taking values in the range of 0 to 20.5 years (0–1994 NATSIS) (reference)XXX
Survey reference point after 1 January 20081AfterXXX
2Before (reference)   

a. Education variable that was not available in 1995 NHS was categorised as ‘not known’.
 

Variables used when estimating average annual absolute change

14 The following table indicates which variables were used when estimating average annual absolute change in smoking prevalence, smoking initiation and smoking cessation and the reference categories used for comparisons.

Table 2.3 ABS national health and social surveys, variables used for estimating average annual absolute change

VariableCategoriesModel
Smoking prevalenceSmoking initiationSmoking cessation
Sex0Female (reference)XXX
1Male   
Age group218–24 years (reference)X X
325–44 years   
445 years and over   
Age115 years (reference) X 
216 years   
317 years   
Indigenous status1Non-Indigenous (reference)XXX
2Aboriginal and/or Torres Strait Islander   
   
Smoking prevalence – 18 years and overContinuous variable X  
Smoking prevalence – 15–17 yearsContinuous variable  X 
Quit ratioContinuous variable   X
Remoteness1Non-remote (reference)XXX
2Remote   
Number of years between survey reference point and 1 July 1994Continuous variable taking values in the range of 0 to 20.5 years (0–1994 NATSIS) (reference) XXX
Survey reference point after 1 January 20081AfterXXX
2Before (reference)   

Acknowledgment

15 The data used in this publication was dependent on the high level of cooperation received from Aboriginal and Torres Strait Islander peoples and their communities. Without their continued cooperation, the wide range of Aboriginal and Torres Strait Islander statistics published by the ABS would not be available. Information received by the ABS is treated in strict confidence as required by the Census and Statistics Act 1905.

Appendix 1 - smoking questions by survey

Show all

Appendix 2 - results of modelling

Show all

Technical note - modelling average annual change

Introduction

1 Regression modelling was used to produce estimates of average annual relative and absolute change. This Technical Note provides further information about the modelling processes that were undertaken. Information about the dataset and variables used in the modelling is provided in the Explanatory Notes.

Modelling average annual relative change

2 Logistic regression is widely used in many fields, including the medical and social sciences, as a statistical method for modelling categorical outcomes. Binary logistic regression is used for modelling a dichotomous outcome (for example, modelling 1 and 0, where 1 = event/outcome of interest occurs and 0 = event/outcome of interest does not occur). In this publication, this method is used to estimate average annual relative change, as people are either smokers or not, or have successfully quit or not. The output of the logistic regression model is usually presented in terms of the odds that the event will occur. For each categorical characteristic in the model a comparison group is selected.

3 The binary logistic model for estimating the likelihood of an event can be expressed as follows:

\(\Large{log \left(\frac{p_{event}}{p_{non \ event}}\right)=\alpha+\beta_{1} x_{1}+\cdots+\beta_{k} x_{k}}\)

where \(\alpha\) is the intercept, the \(\beta_{1}, \ldots, \beta_{k}\) are \(k\) regression coefficients, and the \(x_{1}, \ldots, x_{k}\) are a set of \(k\) explanatory variables for each individual in the data.

4 This model indicates that the log of the odds of an event occurring is a linear function of the explanatory variables. The coefficients of the model can be estimated using standard maximum likelihood techniques.

5 To take into account the sample design of the surveys, sampling weights were applied in the models. To adjust for possible dependence of units within clusters due to clustering sample design, clustering was also implemented into the modelling process.

6 The time variable is the primary explanatory variable of interest in all models as it describes the annual relative trend in the outcome variable (smoking prevalence for people aged 18 years and over, smoking initiation among people aged 15-17 years, or the quit ratio for people aged 18 years and over).

7 The logic of the modelling approach is as follows.

  • Firstly, models were built for each outcome variable with time, age (or age group), sex and Indigenous status and an interaction term was added between Indigenous status and time. The interaction was then tested to see whether it was significant or not. If the interaction is significant then it can be concluded that there is a significantly different relative trend for the Aboriginal and Torres Strait Islander population and the non-Indigenous population. More covariates were then added into the model to understand whether controlling for remoteness, state/territory and/or education changed the interaction.
     
  • The next step was to model outcomes separately for the Aboriginal and Torres Strait Islander and non-Indigenous populations. Sex and age were tested as covariates in the model, including whether there were significant interactions between these covariates and the time variable. A significant interaction means significantly different relative trends for the different categories of the covariate. If any interactions were not significant then these were removed sequentially until all remaining interactions were significant (p<0.05), removing the least significant first. More covariates were then added and their interactions with time assessed to determine whether there were also significantly different trends by remoteness, state/territory or education.
     
  • Finally, an assessment was undertaken to determine whether relative trends were different before and after the increased attention to and funding for tobacco control programs targeted to the Aboriginal and Torres Strait Islander population since 2008. This involved creating a new variable with a value of 0 for before 1 January 2008 and a value of 1 for after 1 January 2008. The interaction of this variable with time was added to the models for the Aboriginal and Torres Strait Islander and non-Indigenous populations, along with the other covariates. If the interaction between this new variable with time is significant, this indicates that after 2008 the relative trend was different from what it was before 2008. Note that these models do not include any of the other significant interactions with time that were detected in the initial subgroup analysis
     

8 It should be noted that the power to detect a significant interaction for this research is low, especially for the smoking initiation analysis. This is due to the smaller number of surveys before and after 2008. This implies that if a significant effect was not detected, a true effect may still exist; however, it also means that there is a high level of confidence in the significant effects found.

Modelling average annual absolute change

9 Linear regression is the most widely used of all statistical techniques. It is typically used for modelling linear, additive relationships between a scalar dependent variable and one or more explanatory variables that can be either numerical or categorical. For each categorical characteristic in the model a comparison group is selected. In this publication, this method is used to estimate average annual absolute change, assuming the percentage point gap between two estimates is a linear relationship.

10 The linear regression model can be expressed as follows:

\(\Large{y=\alpha+\beta_{1} x_{1}+\dots+\beta_{k} x_{k}+\varepsilon}\)

where \(\alpha\) is the intercept, the \(\beta_{1}, \ldots, \beta_{k}\) are \(k\) regression coefficients, and the \(x_{1}, \dots, x_{k}\) are a set of \(k\) explanatory variables. The coefficients of the model can be estimated using standard least squares or maximum likelihood techniques.

11 The average absolute (percentage point) changes were analysed at the aggregated level rather than at the individual level.

12 Estimates of smoking prevalence, initiation and quitting were calculated for age-sex-remoteness groups in each survey. To take into account the sample design of the surveys, sampling weights were applied to individual data.

13 Regression models were then built for each outcome variable with time, age (or age group), sex, remoteness and Indigenous status and an interaction term between Indigenous status and time was added. The interaction was then tested to see whether it was significant or not. If the interaction is significant then it can be concluded that there is a significantly different absolute trend for the Aboriginal and Torres Strait Islander and the non-Indigenous populations. To take into account the different levels of uncertainty of estimates due to sampling error, all regression models used the weighted least squares method with the inverse of the estimate variance as the weight. The estimated variances were calculated from the individual data using the Jackknife method.

14 The next step was modelling of outcomes separately for the Aboriginal and Torres Strait Islander and non-Indigenous populations. Age, sex and remoteness were tested as covariates in the model and significant interactions between these covariates and the time variable were checked. A significant interaction means that there are significantly different trends for the different categories of the covariate. If any interactions were not significant then these were removed sequentially until all remaining interactions were significant (p<0.05), removing the least significant first.

15 Finally, an assessment was undertaken to determine whether absolute trends were different before and after the increased attention to and funding for tobacco control programs targeted to the Aboriginal and Torres Strait Islander population since 2008. This involved creating a new variable with a value of 0 for before 1 January 2008 and a value of 1 for after 1 January 2008. The interaction of this variable with time was added to the models for the Aboriginal and Torres Strait Islander and non-Indigenous populations along with the other covariates. If the interaction between this new variable with time is significant, this indicates that after 2008 the relative trend was different from what it was before 2008. Note that these models do not include any of the other significant interactions with time that were detected in the initial subgroup analysis

16 It should be noted that the power to detect a significant effect for this research is lower than for modelling relative declines due to the model’s use of aggregated data rather than individual data. This implies that if a significant effect was not detected, a true effect may still exist; however, it also means that there is a high level of confidence in the significant effects found.

Glossary

Show all

Quality declaration

​​​​​​​Institutional environment

Relevance

Timeliness

Accuracy

Coherence

Interpretability

Accessibility

Abbreviations

Show all

Back to top of the page