4 | 109 The transition to adulthood: A game changer!? highest schooling level completed or currently in progress in years, defined as the nominal duration in years of schooling for each type of education, varying from 4 years for not completed primary school, to 16.5 years for a university degree. When respondents reported a diminishment of their educational status between the waves (most likely because they dropped out), we used the educational level reported in wave 1. Respondents were removed from our analyses if they reported a decrease of more than 2.5 years, because this was likely indicative of measurement error. Table 4.1 presents the variables. Analytical strategy We performed different types of analyses to gain insight into the role of major life events in respondents’ sport participation. First, information on the frequency and number of sports were obviously count data. Both aspects were characterised by a highly right-skewed frequency distribution. We performed statistical tests to assess the degree of overdispersion; this appeared to be insignificant (p=0.85) for the number of sports, but highly significant (p=0.00) for the more skewed sport frequency. Therefore, we used the Poisson distribution to model the number of sports and a negative binomial distribution for sport frequency. Second, to assess changes in these variables between waves, we applied multilevel modelling for longitudinal data (Singer & Willett, 2003; Snijders & Berkhof, 2007). In such models, wave 1 and wave 2 observations constitute the lower-level units that are nested within persons. Snijders & Berkhof (2007) showed that estimates of a regression coefficient for a lower-level predictor, as for instance “being a parent”, may reflect not only a within-person effect (becoming a parent), but also a between-person effect (difference between parents and non-parents). Such confounding is undesired, as we are interested in both the influence of change experienced by individuals and interpersonal differences. Using a “between-within” model (Neuhaus & Kalbfleisch, 1998), or “hybrid” method as Allison (2009) calls it, a corrected estimate of the withinperson effect of a predictor can be obtained by centring the predictor around its cluster mean. In our study this was the mean of both observations for each respondent. Including these centred predictors instead of the original ones as “within variables”, as well as the personal means as “between variables” in the multilevel models, yielded estimates of the within-person changes and the