Employment Vulnerability and Earnings in Kyrgyzstan

Abstract Employment vulnerability is considered as working under inadequate conditions. This research examined the impact of employment vulnerability on earnings with special reference to gender-based differences. Analyses were based on panel data for 2010–2013 and 2016 from the household survey. A panel data fixed-effects model with instrumental variable within the Lewbel (2012) method was applied to estimate an earnings equation. Results indicates negative impact of employment vulnerability on earnings. Women experienced this negative effect more severely. Given these empirical findings, government labour-market policy should not focus on increasing employment alone, but should also address the issue of vulnerability of employment, improvement of working conditions, and expanding employment opportunities for women.


Introduction
Most labour market analyses in the context of youth and women empowerment focus on reducing unemployment and increasing the labour supply (Blundell, Costa Dias, Meghir, & Shaw, 2016;Kluve et al., 2019;Mammen & Paxson, 2000;O'Reilly et al., 2015). Both in low-and high-income countries, however, employment arrangements vary by stability, security, and contract types. Under poor economic conditions and limited employment opportunities, individuals may face unstable remuneration, under-employment and part-time work, and informal employment and, consequently, look for additional employment (Bick, Fuchs-Schundeln, & Lagakos, 2018;McCaig & Pavcnik, 2015). Therefore, the fact of being employed did not guarantee long term job stability and security, and most of the workers in developing countries face vulnerable employment (Gammarano, 2018). This tendency has direct implications for poverty and the vulnerability of households (Chen, Vanek, & Heintz, 2006). Moreover, employment vulnerability was reflected in differences in wages. Employment vulnerability in general can be defined 'as the risk of working under inadequate conditions' (Bazillier, Boboc, & Calavrezo, 2016, p. 267). The empirical literature has focused on such aspects of employment vulnerability as informality, job satisfaction, and adverse treatment by employers (Bargain & Kwenda, 2014;B€ ockerman & Ilmakunnas, 2006;Botelho & Ponczek, 2011;Forth & Bewley, 2010;Maloney, 2004;Nguyen, Nordman, & Roubaud, 2013).
A limited number of studies, however, have analysed employment vulnerability as a multidimensional concept and its effects in developing countries (Bocquier, Nordman, & Vescovo, 2010;Gokhool, Kasseeah, & Tandrayen-Ragoobur, 2018). These studies were focused on determinants of employment vulnerability and its effect on earnings, though differential effects by gender in the labour market of developing countries were one important issue for labour market policy. In response, this paper analyses the effect of employment vulnerability on earnings in Kyrgyzstan, exploring for differential effects by gender. Within the theoretical and empirical literature, the effect of employment vulnerability on women has been ambiguous. On the one hand, women may have different preferences for working conditions and wages than do men. Non-pecuniary job characteristics may dominate in the utility function of women (Bertrand, 2011;Bonin, Dohmen, Falk, Huffman, & Sunde, 2007;Nguyen et al., 2013). Therefore, women may be less likely to be in vulnerable employment. On the other hand, increasing women's employment, for example, may increase the burden of combining labour market work with household work (Bertrand, Duflo, & Mullainathan, 2004;Van den Broeck & Maertens, 2017). At the same time, empirical findings about the gender-wage gap and other forms of gender discrimination in developing countries suggest that women may well be negatively affected by employment vulnerability (Bue, Le, Silva, & Sen, 2022).
Kyrgyzstan is a landlocked and lower-middle-income country in the Europe and Central Asia region. Economic transition since independence in 1991 has led to structural changes in the economy that have not generated adequate job opportunities for a growing labour supply (Schwegler-Rohmeis, Mummert, & Jarck, 2013). Analytical reports, however, have noted that employment in Kyrgyzstan corresponded with poverty. In rural areas, higher poverty and lower unemployment were explained by seasonal employment and high under-employment (International Labour Office, 2016;World Bank, 2007). Moreover, the share of underutilized labour potential accounted for 38.7% of youth (International Labour Office, 2016). Official statistics indicate that more than 66% of the active labour force in Kyrgyzstan is employed informally (National Statistical Committee of the Kyrgyz Republic, 2022). Therefore, although job places were generated in the Kyrgyzstan labour market, vulnerability remained an important issue. These general economic and labour-market conditions offer an interesting context in which to explore the effects of employment vulnerability on earnings through gender differences.
For the investigation of employment-vulnerability effect on earnings, panel data for 2010-2013 and 2016 from the 'Life in Kyrgyzstan' household survey were used. A panel-data-fixedeffect model with instrumental variable was applied within the (Lewbel, 2012) method for estimating an earnings equation. Our findings indicated a negative impact of employment vulnerability on earnings. Employment in vulnerable jobs was not compensated at a higher rate.
Therefore, this study contributes in several ways to the literature on employment vulnerability. First, it contributes to the scarce literature on employment vulnerability in a developing country context. Second, this study uses longitudinal household-survey data for 2010-2016. This gave us an opportunity to collect comparable observations over the survey period and deal with endogeneity and selection bias caused by unobserved time-invariant characteristics. For this the instrumental variable approach suggested by Lewbel (2012) was applied. Third, we measured employment vulnerability through a multidimensional approach. This allowed assessment not only of a particular segment of the labour market, such as informality or unemployment, but also discussion of labour market fragility within the broader concept including employment stability, security, and so forth. Fourth, this study extends analysis of the effect of employment vulnerability with a specific focus on differential effects by gender.

Employment vulnerability concept
Creation of jobs and a decrease in unemployment were not the government's only labour-market goal. Macroeconomic instability and weak labour-market institutions may result in different forms of employment fragility, which may include atypical job contracts, increase in job turnover, and jobs that are unstable and insecure. Working under these conditions poses economic and psychological difficulties for individuals. These trends in multiple aspects of the labour market can be described within the 'employment vulnerability' concept.
Vulnerability can be defined as 'how hard it is for individuals to manage the risks or cope with the losses and costs associated with the occurrence of risky events or situations' (Bocquier et al., 2010(Bocquier et al., , p. 1297. Bazillier et al. (2016) noted three approaches to define employment vulnerability. The first was developed by the International Labour Office (2010) and defined workers in vulnerable employment simply as the sum of self-employed and unpaid family workers. However, this definition has some limitations. For instance, wage-earning individuals may face significant risk in employment, while self-employed workers may not necessarily be vulnerable. The second approach used income level and membership in trade unions to indicate employment vulnerability. Workers at lower income levels and whose working conditions had not been the result of agreements arranged by trade unions were considered vulnerable (Pollert & Charlwood, 2009;Saunders, 2003). However, this approach was more focused on income level than on working conditions. The third approach suggested a multidimensional measurement of employment vulnerability (Bocquier et al., 2010;Forth & Bewley, 2010;Saunders, 2003). Bocquier et al. (2010) proposed several indicators to construct an employment vulnerability variable. These variables included job security, job stability, under-employment, unwanted job, and unstable remuneration, among others. We used a multidimensional approach and followed Bocquier et al. (2010) in constructing our employment-vulnerability variable.

Employment vulnerability and earnings
Working conditions are considered as one of the fundamental factors explaining the gap in earnings (Duncan & Holmlund, 1983;Smith, 1979). The relationship between job characteristics and earnings is explained within the theory of compensating wage differentials. This theory suggests that nonpecuniary job characteristics such as risk to health, differences associated with the pollution, crime, flexibility of work schedules, fringe benefits within pay packages, and so forth are important for labour market equilibrium and can be reflected in the wages for any job. According to this theory, holding individual characteristics as a constant, a worker must be offered additional wages in order to accept the undesirable characteristics of a job. Rosen (1986) explained this as the supply and demand of the labour market, which took job characteristics into account. On the supply side, a worker sells their labour and buys the characteristics of their jobs while, on the demand side, employers buy the labour and characteristics of workers and sell the characteristics of jobs they offer to the market. At the market equilibrium, the worker finds job characteristics desirable, and the employer find the worker's characteristics desirable. Therefore, at the equilibrium level, wages consist of two distinct aspects: first, labour service and worker characteristics and, second, job characteristics. Under the assumption of a competitive market structure that includes unconstrained labour mobility, labour-market clearing, and perfect information, wages differ among workers (Purse, 2004). In order to induce workers to accept unfavourable job characteristics, an employer pays premium wages.
Therefore, the certain trade-off between wages and undesirable job characteristics is expected those who hold jobs with unpleasant characteristics can be expected to earn comparatively better remuneration (Brown, 1980). As a consequence of the different preferences of workers for working conditions and wages, different wages among workers occur, and individuals with different employment conditions appear. Undesirable job conditions reflect employment vulnerability and, hence, this theoretical framework can be used for understanding wage differentials within employment vulnerability. Although this study did not aim to test this theory, whether workers employed in more vulnerable jobs are paid more than workers in less vulnerable jobs is a question that can be examined.
Along with this general expectation, differences in earnings within the employment-vulnerability concept may appear because of the gender dimension. Filer (1985), further elaborating the compensating-wage-differential concept, argued that workers' gen der caused different preferences for working conditions and wages. The utility function of women workers was less explained by income incentives, while other non-pecuniary conditions were more pronounced (Nguyen et al., 2013). Relatively speaking, women tend to value non-wage conditions of the job more highly (flexibility in working hours, stability, better interpersonal relations, and less exposure to hazards, for example: see Bonin et al., 2007;Bertrand et al., 2004). Therefore, women may be less likely to accept vulnerable employment (Gokhool et al., 2018). The expected effect of differences in preferences for job characteristics had implications for differences in earnings by gender. Men and women with the same level of other characteristics, but with different preferences for working conditions, earn at different levels. Taking into consideration that the jobs preferred by men are less preferred by women because of unfavourable work conditions, one may expect that men's jobs require higher earnings than women's.
Thus, within the theory of compensating wage differentials, it can be argued that higher employment vulnerability increases earnings because of unfavourable job conditions. Moreover, differences in preferences for job characteristics between men and women lead to different choices regarding vulnerable employment and, hence, to different earnings. Apart from the theory of compensating wage differentials, however, the economic literature suggests that other factors in the labour market also affect earnings. The dual-market theory describes a labour-market structure that has implications for the different workers' groups with varying wages. This theory divides jobs into primary and secondary groups. Primary jobs are characterized by secure employment, good working conditions, and relatively higher earnings, while secondary jobs are unstable, afford little opportunity for advancement, and provide low earnings (Dickens & Lang, 1985).
Models incorporating labour-market searches and frictions in matching workers with firms have argued that wages do not increase with poor working conditions (Hwang, Mortensen, & Reed, 1998;Lang & Majumdar, 2004). Moreover, labour market frictions associated with the inability of workers to change jobs quickly and the monopsony of employers suggest that working conditions may not play a substantial role in variations of earnings (Manning, 2003).
Women's preferences were not the only reason for lower earnings, however. Discrimination in the labour market and labour-market segmentation may result in both vulnerable employment and lower earnings for women (England, Farkas, Kilbourne, & Dou, 1988). Several studies indicate that vulnerable employment is more widespread among women (Sparreboom & De Gier, 2008;International Labour Office, 2017). Therefore, the net effect of employment vulnerability on earnings both in general and by gender is ambiguous.
Most empirical studies have examined the effect of different aspects of employment vulnerability on earnings (Bargain & Kwenda, 2014;B€ ockerman & Ilmakunnas, 2006;Botelho & Ponczek, 2011;Forth & Bewley, 2010;Maloney, 2004;Nguyen et al., 2013;Nordman, Robilliard, & Roubaud, 2011;Perry et al., 2007). These studies have focused on such particular aspects of vulnerability as informality, adverse treatment in the workplace, and discomfort in the workplaces, while other aspects of employment vulnerability were not involved in the analysis. Among them, Nguyen et al. (2013) argued that the informal-sector earnings gap in Vietnam was explained by workers' job status and their relative position in the earnings distribution. Nordman et al. (2011) showed that, in seven West African cities, the gender earnings gap was explained by an education gap, sector-by-sector differences, and job informality. Only Employment vulnerability and earnings in Kyrgyzstan 1079 a few studies have analysed employment vulnerability as a multi-factorial concept and these have largely measured vulnerability and its demographic and labour-market characteristics without quantitative assessment of the consequences of vulnerability (Bazillier et al., 2016;Gokhool et al., 2018;Mowla & Somaya, 2011;Sparreboom & De Gier, 2008;Sparreboom & Shahnaz, 2007). To the best of our knowledge, only a few studies have evaluated the effect of employment vulnerability, as a multi-factorial concept, on earnings in developing countries. Thus, Bocquier et al. (2010) using cross-sectional data for seven capital cities of West African countries, investigated the impact of employment vulnerability on earnings and found a negative impact. They also studied non-linear relationship and found that, in some of the distribution, the effect of employment vulnerability on earnings was consistent with the theory of compensating wage differentials. This study was based on cross-sectional data and employed an instrumental variable approach. Although the use of a cross-sectional dataset within the instrumental variable approach was consistent with the general requirements of empirical analysis, it may not exclude fully unobserved heterogeneities. From this standpoint, the use of panel data takes into consideration time variations of the effect and gives an opportunity to control for time-unobserved heterogeneity.

Source of data and sample selection
This study is based on panel data from the 'Life in Kyrgyzstan' survey conducted annually from 2010 to 2013 and again in 2016. This is a multi-topic longitudinal survey of households and individuals in Kyrgyzstan, conducted by a consortium of institutions with the support, in various years, of the Volkswagen Foundation, DFID, and FAO. Adult members of the households from the initial sample were tracked over time. The first-wave sample included 3000 households. In general, the survey included 3000 households and more than 7000 individuals for each wave. The data is representative nationally and regionally.
The survey included a wide range of data including information on household characteristics (composition, education, child education, health, for example, employment, assets, shocks, social networks, income, and spending). At the individual level, the survey inquired about labour supply, education, health, security, social life, worries, and other factors. Availability of information at the individual level makes this dataset most relevant for our analysis of employment vulnerability. Therefore, we use survey data for 5 years (2010-2013 and 2016) and focus on salaried and self-employed workers.
To construct the panel dataset, a common individual identifier across five waves of the survey was used. Following the main objectivean investigation of the impact of employment vulnerability on earnings, we focused on salaried and self-employed individuals only, the 3940 observations for 5 survey years, in each year 788 observations. The loss of observations in a panel data set may raise concern about attrition bias in empirical analyses. However, one of the main reasons for such loss of observations was the refusal of respondents to participate in the survey. According to the survey information, refusal to participate was relatively high. Thus, out of 8,160 individuals interviewed in the first wave of the survey, 7364 were interviewed in the second wave (attrition of 796 individuals). Attrition in our sample paralleled this tendency. A loss of observations as a result of economic factors (such as the inability to find a job) was limited, while a relatively larger part was explained by other reasons (birth and childcare, retirement, disability, and so forth). 1 However, samples' attrition may potentially add bias to our estimates, thus limiting the interpretation of results.
3.2. Construction of the key variables 3.2.1. The employment-vulnerability index. For the construction of the employment-vulnerability variable, we followed Bocquier et al. (2010). In this approach, the use of a number of indicators of employment conditions was believed to better describe employment vulnerability. Based on the scope of the data used in our study, employment vulnerability was measured through these seven dichotomous variables: Contractual insecurity and unregistered self-employment. This variable concerned contractual insecurity of wage workers and the business informality of self-employed workers. This indicator reflected potential difficulty for both type of workers to secure employment as a result of the absence of written contracts or the informality of the self-employment. Moreover, those without written contracts and who were registered as self-employed were not always eligible for social-security services. This indicator therefore encompassed workers' social-security status as well. This dummy variable took the value of 1 if wage workers had no written contract for their current job or did not use the workbook in the current employment and 0 if otherwise. For self-employed workers, it took the value of 1 if the business was not registered officially and 0 if otherwise. Bocquier et al. (2010) used this variable only for wage workers and defined it as contractual security. However, the vulnerability of self-employed workers may have negative implications for their employment, too, including the inability to protect employment rights and limited access to some government services. Therefore, inclusion of this aspect extended the Employment-Vulnerability Index to reflect the employment conditions of self-employed workers. Other employees. This variable concerned self-employed workers only and equalled 1 if they had no employees. Thus, self-employed workers without employees were considered vulnerable. The presence of other employees showed that independent workers acted as employers and offered positions to others. Thus, such workers were not vulnerable worker, but rather could be considered as entrepreneurs with managerial positions. Under-employment. Under-employment indicated a situation where working hours were less than the general statutory work week. This variable took the value of 1 if the individual worked less than thirty-five hours per week and 0 otherwise. This variable encompassed the situations of workers who would have preferred to work more hours. Additional employment. Holding a second (additional) job reflected the instability of or insufficient income from the main job, which may have reflected under-employment and attempt to diffuse the risk of income loss in the main job. It took the value of 1 if the respondent had an additional vulnerable job. Additional jobs at the managerial, professional, and associate professional levels were excluded from the category of vulnerable jobs. Job instability. Instability in one's job was another aspect of employment vulnerability and could cause quick changes in individuals' workplaces. Bocquier et al. (2010) defined this as job change without improvement in job position or with a drop in position. However, the survey data used in our study did not specify job-position transitions resulting from each job change. Therefore, we derived this variable from the description of the activity in the previous 12 months. Any job changes during the 12-month for any reasons was considered evidence of instability, except in the cases of maternity leave, education, internship, retirement, or military service. If there was a change in job, it took the value of 1. Wage arrears. Vulnerable employment is mostly reflected in wage arrears of workers. This variable was derived from the question on if worker had any wage arrears during the last 12 months. This criterion is used only for wage employee.
The vulnerability dummy variables given above were summed up and reflect employment vulnerability intensity. Therefore, our Employment-Vulnerability Index is continuous variable and Employment vulnerability and earnings in Kyrgyzstan 1081 ranges from 0 to 5 for each individual. If the index equals to 0, then there was no employment vulnerability. One may argue that this approach for construction of the Employment-Vulnerability Index assumes that each dimension was contributing equally to employment vulnerability. Although, importance of one dimension for the aggregated index may vary. However, corresponding literature did not provide with the conceptual framework to incorporate potential differential weights for the Employment-Vulnerability Index. Therefore, in this study following earlier research by Bocquier et al. (2010) Employment-Vulnerability Index was constructed by summing up dimensions. However, we also analyse the impact of employment vulnerability via the Principal Component Analysis to allow for the differential weights of components in the index.
3.2.1. Earnings. To examine the effects of employment vulnerability on earnings, the log value of monthly earnings was estimated. Earnings consisted of the monthly wages of salaried employees and the monthly profit of self-employed workers. In the survey, wages and profits could be indicated on a daily, weekly, or monthly basis; these were all converted to monthly values and expressed in units of the national currency. Earnings were deflated based on the regional Consumer Price Index.

Descriptive evidence
Average employment-vulnerability index for survey years were given in the Table 1. The employment-vulnerability index varies from 0.82 to 1.1.074. Average value of employment vulnerability for women increases by survey years, whereas for men it has slightly decreasing tendency. Though in absolute terms employment vulnerability is higher for men. Components of the employment-vulnerability index show that the main source of employment vulnerability was the contractual security and business informality. Thus, about 30% salaried employees and self-self-employed workers did not have employment contracts and their business was not registered officially. Also, under-employment and job stability were another important aspect of employment vulnerability. Almost 87% of self-employed workers did not have any other employees. Remaining components: additional employment, job stability and overdue wages demonstrates relatively lower mean values. Table 2 presents summary statistics of variables used in the empirical analysis both for the total sample and by survey years. For the 2010-2016 period on average, earnings were about 7536 soms (the national currency of Kyrgyzstan, abbreviated KGS). Out of the total sample, 32% individuals were self-employed workers, while the remainder wage workers. Average age was 41 years and 65% were men. The sample of women consisted of 1390 individuals and accounted for 35%. Almost all of the workers were educated (95%), and one-third of these had a tertiary education (34%). Most were in low-skilled jobs, while professional and associated professional positions accounted for 30%. Household size in average was more than four persons per household. Table 2 presents summary statistics of variables used in the empirical analysis both for the total sample and by survey years. For the 2010-2016 period on average, earnings were about 7536 soms (the national currency of Kyrgyzstan, abbreviated KGS). Out of the total sample, 32% individuals were self-employed workers, while the remainder wage workers. Average age was 41 years and 65% were men. The sample of women consisted of 1390 individuals and accounted for 35%. Almost all of the workers were educated (95%), and one-third of these had a tertiary education (34%). Most were in low-skilled jobs, while professional and associated professional positions accounted for 30%. Household size in average was more than four persons per household.

Employment vulnerability and earnings in Kyrgyzstan 1083
Correlations between earnings and employment vulnerability are given in the distribution in Figures S1 and S2 of the supplementary materials. Inspection of the figure indicates that employment vulnerability for men has intensity up to the index value at 4. Along with this, trendline decrease as values as the vulnerability index grew. In the subsample of women most observations are grouped around the Employment-Vulnerability Index 3, though it also has decreasing earnings at higher employment vulnerability. These descriptive trends show negative correlation between earnings and the Employment vulnerability index.

Fixed effect model
Most empirical studies on aspects of employment vulnerability have used cross-sectional variation across individuals to estimate earnings. By using panel data in this research, however, we are able to control for time-invariant individual unobserved heterogeneity through the inclusion of individual fixed effects. These allow us to compare earnings for the same individual over the 2010-2016 period covered by the data. The benchmark equation for estimation is: where y it denotes the dependent variable, which is in our case the earnings of the individual; X it is the vector of characteristics of individual i observed at time t (which included a constant term); and I it represents the employment vulnerability of individual i at time t. v i is the individual fixed effect, d t the time specific effects, and an it i.i.d. normally distributed stochastic term accounting for possible measurement error E( it jX it , I it , v i ) ¼ 0, for all individuals i and periods t. Coefficient c measures the impact of employment vulnerability on earnings. The vector X it typically includes variables underlining individual and household characteristics. Hence the set of explanatory variables includes age of individual, dummy variables for education levels, household size, number of children in the household up to 6 ages, and dummy variables for individual's sector of employment and job rank. Senior manager is considered as the highestranking position.
A potential threat to the identification strategy used in Equation 1 is the presence of unobserved individual characteristics that varied over time, affecting both employment vulnerability and outcomes, and thus not captured by the individual fixed effects. This, in turn, causes an endogeneity issue. While the fixed-effects model eliminates the effect of unobservable timeinvariant characteristics of employment vulnerability and earnings, time-varying unobservables affecting individuals' employment vulnerability and earnings could remain in the error term. To deal with this issue, we used an instrumental-variable approach in addition to individual fixed effects (IV-FE).
Another methodological concern was self-selection into the labour market. In cross-sectional models, Heckman's correction has mostly been used to address selection bias. In a panel-data setting, however, correction for the selection bias was non-trivial and required an instrument for the selection equation (Semykina & Wooldridge, 2010). Moreover, estimation of this type of equation was complicated when the explanatory variable was endogenous (Semykina, 2018). Taking into account non-trivial solutions, our empirical strategy did not fully address potential selection bias. Therefore, this may represent a limitation in interpreting estimation results.

Instrumental variable model and Lewbel's approach
The instrumental variable used to correct for endogeneity should be correlated to employment vulnerability and meet the exclusion restrictionthat is, it should not affect the outcome variable except through its effect on employment vulnerability. Finding suitable instruments that fulfilled these requirements for panel data, which should involve time-variant variables, was difficult. Moreover, one instrumental variable was not necessarily valid for both outcome variables. Exclusion restrictions necessitated the use of different instrumental variables for the earnings and subjective-well-being equations. On the one hand, the difficulty in finding good instrument was conditioned by the scarce literature on employment vulnerability. Most previous studies have been based on cross-sectional data, suggesting time-invariant variables as an instrumental variable. Thus, among the literature Bocquier et al. (2010) used dummy variables related to the status of the head of household and father when the individual was 15 years old. However, such instrument was not valid within our dataset for main two rea sons. First, the structure of panel data necessitated the use of time-variant instrumental variables. Second, in our study, employment vulnerability was a multidimensional concept and, hence, such aspects as unemployment would have explained only a limited part of variation in employment vulnerability.
We propose shocks experienced by households as an instrumental variable for the earnings equation. This variable was generated by asking households whether they had experienced shocks during the previous year related to the inability of the household to sell agricultural products, loss of job of household member, death of breadwinner, death of household member, illness of breadwinner, illness, divorce, or accident experienced by household member. Generally, this variable has not been used extensively in previous studies as an instrumental variable. The rationale for this variable is that shocks experienced by households during the previous year may have created incentives for individuals to accept vulnerable jobs, thus increasing employment vulnerability. Because shocks were measured at the household level, however, they may not necessarily be correlated to the level of earnings of any individual.
There may be concern about the validity of household shocks as an instrumental variable for the earnings equation. Given the scarce literature on employment vulnerability, this concern cannot be neglected. Availability of more than one instrumental variable would, to some extent, ensure the robustness of estimation results. As mentioned above, however, determination of time-variable instrumental variables was a non-trivial approach. As an alternative strategy, therefore, we estimated the earnings equation using the Lewbel (2012) two-stage heteroscedasticity-based instrument approach. This approach may be applied when no external instruments are available or can be used along with external instruments to improve the efficiency of the instrumental variable's estimator, in particular when instrumental variables are weak. Identification was implemented through the estimation of regressors which were uncorrelated to the product of heteroskedastic errors. Equation 1 was modified according to Lewbel method as a linear triangular model, as follows: (2) where I it is the observed endogenous variable. In our case, the employment-vulnerability index, X it is a vector of observed exogenous regressors, and it ¼ ( 1it , 2it ) was the vector of unobserved errors. In such models, where no instruments are available, the Lewbel method bases identification on higher moments, restricting correlations of 2it 0 it with X and assuming the heteroskedasticity of 2it . To generate instruments, the residuals of the first-stage regression were multiplied by each mean-centered control variable presented in the X it vector.
The other feature of Lewbel method was to increase the efficiency of exogenous instruments in the model (Lewbel, 2012) and equations (2) and (3) could be augmented as follows, correspondingly: Employment vulnerability and earnings in Kyrgyzstan 1085 The quality of generated instruments and proposed exogenous instruments could be checked with corresponding tests on under-identification, over-identification, or weak identification.1 For the purposes of comparison, the earnings equation was estimated by fixed effect and fixed effect with instrumental variable (hereafter, IV) within the Lewbel (2012) method.
4.2.1. Validity of the instrument. The literature on earnings and its relationship to employment conditions is discordant regarding the considerable differences between genders. To take account this heterogeneous impact, all estimations were further implemented by drawing subsamples of men and women.
First-stage results for the earnings equation within the Lewbel (2012) approach are given in Table S2 of the supplementary materials. The first column of the table refers to the use of solely an external instrument (or household shock in our case). The, second column includes only internally generated variables using Lewbel's (2012) approach, and the third column shows the use of both internally generated instrumental variables and of the house hold shock variable as an instrumental variable. In general, the first and third columns indicate that household shocks had a significant positive impact on employment vulnerability and greater shocks more increased individuals' employment vulnerability. This result was consistent with our expectation that, under household-shocks conditions, individuals would be more inclined to accept vulnerable jobs.
The under-identification test reports the p-value of the Kleibergen-Paap LM statistic on the null hypothesis that there was no correlation between the tested instruments and the endogenous regressors. Though, results do not validate fully weak identification test, suggesting potential weakness of the IV. The over-identification test reports the p-value of the Hansen J statistic. This test did not reject the instruments' validity, which means that one cannot jointly reject lack of correlation to the error term. Following these validity analyses, estimation results are presented with exogenous IV, internally generated IV and both IV with Generated Instruments and External Instruments within the Lewbel (2012) method outputs.

Estimation results
For the purposes of comparison, estimation results for the earnings equation are re-ported based panel fixed effects, and IV approach within the Lewbel method, which included exogenous IV, internally generated instrument results and both internally generated and exogenous instrument results. In Table 3, only coefficients of the main variable of interestemployment vulnerability was presented by panel fixed effects, and Lewbel method estimation outputs. Estimations were based on total sample and by gender subsamples. Full estimation results with other explanatory covariates are available in Tables S3-S5 of the supplementary materials.
All estimation methods for samples revealed that employment vulnerability had a negative impact on earnings. While FE estimates showed statistically insignificant results, IV approach reveals significant negative effect for total sample and particularly negative higher impact on men.
Hence, disregarding selection bias and the endogeneity of employment vulnerability would result in biased estimation outcomes. Additional unit of employment vulnerability reduced individuals' earnings by 22%, while for men it was 23%.
In general, this finding is in line with previous empirical studies by Bocquier et al. (2010) that employment vulnerability decreased earnings. Indeed, the negative effect of employment vulnerability on earnings demonstrates that propositions of the theory of compensating differentials were not applicable in this case. Results show that individuals did not receive greater compensation for accepting vulnerable employment. This was probably related to tight conditions and limited opportunities in the labour market. Under these conditions, individuals did not have a wide range of choice to select jobs according to vulnerability level or to bargain on earnings. Moreover, difference in negative effects by gender with more severe negative impact on men falls in line with previous empirical studies that women may be less involved in vulnerable employment (Gokhool et al., 2018).
Other explanatory variables revealed several interesting findings (see Tables S3-S5 of the supplementary materials). Generally, explanatory covariates demonstrated different effects on gender subsamples. Household composition showed significant impact on earnings for women. Household size had a positive effect on their earnings, while with increase of children number in the household the men are more likely to earn more. These results may indicate that a higher number of children in a household created pressure to seek employment opportunities and that effect was reflected in the behaviour of adult men in the household.
For women, the one of main significant variable was education level. Women with secondary technical education earned more than other counterparts, while for men both secondary technical and tertiary education is highly significant. These results point out two important arguments. First, education is an important factor for earnings potential. Higher education provides better opportunities in the labour market for all. Second, it may indicate differential selection conditions in the labour market by gender.
Despite our expectations, number of children had no impact on women's earnings. This may have been the result of the fact that women with children up to age 5 did not show active labour-market participation. Among other results from explanatory variables was that both men and women with higher-ranking job positions were more likely to earn more than their counterparts with lower-ranking job positions.
The Employment-Vulnerability Index constructed by summing up all aspects of vulnerability assumed that all aspects had the same weight. However, one may argue that factors influencing vulnerability have different weights. Therefore, for a robustness analysis Principal Component Analysis (PCA) has been applied to Employment-Vulnerability Index for wage employee and self-employed samples separately. For this wage employee and self-employed samples are separately estimated. The first four orthogonal axes from the PCA for both samples are taken into consideration as they sufficiently explain large proportion of variance and represent components of the Employment-Vulnerability Index (Table S6 of the supplementary materials). These axes are used in explaining differences in earnings based on the Lewbel approach (Table 4).
Results indicate that the negative effect of employment vulnerability on earnings was statistically significant for self-employed women. This evidence is confirmed in three axes of the PCA. From Table S6 of the supplementary materials it can be noted that the first and fourth axes of self-employed sample mostly include unregistered business and other employees' criteria, whereas the third axes is mainly represented by under-employment and job instability. Negative effects for men are observed in both sub-samples, though at less statistical significance level. These findings confirm the negative effect of the employment vulnerability on men. However, also it reveals the strong negative impact of employment vulnerability on selfemployed women. In particular, self-employed women with unregistered business, job instability and underemployment face lower earnings.

Conclusions and policy implications
This study aimed to investigate the impact of employment vulnerability in Kyrgyzstan on earnings and subjective well-being. Research was based on the panel data from the household survey for 2010-2013 and 2016 years. Following the previous studies employment-vulnerability index and subjective well-being index were constructed. In order to deal with the potential timevariant unobserved characteristics instrumental variable approach suggested by Lewbel (2012) was applied for the analysis of the impact of employment vulnerability on earnings. This method provides with generated instruments with additional external instruments for estimation of the model under the condition of heteroscedasticity of the error terms of the endogenous covariates.
Our findings suggest that employment vulnerability had a negative effect on earnings. In general, this was in line with the observation of Bocquier et al. (2010) that employment vulnerability decreased earnings. Individuals did not receive compensating earnings for being involved in vulnerable employment. Moreover, the negative impact was severe for self-employed women, reflecting the fact that employment vulnerability decreased earnings of self-employed women at higher rates than it did for men. The gender bias of employment vulnerability on earnings had not been observed in previous empirical studies.
One potential limitation of this study is that we did not fully address potential selection bias. Hence, possible time-varying unobserved heterogeneities may have influenced estimation results. However, solution of selection-bias and endogeneity problems in a panel data setting is non-trivial. Moreover, given the multidimensional context of employment vulnerability, use of panel data fixed effects represented better empirical evidence.
These findings have important policy implications. The importance of employment vulnerability for earnings leads to the necessity to conceptualize labour-market policy that focuses not only on increasing employment but also on addressing vulnerability of employment and improvement of working conditions. Such a policy would contribute both to improvement of labour productivity and to the reduction of earnings inequality in the long-run. One important direction for policy in this field is contractual security. Most employment vulnerability is related to this issue, which may have long-term negative effects through limiting access to social-security services. From this standpoint, government policy should take into account the sector-by-sector specifics of vulnerable jobs and up-date social-security policy by making enrolment in social-security less costly. The most effective mechanism for this might be revision of the social-security contribution rate and simplification of procedures for vulnerable workers.
Policy should also address the significant negative effects of employment vulnerability on women. An action plan for empowering women in the labour market could be implemented with special focus on those whose earnings are lowest. Such a plan could aim to increase earning capacity by incorporating multidimensional education programs (development of entrepreneurial skills, raising awareness about laws regulating working conditions, and financial resources for business development, among others). Note 1. Estimation of the difference of means between our final sample and the attritted sample showed statistical significance for several covariates. Results are available from authors upon request.