Empirical Specifications and Estimation Methods

Equation (2.1) is a fixed effects estimator, accounting for time invariant unobserved heterogeneity as follows:

𝑆𝑊𝐵_𝑖𝑡 = 𝛽₀+ 𝛽₁𝐵𝑢𝑙𝑙_𝑖11+ 𝜷𝑿_𝑖𝑡𝑎 _{+ 𝑎}

𝑖 + 𝜀𝑖𝑡 (2.1)

𝑆𝑊𝐵_𝑖𝑡 denotes individual 𝑖’s SWB at age 𝑡, as measured by either the measure of life satisfaction or total malaise. 𝐵𝑢𝑙𝑙𝑖11 is a binary variable taking the value 1 if individual

𝑖 was sometimes or frequently bullied at age 11, and 0 if they were never bullied.30

𝑿_𝑖𝑡𝑎 indicates a vector of the adult characteristics of individual 𝑖. 𝑎𝑖 denotes an individual

fixed effect: the characteristics of the individual that have a constant effect on their SWB at ages 33, 42, 46 and 50. 𝜀𝑖𝑡 represents a time and individual specific error term

and 𝛽0 denotes a constant.

The SWB variables, which were recorded from age 33 onwards, are predated by the measurement of the bullying variable (recorded at age 11). As a result, if a fixed effect estimator is used, the effects of being bullied at age 11 on SWB is absorbed into 𝑎𝑖. For

this reason, the fixed effects estimator cannot be used to estimate the effect of bullying victimisation as a child on adult SWB. A random effects estimator is used instead, as shown by equation (2.2):

𝑆𝑊𝐵𝑖𝑡 = 𝛽0+ 𝛽1𝐵𝑢𝑙𝑙𝑖11+ 𝜷𝑿𝑖𝑡𝑎+ 𝑢𝑖 + 𝜀𝑖𝑡 (2.2)

𝑢_𝑖 indicates a randomly generated, time invariant individual effect. Unlike 𝑎𝑖, which

may be correlated with the explanatory variables, 𝑢𝑖 is assumed to be randomly

distributed across individuals, and hence uncorrelated with the explanatory variables.

30 _{For robustness, the results have also been replicated using binary variables denoting whether the} individual was sometimes bullied or frequently bullied at age 11. These results are summarised in Section 2.5.3.

27 In the context of the effects of being bullied at age 11 on adult SWB, the random effects estimator assumes that the stable unobserved individual characteristics that affect SWB are uncorrelated with whether an individual was bullied at age 11. However, as previously discussed, the mother’s report of whether the child was bullied is not exogenous: a child’s characteristics may affect their likelihood of bullying victimisation. If these characteristics also affect SWB, then equation (2.2) is likely to suffer from omitted variable bias. To reduce this bias, Mundlak averages are included in the following analysis, see Mundlak (1978). Mundlak averages account for fixed individual heterogeneity that may affect both SWB and the likelihood of being bullied. The inclusion of the Mundlak averages is an alternative to using a fixed effect estimator.

Due to the nature of the dependent variable, the results from random effects ordered probit models are presented. Random effects ordered probit estimators have been used to model SWB and self-reported health data in the existing literature, (see, for example Winkelmann, 2005; Contoyannis et al., 2004). The measures of life satisfaction and total malaise are ordinal and take integer values. An individual reporting a score of 7 is assumed to have a higher level of well-being than if the same individual reports a score of 6, as argued by Ferrer-i-Carbonell and Frijters (2004). However, SWB is not measured on a cardinal scale: the difference between a score of 7 and a score of 8 may not be the same as the difference between a score of 5 and a score of 631. The random effects ordered probit model applies equation (2.2) to an ordered probit specification. Due to the ordered, rather than cardinal nature of the SWB variable, a latent variable approach is used. The latent variable, 𝑆𝑊𝐵𝑖𝑡∗, is shown

by equation (2.3):

𝑆𝑊𝐵_𝑖𝑡∗ _{= 𝛽}

0+ 𝛽1𝐵𝑢𝑙𝑙𝑖11+ 𝜷𝑿𝑖𝑡𝑎 + 𝑢𝑖 + 𝜀𝑖𝑡 (2.3)

Assume that the measure of SWB has 𝐽 potential responses. The probability of each single response to the SWB question (denoted 𝑗) being reported by individual 𝑖 is thus determined by the value of the latent SWB variable, 𝑆𝑊𝐵_𝑖𝑡∗, relative to the values of the thresholds (𝜇𝑗). This is demonstrated by equation (2.4):

𝑆𝑊𝐵𝑖𝑡 = 𝑗 if 𝜇𝑗−1< 𝑆𝑊𝐵𝑖𝑡∗ ≤ 𝜇𝑗 ∀𝑗= 0 … 𝐽 ∀𝑗denotes for all 𝑗 (2.4)

Hence, an individual's reported SWB score equals 𝑗 if 𝑆𝑊𝐵_𝑖𝑡∗, the latent variable, lies

between the threshold values of 𝜇𝑗−1 and 𝜇𝑗. Therefore, assuming that the error terms

(𝜀𝑖𝑡) follow a cumulative normal distribution, the probability that at time 𝑡, individual 𝑖

reports a SWB score of 𝑗 is given by equation (2.5):

31 _{In contrast to measures of SWB, height is measured on a cardinal scale because the difference} between a height of 150 cm and 160 cm equals the difference between heights of 160 cm and 170 cm.

28 𝑃_𝑖𝑡 = 𝑃(𝑆𝑊𝐵_𝑖𝑡 = 𝑗) = 𝜙(𝜇_𝑗− 𝛽₀− 𝛽₁𝐵𝑢𝑙𝑙_𝑖11− 𝜷𝑿_𝑖𝑡𝑎 _{− 𝑢}

𝑖) (2.5)

−𝜙 (𝜇_𝑗−1− 𝛽₀− 𝛽₁𝐵𝑢𝑙𝑙_𝑖11− 𝜷𝑿_𝑖𝑡𝑎 _{− 𝑢} 𝑖)

Where 𝜙 denotes the cumulative normal distribution. Initially, the analysis makes use of a random effects ordered probit model to estimate the effect of bullying victimisation as a child on adult SWB. Following this, for robustness, the models are re-estimated using a random effects model, which assumes that the measures of SWB are continuous. In this chapter, two specifications are estimated to cast light on the effects of being bullied as a child on adult SWB; the first is outlined by equation (2.6):

𝑆𝑊𝐵𝑖𝑡 = 𝛽0+ 𝛽1𝐵𝑢𝑙𝑙𝑖11+ 𝜷𝑿𝑖𝑡𝑎 + 𝛂𝐗itp +λ𝑿̅𝒊𝒕𝑎 + 𝑢𝑖 + 𝜀𝑖𝑡 (2.6)

𝑿_𝑖𝑡𝑎 _{indicates a vector of individual 𝑖’s time-varying adult characteristics at age 𝑡:}

household income, marital status, labour force status, disability status, and homeownership status32. The existing literature suggests that these characteristics affect SWB. Powdthavee (2010) finds that income has a positive effect on happiness. Thus, the log of monthly household income is included as a control variable. Stutzer and Frey (2006) find that married individuals report to be happier than unmarried individuals. For this reason, binary variables indicating whether the individual is married or cohabiting; and separated, divorced, or widowed, are included in the analysis. Single individuals form the base category. Unemployment is found by Clark and Oswald (1994) to reduce SWB. Therefore, binary variables indicating that the individual is unemployed or out of the labour force are included as conditioning variables. Employees form the base category. Also, Hu (2013) finds that homeowners tend to report higher SWB. Consequently, binary variables indicating whether the individual owns their home outright, owns their home with a mortgage, or rents privately are included as control variables. Individuals who live in "other" housing or social housing form the base category.

𝐗_itp denotes a vector of time invariant (permanent) and rarely changing adult characteristics: gender, ethnicity, and highest educational attainment33. For instance,

Stevenson and Wolfers (2009) find that males report higher SWB than females and therefore a binary variable taking the value 1 if the individual is male, and 0 if female is included in the analysis. Oreopoulos (2007) finds that high school dropouts report lower SWB, relative to individuals who did not drop out of high school. As a result, binary variables indicating that the individual’s highest educational attainment is a degree, A-levels, or O-levels are included in analysis. Individuals without qualifications form the base category. Oswald and Powdthavee (2008) find that individuals who have 32 _{The proportion of individuals in the estimation samples for which these variables vary from age 33 to} age 50 is as follows: household income (100%), marital status (23%), disability status (19%), homeownership status (45%), and labour force status (28%).

33 _{The individual’s highest educational attainment does not vary between age 33 and 50 for} approximately 90% of the individuals in the estimation samples.

29 health problems which limit their daily activities report lower SWB. Unfortunately, the NCDS contains no variables detailing how an individual’s health affects their ability to carry out day to day activities. Thus, as an alternative, a binary variable is included which is equal to 1 if the individual is registered disabled, and 0 if the individual is not registered disabled.

Stevenson and Wolfers (2012) find that Blacks report lower SWB than Whites. For this reason, binary variables indicating whether the individual is African, Asian or another ethnicity are used as conditioning variables. Whites form the base category. Table 2.4 contains full definitions of the variables used in this chapter and Table 2.5 presents summary statistics.

𝑿̅_𝑖 denotes a vector of Mundlak averages. The Mundlak averages are the "within-person" averages of the time-varying adult control variables. These variables are disability status, the log of monthly household income, marital status, homeownership status, disability status, and employment status. The second specification is shown by equation (2.7):

𝑆𝑊𝐵𝑖𝑡 = 𝛽0+ 𝛽1𝐵𝑢𝑙𝑙𝑖11+𝜷𝑿𝑖𝑡𝑎 + 𝛂𝐗itp+λ𝑿̅𝒊𝒕𝑎+ 𝜸𝑿𝑖𝑐+ 𝑢𝑖+ 𝜀𝑖𝑡 (2.7)

𝑿_𝑖𝑐_{indicates a vector of childhood characteristics: these are the individual’s BSAG score}

(see below) at age 7 and their birth-weight. The BSAG and birth-weight variables are included to account for pre-existing differences between children who are and are not bullied. Black et al. (2007) suggest that birth-weight predicts a range of adult socioeconomic outcomes, such as educational attainment and earnings. In addition, Frijters et al. (2014) find that higher birth-weight predicts higher reported adult life satisfaction.

Frijters et al. (2014) suggest that poorer mental health as a child predicts lower adult life satisfaction. As a result, we condition upon the individual's mental health as a child, as measured by the BSAG at age 7.34 We control for the BSAG at age 7, rather than at age 11, because bullying victimisation at age 11 cannot affect mental health problems at age 7. The BSAG questionnaire contains 11 phrases that describe a child’s mental health problems and their teacher is asked to underline the phrases which best characterise the child.35 The scores for each of these 11 "syndromes" are summed to give a total "syndrome" score, a measure of the number of mental health problems

34

For further information on the BSAG, see Stott (1987). For other uses of the BSAG, (see Delaney and Doyle, 2012; Goodman and Sianesi, 2005; and McAllister et al., 2012).

35

Each "syndrome score" measures the extent to which the child has a specific problem or "syndrome". Clark et al. (2007) argue that 6 of the "syndrome" scores measure externalising problems and 4 measure internalising problems. "Hostility towards children", "hostility towards adults", "inconsequential behaviour", "restlessness", "anxiety for acceptance by children" and "anxiety for acceptance by adults" are measures of the propensity to externalise. Whereas, the measures of the propensity to internalise are "depression", "withdrawal", "unforthcomingness", and the "writing off of adults and adult standards". The final "syndrome" score is "miscellaneous symptoms" which is a measure of neither internalising nor externalising problems.

30 associated with the child. The index is normalised so that the highest observed total "syndrome" score is equal to 1 and the lowest observed score equals 0, an approach used by Frijters et al. (2014). A value of 1 (0) denotes the individual with the most (fewest) mental health problems.

By construction, the estimate of 𝛽1 in equation (2.7) cannot be characterised by

reverse causality because the bullying variable clearly predates the SWB variables. However, equation (2.7) may still suffer from omitted variable bias. The results will suffer from omitted variable bias if the error terms of the SWB models, conditioning on a rich set of controls as well as their adult averages, still contain unobservables that affect the likelihood of being bullied at age 11.