Why Do Arabs Earn Less than Jews in Israel?

Why Do Arabs Earn Less than Jews in Israel? 1 Introduction Israel is a multicultural, multiethnic society. Its population brings together Western and Eastern Jews, foreign- and locally-born citizens, and Arabs from a range of Muslim, Christian, and Druze religious and ethnic backgrounds. Arab Israeli citizens constitute about 20% of the total population; 1 yet despite extensive studies of ethnic wage disparities in Israel (e.g., Neuman and Silber (1996), Neuman and Oaxaca (1998), and Neuman and Oaxaca (2004b)), very little attention has been paid speci cally to the characteristics of this group as workforce participants. The objective of this paper is to measure and document the evolution of wage gaps between Arabs and Jews in the Israeli labor market in the years 1990 2003, aiming to characterize and evaluate the di erent mechanisms according to which these gaps may be said to have developed. The existence of an observable wage gap in itself, though, does not count as su cient proof that a certain labor market is marked by discrimination. In order to separate out the e ects of discrimination from those of potentially unrelated factors, this study used a modi ed form 0 I owe special thanks to Dan O Flaherty, Janet Currie, and Nachum Sicherman for their continuous support and invaluable comments. The helpful comments of an anonymous referee are also gratefully acknowledged. Michael Beenstock, Joseph Zeira, Lena Edlund, and Till von Wachter o ered helpful comments and discussions. Thanks must also go to seminar participants at Columbia University, and to the Falk Institute for Economic Research in Israel and the Israel Social Sciences Data Center, at the Hebrew University of Jerusalem, for providing this study s data. 1 Arabs mentioned here are citizens and residents of Israel. They live and work in Israel, have Israeli citizenship, and share the same national institutions with the Jewish citizens. Palestinians, living in the West Bank and Gaza Strip, are not the subject of this paper. 1

of the Oaxaca-Blinder decomposition in order to disaggregate observed wage gaps into three components: those owing to human capital di erences, to occupational segregation, and to discrimination (or unexplained gap). As well as recording the Jewish-Arab income gap in gross terms, this study also set out to analyze patterns of change between the di erent wage gap components over the years of the study s data. It was found that the overall shape or trend of the changes between the study s explanatory categories for the Jewish-Arab wage di erential were robust to di erent model speci cations and underlying assumptions, even when speci c levels of income gap components varied in more unexpected ways. The Jewish-Arab male hourly wage gap hovered at around 45% (of Arab hourly wage) in the years 1990 1994, going on to peak at 77% in 1999. Since then, the hourly wage gap has started to decrease, falling to a level of 56% by the end of 2003. The unexplained component of the gap (regularly interpreted as discriminatory) accounted only for 5% 10%, or less in other instances, of the overall wage gap in the years 1990 1991. However, since 1992, this component has accounted for an increasing portion (about 25% 38%) of the overall wage gap. Occupational segregation explained a portion of 35% 80% of the overall wage gap over the entire period. It is important to seek to understand these uctuations in the Jewish-Arab income gap against the background of the comprehensive changes undergone by the Israeli economy as a whole between the years 1990 2003. 1990 saw the arrival in the country of some 200,000 immigrants, with 176,000 arriving in 1991. This massive in ow of immigrants continued at a 2

yearly rate of about 77,000 until 1995; with numbers subsequently starting to fall to a point where only 23,000 immigrants entered the country in 2003. 2 Besides being boosted by this new pool of available and sometimes skilled labor, the Israeli economy was also fed over the 90s by a large in ux of temporary foreign workers. Those workers were imported by the state as a solution to the immigration-driven demand for construction workers and the shortage of Palestinian workers due to closures i.e., restrictions on Palestinian workers freedom of movement imposed by Israel. By 1995 there were 92,000 non-israeli nationals working in the country (or about 4.7% of the total employed workforce). This number only ceased rising subsequently in 2002 at 232,000 counted foreign employees (or 10.2% of the total employed workforce). The wider socio-political context of changes in Israel s labor force composition over the study period is of the Jewish-Arab peace process, set in train with the 1993 signing of the Oslo accords but then dealt a fatal blow by the assassination of the Prime Minister Yitzhak Rabin in 1995 and the onset of second Intifada in September 2000. Within pre-1967 Israel, the so-called 2000-events (referring to the killing of 13 Israeli-Arab rioters by Israeli soldiers) also in uenced the climate in which employment (hiring, ring, and training decisions) were made with respect to Jews and Arabs. It has to be supposed that events in this socio-political register will have had a direct e ect on labor market outcomes, especially as these pertain to Jewish-Arab (wage) disparities. Given the inseparability of social, political, and economic factors, it is di cult to come 2 Israel s total population (excluding the Gaza Strip and West Bank) was estimated at around 5 million in the early 1990s and has steadily increased since then. In 2003 the estimated population was 6.7 million. See the Statistical Abstract of Israel, Central Bureau of Statistics (2005). 3

to conclusive explanations as to the changing pattern of income inequalities between Arabs and Jews over the study period. While not claiming to be incontrovertible, this paper o ers a series of interpretations of the phenomenon in terms of the changing demographic composition of the workforce, changing perceptions of ethnicity on the part of both Arabs and Jews, and shifting skills shortages in the broader Israeli economy and speci cally in the Arab population compared to the Jewish (Arabs human capital de cit). It is hoped that the paper s terms and explanations will stimulate further research examining the study s issues from di erent perspectives and in greater depth. The paper is organized as follows: the next section describes the methodology used to measure wage gaps and their decompositions. Section 3 features a detailed description of the study data and de nes the paper s terms as regards its explanatory categories for the causes of income gaps. Summary statistics for the relevant variables are found in section 3.1. The main results of the study are in turn reported in section 4, after which section 5 spells out the paper s analysis of the changing shares in the Jewish-Arab income gap attributable to the decomposed causal mechanisms. Policy implications are discussed in this section as well. 2 Methodology In an attempt to quantify the coe cient of (wage) discrimination [Becker (1957)], Oaxaca (1973) and Blinder (1973) proposed a simple, yet reliable, estimator based on a straightfor- 4

ward OLS estimation. This paper adopts the Oaxaca-Blinder decomposition method in the context of an analysis of wage gaps between Arab and Jewish workers, which are broken down into rst, a component representing more neutral di erentials in human capital, and second into a component normatively representing the e ect of discrimination. My approach in e ecting this decomposition broadly follows the more general treatment of wage decomposition proposed by Neumark (1988) and Oaxaca and Ransom (1988, 1994). Simply put, Oaxaca-Blinder decomposition entails running separate OLS wage regressions, for each of the two groups under consideration, and comparing the means and estimated coe cients of the variables from those regressions. A di erence in the average productivity-related variables, weighted by the estimated coe cients of the nondiscriminatory wage standard, is attributed to the human capital (or explained) portion of the overall wage gap. Any remaining (unexplained) wage gap, measured by the di erence in the estimated coe cients, evaluated at some level of average characteristics, may then be referred to discrimination. We can represent this mathematically as follows: Let the wage equation for any individual i in the ethnic group j (j =Arab, Jew) be ln W ij = X 0 ij j + " ij (1) where W ij is the hourly wage, X ij is a vector of worker characteristics, and " ij a zero-mean, constant-variance error term. Then, the wage equations, estimated by OLS at the mean point, will be 5

ln W j = X 0 j ^ j ; for j = A; J (2) where A and J subscripts stand for Arabs and Jews and upper bars signify averages of the di erent variables. The regressors vector X ij in this paper includes variables for years of schooling, experience, experience squared, a marital status dummy, a full-time employment dummy, a large city dummy, one- and two-digits occupational dummies, and one- and twodigits industrial dummies. 3 Income survey for the year 1990 includes a categorial, rather than a continuous, schooling variable. Consequently, analyzing the data for this year, I use a set of schooling dummy variables, and use age and age-squared in lieu of experience and experience squared. Nonetheless, the results of the decomposition and wage equations remain meaningful and robust to comparison, so far as wage di erentials are concerned. Writing the wage equation in 2 for each group and di erencing those, with mild arithmetic transformation, 4 yields the following wage gap decomposition: ln W J ln W A = ln (1 + G) = X J XA 0 ^ {z } Q + X J ^J 0 ^ + X A ^ 0 ^ A {z } D (3) where ^ is the estimate of the nondiscriminatory wage coe cients, and G is the gross (geometric) wage gap. The rst term in equation 3, Q, represents the human capital component of the overall wage gap; D, the sum of the second and third terms, is the discriminatory 3 The Income Surveys do not provide a direct measure of labor market experience. Consequently, I use potential experience, de ned as: Experience = Age Schooling 5: 4 Namely adding and subtracting the term X J XA 0 ^ : 6

component. The decomposition method performed in equation 3 is the general Oaxaca decomposition, Oaxaca and Ransom (1994). If we assume that ^ = ^ J or ^ = ^ A ; then the general Oaxaca decomposition reduces to the classical Oaxaca-Blinder decomposition. I carry out these analyses under two di erent assumptions about the nondiscriminatory wage standard. First, as in the classical Oaxaca-Blinder decomposition, I adopt the estimated wage structure of the dominant group as the nondiscriminatory standard, i.e., ^ = ^ J : That is, in the absence of labor market discrimination, it is taken that the current Jewish wage structure would apply to both Jews and Arabs. Alternatively, as in the general Oaxaca decomposition, I assume that ^ is equal to the characteristics-weighted wage coe cients, as this is shown to be equal to the estimated wage coe cients from a simple pooled regression that includes both Arabs and Jews (see Oaxaca and Ransom (1994)). For the dataset in question, the average income varies widely across occupations (as well as across ethnic groups). Therefore, even in the absence of unexplained wage gaps within occupations, wage di erences could still exist according to di erent distributions of Arab and Jewish workers across employment sectors. While controlling for occupational and industrial a liation in the wage regressions would eliminate inter-occupational wage gaps, it would also have the e ect of underestimating the discriminatory component of the overall wage gap, to the extent that occupational segregation, at least in part, derives from discrimination. This could be through di erent barriers to entry o ered to representatives of the two groups. To show this, let C A i C J i be the proportion of Arabs (Jews) employed in occupation i; 5 and W A i 5 Note that this proportion is equal to the average of the corresponding dummy variable in the wage equation. 7

Wi J the mean (log) hourly-wage of an Arab (Jewish) worker in occupation i (i = 1; 2; :::; I). It then follows that: W j = P i Cj i W j i for j = A; J and ln(g + 1) = IX i=1 C J i W J i Ci A Wi A (4) Therefore, even if we assume that there are no wage di erences between workers within each occupation (i.e., W J i = W A i = W i 8i), wage gaps may still arise if the two populations have di erent occupational distributions (i.e., C J i 6= C A i for some i), since, according to equation 4, the wage gap will be equal to P i C J i C A i Wi which will not be zero other than in the cases Ci A = Ci J for all i (i.e., identical occupational distribution), or W i = W for all i (i.e., equivalent average wages in all occupations for members of Arab and Jewish groups). The above illustration indicates the possibility that including occupational dummies in the wage regressions may result in underestimating the discriminatory component of the overall wage gap, in this sense that the added variables would disguise labor market discrimination as a human capital component. In seeking to account for i.e. decompose the wage gap, this paper assesses the above possibility in two ways. First, I compare the estimated coe cient of the Jewish dummy in a pooled wage regression with and without occupational dummies. Second, regression analysis includes occupational (and industrial) dummy variables in Oaxaca decompositions, comparing these results with those obtained without the occupational dummies. Oaxaca decomposition does not account for group di erences in group members occupational distribution. Di erent methods were suggested to measure the occupational segre- 8

gation component of the wage gap. For example, Brown et al. (1980) and Miller (1987) use a multinomial logit model to estimate the distribution of one group across occupations, and compare this with the observed distribution of the other group. Neuman and Silber (1996), alternatively, estimate the occupational segregation component of a wage gap by comparing each group s share in a certain occupation with the share of that occupation in the total employed labor force, and sum the di erences over all occupations. This paper introduces a new, yet closely related, method to measure the component of the wage gap attributable to occupational (and industrial) segregation. I include a set of occupational dummies in the wage equations, and modify the Oaxaca decomposition by dissociating the part explained by those dummies from the human capital component. Let the estimated wage equation be ln W j = X 0 j ^ j + C 0 j^ j ; where C j is a vector of average occupational dummies for the group j = A; J: De ning Z 0 j := [X 0 j; C 0 j] and ^ := [^ 0 ; ^ 0 ] 0 I arrive at the following decomposition: ln (1 + G) = X 0 h J XA ^ + Z 0 J ^J ^ + {z } i ZA ^ 0 ^A {z } Q D + C J C A 0 ^ {z } S (5) Equation 5 accommodates the e ect that di erent occupational distributions have on wage gaps. The rst two terms in this equation, Q and D, are the familiar human capital and discrimination components. The last term, S, representing di erences in the occupational 9

distribution weighted by a nondiscriminatory norm, takes its place as the occupational segregation component. 6 Although estimating the individual contributions of sets of dummy variables to the unexplained component of the wage gap may produce arbitrary results, since those depend on the dropped category, it is true that the overall decomposition and estimated separate contributions of dummy variables to the explained component are consistent and invariant to any choice of the dropped category, Oaxaca and Ransom (1999). This fact makes possible the isolation of the occupational distribution e ect from the human capital component. Note that, in the absence of occupational segregation (i.e., C J = C A ), S = 0 and that the only way in which occupations may a ect the decomposition is by adding the term C 0 (^ J ^ A ) (the within-occupation di erences) to the discriminatory segment D. Also, if all occupations share the same wage structure for Arabs and Jews (i.e., ^ A = ^ J ), then the discriminatory component due to within-occupation gaps disappears, although the segregation component, manifesting di erent occupational distributions, will remain. Lastly, if all occupations pay the same wage (as the dropped category, i.e., ^ = 0), then equation 5 reduces exactly to equation 3, in which case the estimates of the human capital component and the discriminatory component are not a ected by group di erences in occupational distribution. In the appendix, section 6.2, I refer to the bias, in the wage gap decomposition, resulting from possible self-selection into employment. It is possible for self-selection to have only 6 Industrial dummies, when included in the wage equations, are treated in the same way as occupational dummies. C will be a vector of not only occupational but also industrial dummies, and its respective vector of coe cients. The S component will represent both the occupational segregation and industrial segregation components. Added together, S will be called the labor market segregation measure throughout the paper, for simplicity. 10

a mild e ect, if any, on the human capital or occupational segregation components of the wage gap, since productivity-related variables are not a ected by the measurement method. On the other hand, the unexplained component may be greatly a ected by the selectivity correction, since this whole component relies on our estimators of wage regressions, which are liable to change markedly on account of this correction. However, correcting for selectivity, in all its variations and e ects, lies beyond the scope of this paper. While selection into employment may result in bias in the decomposition estimates, it is at any rate not the only source of disturbance. Selection into a speci c occupation, for those who are already in the labor market, represents another dimension of the problem. Although the occupational segregation component is estimated consistently in this paper, without knowing the e ect of occupational self-selection e ect we cannot identify that part of occupational segregation which represents labor market discrimination (or barriers to entry in certain occupations for members of the Arab ethnic group). Two important points are worth emphasizing before we leave this section. First, the paper refers to the unexplained component of the wage gap as discriminatory. However, since the choice of explanatory variables can greatly a ect the results of decomposition, the unintentional exclusion of certain relevant variables from the wage regressions may bias ndings related to the unexplained (or discriminatory) component. 7 While it seems beyond 7 The possibility of decomposition results being heavily in uenced by the choice of regressors to be included was originally raised by Oaxaca: It is clear that the magnitude of the estimated e ects of discrimination crucially depends upon the choice of control variables for the wage regressions. A researcher s choice of control variables implicitly reveals his or her attitude toward what constitutes discrimination in the labor market. Oaxaca (1973), p. 699. 11

reasonable doubt that some fraction of the unexplained component derives from labor market discrimination, it is equally the case that some other fraction will be owing to other factors. Such a fraction may merely signify the e ect of uncontrolled-for variables. Therefore, my terminology refers to the D component as the unexplained component of the wage gap, rather than the discriminatory component. In this sense, this component may be understood as an upper bound for labor market discrimination. Secondly, it has been noted that occupational segregation can be the result of labor market discrimination. 8 Now, this labor market segregation may re ect barriers to entry into well-paying jobs, but can only exist more neutrally as a manifestation of di erent preferences. Some people may prefer to work in low-paying occupations. In this paper I do not further decompose the occupational segregation component into self-selection on the one hand and discrimination on the other. Therefore, the labor market segregation component as discussed in this paper should be viewed as a compound of the two e ects. The issues of occupational selection, and its e ect on measuring discrimination in the terms of economics, represent a fertile area to be addressed in future research. 8 As noted by Neuman and Silber (1996) (p.651, n.3), occupational segregation represents another dimension of labor market discrimination. Segregation and barriers to entrance based solely on ethnic a liation, other things being equal, should be viewed as discriminatory. Neumark (1988) expresses the same concern in other words: The question of whether industry or occupation dummy variables should be included in regressions to estimate wage discrimination hinges on the extent to which the distribution of men and women across industries and occupations is itself a result of discrimination. p. 291. 12

3 Data This paper s data are drawn from the yearly income surveys conducted by the Central Bureau of Statistics in Israel for the years 1990 2003. Income surveys are based on questionnaires conducted on household and individual levels and cover information on demographic, personal, and labor market characteristics. The samples include Jewish and non-jewish respondents living exclusively in Israel (latterly including disputed East Jerusalem, but not the Israeli occupied territories of the West Bank and Gaza Strip). Hence, all respondents in the income survey, both Arabs and Jews, are residents and citizens of Israel. Among the variables included in the analyses of the study data are: full-time employment (in the form of a dummy which takes on the value 1 if the worker works at least 35 hours a week, and zero otherwise); marital status (as a dummy which takes on the value zero if the worker is single and 1 otherwise); and urban/non-urban location (a large city dummy takes on the value 1 if the city of residence is Tel-Aviv, Haifa, or Jerusalem, and zero otherwise). Occupational and industrial a liation are coded according to the one- and twodigits classi cation variable. In investigating the Jewish-Arab wage di erentials in the Israeli labor market I limit my analysis to salaried, prime-aged (25 65) male workers. Worker s hourly wage is calculated by dividing monthly income by the product of hours worked per week and working weeks per month. I deal with outliers, in terms of hourly wage, by dropping observations below the 1st and above the 99th percentile of the log hourly wage distribution for each year. This procedure is more robust and meaningful than dropping observations on a given currency 13

(New Israeli Shekel, NIS) cuto point, since the analyses involve di erent years, from 1990 to 2003, over which the currency value is not comparable. Moreover, this procedure circumvents the problem of an a priori imposed NIS cuto point by accommodating changes in the wage distribution over period years (see Chandra (2000)). Israel remains an immigrant society in the sense that incomers, including migrant workers, constitute a large portion of its population. Therefore, it is important to include in any analysis of labor market discrimination a treatment of the working participation in the Israeli economy of foreign-born citizens. Since it is conceivable that labor markets outcomes may re ect, for instance, the displacement of resident workers by immigrants, it is correct for analysis to consider incoming Jews wages integrally with the other members of their cohort. At the same time, though, it is also the case that for some Arabs and Jews, local birth or longstanding assimilation can o er labor market participants advantages over recent arrivals. To integrate these opposing considerations in my analyses, I exclude from certain of my analyses recent immigrants (who arrived in the last 10 19 years). 9 Hereafter, I refer to the sample excluding these later immigrants as the reduced sample. As a benchmark, and providing a basis for comparison, I carry out parallel analyses on the basis of the full sample, excluding no worker or recent immigrant. 9 I use an exclusion rule for immigrants based on a range of years (10 19) since arrival, rather than a xed cuto point, since the exact year of arrival is not always provided in the data. In the reduced samples, for each of the years 1990 1999, I excluded immigrants who arrived after 1980; and for each of the years 2000 2003 I excluded immigrants who arrived after 1989. 14

3.1 Descriptive Statistics Averages of the most important variables, for each investigated year, are given in the Tables 1 and 2. Table 1 describes the sample means after applying all the selection rules described above but before excluding immigrants; in other words, it pertains to the full sample. Table 2 describes the mean characteristics of the reduced sample, that is, after excluding all the newly arriving immigrants. 10 In some of the years the omission of immigrants from the samples reduces (by very little) the number of observations of Arabs. This may happen due to the non-one-to-one relation between the non-jews, as de ned in the data, and Arabs. This non-coincidence makes the interpretation of comparisons between the full and reduced samples more problematic, though not to any great extent. It remains true that results from the reduced sample are more robust and straightforward to interpret, since excluding new immigrants results in an (almost) perfect match between the two de nitions. Another peculiarity of the data should be noted. In 1997 the Income Survey sample design changed to cover the rural population as well as urban households. Respondents were thereafter also drawn from East Jerusalem, while Kibbutzim, institutions, and groups of Bedouins living outside localities are still absent from income surveys. The dataset for the year 1997, serving as the linking year, has been produced in two versions: an old version 10 The 1990 dataset includes a categorial, rather than a continuous, schooling variable. This means that relevant gures are not provided in Tables 1 and 2. In this year, 44.5% of the Arab workers fell into the 0 8 years of schooling group, as opposed to 14.6% of the Jewish workers (or 14.8% in the reduced sample). 37.8% of the Arab workers had 9 12 years of schooling, versus 46.9% 47.9% of the Jewish workers. Only 17.5% of the Arab workers had more than 12 years of schooling, while among the Jewish workers this gure was 37.3% 38.4%. 15

according to the old estimation method and sample design, and a new version based on a wider population and deploying the new method of estimation. In Tables 1 and 2 (and in all subsequent tables) I refer to the old 1997 version as 1997a, and to the new as 1997b. We may note also that, after applying the new sample design in 1997, samples doubled in size. 4 Results In Table 3, I provide estimates for the Jewish dummy coe cient in the pooled wage equations, from both the full and reduced sample. This is a measure of the overall wage gap after controlling for di erent variables. The dependent variable in all speci cations is the logarithm of individual hourly wage. In column 1 the only regressor is the Jewish dummy, hence, the reported coe cient measures simply the overall unadjusted (logarithmic) wage gap. 11 The speci cation in column 2 includes, beside the Jewish dummy, a years-of-schooling variable, a potential experience variable, a squared potential experience variable, a marital status dummy (0 if the individual is single and 1 otherwise), and a full-time employment dummy. 12 Therefore, in speci cation 2, the given coe cient measures the adjusted wage gap, controlling for these variables insofar 11 In this paper, the term wage gap designates the di erence in the average of logarithmic hourly wage between Jewish and Arab workers. However, it is important to note that this gap is no more than an approximation to (and is less than) the geometric wage gap, which, in general, is again less than the observed hourly wage gap. For example, in the reduced sample of 2003 (see Table 3) the gross wage gap is 0.4056; however, this translates to a 0.5002 wage gap in geometric means. In Table 2, further, we see that the actual wage gap is 0.5559, or 55.59%, which is far higher than the initially reported 40.56% gure. 12 For the year 1990, since the data only provide a categorial schooling variable, I use a set of schooling dummies replacing the years-of-schooling variable. Likewise, analysis could not rely on experience and experience squared gures, which were replaced by variables for age and age squared. 16

as they are deemed relevant to productivity. This coe cient is not intended as a measure of discrimination since, by construction, it imposes the assumption that the individual characteristics of Arabs and Jews are similarly rewarded in the labor market. It is important to recognize this as an (unlikely) assumption. The speci cation in column 3 extends that in 2 by adding a set of one-digit occupational dummies. Two facts are evident from Table 3. First, the overall wage gap, as measured by the Jewish dummy coe cient, is higher in the reduced sample than in the full sample, under all the di erent speci cations. This supports the claim that the relatively well-absorbed immigrants in the (Israeli) labor market perform better than more recent arrivals. Hence, excluding the recent immigrants as in the reduced sample yields a higher measure of the inter-ethnic wage gap. At issue here are locally-born and comparatively assimilated Jews, as against these cohorts plus Jewish immigrants. Second, it is apparent that introducing our productivity-related variables into the wage equations greatly reduces the measured wage gap (by 59% 98%). It is also evident from Table 3 that wage gaps increased vastly and monotonically up until 1999, when they started to decline. The new sample design, applied in 1997, resulted in a higher wage gap in all speci cations, as is evident from Table 3. The change in the sample design, in that it began counting inhabitants of rural areas and East Jerusalem, was expected to increase the measured wage gap, since it made available a comparison with a greater proportion of less advantaged Arabs. Tables 4 8 document the main ndings of the paper. They present the overall wage 17

gap decompositions, according to the techniques discussed earlier in the paper, and under di erent assumptions. 13 For each sample, the results proved robust to di erent model speci- cations and assumptions as to the nondiscriminatory wage structure. Table 4 is predicated on an equation between the estimated wage structure of Jewish workers and a nondiscriminatory wage norm, i.e., ^ = ^ J ; analyses in this table pertain to the full sample. Table 6 presents similar analyses as performed on the reduced sample. In Tables 5 and 7 analyses are performed, for the full and reduced samples, under the assumption that ^ = ^ pooled, i.e., that the nondiscriminatory wage norm is equal to the estimated wage structure taken from the pooled regression, including both Arabs and Jews. All the estimates in those tables are signi cant at the 5% signi cance level. Results from both full and reduced samples are reported for the sake of completeness. Nonetheless, I focus attention on results from the reduced sample, which, due to its special composition, better serves the analyses of wage gaps, as yielding a less ambiguous picture of di erences in income between (native or assimilated) Arab and Jewish labor market participants. My description of data ndings, and subsequent inferences, will be con ned to the assumption that ^ = ^ J : To summarize our ndings, in the early to mid 1990s, the hourly wage gap hovered between 40% and 50%. 14 In 1995, the hourly wage gap recorded a level of 50% beneath 13 Note that any di erence in G across di erent tables, when pertaining to the same sample, derives from the absence of available values for the added explanatory variables. For example, in Table 6, we have G = 0:4057 for the year 2003. However, when we add a richer set of variables, as in Table 8, G becomes 0.4017 for the very same year. 14 Figures are calculated from Table 2. For example, the hourly wage gap in 1990 is: 36:9=26:2 1 = 0:408 4; or 40.8%. 18

which it has yet to dip. The hourly wage gap peaked in 1999 at a level of 77%; since then, it has followed a steady path downwards, reaching a level of 56% in 2003. While trends in the human capital and unexplained components of the gap are similar under di erent assumptions as to the discriminatoriness of wage structure for the two groups, it is worth noting that di erences in human capital explain a higher portion of the wage gap when the pooled wage structure is assumed to be nondiscriminatory. Bearing this in mind, my account of the evolution of wage gaps, and their decomposition, assumes an approximation of nondiscriminatory wage structure to the wage structure manifested by the dominant group (Jews). Further, I con ne my attention to the reduced sample excluding recently arriving immigrants. Not controlling for occupational and industrial a liation, productivity-related variables could explain 0.244 of the gross wage gap in 1990 (0.269 see Table 6). That is, only a minor gap of 0.025 remains unexplained when di erences in those variables are taken into account. When one-digit occupational dummies are included in the wage regressions, this portion declines to 0.018. When regressions are further extended to include more speci c occupational and industrial dummies (two-digits), and a large city dummy, as in Table 8, this portion becomes -0.03, i.e., the factoring-in of the extended form of these productivityrelated variables not only explains the wage gap, but suggests that Arabs are favored, given their characteristics and occupational choice. Nonetheless, within this extended analysis, a new component emerges, that pertaining to labor market segregation. This accounted for a wage gap of 0.05 and 0.134 in the short and general speci cations respectively. As argued 19

in the Methodology section, this component may well represent a form of labor market discrimination, manifested in barriers to entry for higher-paying professions; on the other hand, it is likely that part of this component is attributable to the di erent self-selection of groups into low paying occupations. The same sets of analyses are then carried out for the other years of the study data, with a similar pattern of results emerging. A greater measure of speci city or detail in terms of the productivity-related variables tends to explain a greater portion of the wage gap, and therefore reduces its unexplained component. Given that axiomatically only part of the unexplained wage gap measures discrimination in the labor market, it then becomes reasonable to regard the unexplained wage gap in the most general or multivariate speci cation (as in Table 8) as a ceiling on the extent of within-occupation wage discrimination. Over the whole period, there was a noticed convergence in some of the important productivityrelated variables, such as schooling, age, and marital status. The schooling gap between the groups declined from 3.1 in 1991 to 2.4 in 2003; the di erence in the average age of the workers declined from 5.6 in 1990 to 4.5 in 2003; and the di erence in the marriage rate among workers declined from 10% (with a greater proportion of Jews married) to -2% (with a greater proportion of Arabs married). Despite this convergence, human capital di erences contributed a relatively unchanging portion to the wage gap (0.08 0.17). A higher weight (or set of coe cients) for the productivity-related variables, as assigned primarily to the dominant group, can serve to reconcile these facts. Nonetheless, comparison on a year-to-year basis tends to bear out the result that human capital di erences contributed a proportional 20

part to wage di erences. The representation of labor market segregation becomes ner in proportion to one s increasing speci city about occupational a liation. In other words, using the two-digit occupational and industrial classi cation, as in Table 8, works out as more precise than accounting only for the one-digit occupational a liation, as in Table 6. It is expected, though, this greater measure of precision will lead to higher estimates of labor market segregation. This is readily seen from a comparison of the results from Tables 6 and 8. Labor market segregation contributed 0.13 0.25 to the wage gap over the period in question (and 0.15 0.20 in the second half of the period, from 1997 2003). If we manage to control for all the observables, and assume that any unobservable wagerelevant variables vary only to a negligible degree between consecutive years, then we would be bound to attribute any change in the unexplained wage gap between years to labor market discrimination. These assumptions seem reasonable for all years after 1992, judging from the results of the general speci cation, Table 8. In the years 1993 1997, the unexplained wage gap was 0.04 0.10. In the second half of the sample (1997 2003), this component contributed about 0.10 0.19 to the gross wage gap. The sharp change in the unexplained wage gap between the old and new sample styles around the year 1997 seems to follow from the change in the sample design. If anything, the (higher) gures in the second half of the period are, statistically, more accurate, since those samples are more representative and statistically twice the size. An explanation of this trend of apparently rising discrimination must seek socioeconomic 21

factors in seeking to rationalize uctuations in the unexplained portion of the wage gap. The last decade began with a major in ux to Israel of Jews from the former Soviet Union, with about half million immigrants arriving just in the rst three years of the decade. (This represents an increase of some 11% in the country s population.) These immigrants were, at the beginning of the decade, very educated, with about 14 15 years of schooling on average. (See Locher (2004) for a description of the trend of immigrants education.) In one sense, the incorporation of such a large number of people might be seen as a good stimulus to economic recovery in the sense of strengthening demand. The revival of the construction industry stands as a clear indicator of this. Educated immigrants, on the other hand, compete with locals for jobs. Supposing that some Jewish migrants won jobs at the expense of skilled Arabs, we would expect the gross wage gap among skilled workers to be high around this period. Further, since the cross-ethnic skilled group is homogenous in an important dimension, schooling, we would expect the human capital component of the wage gap to be very low i.e. for speci c non-productivity-related di erences to explain the wage gap in large measure. Consequently, we would expect the unexplained component of the wage gap closely to follow the pattern of the overall gross gap. Table 9 documents the decomposition of the wage gap over the whole period, for skilled and unskilled labor. Results from Table 9 largely con rm this prediction of higher gross wage gaps whose most part is unexplained (or attributable to discrimination). The gross wage gap among skilled workers reached, in 1992 1993, its highest level. (Due to the sample design change in 1997, we can add 0.203=.319-.116 to the gross gures before 1997 to make them 22

comparable to post-1997 gures.) Moreover, the human capital component contributed, for most of the time, less than 0.04 to the overall wage gap. The unexplained gap among skilled workers reached its highest level of 0.222 in 1992, around the time of the in ux of skilled Jewish labor. Though measures of the unexplained portion were volatile afterwards, in many years this value recorded a decline. At the same time as the migration to Israel of large numbers of mostly Russian Jews, the Israeli government imposed extremely stringent restrictions on the movement of the Arab inhabitants (and workers) of Israel s administered territories. A surging demand for construction workers, following the mass immigration, was met with a sudden shortage of Palestinian workers, whom the closures prohibited from reaching their workplaces. Israel responded to this labor crisis by importing foreign workers. About 30,000 foreign workers were employed in 1993; this gure was on the rise till 2002, when the number of employed foreign workers reached an unprecedented apogee of 232,000. 15 Substituting Palestinian workers, and unskilled Israeli Arabs, with foreign workers had the e ect of increasing the wage gap immensely. The unexplained component of the wage gap increased, between 1992 and 1993, by 0.143, a drastic change for a single year. The replacement of Arab by foreign employees would seem su cient by itself to explain this yearly change in the unexplained segment of the gap. When it comes, however, to explaining wage disparities among skilled labor, labor market segregation played a more important role that reduced opportunities for Arabs. This might 15 Estimated numbers of foreign workers employed in Israel are taken from the Bank of Israel web site. They may be accessed online at http://www.bankisrael.gov.il/series/export/html/?series=na.em_frn.a&start=1990&end=2004 23

be attributed to the fact that a wider variety of occupations is open to skilled workers, while unskilled workers have access to low-skilled occupations only. The expectation before analysis was that those industries into which foreign workers were imported, agriculture and construction, would register the greatest change in the wage gap (and in the unexplained portion of the gap) across Israel s economy. Table 10 presents the wage gap and its decomposition over the whole investigated period, rst for workers in agriculture and construction only, then for workers in all remaining industries. The large measure of volatility in the unexplained component of the wage gap among agriculture and construction workers, at least in the rst half of the 1990s, supports the assumption that these industries were especially a ected by changing employment patterns, as the sectors recorded an unexplained wage gap increase of 0.23 between the years 1992 and 1993. These sectoral results, combined with the previous skilled-unskilled comparison, serve to explain the huge change in the overall unexplained wage gap between those years as captured in the general speci cation (Table 8). It is evident from Table 10 that the composition of agriculture and construction workers (or similarly unskilled workers from Table 9) is not constant, but uctuates with labor market and wider socioeconomic circumstances. Figures also suggest that labor market segregation obtains less among relatively unskilled workers in the agriculture and construction sectors than among workers in other industries. This is not surprising, since the occupational classi cation is narrower in these two industries than in the economy as a whole. Lastly, in many cases, the unexplained wage gap among unskilled workers (agriculture and 24

construction workers) is higher than that among skilled workers. Workers heterogeneity in the former group (with some Jews, alongside foreign and Arab workers) possibly accounts for this observation. Pooling all the years from 1991 2003 together, analysis further logged the gross (unadjusted) wage gap between Jewish and Arab workers in every main occupational group and for six schooling categories. Results appear in Table 11. The overall (logarithmic) wage gap, for all occupations and schooling categories, is 0.392. The gross wage gap for every known occupational category comes in at a lower level than the general gap. The highest gap, including all schooling categories, was among skilled workers in the construction and industry sectors (0.303). The wage gap for this occupational group increases with education (from a negative gap of 0.041 for the uneducated to a wage gap of 0.422 for the highly educated workers or those with 16+ years of schooling). A similar pattern is observed among unskilled and service sector workers: the greater the years of schooling, the higher the gross wage gap (excluding the zero schooling group). More schooling does apparently bene t Arab managers, although the most highly educated cohort of managers is excluded, where the wage gap climbs very fast, from 0.064 to 0.441 after getting a Masters degree. As expected, controlling for all the (available) productivity-related variables reduces the measured gross wage gap. occupational-schooling category. Table 12 reports duly adjusted wage gaps within each In contrast with the previous table (11), here I control for schooling level within schooling category (when that is not constant), as well as for experience, experience squared, marital status, full-time employment, size of city of residence, 25

one-digit occupational a liation (when relevant), one-digit industrial a liation, a yearly time e ect, and industry-year interaction terms. The overall adjusted wage gap, over the whole fourteen year period, comes out at 0.15. The trend of changes in the wage gap for managers and skilled workers in industry is similar to that manifested in the unadjusted gap. In general, the wage gap declines with education in high-education occupations, such as academic professionals and managers, and increases with education in low-education occupations, such as industry, construction, and service sectors. Table 13 further speci es groups employment distribution in some important regards. It reports the sum of observations, over the whole period (1991 2003), for each occupationschooling cell. Overall, the occupational a liation of about 51% of the Arab workers is industry and construction (which are among the least remunerative jobs); this contrasts with 29% of the Jewish workers. 2.5% of the Arab workers are managers (the second highest paying job-category), as opposed to 11% of the Jewish workers. Moreover, among workers with 12 years of schooling, 57% of the Arabs work in industry and construction, as opposed to 40% of the Jewish workers. More importantly, high-paying jobs are not open to Arabs even when they match Jews level of schooling: 7% of the Jewish workers are associate professionals, 8% managers, and 16% clerical workers (the highest paying jobs) as opposed to, respectively, 2.6%, 2.6%, and 8% of the Arab workers. For sure, self-selection goes some way towards explaining di erences in groups employment distribution. However, with the table bringing together fourteen years of data, the possibility of group representatives choosing to perpetuate outcomes that they have expe- 26

rienced as unequal becomes increasingly remote. Barriers to entry then represents a more likely explanation of occupational distribution meaning that much of the labor market segregation, as calculated in previous tables, should be associated with this second explanation of persistent wage gaps. It is worth noting that barriers to entry does not imply in the exclusion of Arabs from certain occupations (though discriminatory practices might occur), but rather that Arab workers need a much higher level of credentials than their Jewish counterparts to be admitted through the door to comparable jobs. Both the overall, and the unexplained, wage gap peaked at levels of 0.506 and 0.192, respectively, in 1999 (see Table 8). Under all assumptions, and all model speci cations, 1999 saw the high point in these measures. A range of pecuniary (human capital di erentials) and non-pecuniary (discriminatory) factors can be understood as combining to drive the wage gap to its peak. However, if labor market discrimination, re ecting other political hostilities, is taken as part of the explanation of the wage gap, then it seems that the subsequent decrease in the wage gap itself, as much as in its unexplained portion, represents the triumph of (rational and pecuniary) economics over discrimination. A consideration of historical factors in part bears out this hypothesis. In 2000, with the eruption of the second Intifada, the number of Palestinian workers available for work in Israel dramatically declined from 113,000 in 1999 to 30,000 in 2002. Concurrently, the pace of foreign workers entering the country decelerated (from 28,000 new foreign workers in 2001 to -25,000 i.e. a net out ow in 2003). 16 These facts led to a higher demand for unskilled Arab Israelis. The resulting, 16 Source: Bank of Israel web site. See: http://www.bankisrael.gov.il/series/export/html/?series=na.em_frn.a&start=1990&end=2004 and http://www.bankisrael.gov.il/series/export/html/?series=na.em_ter.a&start=1990&end=2004 27

and theoretically predictable, higher wage for Arab workers is consistent with the continuous decline in the gross, and unexplained, wage gap in, and after, the year 2000. Knowledge of Hebrew, which is an important component of human capital, might have direct bearing on the issue of wage gaps. Nonetheless, the Income Survey data do not provide information about knowledge of Hebrew, thus I could not control for this variable in my analyses. If knowledge of Hebrew has a signi cant e ect on wage gaps, i.e., lack of knowledge of Hebrew leads to lower wage, then the unexplained component of the wage gap, as presented in this paper, might be overestimated. Chiswick (1998) nds that Hebrew, being the worker s primary or sole language, increases the worker s earnings by 11 35%. He uses data about foreign-born men from the 1983 Census of Israel. Note that he uses a variable about usage of Hebrew and not knowledge of Hebrew; so it is not a direct assessment of the e ect of knowledge of Hebrew on earnings. Secondly, he nds that Hebrew-speaking usage increases with duration of residence. Israeli Arabs, living in Israel since birth, are expected, accordingly, to have a high knowledge and usage of Hebrew. Moreover, Hebrew language is a compulsory subject in Arab schools, taught from the second grade until the twelfth grade. Put together, it is suggested that Israeli-Arabs know the Hebrew language very well. While the e ect that knowledge of Hebrew have on earnings is acknowledged, the di erence in Hebrew knowledge between Arab and Jewish workers is likely to be negligible. So is its e ect on this study s results. In a study by Lecker (1997) it has been shown that the usage of Hebrew explained only a tiny part of the wage gap (about 0.02). The comparison was made between Arab groups 28