Was Pro-Poor Economic Growth in Australia for the Income-Poor? And for the Multidimensionally-Poor?

Abstract

We investigate the pro-poorness of Australia’s strong economic growth in the first decade of the twenty-first century using anonymous and non-anonymous approaches to the measurement of pro-poor growth. The sensitivity of pro-poor growth evaluations to the definition of poverty is evaluated by comparing the results for the standard income-poverty measure with those based on a multidimensional definition of poverty. We find that Australian growth in this period can be only categorized as pro-poor according to the weakest concept of pro-poorness that does not require any bias of growth towards the poor. In addition, our results indicate that growth was clearly more pro-income poor than pro-multidimensionally poor. Counterfactual distribution analysis reveals that differences in the distribution of health between these two groups is the non-income factor that most contributes to explain this result.

1 Introduction

After two decades of economic growth Australia is now viewed internationally as the paradigm case of a dynamic economy capable of sustaining strong economic growth. In the period 2000–2010, Australia outperformed most economies in the developed world with an average GDP per capita annual growth above 2 %. This was the largest output growth among the rich OECD economies, which made Australia the sixth richest country within this group, only behind Luxembourg, Norway, US, Switzerland, and Netherlands. ¹ The increase in output came alongside a significant rise in employment. Thus, in 2008 Australia recorded its lowest level of unemployment since 1978, with an unemployment rate slightly above 4 %. Much has been written on the Australian economic miracle, however, yet little is known about the extent to which it has benefited the most disadvantaged groups in this country.

The main aim of this paper is to fill this gap by investigating the pro-poorness of Australia’s economic growth using alternative concepts and approaches to the measurement of pro-poor growth. ² Recent evidence suggests that Australia’s economic growth was not distributionally neutral. Similarly to other high-income countries (Atkinson 2005, for the UK; Piketty and Saez 2003, for the US; Saez and Veall 2005, for Canada), Australia has witnessed an increase in the concentration of income at the top of the distribution. Official figures from the Australian Bureau of Statistics (ABS) suggest a rise in the P90/P10 ratio, the share of income in the hands of the top quantiles, and the Gini index in the 2000s (Australian Bureau of Statistics 2011). ³ These findings are consistent with earlier results in the Australian literature that point to an upward trend in income differences since the mid-1990s (Saunders and Hill 2008; Saunders and Bradbury 2006). Wilkins (2007) concludes that the failure of the incomes of low-income people to keep pace with the growth of the median income explains the increase in relative poverty over that period. Importantly, pro-poor growth analysis provides valuable insights about the distributional impact of economic growth that cannot not be derived from the study of inequality and poverty measures. Inequality indices inform about the differences in the income distribution while poverty measurement is concerned with the short-fall of those who are below the poverty line. Alternatively, pro-poor growth measures evaluate the impact of growth on poverty by looking at the relative and absolute income gains of the poor. ⁴

The second objective of this paper is to investigate the extent to which pro-poor growth evaluations depend on the definition of poverty considered. In recent years there has been in Australia an intense debate on how to measure poverty and the need to move beyond standard income-based measures. Following this debate, the Australian government has decided to take on a new approach to the measurement of poverty based on a notion of social exclusion consistent with Sen’s idea of capability deprivation (Sen 2000). ⁵ We compare the results based on the standard income-poverty definition with those derived using a multidimensional framework recently proposed by the University of Melbourne and the Brotherhood of St Laurence to measure deprivation in Australia (Scutella et al. 2009a). This exercise is interesting for various reasons. First, it will serve to evaluate the capacity of different poverty definitions to identify those individuals that are most likely to be left behind in the process of economic growth. Most importantly, the comparison between growth evaluations based on multidimensional and income-poverty measures will allow us to investigate the importance non-income dimensions of welfare when measuring the income gains of those identified as poor, as well as, to determine the non-income attributes that are likely to shape the conclusions about the pro-poorness of growth.

To evaluate Australia’s growth we use data from the Household, Income and Labour Dynamics in Australia (HILDA) Survey. This is a nationally representative survey that is particularly suitable for pro-poor growth analysis as it provides longitudinal and cross-sectional information on households’ incomes. Although the data is available up to 2010, we focus our analysis on the period 2001–2008 to avoid the influence of the global financial crisis on the results. We find that the income gains from economic growth were highly concentrated at the upper end of the distribution so that growth can be deemed to be pro-poor only according to the weakest concept of pro-poorness. This was largely due to the growth of income from businesses, investments, and private pensions among those at the top. This result is consistent with previous research for Anglo-Saxon countries that identifies the changes in the distribution of these components as the major factor for the upward trend of top income shares observed in these countries since the 1980s (Atkinson and Leigh 2007, 2013; Piketty and Saez 2003; Atkinson 2005). Further, we find that the evaluation of growth critically depends on the concept of poverty adopted. While growth clearly benefited the income-poor, the income gain of those who were multidimensionally-poor was well below that of the mean. We apply the Oaxaca–Blinder and DiNardo–Fortin–Lemieux decomposition techniques to investigate the contribution of the different dimensions of poverty to the explain the gap between the two groups. We find that differences in the distribution of health and the larger incidence of people with disabilities or long-term health conditions among those who were poor in multiple dimensions are the non-income attributes that contribute the most to explain why growth was more favorable for the income-poor than for those facing multidimensional poverty.

The rest of the paper is organized as follows. Section 2 discusses the various concepts of pro-poorness, as well as, the different approaches to the measurement of pro-poor growth. Also in this section, we present the pro-poor growth measures we use in the analysis. Section 3 describes the data sources and definitions used in the paper. In Section 4, we describe the Australian social policy context and the main reforms in the last decade. Section 5 presents the main results on the pro-poorness of Australia’s growth for the different approaches and poverty definitions. We complete this section presenting a decomposition of the growth gap between the income and the multidimensionally-poor. Finally, Section 6 summarizes our main conclusions.

2 Concepts and Measures

2.1 The Concept of Pro-poor Growth

The impact of growth on poverty is a function of two factors: the magnitude of growth, i.e., the change in the mean income, and how the income gains are distributed among different groups (Datt and Ravallion 1992). At present, however, no consensus has been reached on how to integrate these two elements into an appropriate definition of pro-poor growth (Kakwani and Son 2008; Klasen 2008; Duclos 2009; Ravallion and Chen 2003; Kakwani and Pernia 2000). In this analysis we make use of the three concepts that have received the greatest attention in the literature, namely, the poverty reducing, the relative, and the absolute concepts of pro-poor growth. Proposed by Ravallion and Chen (2003), the first of these concepts identifies growth as pro-poor whenever it leads to a reduction in poverty. By looking only at the change in poverty, this definition fails to capture whether growth has a bias in favor of the poor as it characterizes growth patterns without accounting for how the benefits from growth are distributed among the population. The relative and absolute definitions of pro-poorness proposed by Kakwani and Pernia (2000) are stronger as they require a particular distribution of benefits between the poor and non-poor. In the relative case, growth can be characterized as pro-poor only when it increases the share of total income accumulated by the poor by benefiting the poor proportionally more than the non-poor. The absolute concept requires an absolute bias of growth in favor of the poor. Thus, for growth to be considered absolutely pro-poor, the income gain for the poor needs to exceed that of the non-poor so that absolute differences in income between these two groups are reduced as a consequence of growth. Importantly, the relative and absolute concepts both stress the distributional component of growth while omitting any reference to the absolute magnitude of poverty reduction. Osmani (2005) proposes a reformulation of these definitions in which the bias in favor of poor is expressed as a function of the difference between the actual reduction of poverty and the reduction that could be achieved in a distributionally neutral growth scenario. Within this framework, economic growth is relatively pro-poor if it leads to a reduction of poverty greater than the one observed if the benefits from growth were distributed in order to leave relative inequality unchanged. Similarly, growth is pro-poor in the absolute sense when it reduces poverty by more than a equally distributed growth pattern would.

Note that in a context of positive growth, the absolute definition imposes the strongest conditions as it requires that growth benefits the poor more than the non-poor in both absolute and relative terms. Further, the poverty reducing definition is the weakest of the three concepts as it focuses only on the effect of growth on poverty without incorporating any value judgment on inequality. However, as Kakwani and Son (2008) rightly point out, the ranking of concepts reverses when growth is negative. Indeed, when this is the case, the poverty reducing concept becomes the strongest one as it requires a increase in the income of the poor even when there is decline in aggregate income.

2.2 Measuring Pro-poor Growth

Different approaches and measures aimed to articulate the different concepts of pro-poorness have been proposed in the literature. These approaches fall into two broad categories depending on whether the anonymity axiom is satisfied or not. This axiom, otherwise called the ‘symmetry’ axiom, is one of the core axioms in welfare economics and it is generally invoked for the measurement of income inequality and poverty. Social evaluations consistent with this axiom use exclusively information on the income variable excluding any other people’s attributes from the social choice problem. In the context of pro-poor growth measurement, anonymity implies that growth assessments are based on cross-sectional comparison of the marginal distributions of income before and after economic growth (Kakwani and Son 2008; Ravallion and Chen 2003; Son 2004). Importantly, by focusing only in the income changes at different positions of the income distribution, cross-sectional measures disregard the issue of income mobility from the growth evaluation. As Grimm (2007) and Bourguignon (2010) argue, however, by excluding economic mobility from growth evaluations, cross-sectional measures may provide an incomplete picture of the pro-poorness of growth as they are not sensitive to the impact of growth on those who were initially poor. Clearly, growth evaluations that take into account the income change experienced by the initially poor need to incorporate information on the initial status of individuals and consequently they would fail to satisfy the anonymity axiom. Next we discuss the main features of these two approaches and the measures derived from them that we use in our empirical analysis.

2.2.1 Cross-Sectional Measures Based on the Anonymity Axiom

Let y be the relevant income variable and let μ stand for its mean value. We denote by γ and $\Updelta$ the growth rate and the absolute change in the mean income between dates t − 1 and t. Let F _t−1(y) and F _t (y) be the initial and final cumulative distribution functions of income informing about the proportion of the population with income less than y at t − 1 and t. Pro-poor growth evaluations consistent with the symmetry axiom are based exclusively on the information contained in these two functions. Within this approach, the most popular instrument for the measurement of pro-poor is the ‘growth incidence curve’ (GIC) proposed by Ravallion and Chen (2003). If we denote by y _t (p) = F ⁻¹ _t (p) the pth quantile of the income distribution, then the growth rate g(p) of this quantile can be expressed as:

$$ g(p)=\frac{y_{t}(p)}{y_{t-1}(p)}-1. $$

The GIC shows the growth rates at different positions of the distribution ranging from the lowest quantile to p _max. In the present analysis, pro-poor growth evaluations will be made for a general class of additively decomposable poverty measures that we denote by P. For any poverty line, ⁶ z, any poverty measure in this class can be written as

$$ P=\int\limits_{0}^{z}\theta (y,z)f(y)dx, $$

where θ(y, z) is an individual-poverty function homogeneous of degree zero in both arguments, and f(y) is the density function of income. Importantly, this class includes the most common measures of poverty used in the literature including the Foster et al. (1984) family of indices FGT _α and the Watts (1968) index W. ⁷ Importantly, the GIC can be used to derive dominance results on pro-poorness for the class P of poverty measures. Let H(y) denote the headcount index defined as the proportion of individuals whose income is less than y. Thus, when g(p) > 0 ∀ p < H(z) one can conclude that growth was poverty reducing for any poverty measure within this class (Atkinson 1987; Foster and Shorrocks 1988). Theorem 1 in Essama-Nssah and Lambert (2009) provides sufficient conditions for relative and absolute pro-poorness for every poverty index in P but the headcount ratio, for which these conditions do not apply. ⁸ Thus, if g(p) > γ  ∀ p < H(z) growth can be said to be relative pro-poor for any poverty measure within this group. Further the condition $g(p)>\frac{\Updelta }{y_{t}(p)}\,\forall p<H(z)$ is sufficient to characterize growth as absolute pro-poor for the same group of poverty indices. ⁹

When the dominance conditions are not satisfied we need to rely on partial pro-poor growth measures that allow us to draw conclusions for a particular poverty measure. For the present analysis we will consider the family of poverty equivalent growth rate(PEGR) measures proposed by Kakwani and Son (2008). Defined for the entire class of additively decomposable poverty measures, this is a general family that encompasses other well-known measures of pro-poor growth including the mean growth rate of the poor proposed by Ravallion and Chen (2003). ¹⁰ The PEGR can be used to articulate the different concepts of pro-poor growth as it characterizes growth patterns taking into account both the change in the mean income and how the benefits from growth are distributed among the population. Using the original notation of the authors, the PEGR is given by

$$ PEGR=\left(\frac{\delta }{\eta }\right)\gamma =\varphi \gamma , $$

where $\delta =\frac{dLn(P)}{\gamma }$ is the growth elasticity of poverty, and $\eta =\frac{1}{P}\int_{0}^{H}\frac{\partial P}{\partial y}y_{t}(p)dp$ is the neutral relative growth elasticity of poverty derived by Kakwani (1993), which indicates the percentage change in poverty caused by a 1 % growth in the mean income when all incomes grow at the same rate leaving relative inequality unchanged. ¹¹ Therefore, the PEGR is the growth rate that would bring the actual reduction in poverty, δγ, provided that growth increases all incomes by the same proportion. Importantly, for any additively decomposable poverty measure, the PEGR is consistent with the direction of change in poverty so that it can be used to infer whether growth is poverty-reducing or not: a positive (negative) value of PEGR implies a decline (increase) in the level of poverty. Further, a value of PEGR > γ implies that the actual poverty reduction is greater than the one that would be observed under equiproportional growth, and consequently growth can be classified as relative pro-poor. Lastly, as Kakwani and Son (2008) show, we can say that growth was pro-poor in the absolute sense when $PEGR>\bar{\gamma}>\gamma$, where $\bar{\gamma}=\gamma (1+\delta (\frac{1}{\eta }-\frac{1}{\eta ^{\ast }}))$ and $\eta ^{\ast }$ is the neutral absolute growth elasticity of poverty which tells us the percentage change in poverty when the gains from growth are equally distributed among the population.

2.2.2 Longitudinal Approach Based on Non-anonymous Measures

Pro-poor growth measures based on the anonymity axiom evaluate growth patterns by comparing the cross-section distributions of income without taking into account individuals’ mobility within these distributions. Consequently, social evaluations based on cross-sectional measures are independent of the extent to which growth benefits the initially poor. This, however, is an issue that many would consider as relevant for assessing the pro-poorness of any growth pattern. To measure the pro-poorness of growth in Australia without postulating anonymity we use the measurement framework proposed by Grimm (2007). Within this framework, it is assumed that individuals can be followed over time such that the joint income distribution function F(y _t−1,y _t ) can be inferred for a fixed population. It can also be assumed that individuals can be ranked in ascending order according to some variable, $\Upomega _{t-1}$, reflecting their initial status at t − 1. ¹² Let $p(\Upomega _{t-1})$ denote a variable informing about the absolute rank of individuals according to the indicator $\Upomega _{t-1}$. The income growth rate for the different positions within this rank can then be computed as

$$ g(p(\Upomega _{t-1}))=\frac{y_{t}(p(\Upomega _{t-1}))}{y_{t-1}(p(\Upomega _{t-1}))} -1, $$

where $y(p(\Upomega _{t-1}))$ denotes the income of the individual located in the p-th position of the ranking based on the $\Upomega _{t-1}$ variable. Similarly, the absolute variation for each position is given by

$$ v(p(\Upomega _{t-1}))=y_{t}(p(\Upomega _{t-1}))-y_{t-1}(p(\Upomega _{t-1})). $$

Grimm (2007) proposes the mean growth rate (MGRIP) and the mean income variation (MVIP) of the initially poor as measures of pro-poor growth. These can be expressed in terms of the function $g(p(\Upomega _{t-1}))$ and $v(p(\Upomega _{t-1}))$ as follows

$$ MGRIP=\frac{1}{H}\int\limits_{0}^{H}g(p(\Upomega _{t-1}))dp, $$

and

$$ MVIP=\frac{1}{H}\int\limits_{0}^{H}v(p(\Upomega _{t-1}))dp, $$

where H indicates the percentage of individuals classified as initially most disadvantaged according to the indicator $\Upomega _{t-1}$. It is worth noting the differences between these measures and the measures consistent with the anonymity axiom. Growth evaluations based on the measures proposed by Kakwani and Son (2008) and Ravallion and Chen (2003) look at the income change experienced by those positions in the income distribution below some poverty threshold without taking into account whether the occupants of these positions before and after growth are the same or not. In contrast, both the MGRIP and the MVIP use information on $\left\{ F(y_{t-1},y_{t}),\Upomega _{t-1}\right\}$ to describe transitions between t − 1 and t by linking income growth to the initial conditions of individuals. Given a ranking of individuals at the initial period, $p(\Upomega _{t-1})$, the MGRIP and the MVIP summarize the income change experienced by those characterized as initially poor according to $\Upomega _{t-1}$, omitting any information on those who were initially above the poverty threshold. Importantly, despite their focus on the initially conditions, longitudinal pro-poor growth measures can be used to assess the level of pro-poorness of growth. Following Grimm (2007) we define growth as unambiguously poverty reducing when the MGRIP > 0, i.e., when the average income growth among the initially poor is positive. Also, growth can be deemed to be pro-poor in relative terms when growth benefits relatively more those who are initially poor, i.e., when MGRIP is larger than the growth rate in the overall mean, γ. Lastly, growth can be characterized as absolute pro-poor when $MVIP>\Updelta$. ¹³

3 Data Sources and Definitions

We use data included in first eight waves of the HILDA Survey. This is a nationally representative survey that is particularly suitable for our analysis as it contains longitudinal and cross-sectional information that can be exploited to estimate cross-sectional and longitudinal pro-poor growth measures. The HILDA survey began in 2001 with a sample of 7,682 households containing 19,914 people. Of these, 13,969 individuals who were above 15 years of age in 2001 responded to an individual questionnaire including multiple questions on socioeconomic variables. Subsequent waves of HILDA have collected information from members of the original sample and from other new members of their households related to them. ¹⁴ Information on all members of the responding households from each wave of HILDA is used for the cross-section analysis, whereas longitudinal results are based on the panel data derived from the 13,969 respondents interviewed in the first wave. Importantly, using the appropriate cross-sectional and longitudinal weights, ¹⁵ this information can be used to study the changes in the Australian income distribution between the 2001 and 2008, as well as, the link between the initial conditions and income changes experienced by individuals over this period. To examine possible differences in the growth pattern within this period, in addition to the results for the 2001–2008 period, partial results for the 2001–2005 and 2005–2008 sub-periods are also discussed.

The unit of analysis we use in this paper is the individual. We assume individuals’ income is a function of the total income of the household to which they belong to. Concretely, each individual is assigned the equivalent household income, defined as total income per adult equivalent, where the number of equivalent persons is computed using the parametric specification proposed by Buhmann et al. (1988) given by

$$ e=N^{\theta }, $$

where N is the household size and θ is the measure of economics of scale within the household. Throughout the present analysis, a value for θ equal to 0.5 is assumed. Importantly, the main conclusions of the analysis are robust to the choice of this parameter. ¹⁶ The income variable considered in the analysis is household disposable income. This is defined as the sum of wages and salaries, business and investment income, private pensions, private transfers, and windfall income received by any household member. Further, our income variable includes the value of all public transfers provided by the Australian government, including pensions, parenting payments, scholarships, mobility and carer allowances, and other government benefits. The sum of these income components is reduced by personal income tax payments made by household members during the financial year. Finally, non-positive income values are excluded from the computations and all the pro-poor measures refer to growth in real income for which income values were converted in 2008 Australian dollars using the consumer price index provided by the ABS.

For the longitudinal pro-poor growth analysis, the link between poverty and income growth is studied using panel data for those individuals who were above 15 years of age when first interviewed in 2001. Two different approaches to the measurement of poverty are considered for the analysis. The first is the standard income-poverty approach in which income is the only relevant variable for defining individual’s poverty condition. In recent years there has been in Australia an intense debate on the capacity of income-based indicators to measure disadvantage and the need to move towards broader concepts of deprivation. Following the policy initiatives in the European Union and the UK, the Australian federal government has recently decided to adopt an approach to the measurement of poverty based on the notion of social exclusion. Consistent with Amartya Sen’s notion of capability deprivation (Sen 2000), social exclusion aims to capture the capacity of individuals to fully participate in social, economical, and political life. An important feature of the social exclusion approach is its multidimensionality. It assumes that the ability to engage depends on multiple factors and therefore any framework for measuring social exclusion must be multidimensional. Compared to Europe and the UK, empirical work on social exclusion in Australia has been more limited. With the aim to close this gap the University of Melbourne and the Brotherhood of St Laurence have recently developed a framework to measure multiple disadvantage in Australia (Scutella et al. 2009a, b). This measure builds on the Laeken Indicators and the Bristol Social Exclusion Matrix developed in Europe and the UK, respectively, and recognizes the multidimensionality of disadvantage incorporating information on 21 indicators from seven different domains: material resources; employment; education and skills; health and disability; social; community; and personal safety. A summary measure of poverty is derived from these indicators using a ‘sum-score’ method. This variable takes values in the interval [0,7], where 0 corresponds to the highest level of social exclusion. A complete description of the poverty index and the different indicators is presented in the “Appendix”.

4 The Social Policy Context

Australia has traditionally been described as a liberal welfare regime with modest social insurance where emphasis is placed on the private provision of welfare through market mechanisms. The Australian system, however, has a number of distinguishing features. Unlike other liberal systems like the US where the transfer system is financed by contributions from employers and the size of cash payments depends on individual’s earnings and employment history, the Australian model is characterized by flat-rate benefits unrelated to past earnings and funded from general revenue.

Australia has the most targeted system in the OECD. Underpinned by the principle of self-reliance by which every citizen with capacity to work should do so, the welfare system in Australia is aimed to help only those who are most in need while limiting the tax burden and the overall level of spending in order to minimize the disincentives for self-reliant behaviour. Figures for the mid-2000s indicate that the proportion of income paid in taxes by the average household was around 23 %, well below the OECD average of 29 %, whereas Australia’s spending on cash-transfers was the sixth lowest within the OECD group, spending about 8 % of the GDP (Whiteford 2010). Remarkably, more than 80 % of this spending was on means-tested programs, making Australia the OECD country with the highest spending on this type of benefits.

Over the recent decades, similarly to other high-income countries, Australia’s social security system has been through reforms clearly aimed at reducing welfare dependency and promoting self-reliance through paid work (Goodger and Larose 1999). The Australians Working Together package of 2003, the 2006 Welfare to Work reform, and the more recent Building Australia’s Future Workforce reform in 2011 all introduced policy initiatives to increase the conditionality of welfare payments and to strength the incentives to work. Importantly, these reforms led to a transition towards a two-tier system that offers more support to families and the aged while imposing further obligations on highly disadvantaged groups such as the unemployed, disabled people, and single parents (Mendes 2009). ¹⁷

The changes in social policy during the 2000s involved the tightening of access to unemployment benefits. This was done implementing tougher activity tests and higher penalties for non-compliance, extending the waiting periods for those who have accumulated some savings, and by imposing a 2 year waiting period for new immigrants. Reducing the number of recipients of pension benefits ¹⁸ was also a policy priority for that period. Thus, single-parents who started to receive a Parenting Payment Pension after 2005 would be moved onto unemployment allowance once the youngest kid turned six. Further, the eligibility criterion for the Disability Support Pension was tightened so that only individuals unable to work more than 15 hours per week are eligible for a pension. People with capacity to work between 16 and 30 hours who were eligible before the reform, are now entitled only to a lower unemployment allowance and therefore forced to comply with employment obligations that are satisfied by working for at least 15 hours per week, job-searching, or participating in training programs run by employment services (Harding et al. 2005). Given the difference in payment rates between pensions and allowances, these changes implied an important cut in the income transfer received by those affected by the reforms. However, the impact of these policies on the welfare of the most disadvantaged and their overall distributive consequences are issues that have not been investigated yet which clearly demand further research.

5 Results

5.1 Cross-Sectional Pro-poor Growth Measures

From 2001 to 2008, Australia witnessed strong and continuous economic growth. Based on HILDA data, figures on Table 1 suggest that mean and median income values grew more than 3.2 and 2.8 % per year during this period. Growth was particularly high between 2001 and 2005 where average income rose more than 3.6 % annually, whereas it slightly slowed down after 2005 with both mean and median values growing about 2.6 %. Changes in the mean and the median cannot be used to assess whether the distributional change was pro-poor as they are completely uninformative about the changes that took place at different parts of the distribution.

Table 1 Annual income growth in Australia between 2001 and 2008

Figure 1 presents our estimates of the Australia’s GICs consistent with the anonymity axiom for the periods 2001–2008, 2001–2005, and 2005–2008. ¹⁹ Curves for the whole period and the two sub-periods are remarkably similar. The shape of the three curves indicates that growth affected the income of every position within the income distribution. In particular, the GICs are above zero in the whole domain which means that growth was positive over the whole distribution. Therefore, for a broad class of poverty measures and any poverty line, we can conclude that growth in Australia in the period 2001–2008 was pro-poor according to the poverty reducing definition. However, the sufficient conditions for relative and absolute pro-poor growth are clearly not met. For any period considered, the curves shown in Fig. 1 suggest that growth was highly concentrated at the top end of the distribution with most of the bottom and middle positions growing less than the average. In fact, the GIC for 2001–2008 shows that the only positions that grew more than the mean in this period where those above the 90th percentile which implies that, for any relevant set of poverty lines, growth cannot be unambiguously characterized as relative or absolute pro-poor. For these definitions, therefore, we need to rely on partial results derived using specific combinations of poverty lines and poverty measures.

Fig. 1

Growth incidence curves for Australia, 2001–2008. a 2001–2008. b 2001–2005 and 2005–2008. Notes: Estimates computed using cross-sectional enumerated person weights. Source: Author’s calculation using HILDA data

Table 2 shows the estimates of the partial pro-poor growth measures consistent with the symmetry axiom for different additively decomposable poverty measures and a range of poverty lines. Concretely, we calculate the PEGR for the Watts index and three well-known measures within the FGT _α class of poverty measures: the headcount index, the poverty gap ratio, and the severity of poverty. Note that these three measures differ in terms of the weight assigned to those incomes that fall well below the poverty line. In particular, pro-poor growth evaluations based on the severity index put more weight on the lowest incomes than the headcount measure, with the poverty gap ratio lying somewhere in between. Poverty thresholds are defined using various percentiles of the initial distribution so that the proportion of people identified as poor is known. Consistent with the results from the GICs, we find that for any combination of thresholds and poverty measures the estimates are positive, which means that growth was poverty reducing. Interestingly, however, estimates in Table 2 suggest that, regardless the poverty line and the poverty index, the growth pattern in Australia between 2001 and 2008 cannot be characterized as either relatively or absolutely pro-poor. In fact, for all the periods considered the PEGRs are always below the actual growth rate of the mean. ²⁰ Thus, for instance, for the period 2001–2008 and for a poverty line equal to the 10th percentile, the amount of equally distributed growth that would bring the actual reduction in poverty as measured by the headcount index is 2.21 %, more than one percentage point less than the actual growth rate. Further, the comparison across FGT _α poverty measures suggests that the pro-poorness of Australian growth falls as more weight is assigned to the poorest positions. This comes from the fact that the PEGRs based on the severity index are in general below those for other indices, which means that the lowest incomes benefited from growth less than any other positions within the distribution.

Table 2 Partial pro-poor growth measures for Australia, 2001–2008

What are the factors underlying the observed distributive impact of growth in Australia? While this is a question that certainly requires a deeper investigation, we close this section with a discussion aimed at shedding some light on this issue. To this purpose, Table 3 presents information on the characteristics of different parts of the income distribution for the years 2001 and 2008. ²¹ The age and sex distributions of those at the lower and upper ends of the distribution did not experience any significant change: by 2008 women still outnumbered men at the bottom of the distribution, while men continued to be largest group among those at the top. Interestingly, we find a significant increase in the proportion of people reporting disabilities or any long-term health condition among those in low-income: from 39 % in 2001, the rate of disabled people rose up to 49 % in 2008. Consequently, the number of disability pensioners at the bottom of the distribution doubled in that period (from 6 to 12 %). Further, the larger concentration of disabled people at the lower end would contribute to explain why these positions failed to keep up with the rest of the distribution. In fact, as a recent submission to the Senate inquiry by various government departments concludes, relative to other groups in the population, little progress was made in improving the labour market outcomes of people with disability over the period 1998–2009. In this time, the participation rate of those with disabilities increased slightly more than 1 %, whereas the same rate among non-disabled individuals almost tripled (Australian Senate 2012). Further, the trends in the real value of welfare payments are likely to have contributed to the limited growth of the bottom positions. These payments failed to keep pace with the rise in average income, especially in the case of allowances whose value has fallen 25–35 % relative to community living standards (Gregory 2013). Labour earnings account for most of the increase in the income of the positions below the median. Thus, except for the lowest end, the period 2001–2008 saw a significant increase in the participation and employment rates in the bottom percentiles of the distribution: the proportion of part-time and full-time workers among those between the 5th and the 35th percentiles rose in both cases about 5 %. ²² Overall, the wages and salaries of people below the median grew faster than the average. This finding is consistent with the results in Greenville et al. (2013), who find an equalizing effect of labour earnings on the distribution of income over the last decade. Despite the increase in employment rates and wage income, positions below the median failed to keep up with the wealthiest. This was largely due to the large concentration of income gains from businesses, investments, and private pensions at top of the distribution. For these income factors, the gap between the richest 5 % and the other groups significantly widened over the last decade, especially in the case of investment income where the difference to the overall mean rose up from 3.5 in 2001 to 4.5 in 2008. Thus, the changes in the distribution of these components seem to have contributed to the upward trend of top income shares since the 1980s documented for Australia (Atkinson and Leigh 2007, 2013), but also for other Anglo-Saxon countries including New Zealand (Atkinson and Leigh 2005), the US (Piketty and Saez 2003), Canada (Saez and Veall 2005), and the UK (Atkinson 2005). ²³ Results in Atkinson and Leigh (2007, Fig. 7) suggest that the increase in the top income shares in the early 2000s ²⁴ was driven by non-salary income as the salary component became less concentrated at the top of the distribution. Our results are consistent with this trend. In fact, we find that the share of salary income of the richest 5 % declined between 2001 and 2008, as suggested by the fall in the ratio between the mean for this group and the overall mean (from 2.6 to 2.3).

Table 3 Mean characteristics for different parts of the income distribution, 2001–2008

5.2 Longitudinal Pro-poor Growth Measures

Pro-poor growth evaluations based on the cross-sectional comparison of marginal distributions do not provide any information on the gains experienced by those identified as initially poor. To obtain some insight on this issue we must turn to longitudinal pro-poor growth measures. We study the link between poverty and income growth using the standard income indicator and a multidimensional measure of poverty. For both of these measures we present results for the periods 2001–2008 and 2001–2005. A major issue of concern in the analysis of longitudinal data is the non-randomness of non-respond patterns as this constitutes a potential source of bias in the estimates. As other panel data surveys, HILDA is affected by this problem. Thus, of the 13,969 individuals interviewed in 2001 only 10,392 and 9,354 completed an interview in 2005 and 2008, respectively. Importantly, as the HILDA Annual Reports document, the probability of re-interview is not the same for all groups. In particular, males, individuals between 15 and 24, Indigenous people and individuals from non-English speaking countries, as well as, singles, unemployed, and people working in low-skilled occupations all have lower re-interview rates compared to other groups (HILDA 2011). Fortunately, to overcome this problem every wave of HILDA provides a series of longitudinal weights designed to control for selective non-response (including attrition) which would otherwise bias the population estimates. All the results in this section are derived using this set of weights. Furthermore, we propose an alternative set of weights to check the robustness of the results to the way the non-response process is modeled. Lastly, as it is common in the literature on income dynamics (see Gottschalk and Danziger 2001), to minimize the effect of transitory income variation and measurement error we consider a 2-year income average as our measure of income. Estimates for 2001–2008 are thus based on a sample with 8,700 individuals for whom the 2001–2002 and 2007–2008 average incomes can be compared whereas results for 2001–2005 use information from 9,521 individuals for whom the 2001–2002 and the 2004–2005 averages are available. Importantly, we find that the conclusions from the analysis are robust to the weighting method and the averaging of incomes. ²⁵

Table 4 shows the MGRIP and the MVIP computed for a set of thresholds used to identify the poorest individuals in the base year according to the two poverty measures.

Table 4 Longitudinal pro-poor growth measures for Australia, 2001–2008

In particular, we consider thresholds set equal to different percentiles of the distributions of the poverty indicators. Results in this table suggest that income gains among the initially poor were on average positive regardless of the definition of poverty considered. This implies that growth can be deemed to have been poverty reducing for both the unidimensional and the multidimensional approaches to poverty. However, evaluations based on the relative and absolute concepts of pro-poor growth depend on the definition of poverty adopted. As it is clear from Table 4, those who were on low-incomes particularly benefited from income growth. In fact, we find that for the periods 2001–2008 and 2001–2005 growth in Australia was relative pro-income poor as the average income growth rate of those who were in low-income was above the growth rate in the mean no matter which threshold is used to identify the poor. Also, the absolute income gain of the income-poor between 2001 and 2005 was larger than that of the mean for all poverty lines, which implies that growth in this period can be also characterized as absolute pro-income poor. ²⁶ For the period 2001–2008 this result holds only for income poverty thresholds below the 10th percentile of the initial income distribution. Remarkably, we find that Australia’s growth from 2001 to 2008 was clearly more pro-income poor than pro-multidimensionally poor. In fact, in contrast with the case of income-poverty, we find that growth in this period cannot be considered either relative or absolute pro-poor using a multidimensional measure of poverty. Thus, for any poverty threshold, both the average income gain and income growth rate of those identified as poor according to the multidimensional poverty measure are well below those of the mean. It was hypothesized that the difference between the income and the multidimensionally-poor could be explained by a larger presence of individuals at early stages of the income life-cycle among the former group. The figures on the right hand side of the table computed excluding all those were below 25 years of age in 2001 suggest that this is not the case as the gap between the multidimensional and the income-poor still persists after the exclusion of this group. We exploit the multidimensionality of the social exclusion measure to identify the role of the different dimensions in explaining the difference between the two groups of poor people. Table 5 presents the MGRIP for the income and the multidimensional measures, where the later has also been computed taking into account only one dimension at a time. The results suggest that there exists re-ranking across the different dimensions of exclusion. Thus, those who were most disadvantaged in the material resources, employment, and education domains benefited from growth more than the most deprived in the other dimensions. The rise in labour market opportunities in the form of part-time and casual jobs available for those with lower educational levels who were initially more disengaged from the labour market is a plausible explanation for the relative good performance of these groups. This would be consistent with the increase in employment and participation rates for positions below the median of the income distribution documented in the previous section. In contrast, individuals at the bottom of the health scale are clearly the group who benefited least from growth as suggested by the lower values of the MGRIP measure. Again, this is compatible with the increase in the number of people with disabilities and long-term health conditions at the bottom of the income distribution that took place over the last decade. Finally, the larger prevalence of people with poor health among the multidimensionally-poor could well explain why growth was less beneficial for this group than for the income-poor. The next section is dedicated to investigate the validity of this hypothesis.

Table 5 Longitudinal partial pro-poor growth measures for Australia, 2001–2008

5.3 Accounting for the Difference Between the Income-Poor and the Multidimensionally-Poor

Results from the previous section suggest that on average those who were in low-income benefited from growth more than those who were poor in multiple dimensions. Interestingly, we find that differences between these two groups are not only limited to mean values. Figure 2 shows the gap in the benefits from growth between the two groups across the whole distribution for the period 2001–2008. In particular, the results correspond to the case where poor groups are identified using a poverty threshold equal to the 15th percentile of each poverty index in 2001. ²⁷ Clearly, Australian economic growth in this period was unambiguously more pro-income poor than pro-multidimensionally-poor. In fact, the curves for the income-poor stochastically dominate those of the multidimensionally-poor, although in the case of annual variations the difference is only significant up to the median value. The gap in growth rates is particularly large at the bottom and the top end of the distribution, where the difference between the two groups is above 4 %.

Fig. 2

Differences in income gains: income versus multidimensionally-poor. a Annual variation. b Annual growth rate. Note: The graphs show the differences in the inverse distribution function of the benefits from growth between the income and the multidimensionally-poor for the period 2001–2008. Poor groups defined using thresholds equal to the 15th percentile of each poverty index in 2001. Dashed lines show the bootstrapped confidence intervals based on 1,000 replications. All estimates computed using longitudinal responding person weights. Source: Author’s calculation using HILDA data

Understanding the growth gap between the two groups of poor is important for various reasons. First, it will help us to understand why poverty definitions differ as regards their capacity to identify those individuals who are less likely to participate and benefit from economic growth. Most importantly, understanding the differences between the multidimensional and the income-poverty measures is crucial to determine the non-income dimensions that are key for identifying low-growth groups and that are therefore expected to have a critical role in shaping pro-poor growth evaluations. Table 6 presents the characteristics of the poor in 2001 as well as the average annual growth rates experienced by specific demographic groups between 2001 and 2008. We find that those identified as poor according the multidimensional poverty index in 2001 are on average more than 7 years younger than those in the income-poor group. Remarkably, despite of being a younger population, the multidimensionally-poor have worse health conditions than low-income people. Thus, the incidence of people with poor general, physical, and mental health is respectively about 15, 5, and 19 % points larger among those who are poor in multiple dimensions. Also, the proportion of individuals that report some type of disability or long-term health condition in this group is 10 % points larger than in the income-poor group. Those who are identified as poor by the multidimensional poverty index have lower educational attainment than those who were on low-income: the incidence of individuals with less than Year 12 among the income-poor is about five points lower than among those who are poor in multiple dimensions (71 vs. 76 %).

Table 6 Mean characteristics of the initially poor

As the figures on income growth rates in the right column of Table 6 show, people with poor health, disabilities, and lower educational attainment experienced little income growth compared to other groups. The higher prevalence of these individuals among the multidimensionally-poor could therefore account for the growth gap of this group. To check the validity of this hypothesis we will make use of conterfactual analysis. In particular, we follow the Oaxaca–Blinder and DiNardo–Fortin–Lemieux approaches to investigate the role of observed characteristics in explaining differences in the distribution of income gains between the multidimensionally and the income-poor. Let G _MP and G _IP denote the groups with the poorest 15th per cent as defined by the multidimensional and the income poverty indices in 2001, respectively. Let F _MP (g) and F _IP (g) be the distribution of growth rates (or absolute variations) among these groups. The well-know regression based approach first proposed by Oaxaca (1973) and Blinder (1973) allows us to decompose differences in mean growth rates observed between the two groups of poor people. For each individual i we assume that the income growth rate follows the model

$$ g_{i}=x_{i}\beta +e_{i}, $$

where x _i is a 1xk vector of covariates, β is the vector of parameters, and e _i is the error term satisfying E(e _i |x _i ) = 0. Let b _MP and b _IP be OLS estimates of β derived using observations from the G _MP and G _IP groups. Let $\bar{x}_{IP}b_{MP}$ denote the counterfactual value of the mean growth rate among the multidimensionally-poor if those were given the observed characteristics of the income-poor. Then, the difference between the mean growth rate of the of the multidimensionally-poor, g _MP , and the average growth rate of the income-poor, g _IP , can be expressed as

$$ g_{IP}-g_{MP}=\bar{x}_{IP}(b_{IP}-b_{MP})+ (\bar{x}_{IP}-\bar{x}_{MP})b_{MP}, $$

where the first term on the right-hand side captures the part of the gap caused by differences in coefficients, while the second term measures the expected change in the mean growth rate due to the shift in observed characteristics between the two groups (explained effect).

In contrast to the Oaxaca–Blinder decomposition, the DiNardo et al. (1996)-DFL thereafter- reweighting approach permits evaluation of the contribution of covariates to differentials across the whole distribution instead of focusing only on the mean. Each individual observation is drawn from a common joint density function f(g, x, G), where g, x, and G refer to income growth rate, observed characteristics, and group membership, respectively. The marginal distribution of growth rates for group G _MP is then given by

$$ f_{G_{MP}}(g)=\int\limits_{\Upomega _{x}}f(g|x,G_{MP}) \,f_{x}(x|G_{MP})dx, $$

where $\Upomega _{x}$ is the domain of individual attributes and

$$ f_{x}(x|G_{MP})=\int\limits_{\Upomega _{g}}f(g,x|G_{MP})dg, $$

with $\Upomega _{g}$ being the domain of annual growth rates. The counterfactual distribution for group G _MP is defined as the distribution of income gains that would prevail assuming group G _MP had the same observed characteristics of group G _IP . Following DFL, this can be expressed as

$$ f_{G_{MP}}^{G_{IP}}(g)=\int\limits_{\Upomega _{x}}f(g|x,G_{MP})\, f_{x}(x|G_{IP})dx=\int\limits_{\Upomega _{x}}f(g|x,G_{MP}) \,\Uppsi _{x}(x) f_{x}(x|G_{MP})dx, $$

where $\Uppsi _{x}(x)$ is the ‘reweighting’ function given by

$$ \Uppsi _{x}(x)=\frac{f_{x}(x|G_{IP})}{f_{x}(x|G_{MP})}= \frac{P(G=G_{MP})}{ P(G=G_{IP})}\frac{P(G=G_{IP}|x)}{P(G=G_{MP}|x)}, $$

where the last equality holds from Bayes’ rule. The first ratio is just the relative frequency of each group, which is constant and can therefore be ignored for the reweighting process. For the second term, following DFL, we estimate a probit model for the probability of belonging to each group G _IP and G _MP , given characteristics x. The counterfactual distribution function $F_{G_{MP}}^{G_{IP}}(g)$ can then be used to decompose the differences in the distribution of income gains between both groups as follows

$$ F_{IP}(g)-F_{MP}(g)=[F_{IP}(g)-F_{G_{MP}}^{G_{IP}}(g)]+ [F_{G_{MP}}^{G_{IP}}(g)-F_{MP}(g)]. $$

The second term of the equation represents the explained part of the gap which can be attributed to differences in the distribution of observed characteristics between the two groups. In contrast to the Oaxaca–Blinder approach, this decomposition can be used to evaluate the contribution of covariates to explain differences across the whole distribution. Thus, the differential at any percentile p can be decomposed as

$$ p_{{IP}}(g)-p_{{MP}}(g)=[p_{{IP}}(g)-p_{G_{MP}}^{G_{IP}}(g)]+ [p_{G_{MP}}^{G_{IP}}(g)-p_{{MP}}(g)]. $$

To determine to contribution of each covariate (or set of covariates) to explain the overall gap we apply a Shapley-type decomposition procedure (see Shorrocks 1999 and Sastre and Trannoy 2002). Widely used in inequality decomposition analysis, this decomposition identifies the contribution of each factor with the expected marginal effect on the explained gap of eliminating the covariate when computing the conterfactual estimates. Let K = (1, …, j, …, k) be the set of covariates, and let $S\subset K$ denote any possible subset of covariates. The Shapley contribution of characteristic j is given by

$$ Sh_{j}=\frac{\sum_{S\subset K, j\in S}\frac{(s-1)!(k-1)!}{k!} [e(S)-e(S\backslash \{j\})]}{e(K)} \quad\quad \text{with} \quad \sum_{j=1}^{k}Sh_{j} =1 $$

where s is the size of the subset, and $e(\cdot )$ is the explained effect that depends on the particular set of covariates used to derive the counterfactual estimate. ²⁸

The OLS and the probit regressions used for the counterfactual analysis include, as explanatory variables, multiple socioeconomic variables that are expected to influence individuals’ ability to benefit from economic growth. ²⁹ We group the covariates into five categories. The first one includes demographic information about the household where the individual lived in 2001, including the age and sex of the head; type of family expressed with dummy variables for couples with kids, couples with no children, lone-parent households with and without dependent children, singles, and other family types; and thirteen dummy variables for the major statistical regions reported in HILDA. ³⁰ Details on the initial socioeconomic conditions of the household are considered in a separate category. We include an indicator variable to identify those individuals living in an area which falls into the lowest 20 % most disadvantaged areas in Australia as measured by the index of relative socioeconomic disadvantage for areas (SEIFA); type of housing tenure with dummy variables for owners, renters, and rent-free households; and a dummy variable to indicate whether the individual belongs to a jobless household. Demographic characteristics of individuals in 2001 including age, sex, and an indicator variable identifying those with indigenous backgrounds are grouped in a third category. Information on individuals’ initial labour statuses, educational attainment, and English skills are considered in a separate group. This includes dummies for people working part-time, full-time, unemployed, long-term unemployed, full-time students, and other individuals out of the labour force; indicator variables for those with graduate or postgraduate education, bachelor or advanced diploma, certificate I, II, III or IV, Year 12 or less but still engaged in education,and those with Year 12 or less who were not in education; and a dummy variable taking value one for those who speak a language other than English at home and report that they do not speak English well or does not speak English at all. Lastly, the health category includes details on disabilities and the general, physical, and mental health status of the individual. In particular, for the three health dimensions we define five dummies, one for each of the five quintiles of the corresponding health index reported in HILDA. ³¹ The presence of disabilities is captured by an indicator variable that activates when the individual reports a long-term health condition or disability that restricts everyday activities for at least 6 months. The results of the regressions used for the analysis are presented in the “Appendix”. ³²

Table 7 shows the results of the counterfactual analysis for the case of the annual growth rates. Results for annual variations are quite similar and yield similar conclusions, so they are not discussed here for the sake of brevity. It is clear from this table that differences in observed characteristics contribute to explain why those who were poor according to multidimensional index benefited less from growth than those in low-income. Thus, for the mean, results from the Oaxaca–Blinder decomposition suggest that the average growth rate of the multidimensionally-poor would increase about 30 % (from 2.81 to 3.63) if the distribution of characteristics of the income-poor was assumed. ³³ This implies that differences in characteristics account for more than one quarter of the gap in mean growth rates. Figures from the DFL decomposition indicate that the effect of characteristics is not uniform over the whole distribution. Counterfactual estimates for the 10th and 20th percentiles show that the contribution of characteristics is particularly large at the bottom of the distribution, where differences in characteristics account for more than 50 % of the gap between the two sets of poor people. In contrast, we find that characteristics cannot explain the observed gap in the middle and upper parts of the distribution. Indeed, the gap at the median and the 80th percentile increases when compositional differences are taken into account. The Shapley contributions of each group of covariates to the explained gap in mean are presented in the bottom part of the table. Interestingly, both the Oaxaca–Blinder and DFL methodologies point to differences in health conditions and the incidence of disability as the most explicative factor for the gap between the multidimensionally and the income-poor. Thus, differences in the distribution of health and the larger incidence of people with disabilities or long-term health condition among those who were poor in multiple dimensions jointly account for 98–108 % of the explained difference between the average growth rate of this group and that of the income-poor. The initial socioeconomic conditions of the household is the second most important factor with a contribution that is between 23 and 35 %, depending on the decomposition method adopted. The Shapley value of the demographic characteristics of individuals is negative, which means that the gap in mean growth rates between the two groups widens once differences in age, sex, and indigenous background are controlled for. This could be explained by the larger prevalence of individuals above 65 years of age who had little income growth among the income-poor relative to the multidimensionally-poor (see Table 6). ³⁴ Finally, the Shapley contribution of the initial labour status and skills is also negative but statistically insignificant. In this case, from the figures in Table 4 we know that the income-poor population has higher educational attainment than those who were poor in multiple dimensions. However, this effect could be more than offset by the larger prevalence among the income-poor of individuals who were out of the labour force and benefited relatively little from growth.

Table 7 Oaxaca–Blinder and DiNardo–Fortin–Lemieux counterfactual decompositions

6 Conclusions

In first decade of the twenty-first century Australia consolidated its position as a high-growth economy in the developed world. In the period 2000–2009, Australia experienced one of the largest output growth rates among OECD, only overtaken by a group of countries with lower initial income including Turkey, Hungary, Greece, the Czech Republic, Korea, Poland, and the Slovak Republic. Recent evidence suggests, however, that the benefits from growth in Australia were not evenly distributed. Similarly to other rich economies like the US, the UK, and Canada (Atkinson and Leigh 2013; Piketty and Saez 2003; Saez and Veall 2005) Australia has witnessed a rise in the concentration of incomes at the top of the distribution alongside an increase in partial inequality measures (ABS 2011). To date much has been written about the Australian economic miracle, however, yet little is known on the extent to which the strong economic growth has been pro-poor or not. Our aim in this paper was to fill this gap.

Pro-poor growth analysis contributes to our understanding of the distributional effects of growth by providing insights that cannot be derived from the analysis of standard inequality and poverty measures. Thus, while inequality and poverty measures are concerned with the differences in the income distribution and the income gap of those who are below some threshold, respectively, pro-poor growth measures evaluate the impact of growth on poverty reduction by looking at the extent to which growth benefits the poor. In this paper we have investigated the pro-poorness of Australian growth using cross-sectional and longitudinal pro-poor growth measures. These two approaches complement each other in that they focus on different aspects of the distributional change associated to economic growth. Growth assessments consistent with the anonymity axiom evaluate the distributional impact of growth looking only at the income change experienced by the bottom positions of the income parade without taking into account whether these positions are occupied by the same individuals or not. In contrast, non-anonymous evaluations focus on the mobility aspect of growth looking exclusively at the income change experimented by those who were initially poor. An important issue that arises in this type of evaluation is how to identify those initially in poverty. We compare the results based on the standard income-poverty with those derived using a multidimensional definition of poverty that embraces multiple non-income attributes.

Results for the cross-sectional measures suggest that Australian growth in the last decade was pro-poor only according to the poverty reducing definition of pro-poorness. This is the weakest concept of pro-poor growth as it identifies as pro-poor every growth pattern that increases the income of the poor, regardless of how the benefits from growth are distributed among the different positions in the income distribution. We find Australia’s growth was highly concentrated at the top of the income distribution. Consistent with the evidence for other high-income countries, changes in the distribution of income from businesses, investments, and private pensions seem to be the major factor underlying the upward trend of top income shares observed in Australia. In fact, as in Atkinson and Leigh (2007), we find that the salary component became more evenly distributed and less concentrated at the top of the distribution in the 2000s.

We exploit the longitudinal information in HILDA to study the effect of growth on those who were initially poor. Our results based on longitudinal measures indicate that the pro-poorness of growth in this case critically depends on the definition of poverty considered. Thus, while there exists high income mobility, with those initially in the low-income group growing more than those with high incomes, the income gain of those identified as poor according to the multidimensional poverty measure was far below that of the mean. Therefore, we can conclude that growth was more pro-income-poor than pro-multidimensionally poor. Interestingly, we find that differences in the distribution of health and the larger incidence of people with disabilities or long-term health condition among those who were poor in multiple dimensions explain why growth was less pro-multidimensionally poor. Indeed, the average annual growth rate of those who were poor according to the multidimensional measure would increase about 16–30 % if the health distribution of the income-poor was assumed. This highlights the sensitivity of non-anonymous growth evaluations to the way poverty is defined, in particular, to whether the definition of the poor incorporates information about the health dimension of well-being or not.

Appendices

Appendix

1.1 The Index of Multidimensional Poverty

The poverty index proposed in Scutella et al. (2009a, b) combines information on twenty-one indicators from seven different domains: material resources; employment; education and skills; health and disability; social; community; and personal safety. Table 8 presents a description of the indicators included in each domain. For any individual i the measure of social exclusion, x ^S _i , is defined as seven minus the weighted sum of the level of social exclusion experienced within each domain, x _id , where every domain is assigned equal weight:

$$ x_{i}^{S}=7-\sum_{d=1}^{7}x_{id}. $$

The level of exclusion in any domain is given by the actual proportion of indicators within the domain in which the individual is deprived, which can expressed as follows

$$ x_{id}=\frac{\sum_{k=1}^{K_{d}}x_{id}^{k}}{K_{d}}, $$

where x ^k _id is a binary indicator taking value 1 when the individual is deprived in the indicator k of social exclusion included in the domain d, and K _d refers to the total number of indicators for domain d.

Table 8 Index of multidimensional poverty

1.2 Controlling for Selective Non-response

Every wave of the HILDA survey provides longitudinal weights designed to control for selective non-response including attrition. These weights are constructed adjusting the initial person weights for the probability of non-response in subsequent waves. This probability is derived using a logit specification estimated using data internal to the HILDA survey. Concretely, this model assumes that the probability of response is a function of the characteristics of the individuals and the household they belong to and other factors related to the conditions of the personal interviews. A detailed discussion of the model and how its estimates are used to construct the weights can be found in Watson (2004).

To assess the robustness of the pro-poor growth estimates to the way selective non-response is accounted for we compute our own longitudinal weights following the method outlined in Watson (2004). Concretely we estimate the probability of responding using the model employed in the conterfactual analysis of Section 5.3 Compared to the model used for the original weights, our model does not include any information on the conditions of the initial interview and incorporates more detailed information on the statistical regions, disabilities and the health conditions (general, physical, and mental) of individuals. Estimation results are available upon request. Table 9 compares the original estimates based on HILDA weights with those computed using the new weights and those derived without any weights. Further, the bottom panel shows the estimates obtained using 1 year incomes instead of two-period averages. The figures from this table suggest that pro-poor estimates are robust to the way selective non-response is accounted for and that those estimates are not driven by the averaging of incomes.

Table 9 Robustness of pro-poor growth measures to weighting and averaging, 2001–2008

1.3 Pro-poorness and Income Mobility

We aim to partially eliminate the influence of mobility on our pro-poor growth measures by controlling for two sources of income-mobility: that due to the natural growth of life-cycle earnings and the one explained by the initial income conditions. Following the literature on income dynamics (see, for instance, Gosttchalk and Moffitt 1994), we assume that income gains depend, among other things, on some function of individuals’ initial age and income. We propose the following model

$$ \dot{y}_{i}=Age_{0}\lambda _{1}+(Age_{0})^{2}\lambda _{2}+(Age_{0})^{3}\lambda _{3}+(Age_{0})^{4}\lambda _{4}+y_{0}\lambda _{5}+u_{i}, $$

where the income gains of the individuals, $\dot{y}_{i}$, depend on the life-cycle of earnings captured by a quartic function of age, the initial income, y ₀, and other factors included in the residual term u _i . Under this specification, u _i measures the income variation that is not explained by life-cycle factors and the initial income status. We estimate the parameters of the model by ordinary least squares and we use these estimates to compute the residual for every individual in the panel. Table 10 compares the actual pro-poor growth measures with those derived using the residuals of the model for the period 2001–2008. To differentiate the effect of the two sources of mobility, we compute the pro-poor measures using residuals from two versions of the model: one that controls only for life-cycle earnings (MGRIP*) and a second one that controls also for initial income (MGRIP**). The comparison of MGRIPs suggests that part of the income gains of the poor and, therefore, the pro-poorness of growth can be partially explained by income mobility. Thus, for any combination of poverty measure and threshold, we find that MGRIP* and MGRIP** are always lower than the actual MGRIP. The effect is particularly important in the case of the multidimensional measure, as the MGRIP becomes negative when we eliminate the part of the income growth explained by the initial age and income. Interestingly, even after controlling for these factors, we still find that growth was particularly beneficial for those initially in low-income and that growth was clearly more pro-income poor than pro-multidimensionally poor.

Table 10 Income mobility and pro-poor growth measures, 2001–2008

Counterfactual Analysis: Regressions

See Tables 11 and 12.

Table 11 Oaxaca–Blinder OLS regressions

Table 12 DiNardo–Fortin–Lemieux probit regressions for the probability of being income and multidimensionally-poor