2. Data, measures and analysis
2.1 Data
The data used in these analyses are from the first two waves (2001 and 2002) of the HILDA survey. HILDA has several features that make it particularly useful for investigating financial disadvantage. It is the first large-scale Australian longitudinal survey of adults specifically designed to investigate dynamics; previous studies of poverty have relied on cross-sectional data. Second, it includes other measures of financial disadvantage, subjective poverty and financial stress not found in previous studies. Third, income data was collected from all available (and eligible) household members, which improved the accuracy of income and other variables. Fourth, HILDA Wave 2 data includes wealth, assets and debts, which allows for the examination of their relationships with financial disadvantage. Finally, HILDA includes a range of data on other factors that are not usually collected in Australian surveys on income. The HILDA User Manual (Watson 2005) details the sampling, weighting, imputation and other technical aspects of the survey.
2.2 Measures
Measures of Financial Disadvantage
Four measures of financial disadvantage were used: before-housing income poverty, after-housing income poverty, subjective poverty and financial stress. The development of these measures is described in more detail in Appendix 2.
Households were defined as in before-housing income poverty if their equivalised disposable household income was below 50 per cent of the median. Disposable income is calculated as the income received after adjusting for taxes and government transfers such as pensions, unemployment and other benefits. The equivalence scale used is the modified OECD scale, which assigns a weight of 1.0 to the first adult, a weight of 0.5 to the second and each other adult and a weight of 0.3 to each child under 14 years of age (Förster 2001; Whelan, Layte & Maître 2002).
This measure differs slightly from other measures of before-housing income poverty constructed from the HILDA data. There are two main differences. In the measure used in this paper, the poverty line is drawn at the household with the median income, not the household that includes the individual with the median income. Second, households with negative incomes were not included in the calculation of income poverty. Such households are usually running businesses and are likely to have enough assets to carry on. Nonetheless, the estimates of before-housing income poverty in this paper are only slightly lower than estimates by Headey, Marks and Wooden (2005).
The second measure—after-housing income poverty—is similar to the first; the only difference is that housing costs are deducted from disposable household income. For some households, the cost of housing is large and thus disposable income is much lower than for a household with a similar income and little or no housing costs.
The third measure, subjective poverty, was defined by respondents indicating that they were 'poor' or 'very poor' in response to a question on their level of prosperity.
Financial stress, the final measure of financial disadvantage, was defined in terms of seven behaviours due to a shortage of money, sometimes described as cash flow problems. They are: could not pay utility bills on time, could not pay mortgage or rent on time, pawned or sold something, went without meals, was unable to heat home, asked for financial help from friends or family, and asked for help from welfare or community organisations. Individuals were defined as being in financial stress if they had experienced two or more of these events since the beginning of the year. It was not possible to use Bray's (2003b, p. v) three dimensions of financial stress—'missing out', 'cash flow problems' and 'hardship'—since the HILDA questionnaire did not include the items necessary for these classifications.
Measures of the correlates of financial disadvantage
The measures of many of the correlates of financial stress should be apparent from the tables following, for example, sex, education and marital status. The variables (household type, marital status, highest educational qualifications, labour market experiences, personal gross income and disposable income) are derived variables available in public releases of the HILDA data. The construction of the household income variables is described in Appendix 2. Occupational status was measured using the ANU4 occupational status scale (Jones & McMillan 2001).4 For parental occupational status, the father's occupation at age 14 was used. If this information was missing, then the mother's occupation was used.
The measures of wealth, assets and debts were derived from the questions in the wealth module in Wave 2. Questions covering housing, unincorporated businesses, equity-type investments (for example, shares and managed funds), cash-type investments (for example, bonds and debentures), life insurance policies, vehicles and valuables (for example, jewellery and art works) were asked at the household level and answered by one adult on behalf of the entire household. Questions about superannuation, bank accounts, credit cards, higher education debt and other personal debt, however, were asked directly of individuals. For most questions, respondents were asked to provide exact dollar amounts.
For the income and wealth variables, missing data was handled by imputation. For other variables, the few cases with missing data were excluded from the corresponding analyses. Wealth is simply assets minus debts; details on the construction of the wealth variables are available (Headey 2003; Headey, Marks & Wooden 2004; Marks, Headey & Wooden 2005).
2.3 Analysis
Statistical significance tests require that the units of analysis be independent; however, individuals living in the same household are not independent in the statistical sense. Therefore, most of the analyses in the following sections (Tables 2 to 21, but not including Tables 15 to 17) are based on reference persons randomly selected from each household. The reference person was required to be over 18 years of age and not living with a parent. The results are almost identical for different random draws. This procedure is similar to the random selection of a household member at a survey interview who provides data on behalf of the household, as in the 2000–01 Survey of Income and Housing Costs (SIHC) and the HES and GSS surveys. Separate random selections were made for each wave. Therefore, the relevant unit for these analyses are households rather than individuals. This procedure is preferable to other ways of identifying a reference person. For example, choosing the household member with highest income would bias the results toward higher status individuals. Furthermore, it would not be possible to examine the relationships of poverty with sex and age if the reference person was selected using these characteristics. Similarly, if the characteristics of all household members were used, the results would be biased toward larger households and against households with only one adult.5
Weighting
The analyses reported here were weighted and limited to adults aged 18 years and older. Household weights were used for the bivariate and multivariate analyses.
Logistic regression
The sections on income poverty, subjective poverty and financial stress include logistic regression analyses of the independent effects of demographics, education, socioeconomic factors, labour market experiences and wealth on the respective measure of poverty. Logistic regression is the most appropriate statistical technique for the analysis of dichotomous dependent variables.
Highest educational qualification was entered as a categorical variable since preliminary bivariate analyses revealed that it does not have a truly ordinal relationship with indicators of financial disadvantage. For example, a certificate qualification was not associated with a lower incidence of income poverty than completing school. For the logistic regression analyses, household type was indicated by two variables, marital status and number of children, so that the independent effects of each could be assessed. Present labour force status was not included in the multivariate analyses since it is was measured at the same time as poverty status. It was replaced by measures of the percentages of time spent, since leaving full-time education, in work (full-time or part-time) and in unemployment. Disposable income measures were not included in the analyses of income poverty, but were included in analyses of subjective poverty and financial stress. Assets and debts could not be included in the same regression analysis as wealth because they are the two components of wealth.
Within each section, the logistic regression analyses of the Wave 2 data are presented. The analyses of the Wave 1 data are presented in Appendix 3. The statistical significance of the coefficients is indicated in the standard manner (and described in the table notes). In the text, the logistic regression coefficients in the tables are discussed as odds ratios, which are the exponents of the coefficients. For categorical variables—sex, Indigenous status, language background, type of school attended, education and marital status—the effects are interpreted relative to the appropriate contrast group, that is, females, the non-Indigenous, those with an English-speaking background, those who attended a government school, and those who had never married and were not in de facto relationships. The interpretation of the odds ratio is relative to the contrast group; thus, the odds of men being in poverty are so many times the odds for women, the odds of married people being in poverty are so many times the odds for single people, and so on. Unlike other interpretations of logistic regression coefficients, odds ratios do not change depending on the values of the other independent (predictor) variables.
The interpretation of the logistic regression coefficients for continuous variables depends on the unit of measurement. The coefficient refers to a single unit change in the predictor variable, so its magnitude depends on how the variable is measured. For example, the number of siblings and number of children are continuous variables ranging from zero. Therefore, the coefficients of these variables are the effects on poverty status for a one-unit change, that is, for one additional sibling or child. For two siblings or two children, the effects are doubled and the odds ratios are squared. Similarly, the effects for three siblings or three children are 3.0 times the effects for one sibling or child and the odds ratios are cubed. Age has been divided by 10, so the effects are the change in the odds of being in poverty for a 10-year difference in age. Similarly, parental and respondents' occupational status have each been divided by 10, so the effects are for a 10-unit difference on the zero to 100 occupational status scale. Again, the effect for a 20-year difference is twice that for a 10-year difference and the odds ratio is squared. The percentage of time spent working was divided by 10, so the effects relate to an increase of 10 percentage points in the time spent working since leaving school. For unemployment, the effects are for an increase of 1 percentage point in time spent unemployed. Income was divided by 10,000, so the effects are for a $10,000 difference in income. Similarly, wealth was divided by 100,000, so the effects are for a $100,000 change in wealth.
Where appropriate, the variables were centred about their means so that the estimate for intercept would be meaningful. Parental and respondents' occupational status were centred at their respective means (about 43 and 47 on the 100 point scale). Percentage of time spent working since leaving school was centred at its mean of 73 per cent. Wealth was centred at average household wealth, which was approximately $420,000. Therefore, the estimate for the intercept can be understood as the log odds of being in income poverty (or subjective poverty or financial stress) for an individual who scores zero on all variables. That individual is female and 45 years old, has an average socioeconomic background, attended a government school, completed school (Year 12), is single with no children, works in a job with an average occupational status, since leaving full-time education has spent the average percentage of time working but no time unemployed, and has an average income and average wealth.
For the logistic regression analyses, groups of variables were added sequentially, beginning with a model comprising only demographic and socioeconomic background factors. The first group added was education, followed by marital status and number of children, then occupational status and labour market and unemployment history. The final variables added were wealth and, in the analyses of subjective poverty and financial stress, disposable income. The sequential modelling procedure shows which variables have statistically significant independent effects on financial disadvantage and which variables have effects that are mediated by variables added in later model specifications. For example, if 'first language not English' has a significant negative effect, this may be because 'first language not English' is associated with lower levels of education, different experiences in the labour market and lower levels of wealth. Alternatively, 'first language not English' may increase the likelihood of financial disadvantage, even when differences between first language groups in educational attainment, labour market experiences, wealth and, for analyses of subjective poverty and financial stress, disposable income are taken into account. In other words, education, labour market experiences and wealth may not account for the higher odds of this group being financially disadvantaged.
Included in these tables are the pseudo R square values, which indicate how well the independent variables account for the distribution of households on the respective indicator of financial disadvantage.6