Person-Centered Trajectories of Psychopathology From Early Childhood to Late Adolescence

Key Points Question How can the profiles and trajectories of psychopathology from early life (age 3 years) to late adolescence (age 17 years) be mapped with person-centered methods to accommodate problems of comorbidity and movement between different phases of development? Findings This cohort study of 2 nationally representative cohorts of Irish children developed 4 distinct developmental profiles for person-centered psychopathological trajectories. Over 44% of children transitioned into 1 of the 3 psychopathological profiles during development, with less than 5% of the sample experiencing persistent psychopathology. Meaning These results suggest that psychopathology is common in childhood but mostly limited in duration and that optimizing care in the context of finite resources may require the early identification of those with persistent phenomena.

This supplemental material has been provided by the authors to give readers additional information about their work.

eMethods.
Participants. The Child cohort (Cohort '08) comprises 11,134 infants born in Ireland in [2007][2008], who were randomly sampled from the child benefit register (a universal monthly social welfare payment). In total, 16,136 families were approached to participate in the 9-month wave 1 of the study and 11,134 agreed. Families from this 11,134 were approached for subsequent follow-ups interviews (with limited exceptions such as emigration, death or refusal) and 9793 agreed at age 3, 9793 at age 5 and 8032 at age 9. The final sample for this investigation included 7,507 families who took part in the study at ages 9 months, 3 years, 5 years and 9 years old. The sample was compared to children under the age of 1 reported in the 2006 Irish census on 11 key variables including family social class, child's gender, mother's nationality, family structure. A sampling weight was than applied to Cohort '08 such that the sample mirrored the 2006 Census on these characteristics. Additionally to maintain representation the sample was reweighted to account for attrition between waves. The final sample was representative of children born in Ireland during this time period and who continued to live in Ireland until at least age 9 (see Thornton et al., 2013).
Missing data comparisons are conducted after each wave of the study (see eTable 1 below for further information). Using this information the sample is reweighted to account for nonparticipation at the follow up waves. The non-completion weight was combined with the original census weighting to maintain representation.
For the Adolescent sample (Cohort '98) 910 national primary schools were contacted for participation, of which 82% agreed. Within the participating schools, families of children who were aged 9 were approached for inclusion and 57% agreed to participate (n=8658). Those who participated at age 9 were contacted for the follow up waves (with limited exceptions) and 7423 agreed at age 13 and 6212 agreed at age 17. For this investigation the final sample includes the 6039 families who took part at ages 9, 13 and 17. Similar to Cohort '08, this sample was adjusted using a sampling weight to ensure national representation. The sample was compared with the children aged 9 in the 2006 Irish Census based a number of key variables and readjusted for over/under representation. The sample was also reweighted for attrition between the waves. Thus, the final sample was representative of the children living in Ireland who were 9-years-old in 2006 and who continued to live in Ireland until at least age 17/18 years. Missing data comparisons are conducted after each wave of the study (see eTable 1 below for further information). Using this information the sample is reweighted to account for non-participation at the follow up waves. The non-completion weight was combined with the original census weighting to maintain representation.

Strengths and Difficulties Questionnaire additional information. The strengths and difficulties
questionnaire consists is a twenty-five item questionnaire measuring prosocial behaviour and four psychopathology sub-scales; emotional symptoms, conduct problems, hyperactivity or inattention, peer problems (Goodman, 1997). Each subscale consists of five items with possible responses including "not true", "somewhat true" and "certainly true" and scores ranging from zero to two, respectively. Subscale scores were the sum of the scores in each item of the subscale. At each wave, parents/guardians of the participating child/adolescent were required to respond to each item on the questionnaire. The means and standard deviations of each subscale at each wave are presented in Table 1 and the pairwise linear correlation between the subscale scores at each wave of both cohorts are presented in supplementary eTable 1 and eTable2. There were small to medium significant positive correlations between each of the subscales at all waves of the study (r range: 0.22-0.49, all p <.001). The cross-sectional correlations were mostly stable over time (r range difference: 0.01-0.16).

Statistical Analysis.
Analyses were carried out in MPLUS (Muthén & Muthén, 2009) and Stata (Statacorp, 2017). Both cohorts were analysed in the same way and both were weighted for national representation. The demographic characteristics of each cohort are presented in Table   1. Latent profile transition analysis involved exploratory latent profile analysis at each age of both cohorts (Child sample: 3, 5, 9 years; and Adolescent sample: 9, 13 and 17 years) followed by latent profile transition analysis as per Masyn (2013). Latent profile transition analysis allows for examination changes in identified subgroups of a latent construct (in this case psychopathology) over time based on estimated probabilities of subgroup membership at each time point and the estimated probability of transition between the subgroups over time. Within this analysis these models were estimated with profile invariant diagonal variance-covariance matrix structure such that variance for the same construct (i.e. emotional problems) are constrained to be equal across profiles and residual covariance's were set to zero (Johnson, 2021). This prevented profile items from co-varying above there conditional profile membership. Nylund Asparouhov & Muthén (2007) recommend the use bayesian information criteria and the bootstrapped likelihood ratio test to indicate correct class selection based on monte carlo simulation studies. Due to the representative weighting, it is not possible to calculate a bootstrapped likelihood ratio test. Thus, the optimum number of profiles were determined based on bayesian information criteria and adjusted-bayesian information criteria as well as profile interpretability. Global information criteria do not always arrive at a single lowest value. Thus, in line with Masyn's (2013) suggestion, the optimum model fit was determined based on the diminishing gains in model fit (elbow). The elbow in the model fit metrics refers a point of inflection in the model fit gains. This is reflected by a levelling-off in the improvement in the model fit with the addition of another profile, such that adding the additional profile only marginally improves the fit to the data relative to the improvement seen by the addition of previous profile. Visually this point resembles an elbow in the model fit data points. The global fit metrics were independently reviewed by two of the authors (CH and LS).
Model selection is a parsimonious procedure such that including an additional profile should require notable improve model fit. Examining for an elbow in the fit metrics can assist with model selection. This may be particularly important in the case of latent profile analysis where the items are continuous and thus have a greater quantity of information than latent class analysis where items are binary. Thus, within latent profile analysis there may not always be a clear point at which the model fit stops improving with additional profiles and so the diminishing gains in model fit is used to guide the process via a search for the elbow in the model fit metrics. We examined the number of profiles, and the prevalence of each profile and its transition period. Given the duration and significance of the developmental period being investigated, it is conceivable that the mean and variance of the items loading to each profile may differ across this period and thus the profiles were allowed to vary with time. In addition to these theoretical reason this was statistically examined using Bayesian information criteria, a -2log likelihood ratio test for nested models (see https://www.statmodel.com/chidiff.shtml) and global classification accuracy as assessed by entropy. The profile labelling was based on the mean and variance of the scores in the SDQ. For ease of interpretability this is presented in Figure 1 and 2 in percentiles. Profile assignment was based on the highest designated posterior probability.
Multinomial logistic regression was conducted to examine potential differences in sex distribution between the profiles, transitions and in the persistence of the profile. The results are reported as Incidence Risk Ratios (IRR). This was conducted as an auxiliary analysis based on the profile assignment at each age. These analyses were also weighted for national representation and attrition within the sample.
We conducted a cross-cohort comparison at age 9 years. We compared the prevalence of the profiles and discrepancies in sex distribution of the profiles using multinomial logistic regression. These analysis were adjusted for demographic differences between the sample including primary care givers highest level of education, nationality of the child and nationality of the . We also report the effect sizes (Cohens d) for differences in the interpretation of the profiles between the cohorts at age 9 years. The third profile however had limited distinction from the high psychopathology class at age 3 and 5. By age 9 it fails to distinguish an identifiable subgroup and merely reflects a "middle level difficulties" across the subscales. The 5-profile solution had better model fit then the 4-profile solution (lower BIC) however, it had notably diminished model gains (relative to the 3-4 profile change gains (eFigure 2b and d). This was indicative of an elbow at the 4-profile solution.
Additionally the 5-profile solution was similar to the 4-profile solution re-identifying the same four profiles. The additional profile was very similar to the high psychopathology profile and did not have clear distinguishing features. The lack of distinction added to our decision that a fourprofile solution was optimal and aligned better with the literature. Cross-Cohort Comparison. Both samples measured children at age 9 using the same psychopathology measure allowing for a 10 years cross-cohort comparison. There were few meaningful differences in mean SDQ scores in each profile between the cohorts. The exception was elevated peer and hyperactivity problems in the High Psychopathology profile of the Child sample (medium effect sizes). All other between-cohort differences in SDQ scores were small (absolute range: 0.03-0.31, see Supplementary