Using a large administrative data set, a recent study by Dustmann, Ludsteck, and Schönberg [2009] finds convincing evidence for rising wage inequality in West Germany during the past three decades. Their paper shows that the increase occurred above the median during the 1980s, and was augmented by rising inequality below the median in the 1990s. These results challenge the pervasive conception of the German wage structure being a paragon of stability.
Within the scope of this seminar paper, I replicate parts of their analysis using a different data set, namely the German Socio-Economic Panel (GSOEP) for the period from 1984 to 2009. Using monthly and hourly earnings constructs, I assess the extent to which the results of Dustmann, Ludsteck, and Schönberg [2009] can be recovered from GSOEP data. I do so by exploring the wage distribution along several dimensions. In addition to this, I analyze composition-constant counterfactual wage densities using the kernel density reweighting method advanced by DiNardo, Fortin, and Lemieux [1996]. A provisional evaluation of my results suggests that I can confirm the majority of findings of the original paper along a qualitative dimension. A quantitative assessment, however, reveals considerable deviations.
Abstract
Using a large administrative data set, a recent study by Dustmann, Ludsteck, and Schönberg [2009] finds convincing evidence for rising wage inequality in West Germany during the past three decades. Their paper shows that the increase occurred above the median during the 1980s, and was augmented by rising inequality below the median in the 1990s. These results challenge the pervasive conception of the German wage structure being a paragon of stability.
Within the scope of this seminar paper, I replicate parts of their analysis using a different data set,namely the German Socio-Economic Panel (GSOEP) for the period from 1984 to 2009. Usingmonthly and hourly earnings constructs, I assess the extent to which the results of Dustmann,Ludsteck, and Schönberg [2009] can be recovered from GSOEP data. I do so by exploring the wagedistribution along several dimensions. In addition to this, I analyze composition-constantcounterfactual wage densities using the kernel density reweighting method advanced by DiNardo,Fortin, and Lemieux [1996]. A provisional evaluation of my results suggests that I can confirm themajority of findings of the original paper along a qualitative dimension. A quantitative assessment,however, reveals considerable deviations.
I. INTRODUCTION
Secular changes in the wage and income distribution over the course of the past three decades inthe United States (US), United Kingdom (UK), and other industrialized countries have sparked anoutsized amount of research in the field of labor economics. Until today, the study of labor marketoutcomes belongs to the most vibrant fields of on-going research in the economic profession. It ismarked by new findings constantly undermining old assertions and creating new puzzles. Because ofits broad implications for human and economic welfare it recurrently shapes political debates allaround the globe.
Presumably one of the most thoroughly studied features of the wage distribution is the pervasiveupsurge in wage and income inequality, which has preoccupied numerous scholars since the early1990s (see Lemieux [2008] for a review of this literature). Goldin and Katz [2007a] provide acomprehensive historical account of the trends and developments in wage inequality, and alsohighlight the role of explanations based on supply, demand and institutions. In short, the study of wageinequality has provoked an abundance of diverse theories and motivated a large number ofeconometric techniques whose applications go far beyond the economic profession.
Commencing in the late 1970s and persisting throughout the 1980s, the US witnessed a salient andover the years accelerating increase in wage inequality across the entire distribution of wages andalong several dimensions.[1] Among the first economists to expose distinct empirical evidence wereKatz and Murphy [1992], Bound and Johnson [1992], as well as Levy and Murnane [1992]. Theirfindings clearly revealed a conspicuous increase in wage inequality during the 1980s, driven bygrowing dispersion of wages in both the lower-tail (measured by the 50-10 percentile) as well as theupper-tail (measured by the 90-50 percentile) of the income distribution. The pronounced rise in wagespreads among different income and education clusters on the US labor market also continued duringthe 1990s, though the dynamics somewhat changed compared to the preceding decade. In particular,the top part of the wage distribution continued its rapid spreading, meanwhile inequality at the bottomend stabilized or even decreased. Abundant evidence for the US can be found in Autor, Katz, andKearney [2006, 2008] and Lemieux [2008]. Moreover, Gosling, Machin, and Meghir [2000] andBoudarbat, Lemieux, and Riddel [2006] document similar trends for the UK and Canada, respectively.
These almost universal patterns for the 1980s and 1990s appeared to contrast observations made forcontinental Europe, and the majority of other OECD member states. Empirical results for that periodrepeatedly corroborated the finding that the wage distribution was mostly stable throughout the 1970sand 1980s, while starting to disperse during the 1990s. However, the observed overall increase inwage inequality during the 1990s seemed to be the result of changes in inequality among low wages, and was by far not as dramatic as in the US a decade earlier (see e.g. Katz, Loveman, and Blanchflower [1995], OECD Employment Report [1996]).
In particular, findings for West Germany used to be strongly in line with those made for most other European countries.[2] In fact, until recently, Germany was deemed a paragon of wage structure stability. Empirical evidence for the 1980s can be found in Prasad [2004] and Abraham and Houseman [1995].[3] Increasing dispersion in the lower part of the income distribution during the 1990s has been documented by Möller [2005] and Gernandt and Pfeiffer [2007].[4]
The fact that, in contrast to these findings, inequality at the top of the wage distribution did indeedgrow during the 1980s and continued to rise rapidly during the 1990s, has surfaced only recently. Firstindications for the former can be found in e.g. Möller and Bellmann [1996] and Möller [1996]. Thelatter observations have also been suggested by Kohn [2006], Gernandt and Pfeiffer [2007] andGiesecke and Verwiebe [2009]. Dustmann, Ludsteck, and Schönberg [2009] (DLS hereafter)eventually removed any remaining doubts by uncovering conspicuous evidence for rising inequality inGermany throughout both periods, building their analysis on data from the Institute for EmploymentResearch (henceforth IABS).
Why did previous studies fail to reveal these seemingly obvious trends? DLS argue that the key tothis puzzle lies in the selected data source. Most previous studies that reported stable wage inequalityduring the 1980s and increasing dispersion foremost at the bottom end of the wage distribution reliedupon the German Socio-Economic Panel (see, for instance, OECD [1996], Prasad [2004], as well asGernandt and Pfeiffer [2007] and Giesecke and Verwiebe [2009]). In contrast, authors using earlierversions of the IABS generally draw conclusions that bolster the findings of DLS. Most importantly,Fitzenberger [1999] finds similar trends of rising inequality for the 1980s. However, as his observationperiod ends in 1990, Fitzenberger fails to uncover the pronounced increases in inequality during the1990s. Other studies using the IABS, thus having the potential to foreshadow the findings of DLS,either covered a different time period (see Kohn [2006], focusing exclusively on the years from 1992to 2001 and analyzing differences between East and West Germany) or targeted a different researchquestion (Kohn [2006], Fitzenberger and Kohn [2006]).[5]
The objective of DLS, and at the same time their principal contribution to the existing literature, isto revise the prevailing conception of the stable German wage structure and provide a systematiccomparison to US trends in inequality. They are the first to use the IABS data to jointly analyze the1980s and 1990s. In this sense, the authors unify and extend the analyses of Fitzenberger [1999],Möller [2005] and Kohn [2006]. The authors investigate the effects of de-unionization on the wagestructure, and provide sufficient evidence for the tentative conclusion that technology-induced jobpolarization goes a long way in explaining the evolution of wages in Germany. Therewith, the authorscontribute an important piece to the international evidence for technical change driving wage inequality.
DLS advocate four findings: first, and as highlighted earlier, wage inequality did increase considerably during the 1980s and 1990s. During the 1980s, this rise was concentrated in the uppertail of the wage distribution, while the 1990s featured rising wage expansion throughout the entire distribution of wages. Moreover, this pattern manifests for both genders.
Second, using data on unionization rates and union coverage, DLS document a conspicuous deterioration of unions between 1995 and 2004. In conjunction with the first finding, this readily suggests that the decline in union coverage may have had a stake in the observed rising inequality, especially at the low end of the distribution. Accordingly, the authors document that de-unionization can ‘explain’ up to 28 per cent of the observed increase in lower-tail wage inequality, and around 11 per cent of the change in the upper-tail of the wage distribution.[6]
The third finding regards the evolution of wages between different skill groups. Specifically, DLS report a rise in the wage gap between low- and medium-skilled commencing in the late 1980s and persisting throughout the 1990s. In contrast to this, the authors find that the wage gap between medium- and high-skilled over the same time period was rather stable.
Finally, and perhaps most importantly, they find evidence for technology-driven job polarization.Specifically, they firstly observe that occupations with high median wages in the 1980s required a highshare of non-routine cognitive skills, while occupations in the middle income range primarilydemanded the conduct of routine tasks. Non-routine manual tasks dominated occupations with lowmedian wages during the 1980s. Secondly, they find that occupations with high median wages in the1980s also experienced the highest wage growth during the 1990s, while occupations with mediummedian wages during the 1980s incurred wage declines relative to occupations with initially lowmedian wages. Taken together, these observations reinforce a more nuanced understanding of skill-biased technical change which grasps technological progress as complementing non-routine (cognitiveand manual) tasks, while substituting routine (cognitive and manual) tasks.[7]
Against this backdrop, especially with the summary of descriptive findings freshly in mind, theobjective of my seminar paper is to explore to what extent the results by DLS can be supported by areplication based on survey data from the German Socio-Economic Panel (henceforth GSOEP).
My replication focuses on a descriptive exploration of the wage distribution with the goal to identify the differences and similarities between the different data sets, and to uncover where the results give leeway for an interpretation in either direction. Eventually, I use a kernel density reweighting method to account for mechanical changes in wage inequality caused by changes in the workforce composition over time.
My investigation differs primarily along two dimensions. First, my observation period ranges from 1984 to 2009, whereas DLS concentrate on the period from 1975 to 2004. Second, I use the 90th and 10th percentile instead of the 85th and 15th percentile to represent the top and bottom of the wage distribution, respectively.
Inspired by DLS, my findings suggest that overall inequality, measured by the 90-10 percentile did increase between 1984 and 2009. In line with their findings, my results for men suggest that inequality increased foremost at the top during the 1980s, and started to increase in both tails of the distribution over the course of the 1990s. During the 2000s, lower-tail inequality continued to rise while upper-tail inequality muted. For women, my results differ strongly from those for men, but exhibit important similarities to the findings reported by DLS. In particular, the 1980s indicate wage compression from both the bottom and the top of the distribution, while the 1990s suggest slightly increasing inequality at the top of the wage distribution. These trends are followed by a dramatic increase in wage inequality along the entire distribution during the 2000s. Except for the trends in the 2000s, which are mostly outside the time period analyzed by DLS, my results are qualitatively similar to DLS. Some major differences occur in terms of timing and magnitude of the observations.
This paper unfolds as follows: Section II is structured in three subsections. Part a) introduces and succinctly compares the two data sets, before part b) turns to the sample selection. Part c) describes the construction of important variables for the analysis. Section III presents the empirical investigation and is organized in two subsections. In part a) I present and interpret four different measures of wage inequality. Part b) is devoted to the introduction and application of the kernel density reweighting method. Section IV summarizes and concludes.
II. DATA DESCRIPTION
a. The Data
My empirical investigation builds on data from the student version of the GSOEP.[8] The GSOEP isan annual panel study which surveys private households and individuals. It was launched in 1984 inWest Germany with the primary goal to provide representative micro-data for scientific research.
From 1990 onwards it was extended to include the territory of the former German DemocraticRepublic.
Since the very beginning of the survey the sampling size has been steadily increased, commencingwith approximately 12,000 respondents in around 6,000 households, up to more than 20,000individual respondents in almost 11,000 households in the most recent waves. Over time, specialsamples have been added to enhance data for specific demographic groups such as foreigners, topincome earners or East Germans. Since these groups typically represent minorities, the GSOEPoversamples them in order to sustain reasonable sampling sizes for meaningful empirical analyses.
In my investigation, I extend the analysis of DLS to more recent years by using the survey roundsfrom 1984 to 2009 (in contrast to 1975 to 2004 in DLS). While this comprises the drawback of cuttingshort on the early 1980s, it bears the potential to shed light on trends in more recent years, partly underthe influence of the financial crisis.[9] Being the first crisis that hit the German labor market after itunderwent a series of structural reforms during the 2000s, this is also an interesting undertaking interms of a policy evaluation.[10]
The data set used by DLS is a 2 per cent random sample of (administrative) social security recordsprovided by the German Institute for Employment Research (IAB). The IAB Employment Samples(IABS) are representative for approximately 80 per cent of the German workforce. The remaining 20per cent are civil servants, self-employed, marginally employed, and individuals in generalconscription. As these groups are not subject to social security contributions, they are not covered bythe IABS either. The GSOEP, in contrast, contains these groups and I decided to include them into myanalysis.
A brief juxtaposition exposes some advantages and disadvantages of the IABS compared to the GSOEP.[11] Much to the benefit of the IABS are its sample size and its accuracy. Regarding the former, the IABS contains around 200,000 annual observations, which outnumbers the GSOEP by far. In the student version that I am operating, this issue is even exacerbated.[12] With respect to accuracy, the IABS benefits from the sampling method which draws directly from social security records. Individual wage data is reported by the employer for the purpose of determining the contributions to the social security system. The legal obligation to report in conjunction with the risk of being sued for misreporting renders the IABS data more accurate than the GSOEP.[13]
Much to the disadvantage of the IABS is that it is top-coded and lacks information on working hours. While the IABS is representative for 80 per cent of the German workforce, it fails to represent the entire distribution of wages for this group. Specifically, the IABS is top-censored at the highest wage subject to social security contribution. According to DLS, this entails losses of up to 15 per cent at the top part of the male wage distribution in selected survey years. The authors fabricate the missing part of the distribution by imputation, and restrict their analysis to the distribution of wages between the 15th and 85th percentile.[14] Since the GSOEP does not feature the drawback of top-censoring, my replication builds on the 10th and 90th percentile instead.
Finally, the IABS lacks information on working hours. Instead, it contains information on whethera worker is full- or part-time employed with the threshold between the two categories at 30 weeklyworking hours. The GSOEP is considerably richer in this respect. Not only does it contain informationon working hours, it also differentiates between actual and contracted weekly working hours, therebypermitting to compute a broader variety of earnings constructs as a basis for wage structure analyses.
In sum, while the choice of the data set is crucial to the success of an empirical investigation, thereis no general answer as to which one is better suited. On the one hand, the GSOEP bears the advantageof providing individual information along various dimensions, but is at the same time prone tomeasurement error due to misreporting and non-response. On the other hand, the IABS administrativedata capitalizes on a large sample size and high data accuracy, but falls short on detailed personalinformation along dimensions that go beyond pure wage data and basic demographic information.
b. Sample Selection
From the GSOEP database, I select my sample in a specification that resembles the sample characteristics of DLS as closely as possible. Specifically, I draw annual cross-sections (corresponding to choosing the option ‘unbalanced panel’) for the GSOEP samples A (‘German West’) and B (‘Foreigner West’), since my focus is on the wage development in West Germany. I include nonprivate households and both genders, enabling me to discern potential differences in the dynamics and magnitude of wage inequality between genders.
The analysis necessitates further restrictions on the sample. Specifically, DLS only include workersbetween 21 and 60 years of age who are full-time employed (above 30 weekly working hours) and arenot registered in an apprenticeship in the respective sampling year. I sequentially impose theserestrictions while keeping track of the number of observations for men and women with availablewage data. To begin with, my sample contains 132,767 person-year observations. In a first round ofdata manipulation, I delete all observations for which either no information on the attained level ofeducation is available (29,238 observations) or which reported to be still ‘In School’ at the time ofsurvey (1,893 observations). This leaves 101,636 individual observations in the sample, out of whichapproximately 60 per cent (61,292) have non-missing data on wages. Next, I drop workers younger than 21 and older than 60 years of age (reduction by 25,589 observations), as well as workerscurrently in ‘vocational training’ (998), working in ‘marginal / irregular employment’ (2,448), orbeing ‘unemployed’ (19,756) at the time of survey. Finally, only individuals are kept in the samplewhich exhibit actual or contracted weekly working hours above 30 (deletion of 9,050 observations). Inthe end, my sample contains 43,795 person-year observations, 43,785 with readily available wagedata.[15]
It is important to stress that - in contrast to the analysis of DLS - I leave self-employed and civilservants in the dataset. DLS point out that this does not make a notable difference for the comparisonto the IABS results (see Online Appendix, Section 3), and hence I prefer to use the additionalobservations.
The following subsection describes the specification and construction of the relevant variables formy analysis, i.e. wages (and bonus payments), age groups, and education. I also briefly illustrate thereasoning behind and construction of the sample weighting which I utilize in most applications. Thespecification of the wage variables adheres to the description provided in the Online Appendix,Section 3, of DLS. Concerning the definition of the remaining variables, I rely on verbal explanationsand reasoning found throughout the DLS paper as well as Appendix I B (Sample Selection and Variable Description). My goal is to reproduce the measures used in DLS as accurately as possible.
c. Variable Specification Wages
To construct the wage measures I foremost rely on the generated GSOEP-variable ‘LABGRO$$’. I choose this variable because it features three advantages over the standard wage variable. First and foremost, the number of observations is substantially larger because it contains contemporary monthly wage data for all employed individuals in each year. To achieve this, missing values are imputed based on a two-stage procedure combining longitudinal data of the respective individual with crosssectional (trend) data. If an individual was surveyed for the first time, and hence does not yet have a personal history to draw longitudinal information from, the imputation instead builds on a Mincer-type regression that incorporates, among other covariates, net labor income.[16] A second advantage is that it is consistently measured in EUR and hence does not require conversion. Finally, the variable name is constructed in a way that makes it easier to handle in Stata.
Having said this, I additionally include bonus payments to compute the effective (or actual) grosswages of a worker. To this end, I draw the whole set of bonus variables from the GSOEP. The GSOEPquestionnaire offers to choose among six types of bonus payments, and a respondent is requested to classify and quantify the payment he or she received.[17] As opposed to the wage variable, which is measured monthly contemporary, bonus payments are reported in retrospect and for the whole year. I take account of the lag in the reported value by firstly drawing only the waves from 1985 to 2009, and secondly renaming the variables to the previous year before reshaping the dataset.[18]
After merging the reshaped bonus data (ranging from 1984 to 2008) with the wage data, values before and including the year 2000 need to be converted to EUR.[19] Subsequently, I define the variable ‘ boni ’ as the sum of all bonus payments, and set it equal to zero if an individual has not reported a bonus payment in a certain year.[20] Finally, I divide the yearly payment by 12 months.
To make wages comparable across years, I deflate all monetary variables using the Consumer PriceIndex (CPI) provided by the Federal Statistical Office for the period from 1984 to 2009 (Index: 2005 =100).
I trim the distribution at the top as well as the bottom to avoid biases due to misreporting orotherwise extreme outliers. For the bottom truncation I adhere to DLS who drop observations with areported daily wage of less than 20 DM in 1995 prices.[21] At the top, I trim the wage distribution at the 99.0 percentile.[22]
In accordance with DLS, I compute three different earnings constructs: log real monthly wage, effective log real monthly wage, and the effective log real hourly wage.[23] The first measure isstraightforwardly constructed by taking the natural log of the deflated ‘ labgro ’-variable. The secondmeasure includes the average real monthly bonus payment. To assemble this measure, I firstly add thecomputed aggregate real monthly bonus to the real monthly wage, and take the natural log thereafter.
[...]
[1] In fact, this period followed several years of wage compression during the early 1970s, where the skillpremium of young high-skilled workers and recent college graduates exhibited a decline relative to the low-skilled. This observation has been made for the U.S. and the vast majority of other OECD countries alike (e.g.Freeman [1981]). For further details and a selection of potential explanations, I refer to Katz, Loveman, andBlanchflower [1995].
[2] If not explicitly stated differently, I refer to West Germany when using the term Germany. It is worthwhile tonote that the evolution of wage inequality differs largely between East and West Germany. Franz and Steiner[1999] provide a detailed account for East Germany in the 1990’s. Gernandt and Pfeiffer [2007] compare theevolution of West and East German wage inequality for the period from 1984 to 2005. In a nutshell, the EastGerman wage distribution underwent a process of quick convergence to the West German wage structure. Aslabor productivity fell short of wage increases, these developments were chaperoned by strong increases inunemployment.
[3] Prasad [2004] also offers an indicative assessment of potential explanations and has been recited frequently for characterizing the stability of the German wage structure as being “unbearable”. Abraham and Houseman [1995] not only find stability but even a slight compression of overall wage inequality during the 1980s, driven by a decline in lower-tail inequality and stable upper-tail inequality.
[4] Appendix Table A1 in Gernandt and Pfeiffer [2007] summarizes the results of selected studies for different time periods and datasets. In brief, it illustrates that there seems little consensus.
[5] It is noteworthy that the problem of different datasets leading to different conclusions about the evolution of wage inequality is not restricted to Germany, but also surfaced in the US. Lemieux [2006b] and Autor, Katz, Kearney [2008] compare trends in wage inequality between the March/CPS and May/MORG datasets.
[6] It is noteworthy that DLS measure unionization by union coverage (as opposed to workers’ unionization rates),and adhere to this measure even though data for the period before the mid-1990s is missing. They use the factthat unionization rates for this time period were supposedly stable, and only started to decline during the 1990s.
[7] This ‘task-based’ approach to skill-biased technological change was advanced by Autor, Levy and Murnane[2003]. Spitz-Oener [2006] finds empirical support for this hypothesis looking at the German labor market.Autor, Katz, and Kearney [2006, 2008] provide similar evidence for the US, and Goos and Manning [2007] forthe UK.
[8] Compared to the full GSOEP, the student version is restricted along two dimensions: firstly, the final year is not accessible, and secondly all variables only contain half of the total available number of observations. It should be noted that the second restriction is more severe for the present analysis.
[9] The first half of the 1980s is certainly of particular interest when exploring wage trends in Germany with the objective of comparing them to the US, which is one important objective of DLS. As is well-documented, most of the growth in wage inequality on the US labor market occurred during the early 1980s, and meaningful comparison, therefore, require coverage of this period.
[10] It is beyond the scope of this paper to discuss this last aspect, and it is certainly not part of the research question addressed herein. However, I consider it a promising area of follow-up research, especially now where the 2011 wave of the GSOEP is accessible.
[11] For more detailed information on the IABS, I refer to Bender et al. [2000]. Additional information on the GSOEP is provided in Haisken-DeNew and Frick [2005]. A thorough comparison of the two datasets in terms of their respective pros and cons can be found in e.g. DLS and Pfeiffer [2003].
[12] Table I illustrates that my sample contains between 39,000 and 43,000 person-year observations, depending on the wage measure. This number is obtained after the entire sample selection process was performed.
[13] With respect to data on wages and income the GSOEP is prone to misreporting, rounding, and non-response.
[14] They argue that this restriction probably results in an underestimation of the trends in inequality. They undergird this by referring to empirical evidence for the US, which suggests that inequality increased the most beyond the 95th percentile (see Autor, Katz and Kearney [2008]).
[15] The difference of 10 observations is caused by 6 females and 4 males, out of which 8 reported to be in fulltime employment (self-employed) with working hours above 30, but no wages. I decide to drop them.
[16] If net labor income is also missing it is imputed in a first step. The procedure described above then builds on the imputed net labor income. For further (technical) details, I refer the interested reader to the ‘PGENDocumentation’ provided by the DIW Berlin.
[17] These types are: 13th month pay, 14th month pay, Christmas Bonus, Holiday Bonus, Other (unspecified), and Profit Sharing Bonus.
[18] This means that the reported bonus payment in wave 1985 becomes the earned bonus payment in year 1984,1986 becomes 1985 and so on. It also means that I lose the final year 2009 which is the reason for relying onlyon data until 2008 in those parts of the empirical investigation which consider a wage measure that includesbonus payments.
[19] I impose the conversion rate 1 EUR = 1.95583 DM. In 2002, the questionnaire asked the respondent for the bonus payments of the previous year already in EUR. Hence, having dated back all bonus payments by one year earlier on, I only need to convert payments before and including the year 2000.
[20] This is coded by ‘Does not apply’ and was previously set to ‘Missing’.
[21] Assuming 20 DM per day, 5 days a week, 4.33 weeks per month (assumption from DLS), I obtain a monthlywage of 433 DM in 1995 prices. I convert this value to EUR and deflate it to 2005 prices using the CPI. Theresulting monthly wage is approximately 254.18 EUR. Arguably, this is not a reasonable wage for a full-timeemployed person, who is left in the sample after the strict sample selection delineated above. Nonetheless, the
[22] This is somewhat ad hoc and driven by the tradeoff between a larger sample size and a higher risk of measurement error due to outliers. I opted in favor of the sample size and only trim 1 per cent, which is conservative compared to e.g. Gernandt and Pfeiffer [2007] who trim the top and bottom 2 per cent of their sample. As a result, I lose 435 observations.
[23] These measures are described in the Online Appendix, Section 3. Notice that DLS use daily wages in their analysis which include bonus payments only from 1984 on. They correct for this break, which had been stressed to cause spurious increases in inequality, by a similar algorithm as used by Fitzenberger [1999].
8
-
-
-
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X. -
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X. -
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X. -
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X. -
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X. -
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X. -
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X. -
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X. -
-
Téléchargez vos propres textes! Gagnez de l'argent et un iPhone X.