Psychosocial Competencies and Risky Behaviours in Peru

We use a unique longitudinal dataset from Peru to investigate the relationship between psychosocial competencies related to the concepts of self-esteem, self-efficacy, and aspirations, and a number of risky behaviours at a crucial transition period between adolescence and early adulthood. First of all, we document a high prevalence of risky behaviours with 1 out of 2 individuals engaging in at least one risky activity by the age 19 with a dramatic increase between age 15 and 19. Second, we find a pronounced pro-male bias and some differences by area of residence particularly in drinking habits which are more prevalent in urban areas. Third, we find a negative correlation between early self-esteem and later risky behaviours which is robust to a number of specifications. Further, aspiring to higher education at the age of 15 is correlated to a lower probability of drinking and of engaging in criminal behaviours at the age of 19. Similarly, aspirations protect girls from risky sexual behaviours.


Introduction
Risky behaviours are associated with health problems, low productivity and more generally with a decline of individual and collective well-being in the short, medium and long run (see for example Parkes et al. (2010)). The study of the determinants of risky and criminal activities is informed mainly by sociological and psychological literature establishing the link between cognitive skills, psychosocial competencies and risky behaviours (Caspi et al., 1994;Agnew et al., 2002;Pratt and Cullen, 2000). 1 The economic literature on crime and risky behaviours primarily adopts an opportunity cost framework. People choose to commit a crime or to engage in risky behaviours if their expected utility from engaging in that behaviour is greater than the expected utility from their outside options (for example in terms of labour market opportunities). Within this framework, more educated people or people with better cognitive abilities are less likely to be involved in risky behaviours (Travis and Hindelang, 1977;Lochner and Moretti, 2004). However, these models do not acknowledge the role of psychosocial competencies.
More recently, economists have gained an interest in studying the role of soft skills (or non-cognitive skills) as predictors of economic outcomes, such as educational attainments, health and labour market outcomes (see for example Borghans et al. (2008), Heckman et al. (2006), Cobb- Tan (2011), Dohmen et al. (2010), Chiteji (2010), and Jaeger et al. (2010)). Nevertheless, few economic papers analyse the role of soft skills on risky behaviours.
The aim of this study is to get a better understanding of the link between psychosocial competencies and risky behaviours at a crucial transition period between adolescence and early adulthood. Specifically, our analysis has three objectives. First, to document the prevalence of risky behaviours in the context of Peru, and the heterogeneity of these outcomes by gender and area of location. Second, to test the hypothesis that dimensions related to the concepts of self-esteem, self-efficacy, and aspirations have an impact on the occurrence of 1 There is an ongoing debate, and little agreement, on how to refer to those skills which represent the "patterns of thought, feelings and behaviour" (Borghans et al., 2008) and that encompass those traits that are not directly represented by cognitive skills or by formal conceptual understanding. The current list includes such terms as behavioural skills, soft skills, personality traits, non-cognitive skills or abilities, character, life-skills, socio-emotional and psychosocial skills or competencies. In this paper, we use the term "soft skills" and "psychosocial competencies" interchangeably. risky behaviours during adolescence. Third, to test the robustness of this association by applying statistical methods that allow to control for unobservable cofounders.
For this analysis, we exploit the longitudinal nature of the Young Lives data, a unique individual-level panel following a cohort of about 700 children in Peru over four rounds of data collection that took place between 2002 and 2013. The Young Lives data cover a critical phase of the life-cycle for human capital and skills accumulation following the same children between ages 8 and 19. Information on a number of risky behaviours are collected at the age of 15 and 19 which makes the Young Lives data particularly suitable for this analysis. Furthermore, rich information both at the household and individual level are collected which include children's cognitive and psychosocial competencies, school history, parental and children's aspirations and aspirations for education. 2 Based on the data available, we define indicators to measure the prevalence of (i) smoking behaviours; (ii) drinking behaviours; (iii) drinking and violence (engaging in violent or risky activities when drunk); (iv) consumption of illegal drugs; (v) criminal behaviours; (vi) possession of weapons; (vii) unprotected sex; and, (viii) total number of risky and criminal behaviours.
Evidence on psychosocial competencies as a predictor of criminality and delinquency invites questions about the ability to prevent risky behaviours by shaping those skills. Furthermore, while the most 'sensitive' (productive) periods for investment in both cognitive skills and psychosocial competencies occur earlier in people's life, soft skills during adolescence are more malleable than cognitive skills (Knudsen et al., 2006;Cunha and Heckman, 2008;Cunha et al., 2010;Cunha and Heckman, 2007;Carneiro and Heckman, 2003). Of course, the differential plasticity of different skills by age has important implications for the design of effective policies.
There are three recent studies that have looked at the determinants of risky behaviours at age 15 in Peru using the first three rounds of Young Lives data: Cueto et al. (2011), Crookston et al. (2014, and Lavado et al. (2015). The study by Cueto et al. (2011) and colleagues highlights the importance of parents-child 2 It is important to note that information about cognitive and psychosocial competencies are collected for all children regardless of their school enrollment status which avoids any selection problem commonly arising using school-based tests.
relations and peer effects in predicting smoking habits and unprotected sexual relations at early ages. Crookston et al. (2014) document the association between children victimization at school on subsequent risky behaviours. Finally, the study by Lavado et al. (2015) look at the relationship between cognitive and non-cognitive skills and consumption of cigarettes and alcohol and the early initiation of sexual activity. Overall, their results suggest a negative relation between risky behaviours and cognitive and non-cognitive skills.
However, the evidence these studies can provide is limited for two reasons.
First, there is low prevalence of risky behaviours observed at age 15, which authors try to compensate by being very inclusive in the definitions used, particularly in the way smoking and drinking are defined. Second, there is an endogenous relationship between child characteristics and risky behaviour outcomes. There are reasons to think that psychosocial competencies and the outcomes of interest are jointly determined. Therefore, the main challenge is in assessing whether the effect of poor psychological resources on the probability in engaging in risky behaviours is due to potential endogeneity bias; either through reverse causality or uncontrolled confounding variables.
In this analysis we try to overcome both challenges. First (low prevalence), we show that in most cases the frequency of risky behaviours has increased considerably between age 15 and 19 which makes the empirical study more viable.
Furthermore, the use of the last round of data allows us to broaden the scope of risky behaviours observed (at age 19, individuals were asked to report about criminal behaviours in addition to the other risky behaviours collected in previous rounds). Second (potential endogeneity bias), although this paper does not claim any causal relation, we exploit the fact that the data was collected over multiple periods to implement strategies that minimize both sources of endogeneity. To deal with reverse causality we use lagged values of the psychosocial variables of interest, measured three years before the realization of the risky behaviours. To deal with omitted variable bias we estimate a child fixed effects model, which purges bias due to unobservables that are constant over time.
These are our main findings. First, we find that the prevalence of risky behaviours is evident and increases significantly over time: by age 15, two out of 10 individuals had engaged in at least one risky behaviour, whereas by age 4 19 one out of two had. By age 19, the prevalence of smoking and drinking is 19 and 34 percent respectively; 13 percent had consumed illegal drugs, 27 percent had had unprotected sex and 19 percent had engaged in criminal behaviours.
Second, with the exception of unprotected sex, there is a notorious promale bias in the prevalence of most of these behaviours. There are also some differences by area of location, particularly in drinking habits which are more prevalent in urban areas.
Third, and perhaps most importantly, we find a negative correlation between psychosocial competencies and risky behaviours. Keeping everything else constant, an improvement of 1 standard deviation in self-esteem at the age of 15 is associated with a reduction of 6, 7 and 8 percentage points respectively in the probability of smoking, drinking and engaging in violent behaviours while drinking at the age of 19. It is also associated with a reduction in the prevalence of criminal behaviours and in the possession of a weapon by 14 and 5 percentage points respectively. No similar correlation is found with self-efficacy. These results are robust to a large set of controls at the child and household level, and to community characteristics that are fixed over time. Moreover, child fixed effects estimates show that these associations persist once controlling for timeinvariant unobservable characteristics. We note further that early self-esteem, measured at the age of 12 is already a predictor of later drugs consumption, unprotected sex, criminal behaviours and the number of risky behaviours the adolescents engage with at the age of 19.
Finally, we find that aspiring to higher education at the age of 15 reduces the probability of drinking and engaging in criminal behaviours at the age of 19 respectively by 14 and 25 percentage points. Furthermore, while on average girls are more at risk of unprotected sex, girls aspiring to higher education are less likely to have unprotected sex. Nevertheless, once we control for unobservable individual characteristics the correlation between aspirations and risky behaviours is no longer significant.
The remaining of the paper is structured as the following: Section 2 discusses the different strands of theoretical literature on risky behaviours and reviews the empirical economic literature on the determinants of risky behaviours; Section 3 documents recent patterns in risky behaviours in Peru using the Demographic 5 and Health Survey; Section 4 describes the data and the core predictors of risky behaviours used in the present analyses together with some statistics on risky behaviours using the Young Lives data; Section 5 discusses the empirical strategy and specifications adopted and finally Section 6 and 7 report and discuss our findings. The traditional economic approach to youth risk taking is, as mentioned, a utility maximization/opportunity-cost approach. Forward-looking individuals pursue a certain activity if the expected benefits of it exceeds the expected costs. One example of model using this approach is the "rational addition model" developed by Becker and Murphy (1988).
Developmental psychology, although not necessarily in contrast to the traditional economic approach, considers a wider variety of factors determining youth decisions to engage in risky behaviours. As Fischhoff (1992) effectively summarises , according to developmental psychologists, (risk) decision-making depends on three groups of factors: how people 'think' about the world, i.e. their capacity for thinking through problems, examining the alternative available and evaluating their implications ("cognitive" development); how people 'feel' about the world ("affective" development) and the roles that others play in people's choices ("social" development).
Finally, behavioural economists bring the two approaches together and enrich economic models with a number of dimensions suggested by developmental psychologists (O'Donoghue and Rabin, 2001). We highlight here the most relevant features for youth decision making.
The first one is related to the way short run benefits and long run costs are 6 modelled. Empirical evidence suggests that young people are excessively myopic with respect to the future and therefore are more likely to have inconsistent preferences over time (O'Donoghue and Rabin, 2001;Gruber and Koszegi, 2000).
More specifically, they have the tendency to have a higher discount rate in the short run than in the long run. Young people respond to the uncertainty about the future by reducing the importance of the future, an effect known as hyperbolic discounting.
The second feature is related to the youths' tendency to introduce projection bias (O'Donoghue and Rabin, 2001;Loewenstein et al., 2000). They have limited capacity to appreciate the extent to which their preferences may adapt over time and they tend to inappropriately project the current preferences onto their future tastes. For this reason, random changes to their current states affect their long run decision making The third feature is related to the youths' attitude to risk. (O'Donoghue and Rabin, 2001;Gruber, 2001). Youth tend to be less risk averse which is consistent with the myopia and hyperbolic discounting features. Moreover, risky decisions are made in uncertain environments and for many risky activities, the cost is one-time and permanent. Uncertainty and one-time cost with longer term implications might increase risk-taking behaviours and a mistake made in the past becomes permanent in its consequences.

Soft skills as predictors of risky behaviours: evidence from policy and research
Many studies in the economic literature find evidence of contemporaneous correlation between different risky behaviours (DiNardo and Lemieux, 1992;Model, 1993;Chaloupka and Laixuthai, 1997;Farrelly et al., 2001;Dee, 1999;Wiefferink et al., 2006;DuRant et al., 1999). Those evidence support the "bad seed" hypothesis, as described by Gruber (2001). The hypothesis is that there is a certain segment of the youth population that is predisposed towards risky activities, while others are not. In that case, policies targeting the segment of population at risk should work effectively. An alternative hypothesis in psychological literature is that there is a certain amount of risk that youths have the tendency to take ("conservation of risk" hypothesis). Reducing risky activity 7 in one area would have a substitution effect by increasing risky activities in another.
To date, most intervention programmes have been targeting specific groups of the population considered at risk, mainly by targeting single risk behaviours.
Most recently, there are examples of interventions taking a broader approach and target more than one risky behaviour at time. More specifically, they aim to address some underlying determinants of risky behaviours which are believed to protect young people from, or predispose them, to distinct risky behaviours. Therefore a better understanding of which childhood traits predict risky behaviours is crucial from a policy perspectives.
Empirical evidence suggest that interventions focusing on improving cognitive skills or aimed at improving soft skills are effective in reducing risky behaviours. An example of an intervention aimed at improving opportunities for children coming from poor backgrounds is the well-known Perry Preschool Program, an intervention targeting a sample of 3-4 year old African-American children living in poverty and assessed to be at high risk of school failure. Although the literature originally focused on the cognitive impact of the intervention, long-term effects have in fact been more persistent in non-cognitive areas. Heckman et al. (2010) and Conti et al. (2012) show that Perry significantly enhanced adult outcomes including education, employment, earnings, marriage, participation in healthy behaviours, and reduced participation in crime teen pregnancy, and welfare dependency later in life. Interestingly, although the program initially boosted the IQs of participants, this effect soon faded. A persistent effect of the program has been found on improvements in personality skills (e.g, it reduces aggressive, antisocial, and rule-breaking behaviours).
On the other side, Hill et al. (2011) show that several interventions that focus on personality rather than only on cognitive skills were effective at reducing delinquency and traits related to delinquency.
Few economic papers analyse the role of personality traits and non-cognitive skills on criminal activities, or more generally, risky behaviours. Heckman et al. (2006) find that self-esteem and locus of control measured during adolescence are as powerful as cognitive abilities in predicting adult earnings. Moreover, they find that personality factors for men affect the probability of daily smoking more 8 than cognitive factors and the opposite is true for women. Similarly, Cunha et al. (2010) show that personality traits are relatively more important in predicting criminal activity than cognitive traits are. Further, Conti and Heckman (2010) suggest that personality and health status measured during adolescence explain more than 50 percent of the difference in poor health, depression, and obesity at age 30. For males, personality traits and health endowments are more predictive than cognitive skills while for women they are equally predictive.

Patterns of risky behaviours in Peru
Information about the prevalence of risky behaviours among the youth population in Peru is scarce. The National Committee for a Life Without Drugs According to DEVIDA, 12 percent consume cigarettes in the 12-18 age group, and the figure increases to 32 percent in the 19-29 age group. The prevalence of alcohol consumption increases from 32 percent at ages 12-18 to 69 percent at ages 19-29. In terms of gender differentials, there is a clear pro-male bias in the consumption of both cigarettes and alcohol. In the case of illegal drugs the prevalence is much lower, around 3 percent in both age groups. The main drug consumed is marijuana. In this case, there is also a pro-male bias in consumption.
With respect to the prevalence of unprotected sex, this information can be obtained from the Peru Demographic and Health survey, which contains nationally representative information for women in reproductive age, from 14 to 50 years old. We use this survey to construct indicators of sexual behaviours (ever had a sexual relation and age of first sexual intercourse) and unprotected sex for females. These results are reported in Table 1. To resemble the ageperiods observed in the Young Lives study, results are reported separately for adolescents aged 15 to 17, and young females aged 18 to 19. We find that the proportion of females that ever had sex increases from 18% at ages 15-17 to 53% at ages 18-19. On the other hand, the proportion of females that did not 9 use a condom during the last intercourse (unprotected sex) is similar for both age groups, approximately two out of ten.
Although the above is useful as a first diagnosis, the data available presents some limitations and concerns for comparability with Young Lives data. First, there is no available information about the frequency of the consumption of cigarettes and alcohol. Therefore, it is not possible to determine whether the prevalence of cigarettes and alcohol consumption observed in Lima City corresponds to consumption habits as opposed to occasional consumption. Second, the information related to the consumption of legal and illegal drugs is not collected at the national level and, at best, is only informative of urban areas.
Third, the information about sexual risky behaviours is only available for females. Fourth, all the information available was obtained through face-to-face interviews. Therefore, results are likely to be biased, particularly in the case of illegal drugs consumption. Finally, there is no information available related to the prevalence of criminal behaviours.
In the next section we present the Young Lives data for Peru and show how this can be used to have a better understanding of the prevalence and the predictors of risky behaviours among the youth population as well as their determinants. The Young Lives sample for Peru gathers information for approximately 700 older cohort children and 2,000 younger cohort children with an over-sampling of poor areas. The original sample was spread over 20 clusters in different geographical regions. 3 More specifically, the 20 clusters were randomly selected from the universe of districts in 2002, excluding the wealthiest 5%. Each district was given a probability of being selected proportional to its population size.
Then, within each selected district, an area was randomly selected and families with children aged 6 to 18 months and 7 to 8 years were selected to be part of the younger cohort and older cohort respectively. Although Young Lives is not intended to be nationally representative, it is worth highlighting that because of the sampling procedure used, the Young Lives sample for Peru has been found to optimally reflect the diversity of children and families in Peru, excluding the wealthiest 5%. 4 In the present analysis, we use the older cohort data aged 8 years old in 2002, 15 years old in 2009 and 19 in 2013. For this cohort, data on risky behaviours was collected in both round 3 and 4 (ages 15 and 19).
A key challenge involved in the collection of risky behaviours data is the danger of substantial under-reporting, both due to cultural reasons as well as legal reasons in the case of the consumption of illegal drugs. This problem is particularly acute in face-to-face interviews. Although Young Lives administers face-to-face interviews for both the child and her family, the information on risky behaviours comes from a self-administered questionnaire which includes a set of questions about alcohol, cigarettes and drugs consumption together with sexual behaviours, contraceptives use and knowledge about sexual and reproductive health. This questionnaire was applied in Rounds 3 and 4 following a meticulous protocol with the aim of minimising under-reporting.
The protocol of the self-administered questionnaire, which is typically applied at the end of the visit, is as follows. The interviewer explains to the child that she will be asked a number of questions about aspects that might be considered sensitive, such as consumption of alcohol, cigarettes and even drugs, together with sexual behaviours, contraceptives use and knowledge about sexual and reproductive health. She is told that she is free to choose to complete the questionnaire or not, and she is free to leave questions blank if she wishes to do so. Then the interviewer mentions that all her answers will remain confidential and that, once she completes the questionnaire, he will put the paper in a sealed envelope, and that neither the questionnaire nor the envelope will contain her name, but a code. Once the interviewer gives this information, the child is asked whether she wants to complete the questionnaire. If she agrees, she is left alone for 15 minutes. Finally, once she completes the questionnaire, this questionnaire will be sealed in an envelope with the code that corresponds to the child.
The aim of the protocol previously described is to assure the child that her answers will remain confidential (as it is indeed the case). Following this procedure, the percentage of children that decided to answer the questionnaire was very high. In Round 4, only 3.8% of the sample refused to complete the selfadministered questionnaire. From those that agreed to answer, 1.6% left all the self-administered questions blank. Among those that decided to go on, the proportion of missing answers is relatively small, especially for questions related to smoking, drinking, possession of a weapon, and criminal behaviours (1.3% of missing answers in questions related to the consumption of cigarettes; 0.5% for alcohol consumption; up to 1.0% for alcohol consumption and engagement in risky activities; 0.5% for possession of a weapon; and, up to 1.5% for questions related to criminal behaviours).The proportion is slightly larger for sexual relations (4.0% for questions related to use of condoms during last sexual intercourse) and consumption of illegal drugs (up to 5.8%). Although there can be some level of under-reporting hidden in the answers that were left blank, the fact that a small proportion of children answered this way leads us to believe that this is unlikely to cause a meaningful bias in our results, particularly because we do capture a high proportion of adolescents engaged in risky behaviours, as will be shown in the following sections.

Lives data
In this paper we investigate a number of risky behaviours for which information are available at both age 15 and 19. More specifically, we look at smoking participation, alcohol and drugs consumption, unprotected sex and weapons possession. More details on the survey questions administered are available in Table 2. In table A.1 and A.2 we report their distribution by age, gender and rural/urban location.
Based on these survey questions we define seven risky behaviours indicators for the empirical analysis. In the remaining part of this section, we define them and we highlight a number of stylized patterns emerging from our data. In Table 3 we report the prevalence of risky behaviours at the age of 15 and 19, by gender and by rural/urban location alongside tests for statistical significance.
Smoking participation ("smoking" variable) is defined as a dummy variable equal to 1 for those individual who reported to smoke cigarettes at least once per month. At the age of 19, about 19 percent of our sample report to be smoking (Table 3). On average they smoke their first cigarette at the age of 16. Most of them (89 percent) report to smoke only 1 cigarette (or less) per day (see Table   A.1 in the Appendix).
With respect to alcohol consumption, there is growing public concern about how much alcohol young people are drinking in their teenage years. According to our data, drinking is the activity with the highest prevalence at both age 15 and 19. At the age of 15 about 65 percent of the sample report that they never drink alcohol and among those who do, most of them drink exclusively in special occasions or very sporadically. By the age of 19 the number of young people drinking increase tremendously, although most of them do not drink on a regular basis. 5 Although the proportion of 19 year old teenagers who drink regularly is quite low (less than 1 percent drink alcohol on a daily basis and 3 percent at least once per week), alcohol consumption increases tremendously between age 15 and 19 as for the number of times they abuse from alcohol consumption. The excessive alcohol consumption not only puts their own health at risk, but also make them more likely to get involved in anti-social behaviours.
We defined two variables for the excess of alcohol consumption: the "drinking" variable equal to 1 for those adolescents who got drunk at least once in their life (and 0 otherwise) and the "drinking and violence" variable equal to 1 for those who engaged in risky behaviours (either having sex, engaging in a fight, feeling sick or drunk) while drinking. By the age of 19, about 34 percent of the sample report of having been drunk and 40 percent of having engaged in risky behaviours while drinking (Table 3).
With respect to sexual behaviours, about 67 percent of our sample had sex by the age of 19, on average having the first sexual relation at the age of 16 (see Table A.1 in the Appendix). We define a variable to capture those young people at risk of sexual transmitted diseases (STDs). 6 The "unprotected sex" variable is equal to 1 for those who did not use a condom in the last sexual relationship (including also those who used other birth control methods or emergency contraception) and 0 for those who had protected sex (using condoms) or never had sex.
For drugs consumption we define an indicator to identify those who ever tried any drugs; about 13 percent by the age of 19.
Similarly, the dummy variable for weapon possession is equal to 1 for those who during the last 30 days carried a weapon at least once.
Overall, looking at the incidence of risky behaviours by age, we notice that risky behaviours increase significantly between the age of 15 and the age of 19, in correspondence of the transition from childhood to adolescence (Table 3). It is worth to note that by the age of 19 male engagement in risky behaviours is about two to three times that of females in smoking, drinking and taking drugs and criminal behaviours. There is also an urban-rural difference in drinking, where adolescents living in urban areas drink more (9% by the age of 15 and 37% by the age of 19) relative to those in rural areas (5% by the age of 15 and 24% by the age of 19) and by the age of 19 are more likely to engage in risky behaviours while drinking.
However, not only the prevalence but also the intensity of risky behaviours increase over time. We define a variable counting the number of risky activities the young people have been involved in by the age of 15 and 19. 7 By the age of 15, about 22% of young people have engaged with at least 1 risky behaviour.
By the age of 19, slightly more than one out of two had engaged in at least one type of risky behaviour, with a distinctive pro-male bias (65% among males, 43% among females). While 28% of the population engaged in only one risky activity, there is an equally consistent segment of the youth population (28%), mainly male population, that undertakes more than one of these activities.
Our data reports quite a remarkable diffusion of risky behaviours among Peruvian adolescents and a worrisome predisposition towards risky activities for the relevant part of them. Although our data do not provide full support to either the "bad seed" or the "conservation of risk" hypothesis, it is worth to note that there is evidence of a certain persistence (or recidivism) in risky behaviours. Those who engage in risky behaviours at the age of 15 are indeed more likely to engage in risky behaviours at the age of 19. The average "number of risky behaviours" at the age 15 is strongly correlated with the same measured four years later (standardized correlation coefficient of 0.6). Recidivism is more evident in some risky behaviours than others, particularly in drugs consumption, drinking and smoking. In fact, adolescents who consume drugs at age 15 are 64 percentage points more likely to consume drugs at age 19. Similarly, drinking (smoking) at age 15 increases the probability of smoking (drinking) at age 19 by 38 percentage points (39 percentage points).
In the next section we characterise further who are these young people, what is their history, their past experience, their ability and psycho-social well-being and where do they live using a multivariate approach.

Predictors of risky behaviours
In this section we briefly discuss the core predictors of risky behaviours. We focus exclusively on those predictors that are either time-invariant or are available in earlier rounds of the Young Lives survey. 8 As we discussed in the previous section, soft skills have been identified as important factors in predicting risky behaviours. In our data, we capture soft skills (or psychosocial competencies) through two indicators that have been administered in the last three rounds of the Young Lives survey: the self-esteem scale and the self-efficacy scale. In the Young Lives database, these scales are referred as the pride index and the agency index, respectively. The former builds on the self-esteem concept by Rosenberg (1965) and is related to his/her overall evaluation of his/her own worth. The latter builds on the concept of locus of control by Rotter (1966) and self-efficacy by Bandura (1993) and it measures the child's freedom of choice and his/her agency (or power) to influence his/her own life. The full list of survey questions included to compute the two scales are reported in Table 2. 9 Another core predictor for risky behaviours is individual aspirations. There is a considerable body of economic literature investigating the role of aspirations and subjective expectations for contraceptive choices (Delavande, 2008), (sexual) risky behaviour (De Paula et al., 2013;Shapira, 2013) and non-marital childbearing choices (Wolfe et al., 2007). As Dalton et al. (2015) argue, how far people aspire depends on their own beliefs about what they can achieve with effort, i.e. their own expectations. People would not aspire to an outcome that is perceived as inaccessible. Consistently with the "opportunity cost" argument in the risky behaviour literature, if an outcome is perceived as inaccessible, people might believe that they have little to lose by engaging in a risky behaviours.
However, given the endogenous nature of aspirations, the empirical distinction between aspirations and expectations is hard to achieve in a nonexperimental setting. Therefore, the measure of aspirations considered in this study reflects a combination of aspirations and beliefs about the likelihood of achieving the aspired outcomes. Young Lives collects information about educational aspirations by asking to the child the following question: "Imagine you had no constraints and could study for as long as you liked, or go back to school if you have already left. What level of formal education would you like to complete?". In this study, we define a dummy variable equal to 1 for individuals with high aspirations, i.e. for those children who aspire to go to university, and 0 otherwise.
As noticed in previous sections, patterns of risky behaviours may vary by gender and change significantly during adolescence. While there is consistent evidence in the literature of a higher prevalence of risky behaviours among boys, the age pattern is more complex. In fact, while younger teens tend to be both more impatient and subject to peer pressure (Lewis, 1981), which could make them more risk taking than older teens, there are at least three factors which might counterbalance this: biology, income and law (Gruber, 2001). Indeed, some risky activities (e.g. sexual intercourse) become desirable with age (biology). Moreover, some illegal activities for younger teens become legal at older ages (e.g. cigarettes consumption is illegal to under 19 in Peru) (law ). Finally, older teens may have more money available to finance their risky activities (income).
Related with this last point, poverty is often seen as a trigger factor for engaging in risky behaviours. In the present analysis we approximate the socioeconomic status of the natal household by using: father's and mother's education level, an indicator for the rural/urban location where the household resides, and the tercile of wealth index, a composite measure of living standards including housing quality, access to service and a consumer durable index as defined in Table 2.
Finally, we look at two additional sets of potential predictors of risky behaviours related to household composition and education, and cognitive skills.
The first set of characteristics we look at relates to household composition.
In light of the results of past research, we include the number of siblings and whether he is living only with one biological parent (broken home). 10 We also control for whether the young person has an older sibling. Siblings and in particular older siblings might affect the behaviour of our sample child by being a role model to them and by enlarging their peer groups. There is a consistent body of literature investigating the influence of family and peers on behaviour of disadvantaged youths. For example, Clark and Loheac (2007) use the Add Health survey to examine risky behaviour by American adolescents.
They find that the consumption of tobacco, alcohol and marijuana are correlated with peer group behaviour. The correlation is stronger for alcohol use and among young males more than young females.
We also control for a dummy variable to identify movers, i.e. those children/families who are living in a different community as when they were living when they were 8 years old. Although we have no particular prior here, there 10 See Lundberg and Plotnick (1990) for a review of this evidence are some evidence of a positive association between displacement and risky behaviours. For example, Gaviria and Raphael (2001) suggest that recent movers may be more susceptible to peer group pressure.
Finally, we look at a set of predictors relating to education. More specifically we look at school enrollment, delayed enrollment and school achievement.
School achievement, measured either by the Raven test score or the Peabody Picture Vocabulary Test (PPVT) and a Math test, can also be considered as a proxy of the child's cognitive skills. Notably, the two tests have been collected for all children regardless whether they are attending school or not. This feature of the data avoids any selection problem which commonly arises using school-based data.
As an initial exploration of factors that might affect the probability of engaging in risky behaviours at the age of 19, we compare the mean characteristics of the predictors listed above for adolescents "at risk" (engaging in at least one risky behaviour by the age of 19) and adolescents "not at risk". All predictors are measured when the adolescent was 12 and 15 years old. These differences are presented in Table 4 alongside tests for statistical significance.
Looking first at the individual characteristics, young people engaging in risky behaviours by the age of 19 are more likely to be boys and slightly older than those who are not at risk. Young people 'at risk' have lower self-efficacy (slightly lower self-esteem) and are less likely to aspire to university at the age of 15.
Furthermore, risky behaviours are more prevalent among young people having lower cognitive skills (performing worse in the Math test) and those who have already dropped out of school by the age of 15.
Interestingly, risky behaviours is not necessarily a phenomenon prevalent among young people living in poverty. Indeed, young people living in poverty are as likely as young people living in less poor households to engage in risky behaviours.
Additionally, there is no difference in the prevalence of risky behaviours in rural and urban areas and the level of parental education is the same among young people at risk and not at risk. Notably, risky behaviours are more prevalent in single parent households.

Empirical strategy
In this section we define a multivariate set-up, estimating linear probability (OLS) models. Our dependent variables are the risky behaviours as defined in the previous sections. With the exception of the intensity variable (number of risky behaviours) the dependent variable is a variable equal to 1 if the young person engages in risky behaviours at the age of 19, and 0 otherwise.
First of all, we investigate the predictors of risky behaviours looking at the association between risky behaviours measured at the age of 19 and psychosocial competencies measured at the age of 15, controlling for schooling achievement and a broad set of early (or time-invariant) individual and household level characteristics as follows: Similarly, we investigate the correlation between educational aspirations measured at the age of 15 and risky behaviours at the age of 19. According to the "opportunity cost" argument we would expect to find a negative correlation between aspirations and risky behaviours if the perceived cost of engaging in risky behaviours increases with aspirations. The descriptive statistics presented in Table 4 indeed shows that adolescents engaging in at least one risky behaviour at the age of 19 have lower aspirations than "not at risk" adolescents.
Given that aspirations is likely to feed into the child's self-efficacy and selfesteem, we estimate a separate model similar to the one discussed above but including a dummy variable equal to 1 for those children that at the age of 15 aspire to complete higher education (university), and 0 otherwise: In both equation 1 and equation 2 self-efficacy, self-esteem and aspirations are measured at the age of 15. An empirical question is whether the psychosocial competencies measured at younger ages predict later risky behaviours. Young Lives collect self-efficacy, self-esteem and aspirations at both age 12 and 15 which allow us to look at the long term association with risky behaviours. We also report results for this long-term specification. In this case all control variables are either time invariant or measured as early as possible (at age 8). In this case, the Raven score measured at age 8 is used as indicator of school achievement.
Although informative, an estimation of the risky behaviours equations using cross-sectional data would be unbiased only under very strong assumption about the role of unobservable variables. In absence of plausibly exogenous variations in the regressors, their estimation raise endogeneity concerns and might lead to biased interpretations. Therefore, our intention is not to identify causal effects.
Rather, the estimated parameters should be interpreted as partial correlations which may be revelatory about potential drivers of risky behaviours at different ages and the channels through which such effects may be mediated.
Further, we exploit the fact that we have repeated measures of risky behaviours and we estimate the outcome of interest using a child fixed effects 20 model, as follows: In this specification, the role of self-efficacy and self-esteem is identified by exploiting changes between ages 12 and 15 that in turn lead to changes in risky behaviours between ages 15 and 19. In doing so, we implicitly assume the relevant coefficients are age-independent. This strategy has the advantage that it controls for individual unobservable characteristics that are constant over time.

Results
The main results of the analysis are reported in Table 5 and  It is interesting to observe that the role of socio-economic characteristics and schooling achievement is relevant only for certain types of outcomes and that the specific characteristics that matter depend on the nature of the risky behaviours. We find that household wealth (as measured by the wealth index) is not associated to any of the risky behaviours, however other socio-economic dimensions such as parental education (for criminal behaviours), area of location and migration (for drinking behaviours) are. Specifically, we observe a negative association between father´s education and criminal behaviours.
We also observe that living in a rural area reduces drinking by 27 percentage points whereas having migrated between ages 15 and 19 increases the same probability by 17 percentage points (most of the observed migration implies moving from rural to urban areas). Finally, keeping other factors constant, we do not find a statistically significant association between schooling achievement and risky behaviours. However, there is a strong negative association between school enrollment and criminal behaviours.
From the factors previously mentioned, gender, age, self-esteem and living in a single-parent household stand out as factors that systematically predict risky and criminal behaviours. These are also the factors that predict the (overall) number of risky behaviours in which the individual has engaged.
Also, it is interesting to observe that psychosocial competences do not play any role in predicting risky sexual behaviours. This is quite surprising given that previous literature suggest self-efficacy (or self-confidence) to be one of the key factors for contraceptive uses and particularly for the use of condom which, particularly for girls, requires negotiating its use with the partner (see for example Salazar et al. (2005)).
More generally, unprotected sex is the behaviour for which fewer predictors turn out to be statistically significant (only one, coming from a single-parent household) which suggest that other important predictors might have been neglected. Factors such as being born to a teenage mother, knowledge on sexual and reproductive health, access to contraceptive methods, age of the sexual debut and relationship status are some of the factors commonly correlated with teenage pregnancy and motherhood (see for example Azevedo et al. (2001) and Ermisch and Pevalin (2003)). Furthermore, being married or in a stable relationship might influence the decision to use of using contraceptive methods.
Nevertheless, these factors have been not included in the analysis mainly for two reasons: first, to preserve comparability across the different risky behaviours considered; second, some of those variables are only collected at Round 4. 11 In Table 6 we report the results for the risky behaviours models including educational aspirations. Keeping everything else constant, in this model we observe that aspiring for higher education reduced the likelihood of drinking and engaging in criminal behaviours by 14 and 25 percentage points respectively.
Higher aspirations are also negatively correlated with the total number of risky behaviours. The role played by the other predictors (the same as in the previous model) remains very similar. One noticeable difference is that, once aspirations are controlled for, school enrollment does not predict criminal behaviours which suggest that aspirations measured and school enrollment both measured at the age of 15 are strongly correlated.
To further explore the possible differential correlation of psychosocial competencies to risky behaviours by gender, in Tables 7 and 8 we replicate the same results adding an interaction between male gender, self-esteem and self-efficacy, and between male gender and aspirations, respectively. The only noticeable result is that while on average girls are more at risk of unprotected sex than boys, girls aspiring to higher education are less likely to have unprotected sex. Therefore, if the relation between aspirations and risky behaviours would prove to be causal, having higher educational aspirations would be more likely to prevent girls to engage in unprotected sex relatively more than boys.
So far we show that psychosocial competencies and aspirations measured at the age of 15 predict many risky behaviours that occur at the age of 19. An empirical questions is whether this correlation is constant over time and psychosocial competencies and aspirations measured earlier in life similarly predict later behaviours.
In Table 9 and Table 10 we report the estimates for the risky behaviour model where early psychosocial competencies and aspirations are measured at the age of 12. Analogous to previous results, early self-esteem is negatively correlated with drugs consumption, unprotected sex, criminal behaviours and with intensity of engagement in risky behaviours more generally. Similarly, children aspiring to higher education at the age of 12 are less likely to use carry weapons at the age of 19.
On the contrary, children with higher level of self-efficacy at the age of 12 are relatively more at risk of engaging in risky behaviours. More specifically, an increase of 1 standard deviation in the self-efficacy indicator is correlated with an increase in smoking and engaging in unprotected sex of 4 and 5 percentage points. This seems to suggest that while higher self-esteem during childhood and throughout adolescence might play a protective role against risky behaviours later on in life, self-efficacy during childhood might indeed have the opposite effect (and not a significant effect during adolescence). The reasons to this 24 might reside in the nature of the two soft skills themselves. Self-efficacy which reflects the individual's judgment of their own capacity to act and exert agency primarily, is also intrinsically related to action and behaviour. This is where it differs from 'self-esteem', which is a more passive concept, without a necessary relationship to action, and is about the individual's judgment of self-worth.
Thus, for those adolescents with pre-existing low self-efficacy, the transition to adolescence (and the risky behaviour experimentation as part of this process of development) can be problematic.

Fixed effects estimates
In order to obtain a better identification of the relationship between psychosocial competencies and the outcomes of interest, we report individual fixed effects estimates obtained by differencing risky and criminal behaviours at ages 19 and 15 on differences in psychosocial competencies at ages 15 and 12, as well as on differences in all the other control variables at ages 15 and 12. These results are reported in Tables 11 and 12. For this part of the analysis, the criminal behaviours variable is dropped because it is not observed at age 15. Gender and parental education do not vary over time and age varies uniformly across all children between survey waves, thus they are also dropped.
In Table 11, we report the results for the individual fixed effects estimates using self-esteem and self-efficacy as predictors of risky behaviours. The results are qualitatively similar to the ones discussed above. An increase in self-esteem is negatively correlated to the the prevalence of risky behaviours over time.
More specifically, one standard deviation increase in self-esteem reduces smoking, drinking and engaging in drinking and violence by 5, 10 and 9 percentage points respectively; it does not predict the likelihood of carrying a weapon, but the point estimate is very similar (3 percentage points). In contrast to the results in the previous model, self-efficacy is predictive of carrying a weapon. One standard deviation increase in self-efficacy reduces the probability of carrying a weapon by 5 percentage points. Besides this, in this set of estimations schooling achievement is found to play a more prominent role. School enrollment reduces the likelihood of drinking, drugs consumption, and risky sex. A similar role is played by vocabulary and math achievement. In addition, coming from a 25 single-parent household, area of location and migration remain as important predictors of risky behaviours.
In Table 12, we report the results for the individual fixed effect model including aspirations among the predictors. However, in this case we are not able to detect a relationship between aspirations and the outcomes of interest.
To summarise, the fixed effects estimates show that the relationship between self-esteem and risky behaviours is very robust, whereas the relationship between self-efficacy, aspirations and risky behaviours is not. In addition, there seems to be a lot of meaningful variation over time in the control variables, which allows us to show that coming from a single-parent household, area of location, migration and schooling achievement are also important factors that play a role in the determination of risky and criminal behaviours.

Conclusions and discussion
There is a growing concern about the prevalence of risky behaviours among the youth population, which ultimately leads to worse outcomes later in life, including lower salaries and worse socio-economic and life outcomes. On the other hand, there is little evidence about the prevalence of these behaviours and their determinants in the context of developing countries. Our aim is to try to fill this gap using a unique individual-level panel data from Peru following a cohort of children for over a decade between the ages of 8 and 19.
We constructed indicators to measure the prevalence of smoking and drinking; engaging in risky behaviours when drunk; consumption of illegal drugs; unprotected sex; criminal behaviours; possession of weapons; and total number of risky behaviours. While we do not claim any causal relation, the methods used allow us to deal with bias arising from reverse causality and omitted variables that are constant over time.
-From this analysis we identify a number of drivers of risky behaviours. In particular, there is a specific group of the youth at risk; boys, living in urban areas and growing up in single-parents households. In the case of girls, they are more likely to be exposed to unprotected sex. Although these groups are identified for the Peruvian context, similar patterns are likely to be observed in countries with similar characteristics (middle-income countries with relatively high-levels of poverty and low levels of secondary school attainment).
We also observe a dramatic increases in risky behaviours between age 15 and 19 which suggests that policy interventions aiming at preventing risky behaviour should be put in place at age 15 or earlier, when risky behaviours only manifest in a small part of the population.
Although the present analysis is not sufficient to claim any causal relation between socio-emotional competencies and risky behaviours provides some interesting hints. Our results suggest that psychosocial competencies, and selfesteem and high aspirations in particular, might play a role in reducing risky behaviours. To our knowledge this evidence is unique in the developing countries context. Policies aimed at promoting soft skills during childhood and adolescence can play an important role as a mechanism to reduce risky and criminal activities among the youth.
From a policy perspective, considering the age range analysed as well as the fact that, by age 15, most Peruvian adolescents are still attending school, we argue that it is worth to explore whether interventions designed to take place at secondary-level schools can reduce the engagement of adolescents in risky behaviours.
In This encouraging findings from Chile suggests that it is worthwhile to explore the potential effects of this type of reforms and risky behaviours. Further research on JEC in Peru and its effect on risky behaviours will be done using the next round of data.

Risky Behaviours
Smoking Dummy variable equal to 1 if she/he is "smoking at least once a month" (or more frequently), and 0 otherwise. The survey question used is the following: "How often do you smoke cigarettes now?". Drinking Dummy variable equal to 1 if she/he has been drunk at least once in his/her life, and 0 otherwise. The survey question used is the following: "Have you ever been drunk from too much alcohol?".

Drinking & Violence
Dummy variable equal to 1 if she/he engaged in risky behaviours (got into fights/caused trouble, felt sick or fell over, have sex) while drinking and 0 otherwise (including also those who never drunk alcohol before). The survey questions used are the following: "During the past 12 months, how many of these things happened to you because you had been drinking alcohol?" and "During your life, have you ever been drunk from alcohol while having sex?".

Drugs Consumption
Dummy variable equal to 1 if she/he has ever consumed any of the drugs listed before, 0 otherwise: "Have you ever tried any of the following drugs?" An index measuring the intensity of criminal behaviour, defined by the sum of the following dummy variables: whether the YL child has carried a weapon in the last 30 days, ever been arrested by the police or taken into custody for an illegal or delinquent offense, ever been a member of a gang, or ever sentenced to spend time in a corrections institution such as a jail/prison/youth institution (juvenile hall, reform school, training school).

Carrying a Weapon
Dummy variable equal to 1 if she/he carried a weapon during the last 30 days and 0 otherwise.The survey question used is the following:"During the last 30 days, on how many days did you carry a weapon such as a knife, machete or gun to be able to protect yourself?" No. of Risky Behaviours An index created to measure the intensity of risky behaviour and equal to the sum of all the dummy variables defined above (smoking, alcohol and drugs consumption, carrying weapons and having unprotected sex). Child's Educational Aspirations "Imagine you had no constraints and could study for as long as you liked, or go back to school if you have already left. What level of formal education would you like to complete?". Child's educational aspirations are collected at the age of 12, 15 and 18. We define a dummy variable equal to 1 for those children with high aspirations (aspiring to university) and 0 otherwise.

Variables Description
Psychosocial competencies This is the procedure adopted to compute the self-efficacy and self-esteem indicators: (i) all relevant questions are recoded to be positive outcomes, (ii) relevant questions are all normalized to z-scores (subtract mean and divide by SD) and then (iii) an average of the relevant z-scores is taken across the non-missing values of the questions. All the questions are on Likert-type scales going from 1 to 4 in Round 2 (R2) and from 1 to 5 in Round 3 (R3). The questions differs a little from round to rounds specified below. A composite measure of living standards. The variable takes values between 0 and 1, such that a larger value reflects a wealthier household. The wealth index is the simple average of three sub-indexes: a housing quality index (quality of floor, wall, roof and number of rooms per capita), an access to services index (access to drinking water, electricity, sewage and type of fuel used for cooking) and a consumer durables index (TV, radio, fridge, microwave, computer, etc). In the analysis we use the wealth index segmented in tertiles: bottom, middle and top tertiles Parents' education Father's and mother's education, segmented into three categories for none or primary education (less than grade 8), secondary education (grade 10) and higher education (above grade 10) as their highest level of education completed Single parent Dummy variable equal to 0 if she/he is living with both biological parents, and 1 if he/she is living with only one biological parent, the biological parent and his/her partner, or is an orphan Child has older siblings A dummy equal to 1 if she/he has older siblings Number of siblings Number of siblings Delayed enrollment Dummy variable that indicates 1 if the YL child has ever delayed school enrollment and 0 if not Raven's test score (z-score) Total number of correct responses in the Raven test, standardised by round PPVT (z-score) Standardised score for the Peabody Picture Vocabulary Test by rounds Math (z-score) Standardised score for the Maths test by rounds Enrollment Dummy variable that indicates 1 if the YL child is enrolled in school or not        Note: The table reports the estimates of the linear probability model with standard errors (in parentheses) clustered at cluster level, * p<0.1 ** p<0.05 *** p<0.1. All controls were included as reported together with dummy variables for the cluster that individuals were recruited in the 2002 round; coefficients for these are not reported.