Chapter 4. Establishing a profile of disadvantaged youth (Module 2)

Step 1. Select the data

Employment. Information on employment can be found in labour force surveys or in other household surveys with an employment module. When analysing the school-to-work transition, panel data should be preferred, since they follow individuals over multiple time periods. In the absence of panel data, cross-section data with retrospective information on labour market history could be used, such as the ILO school-to-work transition surveys (SWTSs), which have been widely applied to analyse this phenomenon.

Education and skills. Population censuses and household surveys, including labour force surveys, are generally used to analyse education and skills. Learning achievements and cognitive skills in particular can be analysed through a number of international surveys, including the Organisation for Economic Co-operation and Development (OECD) Programme for International Student Assessment (PISA) and the Young Lives longitudinal study of child poverty.

Health. Health and SRH in particular are best analysed using population censuses and household surveys, such as the demographic and health surveys (DHS) and the Centers for Disease Control and Prevention (CDC)-assisted reproductive health surveys. Information on substance use can be found in specialised surveys, including the following two WHO-supported school-based surveys: the Global school-based student health surveys (GSHS) and the Health Behaviour in School-aged Children surveys (HBSC).

Participation and empowerment. Data on social capital and civic and political engagement can be obtained from international surveys, such as the Gallup World Poll surveys and the World Values Surveys. Similar opinion polls conducted at the regional level, including Afrobarometer, Latinobarometer and Asianbarometer, constitute rich sources of information and hence must be exploited to document issues related to youth participation and empowerment. With regard to youth violence, cross-national information is difficult to retrieve, as it is for other risky behaviours, since it is only available in a few specialised international surveys covering a limited number of countries, such as the GSHS and HBSC surveys.

Step 2. Define the dependent variable

The dependent variables used are qualitative variables that take two or more values. In the case of a binary qualitative variable, the dependent variable takes value 1 for youth experiencing the well-being deficit and 0 for youth not affected by the well-being deficit. This coding is used to distinguish the outcome of interest (experiencing the well-being deficit), which is coded 1, from the reference or base outcome (not being affected by the well-being deficit), which is coded 0.

In the case of a qualitative variable with multiple categories, the dependent variable takes several values corresponding to different well-being outcomes. The dependent variable is not ordinal because well-being outcomes cannot be ranked from worst to best in an unquestionable way. However, some well-being outcomes are clearly more favourable than others. The outcome assumed to be the most favourable is considered the reference category and coded 0 accordingly. The way the other outcomes are ordered does not matter, since they are all analysed separately in comparison to the reference category.

Step 3. Select an appropriate estimation model

In the case of a binary qualitative dependent variable, models that can be used include linear and non-linear probability models, such as the logit and probit regression. These models allow deriving the partial effect of the independent variables – i.e. the effect of each independent variable, holding all the others fixed – on the probability that the outcome of interest has occurred.

Of the two, non-probability models are more popular in the empirical literature. The linear probability model, which is estimated using ordinary least square regression, is less adapted when dealing with a binary dependent variable because its predicted values can fall outside the unit interval (be negative or exceed 1) for some values of the explanatory variables. This is actually not very problematic if partial effects are estimated at the mean of the independent variables or if most independent variables are discrete and contain only a few values, such as dummies (Wooldridge, 2002).

Logit and probit models are basically the same: the dependent variable is generally modelled using a cumulative distribution function, and the coefficients are estimated by maximum likelihood. They only differ in that the error term does not have the same distribution (standard logistic in the logit model versus standard normal in the probit model). Both models can be used equally to analyse the determinants of a particular well-being deficit. When the dependent variable is qualitative and has more than two categories, multinomial non-linear probability models, such as the multinomial logit regression, are the most appropriate.

Step 4. Specify the independent variables of the model

Independent variables can be either continuous or categorical. The effect of a continuous independent variable on the dependent variable is not necessarily linear. This is the case, for example, for the variable ‘age’. When the relationship between the two variables is suspected to be non-linear, it is preferable to add in the model the square of the independent variable to evidence whether its effect decreases, remains constant or increases as the variable reaches higher values. The squared variable should be divided over a certain factor (for example, 100 in the case of age) so as to offset any effect of scale.

For interpretational purposes, categorical variables, such as sex and levels of education, must be introduced as dummies in the model (dichotomous variables coded 0/1). When the categorical variable has more than two categories, as is the case for levels of education, one of the categories must be left out to serve as the reference category to which the other categories are compared.

Step 5. Identify the determinants of the well-being deficit

Run the model and display the results. To identify the determinants of the well-being deficit, test the statistical significance of the coefficients of the independent variables and look at their sign. Wald tests can be performed to determine if each coefficient is statistically different from zero. If the results of the Wald tests show that coefficients are statistically significant at a reasonably confidence level (95% or 99%), then we can assume that the corresponding independent variables have an effect on the well-being deficit.

Focus only on independent variables with statistically significant coefficients and look at their sign. If the sign is positive, it means that the independent variable increases the likelihood of experiencing the well-being deficit. In that case, the independent variable can be considered a risk factor. Conversely, if the sign is negative, the likelihood is reduced, and the independent variable is assumed to be a protective factor. From the above, disadvantaged youth can be defined as all youth exhibiting risk factors, i.e. with characteristics that make them more likely to experience the well-being deficit.

In logit and probit models, as well as in other non-linear probability models, such as the multinomial logit, the value of coefficients is not meaningful. It is therefore necessary to compute partial or marginal effects to quantify the effects of risk factors on the well-being deficit. The marginal effect is computed separately for each independent variable and corresponds to the change in the probability of experiencing the well-being deficit induced by a marginal change in the independent variable.

The interpretation of the marginal effect differs depending on the nature of the independent variable. If the independent variable is a dummy, like gender (takes value 1 for female and 0 for male), then the marginal effect corresponds to the change in probability when the independent variable changes from 0 to 1. For example, in the case of gender, the marginal effect is the probability of experiencing the well-being deficit when the young individual is female as compared to male. If the independent variable is instead discrete with more than two values, like age, the marginal effect is then interpreted as the change in probability following a one-unit change in the independent variable, which, in the case of age, corresponds to the situation of being one year older. Lastly, when the independent variable is continuous (takes any value within a specified range), the marginal effect is defined as the change in probability of experiencing the well-being deficit induced by an infinitesimal change of the independent variable (1% change in practice).

Model 1: Determinants of incomplete school-to-work transition

In this model, the ILO definition of school-to-work transition stages can be used to differentiate between youth in transition and youth who completed the transition (see, for instance, ILO, 2013a; Elder and Koné, 2014).

Youth in transition are in one of the following situations:

• unemployed, according to the relaxed definition (not employed and available but not necessarily looking for work)

• employed in a temporary and non-satisfactory job

• employed in non-satisfactory self-employment

• inactive non-students with future plans to work

Youth having completed the transition are in one of the following situations:

• employed in a stable job, whether satisfactory or non-satisfactory

• employed in a satisfactory but temporary job

• employed in satisfactory self-employment.

Stable employment is defined as including all employees with a contract of employment, either written or oral, of more than 12 months’ duration. Employees with a contract of no more than a year are considered in temporary employment. Satisfactory and non-satisfactory employment are defined based on a subjective question asking workers whether they are satisfied (either very satisfied or somewhat satisfied, in the SWTSs) or not with their current jobs.

Individual, family and community characteristics, such as those reported in Table 4.2, have an impact on school-to-work completion. At the individual level, the model should at least include independent variables for age, gender and level of education attained or completed. Migratory and disability statuses should also be considered, among other relevant individual characteristics. If using SWTSs, some retrospective information on individuals’ labour market history should be added, such as the number of past spells of unemployment, temporary employment or self-employment (Shehu and Nilsson, 2014) or whether youth were working while in school (Rosati, Ranzani and Manacorda, 2015).

At the family level, poverty or scarcity needs to be taken into account whenever possible. This can be done by introducing a dummy variable for poverty status or, alternatively, by introducing several dummies identifying the different quantiles of the household income or consumption expenditure distribution in which youth are located. Parental education can be used as a proxy in the absence of information on household income or consumption expenditure. Other relevant family characteristics include household head status and the presence of children.

At the community level, location variables are important, given the large spatial disparities commonly documented in the literature. Dummies for urban/rural area of residence and regions are recommended in this regard.

Note that the literature proposes other interesting analyses and estimation methods relative to the school-to-work transition that will not be covered here. Rosati, Ranzani and Manacorda analyse simultaneously the probability that youth never transit to employment, either to a first job or to stable employment, and the duration of the transition among those who are expected to transition, using a split population model (2015). Elder and Koné propose, in turn, three different models estimated using the logit regression to analyse the probability of making the transition, transitioning into stable employment versus satisfactory temporary work or self-employment, and making a short transition (2014). Similarly, Shehu and Nilsson estimate three different models using the probit regression to analyse the probability of making the transition, transitioning into satisfactory temporary work or self-employment, and transitioning into stable employment (2014).

These last two models, be it estimated with a logit or probit regression, are good alternatives to the one suggested here. In fact, not all jobs held by youth who completed the school-to-work transition are of equal quality. It could be interesting, then, to run two additional models that distinguish jobs of higher quality from those of lower quality, such as stable employment versus satisfactory temporary work or self-employment. Analysing the length of the school-to-work transition provides further useful information on this phenomenon but has the major drawback of excluding youth who never went to school. In some developing countries, the latter amount, in fact, to a significant share of the total young population.

Model 2: Determinants of selection into different school-to-work transition pathways

This model requires panel data with successive samples or rotation groups in which individuals are observed on a monthly basis for a long period (e.g. 48 months) (OECD, 2013). In the absence of panel data, cross-section data with retrospective information on labour market history (e.g. from SWTSs) could be used.

The sample must be restricted to individuals defined as youth and enrolled in education at the beginning of the observation period. In each month of the observation period, these individuals must be classified into one of the following labour market states: i) in employment; ii) in education; iii) in employment and education; iv) not in employment; and v) inactive NEET.

The dependent variable is a categorical variable which has the following values:

• 0 – for early workers: find employment relatively soon

• 1 – for students: remain in education during the observation period

• 2 – for working students: combine school and work

• 3 – for late workers: find employment towards the end of the observation period

• 4 – for job-seekers: enter unemployment

• 5 – for labour force drop-outs: become inactive NEETs

Each of these categories corresponds to a different school-to-work pathway. The most successful pathway is the one followed by early workers who find employment relatively soon after leaving the educational system. This category is therefore considered the reference category and coded 0 accordingly.

The dependent variable is not ordinal because the different pathways cannot be ranked from worst to best in an unquestionable way. However, some pathways are clearly less favourable than others. This is undoubtedly the case for youth who do not access employment after leaving school and become unemployed (job-seekers) or, worse, inactive (labour force drop-outs). The primary focus of this model will be on these two last categories.

These pathways are defined after conducting a cluster analysis on individual trajectories based on the following two elements: i) their labour market state at the end of the trajectory; and ii) their transition paths (the relative frequency of each labour market state and the number of transitions between different states over the observation period). The cluster analysis is performed using the Ward’s hierarchical agglomeration algorithm (1963).

Some of the factors reported in Table 4.2 could be taken into account, since they are very likely to affect the probability of following a negative school-to-work pathway and in particular becoming unemployed or inactive non-students (OECD, 2013; Eurofound, 2012; Bills et al., 2014). As the analysis is conducted using panel data, it is recommended to add dummies to control for year or month effects. If the data allow it, some macro factors disaggregated at the local level could be added, such as fertility and youth unemployment rate (OECD, 2013).

Model 3: Determinants of informal employment

Youth are considered to be working informally if they belong to one of the following categories:

• Formal sector enterprises: contributing family workers and informal employees

• Informal sector enterprises: own-account workers, employers, contributing family workers, informal employees and members of producer co-operatives

• Households: own-account workers and informal employees.

The informal sector is composed of unregistered businesses with a small number of workers (in practice, the threshold is usually set at five or ten), and informal employees are defined as those without access to at least one of the key employment benefits, such as annual paid leave, paid sick leave or social security contributions.

Factors displayed in Table 4.2 are believed to play a significant role in the probability of falling into informal employment (see, for instance, Shehu and Nilsson, 2014). If using SWTSs or similar surveys, information on labour market history could be added to the model, such as the number and length of past employment spells, as well as the length of the transition from school to the labour market (ibid.).

Model 1: Determinants of poor learning achievements

For this model, data from the Young Lives longitudinal study of child poverty can be used. This study covers four countries, namely Ethiopia, India, Peru and Viet Nam. Two age cohorts of children, born in 1994-95 and 2000-01 respectively, are traced in three successive rounds conducted in 2002, 2006-07 and 2009.2

The sample is restricted to the first cohort of the Young Lives study. Children are thus observed at ages 7/8 in 2002, 11/12 in 2006-07 and 14/15 in 2009. Poor learning achievements in literacy, numeracy and general cognitive skills are analysed separately for each of these ages whenever possible. The analysis is performed at different ages in order to determine how differently risk factors – in particular, past conditions and outcomes, such as prior cognitive and non-cognitive skills – can influence poor learning achievements over the childhood phase.

Model 2: Determinants of school drop-out

The analysis of school drop-out must be restricted to youth belonging to the official school-age population. School drop-out refers to youth who are currently out of school but were previously enrolled and left education before completion. If the information available is limited to school attendance at the time of the survey, school drop-out can be proxied by out-of-school youth.

The main risk factors leading to school drop-out are displayed in Table 4.3. Early parenthood, child labour and poverty are likely to be endogenous, i.e. correlated with unobserved factors that also affect the probability of dropping out of school. This can potentially bias the results of the model. A way to deal with this problem is to follow the instrumental variable approach. The literature proposes several instruments for these potential endogenous variables. In Cardoso and Verner, for example, early parenthood is instrumented with the declared ideal age to start sexual activity, child labour with the declared reservation wage, and poverty with self-reported hunger (2006).

It is important to account for child labour when the analysis is performed for developing countries, where child labour is frequent and constitutes one of the main obstacles to education, in particular to children’s persistence in the school system (Guarcello, Lyon and Rosati, 2008). Dropping out of school is also positively associated with early parenthood and poverty (Cardoso and Verner, 2006), as well as with educational factors, such as grade repetition and completion, poor learning achievements and low school quality (Branson, Hofmeyr and Lam, 2013; Hanushek, Lavy and Hitomi, 2006). Poor parental education and occupation, and the presence of siblings, in particular younger ones, are some of the other factors increasing the likelihood of school drop-out (Huisman and Smits, 2014).

Model 1: Determinants of adolescent pregnancy

Factors included in Table 4.4 are often found to be positively correlated with adolescent pregnancy or motherhood and could therefore be included in the model. According to the literature, adolescent pregnancy is positively associated with, among other factors, lack of parental support and connectedness, living with a single parent, residing in unsafe neighbourhoods, having older sexually active siblings or other pregnant or parenting teenaged sisters, and sexual abuse (Miller, Benson and Galbraith et al. 2001). Besides lack of parental support/connectedness, lack of parental regulation (supervision and monitoring) and poor parent/teen communication about sexual issues also increase the probability of adolescent pregnancy.

Model 2: Determinants of substance use

In the GSHS, current alcohol consumers are defined as those who consumed one or more alcoholic drinks in the past 30 days and current drug consumers as those who used illicit drugs at least once in the past 30 days. Illicit drugs include marijuana, amphetamines or methamphetamines, cocaine, solvents or inhalants, ecstasy, heroin and other country specific drugs. The GSHS concerns students aged 13-17.

Table 4.4 lists the main risk factors associated with substance abuse. Parental substance abuse in particular can lead to impaired parental control and child maltreatment and, subsequently, the consumption of substances by the child (Singh, Thornton and Tonmyr, 2011). Substance abuse contributes to risks during adolescence for injury, violence, poor cognitive skills development, unprotected sex and suicide attempts. In adulthood, it plays a role in risks for non-communicable diseases and injuries (WHO, 2014).

Model 1: Determinants of the lack of civic engagement

Civic engagement can be defined using three questions from the Gallup World Poll surveys. Individuals are asked whether, in the past month, they have i) volunteered their time to an organisation; ii) donated money to a charity; or iii) helped a stranger or someone they did not know who needed help. In this model, youth who answered no to all three are considered not civically engaged; those who answered yes to at least one question are considered civically engaged (to some extent).

Table 4.5 presents some factors that could be integrated into the model, since they are very likely to affect youth civic engagement. As regards personal attributes, youth civic engagement can be found deeply rooted in the process of identity formation during adolescence. The psychological literature shows that adolescents who are self-reflective and who actively explore identity alternatives before making decisions about the values, beliefs and goals that they will pursue have more non-cognitive skills and positive personality traits, develop a higher sense of social responsibility and are more engaged in communities (Crocetti, Erentaitė. and Žukauskienė., 2014; Crocetti, Jahromi and Meeus, 2012). This is in contrast to adolescents who more automatically conform to others’ prescriptions and expectations (parents or other reference groups) or who do not deal with identity issues.

At the family level, affluence is shown to be positively related to youth participation in community activities (Lenzi et al., 2012). Empirical studies focussing on the role of parents demonstrate, in addition, that parental civic values and behaviours have an influence on youth civic engagement (Atkins and Hart, 2003). Moreover, lack of parental support and monitoring in childhood is shown to have an adverse impact on future youth civic engagement (ibid.).

Community characteristics also play a significant role in youth civic engagement. For instance, youth are less likely to be civically engaged when living in poor urban neighbourhoods (Atkins and Hart, 2003) or in communities with low levels of social capital (social ties and trust) and school connectedness (Lenzi et al., 2013, 2012).

Model 2: Determinants of youth violence

The last model proposed examines the factors correlated with youth involvement in violence. Using the GSHS, youth can be defined as violent if they reported any of the following behaviours: i) belonged to a violent group (attacks with insults, bullying, hits, assault, robbery, or rape); ii) were repeatedly in a physical fight in the past 12 months; or iii) have carried a weapon (gun, knife, club or other) on several occasions in the past 30 days.

Characteristics displayed in Table 4.5 have been identified by the literature as contributing factors to violent youth behaviours (O’Brien et al., 2013; Burfeind and Batusch, 2016), and hence could be taken into account. Risk for violence among youth tends to increase in particular with mental disorders (e.g. attention-deficit hyperactivity disorder symptoms), poor non-cognitive skills (e.g. emotional distress), low school connectedness, poor academic performance and educational aspirations, and high peer delinquency (Bernat et al., 2012). Substance use is another significant risk factor leading to juvenile delinquency (Simoes, Matos and Batista-Foguet, 2008).

Other proximal determinants of youth violence can be found at the family level. For instance, youth living in broken homes and who lack parental support and control are more susceptible to delinquent peer influences, thereby increasing opportunities for involvement in delinquency (Burfeind and Bartusch, 2016; Alboukordi et al., 2012).