Low income dynamics among ethnic minorities in the UK

Transcription

1 Low income dynamics among ethnic minorities in the UK Victor Perez Ricky Kanabar Alita Nandi Institute for Social and Economic Research (ISER) University of Essex Abstract We investigate low income dynamics among ethnic minorities in the UK, whilst simultaneously controlling for initial conditions and non-random attrition. Using a first order Markov model developed by Cappellari and Jenkins (4) we use data from Understanding Society, a large representative longitudinal household survey comprising a specific boost sample for the study of ethnic minorities to analyse the differences between and within the main ethnic groups in the UK. We model and test the existence of state dependence for each ethnic minority group, and examine how the estimated poverty persistence and poverty entry rates change for individuals with particular characteristics. The results suggest that ignoring the presence of initial conditions and/or nonrandom attrition overestimates the magnitude of poverty persistence, particularly for Pakistani and black African groups. Indeed, with the exception of the black Caribbean group, the hypothesis that conditional poverty status, initial poverty status and non-random attrition are uncorrelated is strongly rejected. Moreover, our results suggest that one cannot reject the absence of genuine state dependence i.e. scarring effects for the Pakistani, Bangladeshi and black Caribbean groups. Stylised examples for individuals with particular characteristics highlight that differences in poverty persistence, poverty entry and duration in poverty or non-poverty arise not only between, but also within members of the same group, highlighting significant within group heterogeneity. Keywords: Poverty persistence, Ethnic minorities, UK JEL classifications: I3, J15 This work was supported by the Economic and Social Research Council. ISER, University of Essex. vhperea@essex.ac.uk ISER, University of Essex. Essex, CO4 3SQ, UK. ISER, University of Essex. Essex, CO4 3SQ, UK.

2 1. Introduction Ethnic minorities in the UK face higher risks of being unemployed, having part-time jobs or working in less well-paid occupations than individuals from the white majority, as attested by several studies (e.g., Berthoud, ; Brynin and Güveli, 1; Hills et al., 1; Longhi et al., 1; Modood et al., 1997; Twomey, 1). In the case of poverty, investigations such as those by Adelman et al. (3), Berthoud (), Platt (7b), Palmer and Kenway (7), Nandi and Platt (1) and Barnard and Turner (11) have documented that the rates of individuals living in households with low income is substantially higher among ethnic minorities, although with substantial differences between and within ethnic groups. 1 In contrast, the evidence on the extent and characteristics of low-income dynamics among ethnic minorities in the UK is still limited, largely due to lack of appropriate data (Barnard and Turner, 11; Platt, 7a,b; Smith and Middleton, 7). While high rates of incidence of poverty may be due to economic fluctuations, high rates of persistent poverty are indicative of systematic causes and is thus of serious concern to policy makers. Higher rates of persistent poverty among ethnic minority groups as compared to the white majority group could be indicative of discrimination in various domains of life, such as employment, housing, education and health which has been outlawed by the Equality Act 1 (and earlier Race Relations Act 1965, 1968, 1976). Thus, it is vital for policy makers to have accurate estimates of poverty persistence. Making use of a new longitudinal household survey with an ethnic minority boost sample, Understanding Society, a recent report by Fisher and Nandi (15) indicates that individuals from most ethnic minorities not only have higher rates of poverty and low-income, but also higher rates of poverty persistence (defined as experiences of poverty lasting at least two out of three consecutive years). The study finds that, between 9 and 1, while 13 percent of individuals from the white majority group were poor over two or more consecutive years, for the Bangladeshi and black Caribbean groups this rate was around percent, and for the Pakistani and black African groups over 3 percent. These higher rates of poverty persistence suggest that previous experiences of poverty can be a strong determinant of future poverty in these populations, highlighting the need for evidence which assesses the extent of scarring effects among these groups. This study also found factors such as education and English language skills were associated with poverty persistence. However, this study did not take into account attrition. As there is evidence of non-random attrition, which may be associated to migration, health problems, age, or other factors correlated with either poverty or ethnicity (Lynn et al., 1; Schneider, 13; Uhrig et al., 8; Watson and Wooden, 9), not accounting for attrition implies that the results may be subject to bias. This study also failed to correct for endogeneity of initial poverty status. In this study, we extend this analysis and model the dynamics of poverty and low-income among ethnic minorities in the UK using an approach initially developed by Cappellari and Jenkins (4) that explicitly address two common problems in this area: initial conditions and non-random survey attrition (Cappellari and Jenkins, 4; Jenkins, 11). Previous studies for the UK, such as those by Cappellari and Jenkins (b) or Cappellari and Jenkins (4), have used multivariate models to control for various sources of observed and unobserved heterogeneity, as well as to test the existence of true or genuine state dependence in poverty status (Jenkins, 11). These studies have found that the 1 See Berthoud (), Kenway and Palmer (7), Platt (7b) and Barnard and Turner (11) for reviews of the research on poverty and ethnicity in the UK. 1

3 higher levels of poverty persistence in ethnic minorities remain, even after controlling for differences in observable characteristics. However, a major limitation of these results is the inclusion of ethnicity as an unobserved effect, so that other individual and household characteristics have the same effect across all individuals and groups. The literature on ethnic minorities has consistently reported the existence of substantial heterogeneity between and within ethnic groups (Berthoud, ; Clark and Drinkwater, 7, 9; Dustmann and Theodoropoulos, 1; Nandi and Platt, 1; Platt, 7b), suggesting that there could be substantial variations in the degree and characteristics of poverty persistence not addressed in the existing studies. A common restriction for this kind of analysis is the small sample of ethnic minorities in household surveys (Devicienti, ; Platt, 9; Smith and Middleton, 7). Although nowadays the inclusion of questions regarding ethnic self-identification is routine in all major surveys (Afkhami, 1), the sample of ethnic minorities is usually inadequate to investigate differences between and within ethnic groups. The small sample problem is particularly important in longitudinal studies, subject to attrition over the years, which is higher for ethnic minorities (Knies, 15). To our knowledge, this is the first study to focus on how the persistence of poverty varies between and within ethnic minorities. In addition, we treat as endogenous the initial poverty status our estimates control for potential differences in the propensity to poverty defined before the survey begins, so that the poor population could be a non-random sample more likely to remain poor over time Andriopoulou and Tsakloglou (15); Heckman (1981a); Jenkins (11). By explicitly modeling attrition, we control for potential non-random differences among ethnic minorities reported in other studies that could introduce similar biases (e.g., if those with the worse economic conditions and more likely to remain poor decided to migrate) (Dercon and Shapiro, 7; Lynn et al., 1; Schneider, 13; Watson and Wooden, 9). Our results suggest that there are important differences in both the magnitude and direction of the association of individual and household characteristics with the transitions into and out of poverty across ethnic groups, so that ignoring these differences introduce considerable biases in the magnitude of poverty persistence, particularly for the Pakistani, Bangladeshi and black African groups. In addition, the hypothesis of joint exogeneity to initial conditions and non-random attrition is strongly rejected for all groups, except black Caribbean. We find evidence of substantial state dependence for the white majority, Indian and black African groups, but we cannot reject the hypothesis of no state dependence for Pakistani, Bangladeshi and black Caribbean individuals. The poverty persistence and poverty entry estimates vary not only between ethnic groups, but also within individuals of the same group. In terms of policy design, this information may provide guidance to identify which sub-populations within each ethnic group are in higher vulnerability to fall or remain in poverty, such as lone parents in the black African group or women and young adults for the Indian and Bangladeshi/Pakistani groups. In the next section, we provide a short review of the literature on poverty dynamics among ethnic minorities in the UK, followed by the description of the econometric model in Section 3. Section 4 introduces our data and the descriptive statistics of the variables and instruments used in the analysis. Section 5 contains the results of the model, starting with a discussion of the estimation procedure, the tests of instrument validity and state dependence, and finalises with a description of the estimated poverty entry and exit probabilities for different types of individuals. Section 6 presents our concluding remarks.

4 . Literature review.1 Low-income among ethnic minority groups The research on the disadvantages of ethnic minorities has paid special attention to the gaps in pay, employment rates and other labour market outcomes of these groups, with respect to the white majority (Blackaby et al., 5, ; Brynin and Güveli, 1; Brynin and Longhi, 15; Clark and Drinkwater, 7, 9). The existing evidence suggests that controlling for differences in observable characteristics explain only a fraction of the disparities between ethnic groups, with a substantial unexplained component (usually known as the ethnic penalty Heath and McMahon, 1997). Individuals from all ethnic minority backgrounds have lower employment rates than their white counterparts, even when differences in educational qualifications (Machin et al., 9) or language skills (Dustmann and Fabbri, 3) have been considered, particularly in the case of women (Hills et al., 1; Nandi and Platt, 1). Longhi et al. (1) and Lindley () find similar patterns for earnings, and for different definitions of minority group (e.g., considering religion in their definition). In addition, these inequalities not only persist over time (Berthoud and Blekesaune, 7; Catney and Sabater, 15), but also across generations (Algan et al., 1; Cheung and Heath, 7; Longhi et al., 1). Moreover, there are substantial disparities between ethnic groups, with those of Indian and Chinese origins having similar or even better outcomes than those of the white majority (specially in education, where these groups performs significantly better than the white majority), and significantly worse among Pakistanis, Bangladeshis and black Africans. In the case of low-income, figures from the Office for National Statistics (ONS) for 14/15 show that 14 percent of the individuals from the white majority group lived in low-income households, while 38 percent of the Pakistani and Bangladeshi, and more than percent of other ethnic minorities were in the same situation (Shale, 16). Despite considerable reductions in poverty rates across all ethnic groups since the 199s, the pattern of greater disadvantage among individuals from Pakistani, Bangladeshi and black African backgrounds still persist (Berthoud, ; Berthoud and Casey, 1998; Modood et al., 1997; Platt, 7b). The worse labour market outcomes of most ethnic minority groups provide a possible explanation for the high poverty rates of this population (Barnes et al., 15; Shale et al., 15), but, as noted in Nandi and Platt (1), even those ethnic groups with similar labour market performance than of the white majority have higher than average poverty rates, suggesting that the existence of additional restrictions of disadvantages in ethnic minority households that makes them prone to experience poverty. Palmer and Kenway (7) indicate that the higher rates of poverty among ethnic minorities in the UK can be observed for all ages, family types, work status and across all regions. Using decomposition methods, these authors find that around half of the ethnic poverty penalty among Bangladeshi, Pakistani and black African groups can be accounted for by differences in demography and employment status (being the first more relevant for the black African group, and the latter among the Pakistani and Bangladeshi), although the other half remained unexplained. Platt (9) provide estimates of the ethnic penalty in poverty status, finding a similar pattern: large penalties for the Pakistani and Bangladeshi groups (over 3 percent), and lower but still considerable penalties for the black African (16 percent), Indian and black Caribbean groups (around 1 percent). Refers to households with income before housing costs below 6 percent of the median. See Section 4 for further details. 3

5 . Low-income transitions and ethnic groups The literature on the dynamics of low-income and poverty in the UK provides limited evidence on the extent and characteristics of poverty persistence among ethnic minorities, largely due to lack of longitudinal data sets with sizeable samples of this population (Cappellari and Jenkins, 4; Platt, 7a,b; Smith and Middleton, 7). 3 The introduction of the British Household Panel Survey (BHPS), which started in 1991 and continued until 8, provided a fertile ground for several studies on the dynamics of poverty with several approaches: endogenous switching models (Cappellari and Jenkins, 4), hazard models (Devicienti, ), structural models (Aassve et al., 6), or variance decomposition models (Blundell and Etheridge, 1). 4 However, many of these studies do not consider differences associated to ethnic background (e.g., Antolin et al., 1999; Barton et al., 1; Jarvis and Jenkins, 1997, 1995; Jenkins and Rigg, 1; Jenkins, ; Jenkins et al., 3; Maggio et al., 4), or provide mainly descriptive evidence (such as Adelman et al., 3; Barnes et al., 8). Adelman et al. (3) suggest that children in households with at least one non-white adult were twice as likely to be in severe and persistent poverty than those where no adult was non-white, while Barnes et al. (8) find that ethnicity is a relevant factor to explain the persistence of poverty among couples with children. In two related studies, Platt (3) and Platt (6) analyse the persistence of benefit receipt as a proxy of poverty in a short panel study in Newham (a borough in East London). These studies found that the Pakistani, Bangladeshi and black African groups presented the higher rates of poverty persistence and material deprivation in the sample. Using a multivariate duration model, this author concluded that the Bangladeshi group in this sample had lower probabilities of leaving benefits and higher of re-entry to benefits that the white majority group, while the other ethnic minorities presented similar levels to the latter group (with exception of the black Caribbeans who had lower probabilities of re-entry). The studies by Cappellari and Jenkins (4), Devicienti () and Tomlinson and Walker (1) estimate multivariate models of low-income transitions using BHPS data. Cappellari and Jenkins (b) and Cappellari and Jenkins (4) introduce a model of transitions into and out of lowincome as a first order Markov process which explicitly controls for initial conditions and non-random attrition (Jenkins, 11). In their application for the UK, these authors find that for households with Pakistani or Bangladeshi backgrounds, being poor in one period imply a probability 16 percent higher of being poor in the next period, with respect to a comparable household from the white majority. The higher persistence of poverty in these groups result in longer spells of poverty and shorter spells of non-poverty. In these studies, the hypothesis of exogeneity of poverty transitions with respect to both initial conditions and attrition was strongly rejected, remarking the relevance of controlling for these elements to provide unbiased estimates. In addition, these authors find evidence of substantial state dependence, suggesting that around 6 percent of it was genuine state dependence. 5 Devicienti () uses discrete-time proportional hazard duration models to estimate the probability of poverty entry and poverty exit, controlling for multiple socio-economic variables and localeconomic conditions. The results of this model suggest that non-white individuals (mostly Afro- 3 Smith and Middleton (7), Platt (7b) and Platt (9) provide reviews of the literature on the dynamics of poverty and low income among ethnic minorities in the UK. 4 See Jenkins (11) and Kanabar (forthcoming) for discussions on the assumptions and characteristics of these approaches. 5 See Section 3 for details. 4

6 Caribbean, Indian and Pakistani) are 35 percent less likely to exit poverty at any given moment than the white majority. Tomlinson and Walker (1) use dynamic probit models to study recurrent poverty, finding that that it is more common among individuals with low education levels, skilled manual and lower-skilled workers, lone parent families and those unemployed or economically inactive. Moreover, these authors find that individuals with previous experiences of poverty are more likely to be poor in the future, even after differences in education, occupation and family composition have been considered. Other authors have focused on the analysis of demographic and employment transitions to explain poverty inflows and outflows, focusing on the behavioural decisions of individuals and not only in the income or poverty status (e.g., Aassve et al., 6; Antolin et al., 1999). The recent introduction of longitudinal surveys with boost samples aimed specifically for the analysis of ethnic minorities, the United Kingdom Household Longitudinal Study (UKHLS, also known as Understanding Society) 6 and the Millennium Cohort Study (MCS) 7, provide additional evidence on the dynamics of poverty in these groups. (Barnes et al., 15) uses the first three waves of Understanding Society to investigate the magnitude and characteristics of poverty entry and exit rates in children, finding that the Pakistani, Bangladeshi and black African groups have higher rates of poverty entry (between 15 to 9 percent) than the white majority (around 5 percent), as well as lower rates of poverty exit (3 percent versus 4 percent for other ethnic groups). Platt (9) uses the MCS to study the life events associated with poverty entries and exits in children from ethnic minority backgrounds, encountering that changes in the household composition and the entry of the main carer into the labour market have strong associations with the dynamics of poverty in this population. This study uses a model poverty transitions to estimate the differences in the probability of entering or exiting poverty for each ethnic minority group with respect to the white majority, finding that children from Pakistani and Bangladeshi origin were less likely to exit and more likely to fall in poverty, while Indian children were only more likely to enter poverty and the black Caribbean and black African children had similar probabilities in both cases. A recent report by Fisher and Nandi (15) uses data on the first three waves of Understanding Society to analyse poverty dynamics among ethnic minorities in the aftermath of the 8/9 recession. This authors indicate that individuals from most ethnic minorities not only have higher rates of poverty and low-income, but also higher rates of poverty persistence (i.e., experiences of poverty lasting two or more consecutive years): between 9 and 1, 13 percent of individuals from the white majority group were poor over two or more consecutive years, while this rate for the Bangladeshi and black Caribbean groups was around percent, and for the Pakistani and black African groups over 3 percent. These authors find that persistently poor households were in average younger, more likely to have no qualifications and to be first generation migrants (i.e., not born in the UK). In the case Pakistani households, those with persistent poverty were larger and more likely to have children. An interesting finding in this report is the role of English language skills in the probability of being persistently poor, showing that, controlling for educational qualifications and type of family, having English as first language decreases the probability of being persistently poor in 5 percent, although 6 In Section 4 we provide further details on this data set. 7 The Millennium Cohort Study is a multi-thematic survey which interviewed a group of approximately 18, children born in /1, who were initially interviewed at age nine months, and with follow up interviews at age 3, 5, 7 and 11, as well as planned interviews at ages 14 and 17 (Hansen, 14). The MCS oversampled children from disadvantaged and ethnic backgrounds, and collected data of around,5 children with ethnic minority backgrounds in the first wave (Hansen, 14). 5

7 this effect could be associated to being born in the UK rather than an effect of language skills per se. In this paper, we use a similar data set to that used by Fisher and Nandi (15) and Barnes et al. (15) to study how the conditional poverty persistence rates vary across ethnic minorities in the UK using an endogenous switching model that controls for both initial conditions and retention status in the longitudinal sample. To our knowledge this is the first analysis of its kind for ethnic minorities in the UK, and one of the first to use these data in the context of poverty dynamics for the UK population. 8 This approach analyses the specific effects of the covariates for either movements in or out of poverty, by interacting the poverty status variable at the initial period (t 1) with all the variables in the model as observed in t 1, providing a relatively straight-forward way to test the existence of state dependence in short panels (Kanabar, forthcoming). However, a crucial assumption is that all relevant elements to determine poverty status at period t are determined at t 1 (Biewen, 9), discarding any possible effect of changes in the covariates between t 1 and t (although this reduces problems of endogeneity with the transitions of other variables, Jenkins,, 11). 3. Econometric model 3.1 Modeling transitions in and out of poverty We use an endogenous switching model developed by Cappellari and Jenkins (4) to analyse transitions in poverty status between two consecutive periods, t 1 and t, as a system of equations consisting of the pooled first-order transitions, the initial poverty status (at t 1) and an additional equation for attrition status (expressed in terms of sample retention) (at t-1), with free correlations between all the equations (i.e., without additional restrictions for their calculation). In the initial period, t 1, each individual i, i = 1,..., N, can be characterised by a latent poverty propensity, p i,t 1, expressed as: p i,t 1 = β x i,t 1 + µ i + δ i,t 1, (1) where x i,t 1 is a vector of explanatory variables at both individual and household level, and the error term u i,t 1 is the combination of an individual specific effect (µ i ) and an orthogonal white noise error (δ i,t 1 ): u i,t 1 = µ i + δ i,t 1. In this model, both µ i and δ i,t 1 are assumed to be normally distributed, so that u i,t 1 N(, 1), i.e., follows a standard normal distribution. Using p i,t 1 we can define P i,t 1, which takes a value of 1 if p i,t 1 > and otherwise. The transitions in and out of poverty can be observed only if the individual provides income data at t and t 1, so the model incorporates the propensity of retention (i.e., that the individual is observed in both periods), ri,t, as: ri,t = ψ w i,t 1 + η i + ξ i,t, () where w i,t 1 is a vector of explanatory variables at both individual and household level, ψ is the vector of parameters, and the error term ν i,t is the combination of an individual specific effect (η i ) and an 8 Kanabar (forthcoming) use this approach to study poverty transitions of pension-age individuals in the UK, rejecting the endogeneity of both initial conditions and attrition in the transitions in and out of poverty. Others authors have used the same model in different contexts, such as in Australia (Buddelmeyer and Verick, 8), Spain (Ayllón, 13), Luxembourg (Fusco and Islam, 1) and Nairobi (Faye et al., 11), as well as similar models for Sweden (Nilsson, 1) and Belgium(Van Kerm et al., 4). 6

8 orthogonal white noise error (ξ i,t ), so that ν i,t = η i + ξ i,t. η i and ξ i,t 1 are assumed to be normally distributed, which implies that ν i,t N(, 1) (follows a standard normal distribution). Just as in the case of P i,t 1, we can define R i,t = I[ri,t > ], which takes a value of 1 if r i,t > and otherwise. For all the individuals observed in both periods, the latent propensity of poverty at t, p i,t, is characterised as: p i,t = [ (P i,t 1 ) γ 1 + (1 P i,t 1 ) γ ] zi,t 1 + τ i + ζ i,t 1, (3) where γ 1 and γ are column vectors of parameters, z i,t 1 is a vector of explanatory variables at both individual and household level as observed at t 1, and the error term ɛ i,t is the combination of an individual specific effect τ i and an orthogonal white noise error ζ i,t 1 : ɛ i,t = τ i + ζ i,t. τ i and ζ i,t are assumed to be normally distributed, so that ɛ i,t follows a standard normal distribution (ɛ i,t N(, 1)). Using p i,t we can define P i,t, which takes a value of 1 if p i,t > and otherwise. Note that ɛ i,t and u i,t 1 are different given that equation 3 is conditional on poverty status in t 1. Moreover, unlike equation 1, equation 3 include two sets of parameters reflecting that the determinants of poverty persistence and poverty entry may be different. In this sense, γ 1 provides the relevant coefficients for poverty persistence (i.e., the probability of being poor in t, given the individual was poor in t 1), while γ for poverty entry (i.e., the probability of being poor in t, given that the individual was not poor in t 1). Thus, this equation will we referred as the transitions equation or the conditional poverty equation. The final component of the model is the definition of the correlations between the error terms, u i,t 1, ν i,t 1 and ɛ i,t 1, which we assume to be jointly distributed as a trivariate normal, so that the unobserved heterogeneity can be parameterised as: ρ 1 corr (u it 1, ν it ) = cov(µ i, η i ) ρ corr (u it 1, ɛ it ) = cov(µ i, τ i ) (4) ρ 3 corr (ν it, ɛ it ) = cov(η i, τ i ) These parameters summarise the association between the unobserved individual-specific factors in each equation. ρ 1 reflects the association between the unobserved individual effects of the initial poverty and the retention equation. A positive (negative) value of ρ 1 would imply that those individuals in poverty at t 1 were more (less) likely to stay in the sample at t, compared to the non-poor at t 1. Analogously, a positive (negative) value of ρ would imply that those individuals in poverty at t 1 were more (less) likely to be poor at t, compared to the non-poor at t 1. A positive (negative) value of ρ 3 would imply that those individuals providing income data in two consecutive periods are more likely to be poor at t, as compared with those who were non-poor at t 1. According to Cappellari and Jenkins (4) a sufficient condition for identification in this model is a set of exclusion restrictions, i.e., the inclusion of covariates associated with initial poverty status or retention, which have no effect on poverty transitions. This implies the inclusion of variables in x i,t 1 and w i,t 1, which can be excluded from z i,t 1 (further details in Section 5). Treating as endogenous the initial poverty status, our estimates take into account the possibility that individuals in poverty at the beginning of the observation period could be a non-random sample, more likely to remain poor over time, which could over-estimate the persistence estimates Heckman (1981a); Jenkins (11). In addition, by explicitly modeling attrition, we control for potential non-random differences potentially 7

9 correlated with poverty status which could as well bias our results. By setting no restrictions on the correlation parameters in equation 4, we are able to test if the transitions into and out of poverty are exogenous to either initial poverty or retention: if ρ 1 = ρ = then there would not be an initial conditions problem, if ρ = ρ 3 = then we could ignore the non-random attrition problem and if ρ 1 = ρ = ρ 3 = both processes would be exogenous and a direct estimation of equation 3 would provide unbiased estimates. 3. Probabilities of transition and poverty durations Using the estimates from equations 1 to 4 we can characterise the poverty transitions process, providing conditional estimates of the probabilities of poverty entry, poverty exit, poverty persistence, and so on. In this paper, we focus on the estimation of the probability of poverty persistence s it (i.e., the probability of an individual to be poor at t, given that she was poor at t 1) and poverty entry e it (i.e., the probability of an individual to be poor at t, given that she was not poor at t 1). Cappellari and Jenkins (4) show that the expressions for each of these probabilities is: s it P r (P i,t = 1 P i,t 1=1 ) = Φ (γ 1 z i,t 1, β x i,t 1, ρ ) Φ (β x i,t 1 ) (5) e it P r (P i,t = 1 P i,t 1= ) = Φ (γ z i,t 1, β x i,t 1, ρ ) Φ ( β x i,t 1 ) where Φ( ) and Φ ( ) correspond to the univariate and bivariate cumulative normal distributions, respectively. Additional probabilities may be estimated taking using the results of the retention process, as detailed in Cappellari and Jenkins (4), but we focus on s it and e it because one of the advantages of using a first order Markov model is the availability of simple expressions to calculate the mean and median duration of the spells (Boskin and Nold, 1975; Cappellari and Jenkins, 4). (6) If we assume an stationary state, the mean duration of poverty spell would be given by 1/(1 s it ), while its median duration by log(.5)/log(s it ). In a similar fashion, the expression for the mean duration of a non-poverty spell is 1/e it and the median duration by log(.5)/log(1 e it ), which will be used in Section 5 to analyse the duration of poverty for different types of individuals. 3.3 State dependence A key element of our analysis is testing the existence of true state dependence in poverty status among ethnic minorities. Cappellari and Jenkins (4) propose two measures based in the estimates of equations 5 and 6, which they denominate aggregate state dependence (ASD) and genuine state dependence GSD. ASD results from subtracting the average probability of not being poor among those who were not poor at t 1, from the average probability of being poor among those who were poor at t 1 : i {P ASD = i,t 1=1 } P r (P i,t = 1 P i,t 1 = 1) i {P i P i,t 1 =} P r (P i,t = 1 P i,t 1 = ) i,t 1 = 1 1 i P. (7) i,t 1 = The GSD measure is defined as the average difference in the probability of poverty persistence and poverty entry, at the individual level, so that same unobserved factors affecting poverty persistence and poverty entry for each individual are removed. The expression for the GSD measure is: 8

10 GSD = 1 [P r (P i,t = 1 P i,t 1 = 1) P r (P i,t = 1 P i,t 1 = )]. (8) N i It is convenient to note that all the estimates for poverty persistence, poverty entry, ASD and GSD can be calculated for the whole sample, and not only for the balanced panel (individuals observed at both t and t 1), which is one of the advantages of this model. 4. Data and definitions In this section we provide a short introduction to our data set and the estimation sample, and then we describe the main patterns of transitions in and out of poverty in our sample, highlighting the potential biases arising from non-random attrition. After that, we introduce the variables used to estimate the model described in the previous section and the estimation procedure. 4.1 Definitions and estimation sample We use data from Understanding Society, the largest longitudinal household survey in the UK, providing information for approximately 4, households in the UK (Buck and McFall, 11; Knies, 15). The first wave of Understanding Society was collected in 9-1, and households are visited every year, so that currently there are five waves available for analysis, covering the period 9 to 14. Understanding Society provides detailed income data for the calculation of disposable household income (total current income net of income tax and social security contributions), but previous studies have reported that in the first wave of this survey income received from benefits and pensions was significantly lower than in other waves, which can affect the estimation of poverty and inequality indices (Barnes et al., 15; Fisher, 15). Thus, we decided to employ information from waves, 3, 4 and 5. The main advantage of Understanding Society for our analysis is the inclusion of an ethnic minority boost sample that collects data for approximately 6, individuals from ethnic minority backgrounds (Berthoud et al., 9; McFall et al., 14). The concept of ethnic minority is associated to multiple factors such as national identity, race, self-identification, religion, migration history, among others (Afkhami, 1; Burton et al., 8). However, for comparability with previous research on this area, we use the standard UK census classification of ethnicity. In spite of the large sample of ethnic minority individuals in Understanding Society, the sample size for each group was inadequate to analyse all groups, so we focus on the main ethnic minority groups in Britain: Indian, Pakistani, Bangladeshi, black African and black Caribbean. 9 In addition, for comparative purposes we provide estimates for the whole sample, the white majority group (English, Welsh, Scottish, Northern Irish or British, excluding other white backgrounds). We include as well estimations for white majority individuals in England as a proxy to the sample used by Cappellari and Jenkins (4), so as to facilitate comparisons with that study. There is not a single measure of poverty in the UK (Platt, 7b), so we use an income poverty concept based on the official low-income statistics (Shale et al., 15). The income variable was calculated summing up the incomes of all household members, net of income taxes and National 9 Additional tests were performed including individuals with mixed backgrounds, but the results were qualitatively similar to those of the main groups. 9

11 Insurance contributions, in constant prices of 15. In order to compare households of different composition and sizes, the total household income is equivalised using the standard OECD scales (DWP, 15). We consider as poor all the individuals in households with net real equivalised income (before housing costs) below 6 percent of the median for all the UK. 1 In this paper, our main focus is on poverty experiences of working-age adults, which are closely associated to labour market outcomes. 11 Thus, the sample is restricted to individuals between 5 to 59 years old in wave, who were not in full-time education. After data cleaning and setting the constraints described before, our sample consists of 67,38 observations (6,3 individuals), mainly from white majority background (58,1 observations,,199 individuals) and ranging across ethnic minority groups from 1,384 observations (57 individuals) for the black Caribbean group, to,76 observations (1,99 individuals) for the Indian group (see Table 4 for sample sizes of all groups). 4. Observed poverty dynamics Figure 1a shows the evolution of the poverty rate for each ethnic group in our sample, as observed at the initial period of observation. As in previous studies, the poverty rates for all the ethnic minority groups are higher than those of the white majority for all periods, being the Pakistani group the most disadvantaged (with over 3 percent of its members in poverty), followed by the Bangladeshi and black African groups. The Indian and black Caribbean groups have lower poverty rates than other ethnic minorities, but still substantially higher than those of the white majority. In addition, the poverty rate of all ethnic minority groups seems to have increased over the period, particularly in 11 and 1, although those of the Bangladeshi and white majority groups changed in less than one percentage point between 1 and 13/14. 1 Figure 1: Differences in poverty levels and trends across ethnic groups /14* White majority Indian Pakistani Bangladeshi Black Caribbean Black African /14* White majority Indian Pakistani Bangladeshi Black Caribbean Black African (a) All the observations (b) Balanced panel Notes: Observations for 13/14 were pooled due to a small sample size. Source: authors calculations using Understanding Society waves, 3, 4 and 5. Figure 1b illustrate how ignoring attrition may bias the poverty rate estimates. In this figure, we can observe the same information of figure 1a, but excluding the observations without income data 1 The monthly net equivalised income used to define a household as poor was 97. in wave, in wave 3, in wave 4, and in wave 5 (constant prices of 15). 11 Although these elements are relevant to explain poverty experiences in other age groups, due to the relevance and particular characteristics of poverty in other age groups (e.g., children or pensioners) a single model for all the population would be inadequate. Thus, we focus on adult poverty, so that future research may extend this analysis for other segments of the population. 1 Due to a small size of 14, we pooled together the observations for 13 and 14. 1

12 in the next period (i.e., considering only the balanced panel). Differences between figures 1a and 1b indicate the existence of non-random attrition. In the case of the Indian and Bangladeshi groups, the poverty rates in figure 1a are higher than in figure 1b, suggesting that poor individuals are more likely to leave the sample in the next period than the rest of the population. In contrast, among the black Caribbean and Pakistani groups poor individuals are less likely to attrit. Furthermore, attrition changes as well the observed trends, as evidenced by the case of the black African and black Caribbean groups (particularly in this last case, the trend change from decreasing to clearly increasing). Table 1: Poverty transition rates with and without attrition, by ethnic group (Row percentages) Balanced panel With attrition Ethnic group Poor Not poor Poor Not poor Attrition at t at t at t at t (1) () (3) (4) (5) Whole Poor at t sample Not poor at t All White Poor at t majority Not poor at t All White Poor at t majority Not poor at t (in England) All Indian Poor at t Not poor at t All Pakistani Poor at t Not poor at t All Bangladeshi Poor at t Not poor at t All Black Poor at t Caribbean Not poor at t All Black Poor at t African Not poor at t All Notes: The whole sample consists of 67,38 observations (6,3 individuals). Table 4 presents the estimation sample by group. Source: Authors calculation using Understanding Society data from waves, 3, 4 and 5. The biases introduced by non-random attrition affect as well the rates of transitions in and out of poverty, as shown in tables 1, and 3. Table 1 contains the conditional poverty rates for the whole sample and each ethnic group in our analysis. Columns (1) and () present the distribution of the 11

13 sample by poverty status in the second period, t, given the poverty status observed in the initial period, t 1, for the balanced panel (row percentages). In contrast, columns (3) to (5) present the same measures, including in column (5) the percentage of observations without data at t (the attrition rate). 13 Table : Poverty transition rates with and without attrition, by year of the initial observation (Row percentages) Balanced panel With attrition Ethnic group Poor Not poor Poor Not poor Attrition at t at t at t at t (1) () (3) (4) (5) 11/1 1 Poor at t Not poor at t All /13 11 Poor at t Not poor at t All /14 1 Poor at t Not poor at t All /15 13/14 Poor at t Not poor at t All Notes: The whole sample consists of 67,38 observations (6,3 individuals) Source: Authors calculation using Understanding Society data from waves to 5. The first element to note in Table 1 is the variation across ethnic groups in the poverty persistence rates, i.e., the proportion of individuals poor at t 1 who remained poor at t (the rows labeled Poor at t 1 in columns 1 and 3). The Pakistani and black African groups have the higher persistence rates, either in the balanced panel or when the attrition is taken into account. In contrast, the Indian group has similar rates to those observed for the whole sample or the white majority, and the black Caribbean population has a mixed pattern: a higher persistence rate than the white majority in the balanced panel and a similar rate when attrition is included. The Bangladeshi group in this sample has an atypical behaviour, with persistence rates significantly lower than the white group. 14 Table 1 also displays the poverty entry rates for each ethnic minority group, i.e., the proportion of individuals non poor at t 1 who were identified as poor at t (rows labeled Non poor at t 1 in columns 1 and 3). In contrast with the poverty persistence rates, all ethnic minority groups exhibit 13 The elements in columns 1 and for each group are known as the poverty transition matrix, which contains which percentage of the individuals poor(non-poor) at t 1 were poor or non-poor at t. Columns 3 and 4 present an analogous version of these quantities, explicitly differentiating those individuals without income data at t. 14 In our exploratory analysis no evident reason was found for this result, but it may possible associated to the lower poverty rate of this group with respect to previous studies, suggesting that the sample of this group consists of individuals with better than average characteristics in this population. 1

14 higher poverty entry rates than the whole population or the white majority group, being the Pakistani and the Bangladeshi groups the ones with the highest rates. The differences between columns (1) and (3) in table 1 are due to the specific attrition rates among poor and non-poor individuals in each ethnic group, which is shown in column (5) of the same table. Table 1 shows that for the whole sample, as well as for the white majority groups, the attrition rate is higher among the poor than for the non-poor. However, in ethnic minorities the pattern is mixed, with the Indian and Bangladeshi groups showing higher attrition rates among the non-poor, while for the Pakistani and black Caribbean it is higher among the poor, and for the black African group the rates for poor and non-poor are similar. The variety of patterns in the attrition rates observed in table 1 let us draw a few preliminary conclusions. First, in most groups attrition seem to have a non-random component, likely to be associated with poverty status and, consequently, affecting the transitions into and out of poverty in most groups. Second, that attrition is more prevalent among ethnic minorities than for the whole population and, particularly than among the white majority. Figure : Observed poverty persistence and poverty entry rates, by year and ethnic group (balanced panel) /11 11/1 1/13 13/14-14/15 White majority Indian Pakistani Bangladeshi Black Caribbean Black African. 1/11 11/1 1/13 13/14-14/15 White majority Indian Pakistani Bangladeshi Black Caribbean Black African (a) Poverty persistence rates (b) Poverty entry rates Notes: *Observations for 13/14 were pooled due to a small sample size. Source: authors calculations using Understanding Society waves, 3, 4 and 5. The last results suggests that the characteristics of individuals in ethnic minority groups groups may make them more prone to leave the survey (as suggested by previous studies, such as Lynn et al., 1; Schneider, 13; Uhrig et al., 8; Watson and Wooden, 9). However, in order to test the validity of these conclusions, we need to take into account the differences associated to both observed or unobserved factors (as shown in Section 5). The transition rates, however, not only change among ethnic groups, but also over time. Tables and 3 show how the poverty persistence and poverty entry rates for the whole sample vary over time (table ) and by wave of the survey (table 3). Column (5) of table present the attrition rates by year of the initial observation. Table shows that the attrition rates have declined after 1, with a minimum in 1. Moreover, as in the result for the whole sample in table 1, these rates have systematically been larger among the poor than the non-poor. The fall in attrition rates is reflected in an increase in the poverty persistence rates and in the percentage of the non poor (at t 1) who remain non poor (at t). In contrast, the poverty entry and poverty exit 15 rates remained at similar 15 The individuals classified as poor at period t 1, who were identified as not poor at t. 13

15 levels throughout the period. Figure present the evolution in the observed poverty persistence and poverty entry rates by ethnic group. In a we can notice that from 1 and 13/14, the poverty persistence rates for all groups have declined substantially (in more than 8 percentage points), 16 with the exception of the Indian group (for which it increased in around 7 percentage points). Conversely, there is not a unique pattern for the poverty entry rates, which increased for the Bangladeshi and Indian groups, decreased for the black African and black Caribbean groups (although only marginally for the latter), and remained in similar levels for all years in the Pakistani and white majority groups. The Indian group is the only one for which both the poverty persistence and poverty entry rates have augmented, which may indicate that the experiences of poverty in this group have become more persistent in recent years, although this change has not been reflected yet in the overall poverty levels (see figure 1a). Table 3: Poverty transition rates with and without attrition, by survey wave (Row percentages) Balanced panel With attrition Ethnic group Poor Not poor Poor Not poor Attrition at t at t at t at t (1) () (3) (4) (5) Wave 3 Wave Poor at t Not poor at t All Wave 4 Wave 3 Poor at t Not poor at t All Wave 5 Wave 4 Poor at t Not poor at t All Notes: The whole sample consists of 67,38 observations (6,3 individuals) Source: Authors calculation using Understanding Society data from waves to 5. Finally, besides the differences by ethnic group and time, table 3 show that we can observe differences in poverty persistence, poverty entry and attrition rates by wave of the survey. Although we would not expect that these rates varied substantially across waves, the overall attrition rates have decreased from 16.4 percent from wave to wave 3, to 9.6 from wave 4 to wave 5, and a similar pattern is observed in poverty persistence rates (which have increased from 4.4 between waves and 3, to 47. between waves 4 and 5). In contrast, just as in the case of table, the poverty entry rates have remained similar (in terms of levels) in all waves. It is unclear if these patterns are due to changes in the strategies adopted in the survey field work to reduce attrition levels, changes in economic context or in the situation of poor households, which makes especially relevant to control for both the year and wave of the survey to discard these potential biases. 16 Note that the trend line for the white majority is parallel to that of the black Caribbean group. 14

16 4.3 Variables We use as covariates a basic set of variables related to the economic and social characteristics of the individual and the household, such as age, sex and educational qualifications, similar to those used in previous studies (Ayllón, 13; Cappellari and Jenkins, 4; Kanabar, forthcoming). Given that the main element in our definition of poverty is household income, most of our variables are defined at this level, particularly for the head of household (HoH). Understanding Society does not explicitly identify a HoH, but instead provides data about family relationships, income, intra-household decision making and ownership of assets to build a definition of HoH appropriate for the research purposes. In this analysis, we define as HoH the person in the household who owns or is responsible for the tenancy of the dwelling. Where more than one individual owns or rent the accommodation, we select as HoH the oldest of these individuals. 17 Table 4 presents the mean of all the variables in our estimation, either for the individual (age and sex), the HoH (age, sex, qualifications) or the household as a whole (type, demographic composition, number of members working). The model for the whole sample includes unobserved effects for each ethnic minority group. All models include controls for year of the initial observation, survey wave and region 18 In the case of region, due to ethnic minorities being concentrated (in terms of their residential status) in particular regions of the country, we consider only four categories: 1) London, ) South and East England, 3) Midlands, Wales and South West England, and 4) North East and North West England, Yorkshire, Scotland and North Ireland. In Appendix A.1 we provide the mean of these variables by conditional poverty status for each ethnic group. In all cases, the information reported for these variables correspond to the initial period (t 1), as mentioned in Section 3. The information in table 4 supports the evidence on how the socio-economic characteristics of ethnic minorities not only substantially differ from those of the rest of the UK population, but also across ethnic groups. Among the most pronounced dissimilarities, we find that the percentage of female HoH in the Indian, Pakistani and Bangladeshi groups is significantly lower than in the white majority, but larger among black African and black Caribbean households. 17 In a few households none of the members owned or rented the dwelling (less than 3 percent in our sample), so we select as HoH the individual who answered the household characteristics questionnaire in the survey, or otherwise the oldest household member. 18 A richer set of characteristics was initially contemplated, but due to insufficient variation within individuals transitioning into and out of poverty in our sample the set of variables was considerably limited. 15