Linguistic Distance and Economic Assimilation of Migrants in Australia

Transcription

1 Linguistic Distance and Economic Assimilation of Migrants in Australia Di Lan* & Stephen Whelan*,^ Abstract This paper investigates how the income of migrants to Australia changes in the years since migration. In particular, the paper considers how the post-migration experience is related to the migrant s language spoken at home. Using a measure of linguistic distance that identifies how close the migrant s language is to English, the analysis examines whether there is evidence for positive or negative income assimilation using the Australian Census Longitudinal dataset. Positive assimilation is observed for non-english speaking immigrants, where assimilation speed depends positively on the linguistic distance to English. Negative assimilation is exhibited by immigrants from English-speaking developed countries (ESDCs). Keywords: Immigrants; assimilation; linguistic distance. JEL codes: F22; J31; J61 * School of Economics H04 University of Sydney NSW 2004 AUSTRALIA. ^ Corresponding author: Stephen.whelan@sydney.edu.au School of Economics H04 University of Sydney NSW 2004 AUSTRALIA 1

2 1. Introduction Around one quarter of the Australian population is a migrant, and the assimilation of migrants into Australian society has been key to the success of the Australian economy over time. Importantly, the background of immigrants in Australia is relatively diverse. Following the Second World War, many immigrants arrived from countries such as Italy and Greece, while the United Kingdom has remained the largest source of immigrants over a prolonged period (Figure 1). More recently Australia has experienced high rates of immigration from countries such as India and China (ABS, 2010). For those immigrants with a non-english-speaking-background (NESB), the role of language in the assimilation process is likely to be particularly important because of its significance for skill transferability and in turn, employment outcomes. In many cases those migrants will have arrived with little or no English language skills and potentially face greater barriers in the labour market. The ability of individuals to assimilate for those of NESB is likely to be influenced by the language spoken and it is this issue that is the focus of the analysis in this paper. FIGURE 1: Immigrants to Australia country of birth, 1961 & 2006 Hongkong(SAR of China ) Lebanon Netherlands Malaysia Germany Greece South Africa Philippines Vietnam Italy India China New Zealand United Kingdom Source: Australian Bureau of Statistics (ABS), Immigrants leaving their home country relinquish networks and skills that are recognized in the home country, and generally have to increase the investment in transferable skills, education and language in the host country. Offsetting these costs are benefits that include better economic opportunities in the host country. In general, the attainment of proficiency in the language of the host country is likely to be an important aspect of the assimilation process. Economic assimilation has been characterised as the rate of wage convergence between immigrants and natives (Borjas, 2012). This process has been examined in a range of studies that have considered the behaviour and outcomes of immigrants across a range of host 2

3 countries (Chiswick, 1978; Baker and Benjamin 1997; Antecol et al., 2006). In general, there is evidence that the wage or income of immigrants approaches those of native workers over time, because new immigrants increase their human capital investment and approach the levels of native workers in a number of dimensions including education, language, non-transferred skills. Chiswick and Miller (2012) note that income assimilation may be positive or negative in nature. Positive assimilation occurs when the wages of immigrants increase as they remain longer in the host country. Positive assimilation has been observed for immigrants from relatively less developed countries as the value of their human capital in the host country increases (Chiswick and Miller, 2012). In contrast, it may be that some immigrants experience negative assimilation. For example, highly skilled migrants might find that despite earning a premium relative to natives initially, the wage outcomes falls relative to natives over time. Such a pattern may be observed for groups with high entry wage and similar cultural background such as immigrants from English speaking developed countries (ESDCs) (Chiswick and Miller, 2011). In this paper, we present evidence on how income among different groups of immigrants changes in the years following migration to Australia. That is, the focus of the analysis is not on the experience of immigrants relative to natives, rather the experience of groups of immigrants with particular language characteristics. Nonetheless, analysis will provide some insight into the broader assimilation question. The set of immigrants that experience positive income assimilation are likely to be those that start with lower income level than natives. Conversely, others who experience negative income assimilation are likely to have started at a high level of income relative to natives. Some limited evidence along these lines is presented in the empirical analysis by showing the average income of Australian native workers and immigrant workers. The contribution of this paper is to consider how the outcomes experienced by different immigrants groups is related to the concept of language distance. This concept captures the idea that some languages are more distant from English so that a native speaker of that language will find it more difficult to learn English and assimilate. Conversely, those individuals who speak a language closer to English will tend to find the assimilation process simpler. Assimilation is measured by examining how the income of immigrants with NESB changes over the time since migration compared to those immigrants with an English speaking background (ESB). The empirical analysis uses the Australian Census Longitudinal Data (ACLD) collected in 2006 and The use of this data offers a number of advantages over alternative data that has been analysed to consider the assimilation process for migrants in Australia, By providing an information on the respondent s language spoken at home, the ACLD provides an opportunity to examine how language distance, or the closeness of an immigrant s home language to English, is related to the economic assimilation process. The remainder of the paper is structured as follows. Section 2 reviews the literature on 3

4 assimilation with an emphasis on those studies that have considered the role of language in the assimilation process. Following this, in section 3 the theory and concept of linguistic distance and assimilation are described. Section 4 and 5 describe the data and methodology used respectively. In section 6, the results of the empirical analysis are set out. The empirical analysis finds evidence on the significant role of language in explaining income assimilation and different pattern of assimilation across language groups defined by linguistic distance. In section 7, conclusions are presented, along with a discussion of the limitations of the present study. 2. Empirical evidence on assimilation There is a long history of studies that have sought to shed light on the experience of newly arrived migrants in English speaking countries such as the United Stated (Chiswick, 1978; Borjas, 2012), the United Kingdom (Lemos, 2013) and Australia (Chiswick and Miller, 1985; MacDonald and Worswick, 1999; Breunig et al., 2013). Such studies often incorporate dichotomous measures of language background such as NESB or non-english Speaking Country (NESC) to capture English proficiency and its relationship with post migration assimilation. For example, Chiswick (1978) examined the experience of migrants to the United States and estimated earnings functions that incorporated various measures such as schooling, experience, years lived in the host countries and some other personal characters. The analysis in that paper finds that for males from NESCs, earnings increase at a decreasing rate with years in the host country. McDonald and Worswick (1999) examine the experience of migrants to Australia and identify a negative wage gap for NESB immigrants relative to natives over the period 1982 to They find this negative wage gap does not narrow as years living in Australia increases. A more recent study by Breunig et al. (2013) using Household Income and Labour Dynamics in Australia (HILDA) data for the period 2001 to 2009, find that the assimilation for NESB immigrants is slowest than other immigrants. It suggests that immigrants from a NESB take the longest time to approach to wage parity with native workers. In comparison to it, the study by Chiswick and Miller (2012) provides evidences on the positive effect of year since migration for NESC migrants, and, negative year since migration effect for ESC. Chiswick and Miller (2012) examines wage assimilation using a measure of linguistic distance. This approach argues that English level proficiency is not adequately captured a simple dichotomous measure of language background, but rather requires a more nuanced measure of how the immigrants language skills are related to an ability to communicate in English. In particular, language distance captures the closeness of a migrant s mother tongue to English. The analysis finds that in the United States, relative to immigrants from ESDCs, immigrants from non-english-speaking developed countries (NESDCs) exhibit positive assimilation. Moreover, the earnings increase associated with duration of residence in the United States is lower for those immigrants with a mother tongue closer to English than those with a mother tongue that is more distant from English. A key limitation of many previous studies is their reliance on cross-sectional data. Borjas 4

5 (1985) argues that cross-section analysis may capture changes over time in the unmeasured dimensions of immigrants skills. In turn, it cannot show accurately identify whether there is convergence in immigrant native earnings over time. Hum and Simpson (2000) also note the concerns around the use of cross-sectional data, and suggest that panel data is a more appropriate means by which to investigate wage growth and economic assimilation. A key limitation of use cross-sectional data is that it is not possible to control for unobserved individual heterogeneity across individuals. Moreover, it is not possible to control for selective out-migration and heterogeneity among cohorts in cross-section data (Cobb-Clark 2003, Breunig et al. 2013). 3. Theoretical framework The approach in this paper follows closely the analysis of the positive and negative assimilation described in Chiswick and Miller (2012). It is assumed that the level and growth of income are closely associated with the language barriers faced by migrants. For the purpose of this paper, the language barrier is assumed to depend on the distance of one s spoken language to English. Immigrants speaking a language similar to English in general find it easier to assimilate in the labor market. Positive and negative assimilation The economic assimilation of immigrants can occur in one of two ways (Chiswick and Miller, 2012; Chiswick and Miller, 2011). Positive assimilation is usually associated with immigrants moving from a low-waged country to a high-waged country and from countries with less similarity in institutions and culture. Due to the cultural, language barriers or the non-recognition of skills, immigrants often experience a lower wage rate compared to natives immediately after migration. While experiencing a lower initial wage, this is expected to be offset by higher economic rewards in the future when moving from a low-wage to high-wage country. For those immigrants who experience relatively low wages in the host country, over time disadvantages associated with language and skill transferability are likely to dissipate and immigrant wages tend to increase. That is, those immigrants experience positive assimilation as the wages or earnings approach those of native workers from a lower base. While many immigrants often experience positive assimilation, Chiswick and Miller (2011) and Chiswick and Miller (2012) highlight the potential of negative wage assimilation for immigrants moving from ESDCs to other English speaking countries. Those immigrants are often observed to have higher wages immediately following migration than similar native workers. Arguably, such workers may not have migrated otherwise considering the costs of migration. Intuitively, for many the motivation to immigrate is related to the positive labour market outcomes immigration presents over time, however, their wage decreases and approach to the native workers when staying longer in the host country as the economic rents disappear. Of course, it is important to emphasise that migration is a decision that reflects a range of 5

6 costs and benefits other than the immediate and long-term impact on income. Immigrants may move from a similarly developed countries for reasons associated with lifestyle and natural environment considerations, in addition to potential economic rewards. At the same time, it may be that the opportunity cost of abandoning the old network and lifestyle for this group is relative higher than immigrants from countries less similar to those of the new host country. For this set of migrants, the immediate economic benefits have to be high enough to motivate the immigration even though over time the positive income gap between themselves and natives tends to narrow. Linguistic distance The contribution of this paper is to use a measure of linguistic distance to help explain the pattern of economic assimilation by new immigrants. The starting point for such an analysis is a consideration of how languages relate to one another. Not all languages in the world develop independently from the indigenous culture. One ancient language might have created branches of new languages that have evolved over time. Thus some languages are closely linked to other languages in terms of their grammar and vocabulary. Linguistic distance is a concept describing the extent of differences across or the distance between languages (Chiswick and Miller, 2005). In this study, linguistic distance is used to describe the distance between English and other languages. It is important to emphasise that linguistic distance is not unique, and different measures of linguistic distance can be used to capture distance in terms of grammar, vocabulary, learning difficulty and ethno-linguistic map. The ethnologue project (ethno-linguistic map) provides the most complete relationship among languages in the world (Grime, 2005). It is based on the historical development of language and categorises existing languages into language families. While useful, such an approach has the disadvantage that it lacks a quantitative measurement of language distance. An alternative measure of language distance, namely Levenshtein distance, considers the vocabulary similarity between each language but ignores the similarity in grammar and context (Greenhill, 2011). This study uses the linguistic score (LS) reported by Hart-Gonzalez and Lindermann (1993) which is obtained from the average language ability after English speakers were trained in a foreign language in a certain period. 1 In particular, LS measures the difficulty of learning a different language by English native speakers. By linguistic symmetry, other speakers of other languages face the same level of difficulty when learning English. In contrast to other linguistics measures, the measure of LS is derived from the evaluation by individuals who are learning new languages, while it also aligns with the ethnologue project. For example, languages that have low LS such as Chinese, Korean and Japanese, they are independent and unrelated to the English, the origin of English and any relatives of English. 1 The linguistic scores are defined by Hart-Gonzalez and Lindermann, but to my knowledge, the bibliography is not available in any librarians, because it is published in Spanish. Reference provided by other studies: Hart-Gonzalez, L., and Stephanie L. (1993), Expected Achievement in Speaking Proficiency, Manuscript, U.S. Dept. State, Foreign Service Inst., School Language Studies. 6

7 The contribution of this paper is to identify the role of language in enabling assimilation. In particular, it uses the measure of linguistic distance defined by Chiswick and Miller (2005) to study the income assimilation of immigrants. Following the approach in that paper, a language is defined to be closer to English when it has a higher LS. That is, languages which are easier for native English speakers to learn are assigned a higher LS. Post-immigration language environment The post migration experience will be shaped by a range of characteristics, not just proficiency in English or the closeness of an individual s mother tongue to English. The language environment is literally defined by the language mostly spoken around an individual. For this reason, the analysis in this paper controls for the language environment the individual experiences. In particular, the empirical analysis incorporates a measured of the percentage of people speaking the same language in the respondent s local government area (LGA). Regardless of the language background, living in a community of ex-pats can accelerate the assimilation process to the extent that economic interaction can occur within an ethnic group to the exclusion of the wider community in the host nation of the immigrant. A priori it is not clear what might be the impact of living in an area in which a large number of individuals speaking the migrants home language. It is possible, for example that immigrants would have higher income when immersed in a familiar language and cultural environment. Immigrants with limited English ability would then find their skills more readily transferable, albeit in a limited local area. In this case, they would earn higher income than if immersed in a community in which there were fewer individuals who spoke the same language. In contrast, we could also expect that immigrants living with people speaking the same language would have less incentives to learn English. In this case, they may not perform as well in the broader labour market. The importance of social environment in the assimilation process is noted in Borjas (2015). In terms of national size effect, some specific nationality group size becoming larger cause less human capital acquisition. Immigrants from the same origin country are able to form a specific community and leads to less incentives to invest in their human capital stocks, such as improving English proficiency. 4. Data The data used in the empirical analysis in this paper is the Australian Census Longitudinal Dataset (ACLD) produced by the Australian Bureau of Statistics (ABS). 2 Recent Australia empirical studies on wage assimilation have tended to use the Household Income and Labour Dynamics in Australia (HILDA) dataset (Breunig et al., 2013; Cai and Liu, 2015). The use of the ACLD offers a number of advantages over the HILDA due to the limited information on the individuals first language or language spoken at home that is available in HILDA. Using the ACLD also provides a number of other advantages. Foremost, it will be possible to 2 The ACLD is accessible only via the ABS secure Data Laboratory. Users are permitted to analyse the data under secure conditions and all output is vetted by the ABS to ensure confidentiality and security is maintained. 7

8 replicate the analysis presented in Chiswick and Miller (2012) for United States. Moreover, unlike the Chiswick and Miller (2012), the ACLD is longitudinal in nature, allowing for estimation of panel models. The key constraint with the use of the ACLD relative to something like HILDA is the limited demographic information that is available. Nonetheless, given the large sample available and the detailed information on the individual s language spoken at home, the ACLD provides a valuable tool with which to analyse the post-migration experience. The Australian Census Longitudinal Dataset (ACLD) The ACLD data consists of linked data for the 2006 and 2011 Census. A random sample from the 2006 census of approximately 800,759 (82%) records were linked to the 2011 Census. The sample used in the empirical analysis in this study consists of working adults aged years in 2006, and therefore years in The choice of age means that the data consists of a set of immigrants who are likely to have completed their formal education and not have entered into retirement by The sample used in the regression analysis uses only those individuals in the census who report being born outside of Australia. That is, the primary focus of the analysis in this paper is on the outcomes experienced by foreign born individuals. The dependent variable is the respondent s total personal income. 3 That is, total income the individual reports receiving received from all source before any deductions and includes wages, dividends and allowances. Summary statistics for the full set of covariates used in the analysis are presented in Table 1. For the 2011 Census data, measured income is deflated to express earnings in real terms using the Consumer Price Index (ABS, 2017). The ACLD is a longitudinal dataset that is produced by linking the 2006 census and 2011 census using a matching algorithm (ABS, 2013). The Census collection process requires that identifying information is destroyed for confidentiality reason so that matching is achieved by using observable characteristics across respondents that appear in both Census. In this paper, in order to obtain a perfectly matched longitudinal dataset, records with unlinked and unstated information were dropped. One key constraint of the ACLD for the empirical analysis is the relatively limited range of socio-economic characteristics. Nonetheless, the information that is available in the ACLD provides an opportunity to construct proxies for labour market experience and the level of education or human capital achieved by the individual. A potential advantage of the use of the ACLD relates to information on location. The data contains detailed information on location including the individual s local government area (LGA) of residence. This makes it possible to consider how individual outcomes are associated with the ethnic mix of those in the individuals local area. LANP identifies the language one speaks at home. The question asked of respondents in the 3 Median income of the income range. 8

9 census is whether one speaks English only at home. If the respondent does not speak English at, then the language spoken at home is reported. The category 1201 English is respondents who speak only English at home. It is important to highlight that the language spoken at home might change over time. This reflects the longitudinal nature of the ACLD and is one of the advantages offered by the use of this data relative to those studies that used cross sectional data. TABLE 1 Description of variables and sample mean of all immigrants Variable Mean (std deviation) ACLD code Annual total income ($) (28636) INCP Experience (years) (10.74) AGEP HEAP 6 Hours worked per annum 1897 (684.66) HRSP Year since migration (14.28) YSM Education attainment (years) (2.56) HEAP Marriage status (Married =1) 0.71 (0.45) MSTP ESDC 0.40 (0.49) English 0.57 (0.49) Linguistic score (LS) 1.30 (1.17) Variable definition Language spoken at home * Language environment (% of same language speaker in LGA) English proficiency Country of birth Gender * State (NSW, VIC, QLD, WA, SA, NT, TSM,ACT, other) LANP LGAUR ENGP BPLP SEXP LGAUR Note: The * items are used for sample specification rather than as direct explanatory variables. HRSP is hours worked per week, and Hours worked per annum = HRSP *50; HEAP is transformed into education attainment years, in the range of 0 and 8 years to 18 years. For each person who speaks a language other than English at home (that is if answers for LANP is not English), language proficiency (ENGP) records their self-assessed proficiency in spoken English ranked from Very well to Not at all. It is important to emphasise limitations associated with responses to the question on proficiency in English. Such responses are inherently subjective, where one respondent may consider that Well' means that they can communicate well enough to do shopping while another may consider Well means being able to hold a social conversation. To alleviate some of the potential mismeasurement, the English proficiency is specified as a dichotomous variable which limits this subjective variation. Speak English Well contains those who identified their English proficiency as Very well and Well. Speak English Not Well contains those who scored their English proficiency as Not very well and not at all. The English speakers are the base level of this categorical variable. The detailed information on location available in the ACLD allows additional information on 9

10 the language environment the respondent is exposed to. In particular, an additional control capturing the percentage of individuals in the local government area (LGAUR) who speak the same language as the respondent is also included in the empirical analysis (Language environment). That is, it measures the percentage of same language speaker (depends on LANP) in the local government area where he/she lives. Linguistic score The full table of linguistic score (LS) are presented in Table 2. It includes the LS of 43 languages which are reported in Chiswick and Miller (2005), and additional languages assigned with LS following ethnologue approach (Grime, 2005). Each language in Table 2 has a LS in the range [1.00, 3.00] with increments of The lower score means that the language is more distant to English. It is also important to note that the LS is not cardinal in nature that LS=2.00 does not indicate twice as difficult to learn as LS=1.00. Any languages population with above 50 records in the sample are assigned with linguistic scores. For those languages with less than 50 speakers in the ACLD, the records are dropped from the sample used in the empirical analysis. 10

11 TABLE 2 Language code and score No. Language Extended group Census code 2011 Linguistic score L3: Close to English 2 Afrikaans 1403 * Norwegian Icelandic 1503 *, Romanian 3904 *, Swedish 1504 * Dutch Dutch,rtd; Frisian 1400,1401 *, Malay 6505 * Swahili 9211 * French 2101*,0006, Italian 2401 *, Portuguese 2302 * 2.50 L2: Medium to English 12 Danish 1501 * German Letzeburgish; Yiddish 1301 *,1302, Russian Byelorussian, Ukrainian, East 3400, 3401,3402 *, Slavic rtd; 15 Spanish Catalan 2301,2303 * 2.25 L1: Far to English 16 Amharic 9214 * Bulgarian 3502 * Cambodian Khmer; Mon 6300,6301, Czech Slovak 3601 *, Dari 4105 * Farsi(Persian) Kurdish, Balochi, Pashto ,4106 * Finnish Estonia; 1601,1602 * Hebrew 4204 * Hungarian 3301 * Indonesian Javanese, Balinese, 6504 *,6514, Mongolian Tibetan 7902 *, Polish 3602 * Serbo-Croatia n South Slavic nfd: Macedonian; Croatian; Serbian; Albanian; Slovene; Serbo Croatian/ Yugoslavian; 29 Tagalog Tagalog, Filipinos, Bikol, Ilonggo(Hiligaynon), Pampanga, Bisaya, ,3901,3505 * *,6512,6514,6515, 6517,6521, Thai Karen 6402 *, Turkish Turkmen, Uygur, Uzbek, 4301 *,4304,4305, Bengali 5201 * Burmese 6101 *, Greek 2201 *, Hindi Gujarati, Marathi, Urdu, 5202,5203 *,5205, 1.75 Romany 5212, Nepali 5206 *, Sinhalese 5211 *, Other Indo-Aryan~ 5200,5204,5207,5208, , Konkani, Punjabi, Sindhi; Assamese; Dhivehi; Kashmiri; Oriya; Fijian Hindustani; Dhivehi

12 TABLE 2 (continued) No. Language Extended group Census code 2011 Linguistic score 39 Dravidian~ Malayalam, Kannada, Tamil, Telugu, Tulu 40 Arabic Maltese, Assyrian Neo Aramaic 4202 *,4203,2501, Lao 6401 *, Mandarin Wu, Min-nan, 7104 *,7106, Vietnamese 6302 *, Cantonese Hakka 7101 *, Japanese 7201 *, Korean 7301 *, English NOTE: 1. *represents the code of the language, to distinct from other extended language. Rtd is abbreviation for related, which indicates languages under the classification. Nfd is abbreviation for not further defined. 2. Albanian, although being one member of the South Slavic republic, the language is not in the south Slavic language family. Instead, it is in the class of Indo-European language. Considering its geographically close to all other South Slavic speaking countries and Greek, hereby I put it to the same class as Greek and the South Slavic language, scored Other Indo-Aryan (~), including Konkani, Punjabi, Sindhi, Assamese, Dhivehi, Kashmiri, Oriya, Fijian Hindustani and Dhivehi, are assigned 1.75 as the same score as other scored languages, such as Hindi, Sinhalese, Bengali and Nepali; Dravidian (~) is a separate large language family in Asia, hence we can only assign it with Hindi, because most Dravidian speakers are in India; Maltese is assigned to the same score as Arabic. This paper extends languages with LS because of the considerable population in Australia. For languages without LS, those un-coded respondents are assigned LS using three methods. For languages that have a related language under the same classification of Ethnologue and with linguistic score, they are assigned to those related language s linguistic score (for example, Maltese to Arabic). For languages that do not have a scored related language (such as Dravidian) under the same classification of Ethnologue, they are assigned to a LS of most used languages in the same country or regions. Indigenous languages (such as Oceanic languages) are assigned with the score of other independent indigenous languages which is independent to English with a LS= In one of our specifications in the empirical analysis, immigrants are classified as ESDC originated or not, regardless of language they speak. In order capture the real effect of ESDC, English speakers in the NESDC are assigned a LS = 3.00, which is the highest score of languages other than English. This differs from the approach in Chiswick and Miller (2012) where they uses the average LS in the respondents country of origin. The approach taken in this paper reflects the likelihood that this English speaking group has more similarity with native workers in Australia than people speaking other languages, so they should be assigned with a linguistic score higher than other language speaker in same country of origin. The reason why English speakers might be observed among those from NESDCs includes: the origin country embodies multiple official languages including English; they move to Australia (or other ESCs at an early age and have adopted English as the language spoken at 12

13 home. 5. Methodology and analysis This study considers how the closeness of language to English would affects the economic assimilation of migrants. The empirical analysis uses linguistic score (LS) to proxy for the linguistic distance between the individual s language spoken at home and English, and the association between this and income assimilation since migration. A number of alternative specifications are used to capture this relationship. First we consider a linear relationship between linguistic distance to English and log annual income of foreign born workers. Workers are classified as those who report speaking English at home and others. Following this, workers are identified as ESDC immigrants and NESDC immigrants. As a simple combination of developed level and language, the ESDC specification respects the difference in institutions, pre-income level and presumably effect of the origin country. Among all those groups classified, the year since migration of immigrants is included in all specifications to identify patterns of income growth associated with YSM. Third, in order to investigate the possible effect of leaving immigrants, the subsample of 2006 in ACLD is used to compare with a cross-sectional analysis of 2006 census. Earning function with English variable Initially, the income of immigrations is modelled as a function of English background (for immigrants that report speaking English only at home), linguistic score (for immigrants that report speaking a language other than English at home) and years since migration in Australia. In all specifications, a range of other factors that are generally identified to be associated with economic assimilation and labour market outcomes more generally are included in the specification. These include measures of human capital such as education and experience. Initially, the earning function is specified as in (1): ln Y i = α ls LS i + α Eng English + δ 1 YSM_ENG + δ 2 YSM_NonENG +γysm_noneng LS i + β i X i + ε (1) where X i is a vector that include a set of observable characteristic including education, marriage, experience, working hours, location, English proficiency and language environment; LS i is the linguistic score of the language (which is not English) spoken at home; English is a dummy variable equal to one for immigrants who report speaking English, and zero otherwise. Specification (1) aims to identify the effect of speaking English. For immigrants speak other languages, they are assigned with LS following Table 2. The interaction terms YSM_ENG and YSM_NonENG, capture income changes with duration in Australia of immigrants who speak English only at home and those who do not report speaking English at home. The interactive term of YSM and linguistic score is also included in the specification. The coefficient on LS, α ls indicates how the closeness of the language to English impacts on the immigrants income. 13

14 An alternative specification allows for the effect of LS to be non-linear by using a series of dummy variables to form a spline function. If the respondent reports speaking a language other than English, the language would be classified into one of three groups reflecting its distance from English, measured as far, intermediate and close to English. The earning function is given by the following specification: ln Y i = α g,3 LS {1,2,3} + δ 1 YSM_ENG + γ g,3 LS {1,2,3} YSM_NonENG +β i X i + ε (2) where LS {1,2,3} is a vector of language group differed in their LS. Three groups {1, 2, 3} represents far to English (LS<=2.0), intermediate (LS between 2.00 and 2.25) and close to English (LS >=2.25)}. Each LS dummy variable is interacted with year since migration. This two specifications investigate the effect of language only rather than collaborated with any country effect when using ESDC. Earning function with ESDC variable Following this, the empirical specification incorporates measures of the country of origin in the form of indicator variables capturing if the if the individual migrated from an ESDC or NESDC. The earnings function estimated is similar to (1), except that immigrants from ESDC are distinguished from immigrants from other countries, irrespective of whether they speak English or not. The earning function will be similar to equation (1) and equation (2) is expressed as follow: ln Y i = α ls LS i + α E ESDC + δ 1 YSM_ESDC + δ 2 YSM_NESDC + γysm_nesdc LS i +β i X i + ε (3) ln Y i = α g,3 LS {1,2,3} + δ 1 YSM_ESDC + γ g,3 LS {1,2,3} YSM_NESDC +β i X i + ε (4) YSM_ESDC and YSM_NESDC are the year since migration of immigrants from ESDCs and NESDCs. ESDC is a dummy variable indicating that immigrants are from ESDCs if it is equal to one. Other immigrants are assigned with LS following Table 2. Those immigrants from NESDC but report speaking English only are assigned with a LS=3.00. Results from these two specifications would show which group experience positive or negative assimilation. In addition, comparing results of specification (1) & (2) and specification (3) & (4) will also indicate to what extent the income assimilation relies on the language distance and developed status of home countries. Possible out-immigration The longitudinal nature of ACLD allows the selective out-migration to be examined. This would be a concern that those people who arrive in Australia and experience negative assimilation leave. Recall that in section 4, those who left Australia after 2006 are not in the sampling frame because the sample is exclude those unlinked and unmatched record between 14

15 2006 and Hence, looking at the longitudinal dataset only will likely underestimate the negative assimilation of this group and overestimate the positive assimilation of the other group because those who remain are likely to be doing better than those who leave. Then the equation (1)-(4) are examined again using an alternative sample, that is the complete sample of The new sample does not drop those reported unlinked in The 2006 complete sample is larger than the matched samples, in fact, it is directly from 2006 census. Recall that the matched sample is based on an algorithm that matches respondents in the 2006 sample to their 2011 records. Immigrants who arrived before 2006 but leave Australia by the 2011 Census (not matched), were still in the 2006 complete sample. By comparing the key contributes of immigrants income, it can provides some implication on the scale of out-immigration issue. FIGURE 2: Potential income changes of positive and negative assimilation model Stayers in 2006, stayer in 2011 Income Include leavers after 2006 Stayers in 2006, stayer in 2011 Include leavers after 2006 Year since migration Complete 2006 subsample Linked 2006 subsample NOTES: The positive and negative assimilation model depicted in solid line follows Chiswick and Miller (2012). The sample of migrants in the ACLD is self-selected. In particular, the matched sample consists of migrants who have chosen to remain in Australia. Those who have left the country will be unmatched across 2006 and For those who are unmatched because they have left the country, it is likely that they have done so because the labour market rewards associated with migration proved to be less than anticipated. That is, we may expect that those who remain are doing better than those who are observed to leave. Hence, when estimating the effect of years since migration for the unmatched sample, we expect it is more likely to experience negative assimilation over time, or lower rates of positive assimilation. This is represented in Figure 2. A key benefit of using longitudinal datasets such as the ACLD is that it allows for these patterns to be explored. 4 In ACLD, 5 percent of 2006 census (979,661) were linked to 2011 census, and 82 percent of them are successfully linked while 18 percent of them are not. Those were not linked to 2011 sample are reported as unlinked, where Unlinked only appears in the column of 2011 in ACLD. 15

16 6. Empirical Results In this section, results from the empirical analysis are discussed. Initially, the empirical results from estimating equation (1) to (4) are discussed. Following this, some implications of the analysis for assimilation of immigrants relative to natives are discussed. Finally, results from a series of robustness checks are presented. Specification 1: English, linguistic score and income Table 3 reports the empirical results from estimating equation (1) and (2). This specification isolates the effect of being an English speaking immigrant. In this specification non-english speaking immigrants are assigned a linguistic score (LS) based on Table 3. A priori, we might expect those immigrants who report being able to speak English to better assimilate into the Australian labour market and to experience relatively higher income and higher income growth over time. The results reported in column (1) and columns (3)-(6) follow specification (1), while the results reported in column (2) use specification (2). The results reported in column (1) suggests that immigrants who report speaking English at home would have income that is 59 percent higher relative to those immigrants who report speaking languages other than English (coefficient α Eng= 0.59, t-value= 7.21), ceteris paribus. There is also evidence that the income of immigrants grows as the time spent in Australia since migration increases. For immigrants speaking a language other than English, the estimates point to an annual increase of 1.8 for each year spent in Australia (coefficient δ2 = 0.018, t-value = 18.24). For immigrants who report speaking English only at home, their income increases with years in Australia significantly though the magnitude is small (coefficient δ2= , t-value= 1.82). For non-english speakers, their income grows as the spoken language gets closer to English. Moreover, the coefficient on interaction terms between linguistic score and year since migration (coefficient γ= , t-value= ) suggests that over time, speaking a language close to English is associated with smaller income growth. In column (2) of Table 3, the results using equation (2) are reported. Recall that in this specification, the linguistic score of non-english speakers is captured by three dummy variables: LS3 (close to English), LS2 (intermediate) and LS1 (far from English). The LS estimators represent the contribution to income relative to English speaking immigrants. The coefficient on LS3 (coefficient α g=3= -0.09, t-value= -1.15) is higher than that for LS2 and LS1 respectively (coefficient α g=2= -0.24, α g=1= -0.28). This implies a similar pattern as suggested by the results reported in column (1), namely that all non-english speaking immigrants in general have less income than English speaking immigrants, but immigrants who speaks a language closer to English (LS3) would have higher income than immigrants speak other. The three dummy variables of LS are also interacted with the variable years since migration (YSM). The coefficient on the interaction between YSM and LS3 is smaller in magnitude than for the other two groups, suggesting that the income growth associated with YSM experienced by immigrants who have a language that is closer to English is lower 16

17 than that for immigrants with languages that are more distant from English. In columns (3) to (6) of Table 3, a series of specifications are presented in which the full sample is split according to their English language score. Column (3) pertains to only English speakers. The estimator of year since migration (YSM_ENG) is our main concern in this specification. Immigrants who report speaking English only at home, experience a slow but positive income growth of approximately 0.05 percent per annum in the years since migration (coefficient δ1= , t-value= 2.51). The small positive income growth associated with YSM provides no evidence for the negative assimilation model for English speaking immigrants. Note, however, that the income growth for English speaking immigrants is smaller than all other language groups presented in column (4)-(6), ceteris paribus. In addition, the coefficients on Education attainment, Experience and Log Hours are in general larger in size than they are reported in all other columns. This result is consistent with a finding that that English speaking immigrants face better skill-transferability, so that their human capital stocks and work effort generate higher income return than immigrants speaking other languages. In columns (4)-(6), the pattern of income growth associated with YSM experienced by immigrants with a language that is close, intermediate or far from English is presented. While all immigrants experience positive income growth in the years since migration, the growth is largest for those immigrants when the language is more distant to English, equal to 0.8 percent per annum for LS1 (far; coefficient δ2= 0.008, t-value= 19.95), 0.6 percent for LS2 (intermediate; coefficient δ2= 0.006, t-value= 11.12) and 0.1 percent for LS3 (close; coefficient δ2= 0.001, t-value= 2.32). Two additional language-related variables are included in each column of Table 3. One is language environment (percentage of same language speakers) and the second is English proficiency ( Speak English very well/ well ; Speak English not well/ no at all ). In columns (4)-(6), the coefficient on the percentage of individuals in the local area who speak the same language that the respondent reports speaking at home is negative for non-english speaking immigrants. That is, living in an area with a large proportion of individuals who speak the same language does not appear to enhance income growth. The possible explanation for this pattern is discussed in the section 3, namely, that such individuals have less incentives to learn English and attain higher assimilation benefits. Speaking English very well/well and speaking English not well/not at all are two categories of the categorical variable, which is ENGP reported in Table 2. In column (1) and (2) of Table 3, the base group of ENGP is those speaking English only. Then estimators of speak English very well/well and speak English not well/ not at all tell the income attributes relative to immigrants speaking English only. For example, in column (1), the estimator on speak English very well/well indicates a 9.1 percent lower income of those immigrants than English speaking only immigrants (coefficient β= -0.09, t-value= -1.16), holding other factors constant. Conversely, those immigrants who speak English not well/not at all have even lower income around 19.4 percent than English speaking only immigrants (coefficient β= -0.19, t-value= -2.45). In column (4)-(6), each sample respectively pertains to LS1 to LS3 17

18 does not include any English-speaking-only immigrants, so the base group of categorical variable ENGP is speak English well/very well. Then the estimators of ENGP in these three columns indicate the income attributes relative to immigrants speak English very well/well. In column (4)-(6), the sample does not contain any English-speaking-only immigrants, so the base group of categorical variable ENGP is speak English well/very well. Then the estimators of ENGP in these three columns indicate the income difference between immigrants with good English speaking ability and those with poor English speaking ability. TABLE 3 Result of pooled OLS ACLD English Variables (1) (2) (3) (4) (5) (6) Pct. same language speakers (-14.62) (-14.88) Speak English very well/ well (-1.16) (-1.42) Speak English not well/ not at all (-2.45) (-2.90) ESDC (11.18) (11.12) Linguistic score 0.18 (15.27) English 0.59 (7.21) YSM_ENG (1.82) (1.71) YSM_nonENG 0.02 (18.24) YSM_nonENG * LS (-12.78) New South Wales (5.49) (5.43) LS3 (close) (-1.15) LS2 (intermediate) (-3.14) LS1 (far) (-3.61) LS3*YSM (5.14) LS2*YSM (14.27) LS1*YSM (24.44) Constant (45.00) (108.84) (-13.00) 0.05 (10.60) (2.53) 0.10 (5.14) (-4.25) (-15.00) 0.10 (4.35) (19.95) 0.04 (0.49) 0.03 (0.35) (-3.17) 0.16 (3.11) (11.12) 0.01 (1.08) (-2.06) (-3.02) (2.66) (2.32) 0.20 (3.25) (81.06) (43.44) (33.06) (28.59) N R Notes: heteroscedasticity-consistent t value reported in the parentheses; Tasmania is removed from the variable list to avoid collinearity. The unreported control variables in each column include education attainment, experience, marriage status and log hours worked and variables of state (Victoria, South Australia, West Australia, Capital Territory, Other State; Tasmania excluded for collinearity). 18

19 Recall that a key benefit of using the ACLD is that it allows for panel data methods to be used. Table 4 presents the results for panel regression under the specification 1. The results from specifications that control for unobserved heterogeneity, namely fixed effects models, are reported in Table 4. Random effect model has also been examined, however since the result is very similar to the pooled OLS and Hausman Test supports the fixed effect, we only present fixed effect results. In the fixed effect model, the coefficients on LS (coefficient α ls= 0.003, t-value= 0.12) and English (coefficient α ls= 0.20, t-value= 2.28) are significantly smaller than when a pooled OLS approach is used (Table 3). In comparison, the coefficients on years since migration in column (3)-(6) is all larger than under the pooled OLS model. As expected the estimated magnitude of the coefficients on the covariates included in the model are substantially larger when using a fixed effect specification. Recall that the fixed effects specification removes the unobserved heterogeneity which is correlated with the included covariates. Hence, the effect of the included covariates becomes substantially larger in the fixed effects specification, even among variables like education and experience that have not been reported. For example, when controlling unobserved fixed effect, the income of immigrants in LS1 (far) grows at the rate of 2.8 percent each year, the income of immigrants in LS3 (close) grows at the rate of 2.3 percent each year, and the income of immigrants in LS2 (intermediate) grows at the rate of 1.7 percent each year. The income of English speaking immigrants still grows at the lowest rate at 1 percent each year (coefficient α ls= 0.01, t-value= 3.08). 19