Dr Itismita Mohanty, Research Fellow, NATSEM, University of Canberra. Dr Yogi Vidyattama, Senior Research Fellow, NATSEM, University of Canberra

Transcription

1 SUBJECTIVE WELLBEING: A SPATIAL MICROSIMULATION OF AUSTRALIAN UNITY QUALITY OF LIFE SURVEY AUTHORS Dr Itismita Mohanty, Research Fellow, NATSEM, University of Canberra Associate Prof Robert Tanton, NATSEM, University of Canberra Dr Yogi Vidyattama, Senior Research Fellow, NATSEM, University of Canberra Dr Marcia Keegan, Research Fellow, NATSEM, University of Canberra Prof Robert Cummins, Personal Chair, School of Psychology, Deakin University ACKNOWLEDGEMENTS Funding for this project was provided by AURIN, which is an initiative of the Australian Government as part of the Super Science Initiative financed from the Education Investment Fund. ABSTRACT This paper uses spatial microsimulation methods to estimate small area (SA2) subjective wellbeing in Australia. The procedure uses the Autralian Quality of Life survey and the 2011 Census data to derive small area estimates of subjective wellbeing. Validation of the results shows that they compare well to another source of subjective wellbeing for areas in the Murray Darling Basin, and aggregate estimates compare well to HILDA estimates of subjective wellbeing at a State level. These estimates are now available from the Australian Urban Research Infrastructure Network (AURIN) at the University of Melbourne. 1

2 1 INTRODUCTION The growth of knowledge measuring human progress along with life satisfaction and happiness has emphasized the focus on both subjective and objective aspects of wellbeing. Subjective wellbeing is an extension of psychological wellbeing or happiness and life satisfaction for individuals. Away from historic, philosophical and religious prospects of life, it is utilitarianism that identified - along with objective aspects, the presence of pleasure and absence of pain are the defining characteristics of a good life (Diener et al., 2012). Research on subjective well-being (SWB) assumes that an essential ingredient of a good life is that the person likes their life. Subjective well-being is defined as a person's cognitive and affective evaluations of his or her life (Diener et al., 20023) as a whole. These affective evaluations include emotional reactions to events that can be either positive or negative. The cognitive judgments of satisfaction and fulfilment are about what one thinks about his or her life satisfaction in global terms (life as a whole) and in domain terms (in specific areas of life such as work, health, relationships, etc.). Thus, subjective well-being is a broad concept that includes experiencing high levels of pleasant emotions and moods, low levels of negative emotions and moods, and high life satisfaction. SWB concerns the study of what lay people might call happiness or satisfaction (Diener et al., 2012; Argyle, 2001; Diener, 1984; Diener et al. 1999; Kahneman et al., 1999). Research on subjective well-being comprises the psychological analysis of how people evaluate their lives both at present and for longer periods such as over the past year. These evaluations include people s emotional reactions to events, their moods, and judgments they form about their life satisfaction, fulfilment, and satisfaction with domains such as marriage and work. Nn individual s moods, emotions, and self-evaluative judgments are not constant over the life time. SWB studies these fluctuations and examines the longer-term mean level differences that exist between individuals and societies. Different components of SWB reflects people s evaluations of what is happening in their lives, yet the facets of SWB such as positive affect, negative affect, and life satisfaction can be studied independently (Diener et al., 2012; Andrews & Withey 1976, Lucas et al. 1996). However, it is often seen that researchers continue to measure a single aspect of well-being or ill-being such as life satisfaction or depression. Research on subjective wellbeing at country and society level is an extension of psychological research on life satisfaction for individuals. This research uses measures of overall satisfaction with life, psychological distress or happiness. For life satisfaction, a question is asked of an individual and they answer based on a scale. The question asked in the Household Income and Labour Dynamics Survey of Australia (HILDA), an annual longitudinal survey, is: All things considered, how satisfied are you with your life? with a rating of 0 to 10. The other survey that asks questions on subjective wellbeing in Australia is the Australian Unity Quality of Life Survey (AUQOL), run by Prof Robert Cummins of 2

3 Deakin University for Australian Unity. The question asked in the Australian Unity Quality of Life survey is: Thinking about your own life and personal circumstances, how satisfied are you with your life as a whole? (0 completely unsatisfied neither unsatisfied nor satisfied completely satisfied) In both HILDA and AUQOL, additional questions are asked about the respondent s satisfaction in a number of domains of life, such as employment opportunities, financial situation, safety, health, neighbourhood, free time and personal relationships. In the AUQOL survey, an additional domain of spirituality is added (so how satisfied are you with your spirituality), making ten domains. All these domains can be ranked from 0 to 10. In both surveys, the results for all domains can then be summed to get a total score out of 90 (HILDA) or 100 (AUQOL). So there are two measures of wellbeing in each survey; a measure of overall wellbeing which is ranked from 0 to 10, and a summary measure of wellbeing for a number of different dimensions in life that can range from 0 to 90 (HILDA) or 0 to 100 (AUQOL). Some international surveys on subjective wellbeing use a scale of 1 to 100 rather than 0 to 10. In this paper, as we use all Australian data, our wellbeing indicators are shown on a scale of 0 to 10. However, we further transform the scale to 0 to 100 to facilitate international comparison. It has been observed that subjective wellbeing can be influenced by a number of factors such as situational factors, the type of scales that are used, the number and the sequence of life events, the time lag since the event, the mood of the respondent at the time of the survey and objective factors such as income and employment. Subjective wellbeing is further believed to have a genetic component. So, some people are inherently predisposed to be happier than their peers. However, personality traits like optimism, internal locus of control and self-esteem along with a rich and fulfilling social life and a network of close social support with family and friends are also strongly correlated with SWB. Some of the research on individual wellbeing is about the stability of subjective measures of wellbeing. Research reveals that SWB has a mean of 7.5 out of 10 (75 out of 100) for Western countries and 7.0 out of 10 (70 out of 100) for non-western countries, with consistent results across different countries (Cummins,1995 and 2003). Further, there has been research in this space about the stability, or the homeostasis of subjective wellbeing over time (see Cummins, 2009; Tanton et al., 2012). This research, using Australian data, suggests that subjective wellbeing is very stable it tends to hold within a narrow range of values. Further, subjective wellbeing is homeostatic. This means that there is a threshold value which, as this value is approached, the person tries to retain control. If this threshold is breached, the person will, over time, regain control and subjective wellbeing will return to it s normal value for that person. So homeostasis is operating as a protective factor for wellbeing, tending the person back to their normal level of wellbeing. Geographical analysis of subjective wellbeing is one important area that has not been covered by other literature. Even though subjective wellbeing is fairly stable or homeostatic, spatial analysis can be used to analyse the variation within a narrow range of values. Realising that SWB is homeostatic it has an inherent tendency to retain control around the threshold value or mean value of 7 or 7.5 out of 10 (70 or 75 out of 100), it would in fact make sense to analyse the geographical pattern of its distribution around this mean value. Such analysis could highlight areas where local services are struggling to provide effective support. It can further 3

4 analyse the areas that are repeatedly suffering from natural disasters ordeclining industry and can help in studying the resilience and adaptive capacity of the people in these areas. It is just as important to obtain spatial small area estimates for a policy perspective. For quite some time now, there has been a demand for spatially detailed information on individual and household level micro data for policy analysis. National Censuses can provide this spatial information, but are typically conducted relatively infrequently and their extensive geographic detail comes at the price of containing only a limited range of information about households. Administrative data are sometimes geo-coded to provide a comparable level of geographic detail, but often only contain information essential to the provision of the services and usually lack socio-economic descriptors like subjective wellbeing. National sample surveys, typically available in unit record formats, contain much richer information about a particular topic than the Census (income, health, expenditure) but usually suppress the geographic detail of respondents to protect privacy and, even if that were not the case, usually provide too small a sample size to produce accurate small area estimates (Cassells et al. 2010). In the recent decades, these difficulties have led to the development of some of very advanced scientific techniques in this space such as synthetic estimation of small area data and spatial microsimulation. Significant research has been undertaken in developing spatial microsimulation models both in Australia and internationally (see Tanton and Edwards, 2013 for a description of these methods). One of the most significant and concerted initiatives has been undertaken by researchers from the National Centre for Social and Economic Modelling (NATSEM) (Tanton et al., 2011). However, while substantial research has been done to obtain information on characteristics of the population such as poverty, unemployment and housing cost/stress, there has been no published attempt to use spatial microsimulation to obtain estimates of subjective wellbeing. Work has been done in the UK, but is so far unpublished (see Ballas, Unpublished). This paper outlines the method for small area estimation of subjective wellbeing in Australia using spatial microsimulation. This modelling is based on the NATSEM s latest spatial microsimulation models SpatialMSM/08 and SpatialMSM/09 - and the process undertaken to prepare sample survey data for matching to 2011 Census benchmarks. This process combines substantial geographic information from the latest 2011 Census with the rich variable detail of Australia s only national level sample survey that is solely dedicated to collecting information on subjective wellbeing - the Australian Unity Quality of Life Survey (AUQOL). The model uses both these sets of data to gain reliable small area estimates of subjective wellbeing. The model is further used to reveal the gender differences, age variations, employment-unemployment variations and differences in wellbeing by family type at a small area level in Australia. 2 SPATIAL MICROSIMULATION The spatial microsimulation technique re-estimates the spatial distribution of survey data based on population characteristics from the census data. This census data covers the entire population in small areas (Vidyattama et al. 2011). There has been significant recent development in this field to construct reliable small area data using microsimulation models, which are constructed on top of the synthetic spatial microdata bases. This rapidly growing field now includes the simulation of the small area impact of changes in income taxes, family payments and social security (Harding et al. 2009b; Chin et al. 2005); development of small 4

5 area measures of poverty and housing stress (Tanton et al. 2009a; Tanton et al. 2009b; McNamara et al. 2007); small area modelling of Activities of Daily Living Status and need for different types of care (Lymer et al. 2009; Lymer et al. 2008); development of the SimObesity model to examine small area obesity among children (Procter, 2007); small area health-related conditions (Ballas et al. 2006a); the socio-economic impacts of major job gain or loss at the local level (Ballas et al. 2006b) and a range of other applications (Ballas et al. 2005a, 2005b; Clarke 1996). Spatial Microsimulation basically refers to the creation of synthetic spatial microdata using a national level sample survey and census population characteristics. Initial research in this field was undertaken by geographers to create small area specific micro data (Voas and Williamson 2000; Williamson et al. 1998) using alternative techniques like data fusion and synthetic reconstruction. The more successful methods reweighted the original sample survey data to match small area population targets from a Census. This has since been the most widely used method to create synthetic spatial micro data. In practice, spatial microsimulation techniques involve matching the characteristics of households contained in the national survey to multiple cross-tabulations of household characteristics derived from the Census for each small area. These characteristics include age, gender, occupation, labour force status, socio-economic group and so on. As Ballas et al. (2006a, p.65) explain, this is basically the merging of Census and survey data to simulate a population of individuals within households (for different geographic units), whose characteristics are as close to the real population as it is possible to estimate. As the literature reveals, various ways of doing this reweighting technique have been developed such as combinatorial optimisation (Voas and Williamson, 2000), iterative proportional fitting (Birkin and Clarke, 1989) and generalised regression (Tanton et al, 2011). Spatial microsimulation modelling is still in its initial stages of development and considered as an emerging branch of knowledge to supply reliable and detailed information for small areas. As Cassells et al points out, one crucial issue has been the validation of the reliability of the small area estimates produced from such models (also see Vidyattama et al., 2013). Voas and Williamson (2000, p. 360) identified that the most challenging aspect of spatial microsimulation has been to model topics that did not feature among the constraints [with constraints being the small area Census benchmarks that the sample survey data have been reweighted to]. They find that the simulated results from variables that have not been included within the Census benchmarks will generally only produce reasonable results if the actual distribution of those variables is close to the national average or if the variables involved are highly correlated with the variables that were included within the Census benchmarks (Voas and Williamson, 2000; Cassells et al., 2010). So far SpatialMSM modelling for small area poverty estimates in Australia (Miranti et al, 2011) has been quite successful, though the same is not true for modelling relatively rare disability status (Lymer et al. 2006). As Chin et al. (2006a) and (2006b) suggested this is possibly because poverty is highly correlated with benchmarks available within the Australian Census and used in the reweighting process (such as income and family type). If the local areas exhibit characteristics that are different to the average then the SpatialMSM model may prove unsuitable for estimating these characteristics. (Ballas et al. 2005a, p. 14). Rahman (2009) finds spatial microsimulation analysis as robust and has particular advantages over other approaches in this space, including the ability to further aggregate or disaggregate data into different spatial units; the capability to easily update and engage in further analysis; and 5

6 through linking the model with a static microsimulation model, the ability to measure policy change at a small area level. 2.1 METHODOLOGY This section describes the procedure for estimating small area wellbeing in Australia using a national level sample survey (AUQOL) and the 2011 Census using the SpatialMSM model. The SpatialMSM model uses a generalised regression reweighting program from the Australian Bureau of Statistics (ABS) called GREGWT. The GREGWT algorithm uses a generalised regression technique to estimate weights for a household or individual from the survey, and iterates until the weighted aggregates of the survey data produces characteristics that closely resemble the census population characteristics for each small area (Bell, 2000; Tanton et al, 2011; Vidyattama et al, 2013). This procedure has been classified as a deterministic method because it uses formulae, similar to the iterative proportional fitting used by Anderson (2007) and Ballas et al (2005). This is different from the probabilistic method that pseudo-randomly selects households to fill an area (Voas and Williamson, 2000; Williamson et al, 1998), even though it is established that the results from different reweighting methods are more or less similar (Tanton et al, 2007). The geographical unit of analysis is the Statistical Area Level 2 (SA2). SA2s are part of the ABS Australian Statistical Geography Standard (ASGS). They replace Statistical Local Areas (SLAs) from the Australian Standard Geographic Classification (ASGC). SA2s are in general smaller than SLAs. There are 2214 SA2s which were part of the 2011 ASGS that are used for this analysis. SA2s generally have a population range of 3,000 to 25,000 persons, and have an average population of about 10,000 persons. SA2s in remote and regional areas generally have smaller populations than those in urban areas (ABS, 2011). The SA2s are a general-purpose medium-sized area built from whole Statistical Area Level 1 (SA1s). While the Statistical Areas Level 1 (SA1s) have been designed as the smallest area of output for the Census of Population and Housing, replacing the Census Collection Districts (CCDs), they are not suitable for such analysis owing to the Australian Bureau of Statistics (ABS) confidentialising small cell counts in tables for smaller areas (such as SA1s). On average SA1s have a population of approximately 400 people, and most are designed to be within the population range people. There are 54,805 SA1s covering the whole of Australia without gaps or overlaps (ABS, 2011). There are some areas where an estimate cannot be produced by SpatialMSM, since the estimation process does not achieve an acceptable error for the estimate. In SpatialMSM, error is measured by the total absolute error (TAE) from all the benchmarks. The TAE has been used in a number of spatial microsimulation models as a criterion for reweighting accuracy (Edwards and Tanton, 2013; Anderson, 2007; Williamson et al, 1998) and has been assessed and supported by the results of other studies, such as Smith et al (2009) and Voas and Williamson (2000). The TAE in SpatialMSM is calculated by summing all the differences between the estimated number from the model and the benchmark number from every benchmark class of each benchmark table. An area is dropped from any further analysis when 6

7 the TAE is higher than the total error threshold that is set for SpatialMSM, which is generally when the TAE from all the benchmarks is greater than the population in that area. This means that if ten benchmark tables are used in the estimation, the average error over all benchmark tables should be less than or equal to 10% of the population. 3 THE SURVEY: AUQOL The Australian Unity Quality of Life Survey has been conducted in Australia since It is a telephone survey of approximately 2000 people, and has been conducted semi-annually and quarterly throughout its existence (currently semi-annually). Certain groups of people are under sampled or not sampled at all due to the nature of the telephone survey. First, it excludes people who do not speak English or who speak it poorly. Second, it excludes people who do not have a landline phone. In the 2006 Census, 2.8 per cent of respondents did not speak English well, or did not speak it at all. Between 12 and 15 per cent of households in Australia do not have a landline phone (Pennay and Bishop 2009), however this is only likely to affect the representativeness of results in certain subgroups (young people living away from home in particular). However, more recent research by Pennay and Bishop (2010) suggests that the number of households with only a mobile phone connection will continue to increase, with one third of year olds opting not to use a landline when moving out of home. Mobile phone only households are younger more likely to rent, live in a group household, have a bachelor degree or higher and are more likely to be employed. 3.1 VARIABLES TO BE USED FOR REWEIGHTING FROM CENSUS AND AUQOL We have examined the variables in previous AUQOL surveys and compared them with questions from the 2011 census. The aim of this exercise was to identify similar questions in AUQOL and the Census for reweighting using SpatialMSM. Along with standard demographic and socioeconomic characteristics of the population such as age, sex, income and household type, other variables in the Census have been analysed because they could potentially have an impact on quality of life and have equivalents in AUQOL. So, the aim is to look for variables or derived variables that are comparable between the survey and the census, and have a demonstrated impact on life satisfaction levels. The next section will compare a number of variables in the Census and AUQOL that are similar, discuss any potential advantages and disadvantages of using these variables, and report on the decisions to include or exclude those. 3.2 DEMOGRAPHIC VARIABLES Age Every person in Australia at the time of the Census is required to report their age. AUQOL also asks respondents for their age. It is helpful that neither database asked for age brackets, so we can create our own age groups for the benchmarks. 7

8 The relationship between age and life satisfaction is U-shaped: people in their late teens/early twenties are very happy, then wellbeing declines as responsibilities of marriage, children, jobs and mortgages start to put pressure on life satisfaction. Wellbeing starts to increase in the 40s to 50s as children grow older and move out, jobs get less stressful and the mortgages get paid off. Happiness continues to increase after retirement age, and as long as physical health and functioning is maintained, can keep increasing into the 80s (Cassells, et al. 2010). This research uses ten year age groups, which is considered narrow enough to take into account the U-shaped life satisfaction profile. The age brackets start from 18-24, and so forth with the final one being 95 and above. These intervals have been chosen as two major milestones in terms of retirement options commence at age 55 and 65 (access to superannuation and access to the age pension, respectively) which allow people to make employment choices that affect their life satisfaction Sex/gender Both the Census and AUQOL ask about gender. Earlier AUQOL studies showed that women had higher satisfaction levels than men, but men have closed the gap in recent years (Cummins et al, 2012) Citizenship, country of birth and ethnic background The Census asks three questions on this issue: In which country was the person born? The options given are Australia, England, New Zealand, Italy, Vietnam, Scotland, Greece, other (specify) What is the person s ancestry? The options given are English, Irish, Italian, German, Chinese, Scottish, Australian, other (specify). Respondents can choose up to two options. Is this person an Australian Citizen, yes or no. AUQOL asks the following questions: In which country were you born? Every possible option is listed. What is your ethnic origin? Every possible option, including Antarctica and the Holy See, are listed; plus inadequately described or not stated What is your citizenship? Again, every possible option is listed. The main way in which a person s citizenship, country of birth or ethnic background can affect life satisfaction is if the respondent comes from a Confucian country (Lau et al, 2005).These countries include Japan, China, Taiwan, Hong Kong and Singapore. Respondents raised in these countries tend to have a mean value of subjective wellbeing ratings of 6.5, instead of the 7.5 by other respondents, see Lau et al (2005). The theory is that the reason for this difference is not because people raised in these countries have lower levels of wellbeing, but the Confucian culture results in people responding more modestly regarding their satisfaction with elements of their lives. It will not be feasible to include all three variables in SpatialMSM as it is highly likely that there will be some unusual combinations of country of birth and ethnic origin, for example a 8

9 respondent might be born in Finland, claim Japanese ethnic origin and have Australian citizenship. These will cause SpatialMSM to have difficulties with reweighting. Using the second question, those relating to ancestry and ethnic origin, are less suitable for inclusion in the model. First of all ethnic origin and ancestry may not be directly comparable respondents may not interpret these terms in the same way. Secondly, the census allows a person to give two ethnic origins, which AUQOL does not Of course, this does not mean that using country of birth as a control variable is perfect a person who was born in a Confucian country and moved to Australia as an infant or small child is more likely to adopt Australian values and perceptions than those of their Confucian parents. Ideally, this research would have preferred to use country of birth as a control variable, however, this question was not consistently asked in all AUQOL surveys Family structure A person s family structure has an impact on their life satisfaction. People in couple relationships tend to be happier than lone persons. Married couples are happier than couples in de facto relationships, unless the de facto couple have a large joint commitment such as buying a house together, in which case their wellbeing is indistinguishable from that of married couples. Couples with children at home tend to be less happy than couples without children (Cummins et al, 2012). The Census has a number of derived variables for each person s relationship in the household, family type and whether a household is a family or non-family household. If a person lives in a family household, they are classified according to their registered marital status, social marital status, relationship to others in the household (spouse, de facto, lone parent, child under 15, dependent student, non-dependent child, other related individual, unrelated individual living in family household, group household member or lone person.) Households are classified as family or non-family households, and families are defined as couple households with no children, couple households with a number of possible combinations of child types, lone parents with a number of possible combinations of child types, or other family types. The living status or home structure question in AUQOL gives the following options: living alone, single parent, living with partner, living with partner plus others, living with non-partner, living with parents, with one or more adults, with one or more children, with one or more unrelated adults, with children and partner. There is some difference between household classifications in AUQOL and the Census, which will require recoding of both before the two can be prepared. It is suggested that individuals within each database are classified according to their relationship in the household: Member of a couple only (marital or de facto) Member of a couple with children (marital or de facto) Single person household Single parent household Adult offspring living with parents 9

10 Group household Urban/regional AUQOL surveys have found that the further people are from major city centres (including Sydney) the happier people are. Some AUQOL surveys have asked respondents if they live in major cities, small cities or country towns. An easier measure to use as an estimate of a person s remoteness is their postcode, and ARIA status (Accessibility/Remoteness Indicator of Australia), which can be readily compared to Census data. 3.3 ECONOMIC VARIABLES Employment Employment has been shown to have an impact on life satisfaction levels. Unemployed people tend to be less satisfied with life than employed people. Hours worked also has an impact on satisfaction levels, however it is not the absolute number of hours worked that impacts satisfaction, but whether a person wants to work more or less hours than they actually work. That is, hours mismatch affects satisfaction more than hours worked. The official ABS classifications of employment as used in the Census are, broadly speaking, fulltime employed (35 or more hours per week); part-time employed (less than 35 hours per week); unemployed and looking for full-time work; unemployed and looking for part-time work; and not in the labour force. Volunteer work is not classified as employment. The AUQOL questionnaire operates differently. One question asks about the respondent s work status, asking the respondent to nominate which options applied to them, namely fulltime paid employment, full-time retired, semi-retired, full-time volunteer, full-time home or family care, full-time study, unemployed and none of these. Similarly, respondents were asked if they are engaged in part-time paid employment, part-time volunteering, part-time study, or casual employment. There are some difficulties in matching AUQOL survey responses with Census labour force classifications. AUQOL does not clarify any formal definition full-time paid employment, parttime paid employment, unemployment or casual work to respondents, allowing the respondents to answer according to their own understanding of what these terms mean. As a result, respondents may have given answers that are quite different to their standard ABS employment classification, meaning that labour force measures in AUQOL may be different from those used in the census. Happiness literature consistently shows a relationship between labour force status, so it is important to examine AUQOL responses to see if they can be reliably weighted against Census data. At a minimum, we can determine if a person is employed or not there is unlikely to be much confusion as to the definition of paid employment (to the extent that there might be confusion, for example family workers, they represent a small enough share of the population that they are unlikely to bias the results). One might consider separating the employed into full-time and part-time employed, but this is probably not necessary for the purposes of this 10

11 project. As discussed earlier, hours mismatch has a detrimental effect on wellbeing, rather than full-time or part-time work. Furthermore, some of the detrimental effect of hours mismatch will be picked up in the income measure. Finally, without some measure of hours worked, it is not possible to reliably classify employed persons as part-time or full-time. The ABS official definition of unemployed can be broadly summarised as a person not in paid employment, who is looking for work and able to start work within four weeks. The unemployment rate is the percentage of the labour force (those people in work and looking for work) who do not have a job. Using the outputs of AUQOL Survey #26 as an example, if we calculate the percentage of people described as unemployed as a percentage of the labour force, the unemployment rate would be 10.5 per cent about double the unemployment rate at the time. Furthermore, of the people who described themselves as unemployed, one person was in full-time employment, and nearly half described themselves as full-time retired. The survey does ask if respondents are currently looking for work. Some of these people are currently employed. However, an estimate of the unemployment rate of respondents can be calculated by dividing the number of people with no job, who are looking for work, by the number of people who responded with No job-looking, Paid employment- looking and Paid employment - not looking. This produces an unemployment rate of 6.5 per cent, which is reasonably within the ballpark of the unemployment rate of just above five per cent. This measure shall be used as an estimate of unemployment. It is also feasible to classify AUQOL respondents as employed or unemployed, and the remainder as not in the labour force. This broad classification of labour force status could possibly be sufficient to capture most of the variation in wellbeing due to employment Household income Literature on the relationship between life satisfaction and income recognises that there is a weak, statistically significant positive relationship between the two, which is stronger at lower levels of income. The Census asks each individual within a household to indicate their total pre-tax income in a series of income brackets. From this, an estimate of total household income (also in brackets) is derived. AUQOL asks each respondent to give an indication of their total household income in brackets (as only one person per household is surveyed). Naturally, these brackets do not line up. First, they will need to be adjusted for Consumer Price Index (CPI) to bring the brackets into 2011 dollar values. Then the brackets from the survey will need to be regrouped so they line up with Census income brackets. AUQOL income brackets go much higher than those from the Census. Their highest income brackets are $500K+ and $250-$500K, because this additional income has a statistically significant effect on wellbeing. Because the Census income brackets only go to a household income of $260K or more, this variation in the very high income brackets could not be controlled for. This is a minor source of variance compared with that derived from the lower income brackets. 11

12 3.3.3 Occupation As with hours worked, occupation can have an impact on satisfaction, but it is occupational mismatch rather than the occupation itself that affects satisfaction. Both the Census and AUQOL have questions on occupation. However, AUQOL s occupation options are broad and do not line up with standard ABS classifications for occupations, so it is not feasible to use occupation as a variable Housing tenure Generally, people who rent their homes are less satisfied with life than people who own their homes, and homeowners are happier if they own their home without a mortgage. The relationship between the total amount of rent or mortgage paid and satisfaction levels is complex; according to the results of AUQOL Survey #16, what matters is a person s comfort with their ability to pay rent or mortgage rather than the absolute dollar amount. The Census first asks if the dwelling is owned outright, owned with a mortgage, being purchased under a rent-buy scheme, rented, occupied rent free, occupied under a life tenure scheme, or other. AUQOL Surveys #15 and #16 ask whether a person owns or has a mortgage on the place where they live, and if they pay rent where they live. The Census asks how much the household pays in rent or mortgage as a continuous variable, AUQOL Survey #16 asks how much rent or mortgage is paid each month in $500 brackets, topping out at $2000. Respondents will be classified into three groups in the Census and AUQOL: pays rent, pays mortgage, pays neither and other tenure type. 3.4 SOCIAL VARIABLES Education People with higher levels of education, on average, tend to be happier than people with lower levels of education. Most of this is due to the higher income and lower risk of unemployment that comes with higher qualifications; when income and unemployment are controlled for, the impact of education on wellbeing is mixed. The Census asks each individual aged 15 and over for the highest level of schooling completed (Year 12 to Year 8 or below). It then asks if each person has completed a post-school qualification and if so, that the nature of their qualification. AUQOL asks respondents what is the highest level of education they achieved, and permissible responses vary from survey to survey. Broadly speaking, they fall into the following categories: primary, secondary, technical/trade or university. Given the breadth of the Census question on education, the relatively narrow focus of the AUQOL survey and the mixed evidence on the impact of education on wellbeing, the education variables are not included in the model. 12

13 3.4.2 Internet usage Both Census and AUQOL contain questions about internet access in the household. However, the wording of the questions is different: the Census asks if the internet can be accessed at the dwelling, AUQOL asks if the respondent uses the internet. Responses will not necessarily match; some people will live in a home with internet access but not personally use it, some people will not have internet at home but will use it at work. Also, there is no evidence that internet access is associated with higher levels of satisfaction, so this will not be used as a variable for the Wellbeing SpatialMSM Religion The Census includes a voluntary question on religion, and simply asks people to nominate their religion. A handful of common options are given, and if people choose other they are asked to specify their religion. AUQOL also asks people about their religion, but in some surveys the wording is slightly different, asking about religious belief or spirituality. This research would not attempt to reweight on the basis of religion for the following three reasons: 1) It appears from the distribution of responses in the Census that many people respond according to their nominal religion rather than what they practice; ie, in the Census people identify themselves as Anglican or Catholic even though they only attend church for weddings and funerals. Seventy per cent of Census respondents reported themselves as having a religion of some sort; that is, they provided an answer to the question that was not atheist, agnostic or no religion. However, AUQOL seeks questions regarding a person s actual faith or practice, and survey results show that substantially less than seventy per cent of respondents reported religious faith. This discrepancy between nominal religion and religious practice means these questions may not be comparable. 2) AUQOL offers respondents the choice of Buddhist, Christian, Hindu, Muslim, Jewish, no religion or other. In the 2011 Census, respondents are prompted to choose their denomination of Christianity, with Catholic, Anglican, Uniting Church, Presbyterian, Greek Orthodox, Baptist and Lutheran provides as options; along with Buddhism, Islam and other, please specify. Thirty one per cent of Census respondents did not profess to religion; and 61 per cent identified as some sort of Christian, leaving only eight per cent of Australians identifying with other religions. 3) Although some research in other countries suggests a relationship between religion and life satisfaction, no such relationship appears to exist in Australia. AUQOL has not found any evidence so far that religious people nominal or practising have significantly different life satisfaction than other Australians. For these reasons religion will not be used as a benchmark to estimate subjective wellbeing. 13

14 3.4.4 Caring and needing care AUQOL Survey 17.1 was specifically focused on carers. It found that carers had substantially lower wellbeing than the general population, so low, in fact, that the average wellbeing score for carers was at a level that suggests moderate depression. The Census has three questions for each respondent on whether they require care: 1) Does the person ever need someone to help with, or be with them for, self-care activities? 2) Does the person ever need someone to help with, or be with them for, body movement activities? 3) Does the person ever need someone to help with, or be with them for, communication activities? For each question, the response options are Yes, always; Yes, sometimes; or no. Similar questions are asked in AUQOL. However, these are not about an individual s caring responsibilities, rather they are about a person s need for care. Despite the importance of caring responsibilities in a person s wellbeing, it is not possible to match AUQOL responses appropriately to Census responses, thus caring responsibilities will not be included Ethnic diversity Previous AUQOL surveys have found that people who live in a neighbourhood with high proportion of non-australian born people are more likely to report low levels of community connectedness. Specifically, there is little relationship between percentage of the neighbourhood that is overseas born and community connectedness when at least 60% of the community are Australian-born. However, above this threshold, community connectedness falls sharply, which has an impact on life satisfaction. Although this is significant, it would be quite difficult to determine in the Census the proportion of the population who are Australian-born. Furthermore, this relationship is uncertain as people who do not speak English well are excluded from the surveys it may be the case that in neighbourhoods with a high percentage of foreign-born residents, non-english speakers are happier than they would have otherwise been in mostly Australian-born neighbourhoods. 4 WHICH SURVEY/SURVEYS TO USE? Although some variables are common to all the surveys, there are a number of questions that are only asked in a handful of surveys. Ideally, we should choose survey(s) that have as many variables as possible that are first, correlated with subjective wellbeing and second, can be readily matched to Census 2011 variables. Initially NATSEM considered simply using the most recent AUQOL survey available, Survey 26, with the option of combining it with earlier surveys such 25, 24 and so forth to increase the sample size if necessary. These surveys have special feature questions on subject matters such 14

15 as chronic health conditions, the impact of natural disasters, trust, climate change and internet usage. While these questions are interesting and are relevant to determining overall wellbeing, these are not matters that are readily comparable to questions on the 2011 Census. Instead, NATSEM decided to use Survey 16 for reweighting (Cummins, 2006). NATSEM did initially consider using Surveys 15 and 16; the advantage of these surveys is that they have questions regarding rent and mortgage obligations, from which one can create a variable on housing tenure. Other research Cummins (2006) has found that homeowners tend to be happier than renters, and homeowners without a mortgage are happier than with a mortgage. However, Survey 16 is the only survey that asks the amount paid on rent and mortgage each month. Since high rents and mortgages can contribute to stress and lower wellbeing, only Survey 16 was used. The variables on home ownership, rent and mortgage obligations can be readily matched to Census 2011 data Census DATA SOURCE AND SCOPE The 2011 ABS Census of Population and Housing data were used for benchmarks - or Constraints, to which the synthetic small estimates produced by the reweighting process must match. In Australia, the Census is conducted by the Australian Bureau of Statistics once in every five years, and information about the personal, family and dwelling characteristics of all Australians is collected. As discussed already, the Census has the advantage (in contrast to a sample survey like AUQOL) of providing data at a high level of spatial disaggregation, however it provides no information on subjective wellbeing and quality of life that is available in AUQOL. The 2011 Census was conducted on the 9 August 2011, and gives details of all people (including visitors) for each dwelling. Census provides information on both counts based on place of enumeration and counts based on place of usual residence (PURP). The Census count for Place of Enumeration is a count of every person in Australia on Census Night, based on where they were located on that night. This may or may not be the place where they usually live. This includes people classified as visitors, or in the Statistical Area level-1 (SA1s) of off-shore, shipping and migratory areas. Off-shore includes persons enumerated on an oil rig/drilling platform etc. Shipping includes persons enumerated on board vessels departing for an overseas port; and migratory covers all people who are in transit on long distance trains, buses and aircraft on Census Night (ABS 2006, p.171). The Census count for PURP is a count of every person in Australia on Census Night, based on the area in which they usually live. Each person is required to state their address of usual residence in a question on the Census form. Where sufficient information is provided, this enables the area in which they usually live to be identified and coded. The count of persons at their usual residence is known as the de jure population count. Census counts compiled on this basis minimise the effects of seasonal factors such as the school holidays and snow season, and provide information about the usual residents of an area as well as internal migration patterns at the state/territory and regional levels. We have used benchmarks based on the counts for the usual address status or PURP of the persons. However in some special cases where we have used benchmarks on the family characteristics and the household dwelling classifications at a person level such as dwelling 15

16 characteristics, household characteristics and household income and expenditure, we had to resort to person counts based on the place of enumeration. In this case we have confined the scope of our benchmarks to only people who were in their usual place of residence on Census night. People classified as visitors, or in the Statistical Areas-1 of off-shore, shipping and migratory areas are excluded. Because of these slight discrepancies, we have inflated the person counts based on the place of enumeration in SA2s in our benchmark tables to match the counts for the usual address status of the persons. Personal characteristics in the Census data relate to the respondent s place of usual residence. Typically in surveys, data is collected from the usual residents of a household, and visitors are excluded. Since the 2006 Census, the standard output tables were produced using usual residence data, and this is what has been generally used for the benchmark tables in the recent versions of SpatialMSM (SpatialMSM/09C). However, for this Wellbeing SpatialMSM we have used both the sources counts adjusted by the person counts based on the place of enumeration in SA2s to match the counts for the usual address status of the persons. Most benchmark tables are multi-dimensional, as they are cross-tabulations of the variables that we want to benchmark to. While for the earlier versions of SpatialMSM, most of the benchmark variables have been sourced via special tables requested from the ABS (see Table 1 for SpatialMSM/09C benchmarks), this difficulty has now been overcome by Census TableBuilder software in which the census data can be sourced directly from the Census files rather than through an ABS data request. TableBuilder allows users the freedom to construct basic to complex tables of data. You can build tables for all geographic areas as defined in the Australian Statistical Geography Standard (ASGS). 4.2 CENSUS DATA AND PERSONS IN NON-PRIVATE DWELLINGS Most surveys, including AUQOL, do not collect information about persons living in non-private dwellings (eg, hospitals, boarding schools, prisons and nursing homes), whereas the Census does include this information. Given this inconsistency, information about non-private dwellings can either be deleted from or added to each data source in order to make them directly comparable. In pre-2008 versions of SpatialMSM, persons in non-private dwellings existed (and were possibly non-excludable) in the Census benchmark table labour force by age by sex (a person-level benchmark table). However, in the recent past special data requests from the ABS and TableBuilder have made it possible to remove the NPD population from this important benchmark table, which means that we can have a direct population match to the survey data we are using. 4.3 OTHER NON-CLASSIFIABLE HOUSEHOLDS The Census data contains information about Other non-classifiable households, which are not included in the AUQOL data. Other non-classifiable households are described as those households that contain no persons aged over 15 years; that the collector deemed occupied but was unable to make contact with any occupants; or where the information supplied on 16

17 the Census form was inadequate (ABS, 2011). This discrepancy between the two data sources has to be corrected to make the data as consistent as possible. In earlier versions of SpatialMSM, there was a practice to create a pseudo non-classifiable population by duplicating all household records on the Survey of Income and Housing (SIH), thereby giving these households exactly the same characteristics as the classifiable households. Later on this was considered to be an inadequate solution, and for the latest models special request benchmark tables from the ABS had excluded non-classifiable households (also see Cassells et al. 2010). However, for this Wellbeing SpatialMSM we have excluded the Other nonclassifiable households along with Visitors only household. Visitors only household are households where all people enumerated are visitors. All households and family classifications in the Census are based on the relationships of people usually residing in the household. This applies when there is at least one person aged 15 years and over present. In these classifications, people temporarily absent are included, and visitors are excluded. The relationship of visitors to one another, or to any resident (including cases where all the people enumerated are visitors) is not further classified and the households containing only visitors are excluded from family variables and the internal migration variables ( Census Dictionary, 2011). 4.4 NOT STATED VALUES Due to the nature of the collection of the Census data (non-interviewer assisted), the data contain fully and partially not-stated values. Since AUQOL is a telephonic household survey non-response households do not exist in the sample. In order to be able to benchmark the AUQOL survey to the Census tables that contain not stated values on different variables, the not stated values were redistributed amongst other known categories. This redistribution was proportionate, based on the relative frequency of the true values of the known categories, so the not stated values were extrapolated out to other valid values (also see Cassells et al. 2010). 4.5 NOT APPLICABLE VALUES In the Census, not applicable values are values where the response to the question does not apply to the person or household, and so no response is required (ABS 2006). An example is mortgage loan repayments for renters. Not applicable categories also include unoccupied private dwellings and migratory, offshore and shipping SA1s. 4.6 PERSONS AGED 18 YEARS AND OVER AUQOL is a national sample of people aged 18 years or over and fluent in English. Interviewers asked to speak to the person in the house who had the most recent birthday and was at least 18 years old. An even geographic and gender split was maintained at all times through the survey. Because of this, the census benchmarks needed to be adjusted so as to match the age profile of the survey. Persons aged between 0 to 17 years for this Wellbeing SpatialMSM modelling were excluded. 17

18 5 BENCHMARKS A full list of benchmark variables used for the spatialmsm is shown in Table 2. Note that Rent and Mortgage were not included in this list of benchmarks as there were technical issues matching the data between the two datasets; but Tenure is included. Table 2 Benchmark variables and categories available for reweighting process Benchmark Variable Name From 2011 Census Categories available AGEP- Age of the person (Concatenated age groups) Age of the person 1=18-24 years 2=25-34 years 3=35-44 years 4=45-54 years 5=55-64years 6=65-74 years 7=75-84 years 8=85 and above LFSP- Labour Force Status of Person 1=Employed full-time and part-time 3=Unemployed 4=not in labour force 99=Not Applicable SEXP- TEND- HIND- Sex of the person 1=male 2=female Tenure type 1= Pays mortgage 2= Pays rent 3= Pays neither 3= Other tenure type Weekly household income 1= below $199 ($1-$10,399) 2= $200-$299 ($10,400-$15,599) 3= $300-$399 ($15,600-$20,799) 4= $400-$599 ($20,800-$31,199) 5= $600-$799 ($31,200-$41,599) 6= $800-$999 ($41,600-$51,999) 7= $1,000-$1,249 ($52,000-$64,999) 8= $1,250-$1,499 ($65,000-$77,999) 9= $1,500-$1,999 ($78,000-$103,999) 10= $2,000-$2,499 ($104,000-$129,999) 18

19 11= $2,500-$2,999 ($130,000-$155,999) 12= $3,000-$3,499 ($156,000-$181,999) 13= $3,500-$3,999 ($182,000-$207,999) 14= $4,000-$4,999 ($208,000-$259,999) 15= $5,000 or more ($260,000 or more) HCFMD - Household family composition 1= Couple family with children 2= Couple family with no children 5= Group household 8= Other family 3= Single parent family 4= Single person household For the Wellbeing SpatialMSM modelling we have used the following five bench marks, which are cross tabulations of the above variables and recognised as highly correlated with the subjective wellbeing. Table 3 Benchmark tables used for Wellbeing SpatialMSM/13 No Benchmark Table Level 1 All household types Household 2 Age by sex by labour force status Person 3 Tenure by weekly household income Household 4 Tenure type by household family composition Household 5 Households by weekly household income Household 7 VALIDATION The final step in spatial microsimulation modelling is to validate the estimates against some already established data source. The subjective wellbeing estimates obtained from our Wellbeing SpatialMSM are validated using a number of methods. The first procedure is to look at the number of areas and the proportion of populations that we need to exclude due to higher TAE than the threshold. The higher proportion of areas and populations that are needed to be excluded is an indication that the survey unit record data may not be a useful representative for most areas in Australia or the model had too many benchmarks. Table 4 shows that overall, there are 105 SA2s excluded by SpatialMSM. This number represents less than 5 per cent of the entire SA2s in Australia. The population in most of the excluded areas is relatively small and this exclusion excludes less than 0.2 per cent of the population. This is a relatively a good indication that the model managed to represent most areas in Australia. 19

20 Table 4 The number and proportion of excluded areas in the model SA2 that cannot meet the threshold Percentage of Percentage of Number Area SA2 population Greater Sydney Rest of NSW Greater Melbourne Rest of Victoria Greater Brisbane Rest of Queensland Greater Adelaide Rest of SA Greater Perth Rest of WA Greater Hobart Rest of Tasmania Greater Darwin Rest of NT Australian Capital Territory Other Territories Australia The next stage in the validation process was compare the aggregate distribution of population around the respective life satisfaction scale values of: Sixty and below, Eighty and more and seventy (as the average values in life satisfaction shows very little variation to validate). These proportions are then compared with original AUQOL survey sample proportions around the same life satisfaction scale values at state level such as NSW, Vitoria, Queensland, Western Australia, South Australia, Tasmania, ACT and NT. We have observed similar pattern in both the data sets, as shown in Table 5. The next procedure that this research has followed for validation is to validate our subjective wellbeing estimates against the Community Indicators Victoria database (Community Indicators Victoria, 2007). Community Indicators Victoria provides a community wellbeing indicator framework with local level data in the state of Victoria in Australia with the purpose of improving citizen engagement, community planning and policy making. Community Indicators Victoria presents data and reports on the wellbeing of Victorians using an integrated set of community wellbeing indicators. These indicators include a broad range of measures designed to identify and communicate economic, social, environmental, democratic and cultural trends and outcomes. We found some issues validating against the CIV data, and our currently following these issues up with CIV staff. Our subjective wellbeing estimates were then validated against a Cotton CRC Sustainable Regions Survey (2012) conducted by the authors in selected areas in the Murray Darling Basin. This survey used a very similar question as used in the AUQOL, the only difference being that 20

21 the rating requested was 1 10, not We have compared the Wellbeing SpatialMSM subjective wellbeing estimates with the Cotton CRC Sustainable Regions Survey data for these specific communities in the MDB and most areas exhibit a similar pattern, as shown in Table 5. The data for the Cotton CRC was available for a number of LGAs, and these were then mapped to Postcodes from the SpatialMSM results. Table 5 Validating Wellbeing SpatialMSM Estimates against the Cotton CRC Survey Cotton CRC Communities Cotton CRC Regions (LGAs) lifesat 6 and below lifesat 8 and above lifesat exactly 7 Corresponding Post Codes in Spatial MSM Post Codes lifesat 6 and below lifesat 8 and above lifesat exactly 7 Tamworth incl Walcha Tamworth incl Walcha Gunnedah Warrumbungles Gunnedah Warrumbungles Liverpool Plains Liverpool Plains Liverpool Plains Liverpool Plains Narrabri Walgetta Balonne 4486, , 4497 St George Waikerie Finally the subjective wellbeing estimates were validated against the Household Income and Labour Dynamics in Australia (HILDA) Waves 8 and 9 the only other national level survey that contains a question on subjective wellbeing, discussed earlier in this paper. The results from this validation are shown in Table 6. 21

22 Table 6 Validating Wellbeing SpatialMSM Estimates against Other Surveys SURVEYS AUS NSW VIC QLD SA WA TAS NT ACT Proportion with life satisfaction level of '6 and below' AUQOL HILDA W HILDA W Wellbeing SpatialMSM Proportion with life satisfaction level of '8 and above' AUQOL HILDA W HILDA W Wellbeing SpatialMSM Proportion with life satisfaction level of 'exactly 7' AUQOL HILDA W HILDA W Wellbeing SpatialMSM DISCUSSION OF RESULTS Once the weights are estimated for each small area (SA2) they are applied to the same AUQOL survey data set to produce the required small are level synthetic estimates of subjective wellbeing. The aim here is to obtain estimates of subjective wellbeing that are not directly available from the census alone but can be measured at a national level from the AUQOL survey after the survey is benchmarked to specific population characteristics from the census. The results for every small area in Australia available are shown in Fig.1. This map shows the population distribution in each small area with life satisfaction scale values of: Sixty and below, Eighty and more and seventy. The "number" presents the number of persons with respective life satisfaction scale values and "proportion" means the percentage of persons in the SA2 with respective life satisfaction scale values. The classification in Figure 1 is based on natural breaks where the biggest gap in the distribution is used for classification. The darker colour in Figure one represents a higher proportion of those who have life satisfaction eighty and more and the lightest areas the lowest proportion of people in that category. 22

23 Figure 1: Map of wellbeing in Australia We have further estimated life satisfaction estimates for each small area that are disaggregated by age groups, labour force status and the family composition of households. This information is available in flat file formats that are amenable for further statistical analysis. It can be seen that in Australia, the estimated level of population wellbeing in small areas remains close to its average value of 7.5. The farther away the area from the major city centres 23