Exploring the differentiation of behavioural and emotional problems across childhood: A prospective longitudinal cohort study

Abstract Background An individual's overall burden of behavioural and emotional problems across childhood is associated with increased likelihood of later mental health conditions. However, the relative extent of behavioural versus emotional problems ‐ that is, the extent to which the domains are differentiated from one another ‐ may provide additional information about who is at risk of developing a mental health condition. Here, we seek to validate differentiation as an independent predictor of later mental health conditions, and to explore its aetiology. Methods We analysed data from ~79,000 children in the population‐based Norwegian Mother, Father, and Child Cohort Study, and linked health‐care registries. In preregistered analyses, we modelled the extent and rate of differentiation of behavioural and emotional problems between ages 1.5–5 years, and estimated associations with later symptoms (age 8) and diagnoses (after age 8). We also explored the aetiology of differentiation by estimating associations with early life exposures and, in a subset of 23,945 full siblings, assessing the impact of accounting for unobserved familial confounding. Results Differentiation of behavioural and emotional problems was associated with later symptoms and diagnoses of mental health conditions, independent of total problems. Maternal at‐risk drinking (β = 0.04 [0.02, 0.06]) and parental relationship problems (β = 0.04 [0.02, 0.05]) were associated with higher behavioural relative to emotional problems at age 5. Maternal prenatal distress (|β| = 0.04 [0.03, 0.06]), concurrent distress (|β| = 0.04 [0.02, 0.06]) and parental education (|β| = 0.05 [0.04, 0.07]) predicted higher emotional relative to behavioural problems at age 5. Estimates for maternal prenatal distress and at‐risk drinking were consistent across both unadjusted and adjusted analyses accounting for unobserved familial risk. Conclusions Differentiation of behavioural and emotional problems in early childhood represents a valid source of inter‐individual variability linked to the later emergence of psychopathology and may be relevant for early detection and prevention strategies for mental health.


INTRODUCTION
Behavioural and emotional problems often co-occur in young children.It is also well established that behavioural problems (i.e., undercontrolled and disruptive behaviour) and emotional problems (i.e., negative mood states and inhibition) in early childhood are associated with an increased risk of mental disorders later in development (Caspi et al., 1996;Mulraney et al., 2021).While common, the co-occurrence of behavioural and emotional problems is not universal.Many children display behavioural problems without concurrent emotional problems, while others exhibit no behavioural problems despite considerable emotional problems.The relative extent of children's difficulties in these domains -that is, the extent to which behavioural and emotional problems are differentiated from one another -may provide additional information about who is at risk of developing specific mental health conditions.
When seeking to understand the relationship between childhood behavioural and emotional problems and later mental disorders, researchers have typically taken one of two approaches.Either behavioural and emotional problems are considered separately (emphasising their distinctiveness, i.e., Lund et al., 2020), or they are combined into a single dimension (emphasising their overlap, i.e., Scott et al., 2009).Comparatively little research has focused explicitly on the extent of differentiation between these domains; that is, inter-individual variability in the concordance between behavioural and emotional problems.There are reasons to believe that this variability -which is, by definition, unrelated to the total burden of problems an individual experiences -may provide important additional information about the risk for specific, future psychiatric problems.First, clinicians rely on distinct profiles of symptoms to assign diagnoses.This means that, while having a high burden of problems across domains may increase an individual's likelihood of presenting for clinical diagnosis, the specific and relative pattern of their symptoms will inform which diagnosis they receive.Second, decreasing correlations among mental health conditions as children grow older may be indicative of differentiation.Previous studies have shown that the largest decreases in correlations among pairs of mental health conditions were between those belonging to the behavioural and emotional domains rather than within either domain (Sterba et al., 2010).Third, previous research has found that when a general psychopathology ("p") factor is accounted for, specific behavioural and emotional problem factors are associated with problems in a wide range of domains (Caspi et al., 2014).Finally, even without explicitly defining differentiation of emotional and behavioural problems, studies showing independent or heterogeneous associations between these domains and later mental health outcomes (Mulraney et al., 2021) provide indirect evidence that the predictive capacity of overall problems in childhood for such outcomes is more than the sum of its parts.
Insofar as differentiation of behavioural and emotional problems in childhood proves to be independently predictive of later mental health outcomes, exploring its aetiological basis will be important for informing early detection and prevention strategies.Observational studies have identified associations between various early environmental risk factors and child behavioural and emotional problems, including-but not limited to-socioeconomic disadvantage (Caspi et al., 2016;Costello et al., 2003;Dearing et al., 2006;Leventhal & Dupéré, 2019), maternal depression (Gjerde et al., 2021;Goodman et al., 2011), adverse childhood experiences (Liming & Grube, 2018), parental disharmony (Auersperg et al., 2019;Jenkins & Smith, 1991), and parental smoking and at-risk drinking (Brion et al., 2010;Lund et al., 2020).It is likely, though not yet established, that similar factors are involved in differentiating behavioural and emotional problems in early childhood.Given the expectation that the extent of differentiation of behavioural and emotional problems may increase over time, it is notable that behavioural genetic evidence provides consistent support for the notion that non-shared environmental factors -to a greater degree than genetic factors -influence developmental change in behavioural and emotional problems (Hannigan et al., 2017).
To explore the aetiological underpinnings of differentiation of behavioural and emotional problems in childhood, it is essential to consider possible routes of confounding.Both childhood behavioural and emotional problems (Polderman et al., 2015) and measures of the childhood environment (Kendler & Baker, 2007) are influenced by genetic factors.Since parents provide both genes and environments to their children, this leads to confounding by gene-environment correlations (Plomin et al., 1977).Therefore, it is crucial to study early environmental risk factors while appropriately accounting for potential confounding due to unobserved familial risk.While recently developed methods offer ways of doing this via the incorporation of parental genetic data (Cheesman et al., 2020;Eilertsen et al., 2022), the most powerful designs for partitioning out covariation attributable to confounding influences on childhood exposures and outcomes continue to rely on structured family data.Foremost among these are

Key Points
� Co-occurring behavioural and emotional problems in early life are strongly associated with later mental health conditions.However, the relative extent of problems in these domains (i.e., differentiation) may provide additional information about who is at risk and for which disorders, essential to detecting and preventing the emergence of psychopathology.
� Findings show that differentiation of behavioural and emotional problems in early life strongly associates with the development of mental health conditions later in childhood and adolescence.
� We identify potentially modifiable environmental factors predicting differentiation in early life, while accounting for confounding by unobserved familial risk.

of 11
- adoption (Kendler et al., 2022) or assisted conception (Rice et al., 2013) designs, wherein the separate influences of biological and rearing parents can be conclusively delineated.However, the sibling comparison design is a powerful alternative.In this design, information from siblings with discordant exposure to putative environmental risks, and/or differential scores on outcomes, can be used to estimate an exposure-outcome association free of unobserved familial confounding.Although models based on sibling comparisons are not without limitations -including risks of biases being induced as well as removed (Frisell, 2021) -comparing unadjusted and adjusted estimates from these models can provide vital context with which to interpret exposure-outcome associations.Sibling comparison models, and other designs equipped to partition out familial confounding, have been widely used to study associations between early life exposures and childhood mental health (for a recent review, see Jami et al., 2021), but not previously with a specific focus on uncovering which exposures may influence how behavioural and emotional problems differentiate from one another across development.
Here, we extend earlier studies of differentiation in childhood by investigating its links to early life exposures in models that account for common sources of shared confounding.We propose a novel operationalisation of differentiation as a developmental process: that is, each person's relative level of, and rate of change in, behavioural versus emotional problems during early childhood.We seek to validate that differentiation operationalised in this way is a relevant factor in the developmental emergence of psychopathology, by establishing whether it is associated with symptom scores in middle childhood, and registry-based diagnoses of mental health conditions in later childhood and adolescence, independently of total problems.
Next, we explore associations between early life exposures and differentiation across childhood, looking for attenuation in effects estimated within sibling pairs as an indicator of the influence of biases from unobserved familial confounding.The data in MoBa was collected by questionnaires from early pregnancy to middle childhood, provided primarily by mothers (around week 17 and 30 of pregnancy, when the child was 6 and 18 months, and at 3, 5, and 8 years).Phenotype data from the fathers were collected by questionnaire around week 17 of the pregnancy.

Differentiation of behavioural and emotional problems
We used the Child Behaviour Checklist (CBCL) to assess behavioural problems (8 items) and emotional problems (5 items), at ages 1.5, 3, and 5 years.Information about psychometric properties of the CBCL subscales are provided in Appendix S1 and Table S1.Differentiation was operationalised as the difference between standardised scores of behavioural and emotional problems (behavioural problemsemotional problems = difference score; see Figure 1, panels A/B for an illustration).This means that individuals with high difference scores have relatively more behavioural than emotional problems, while individuals with low scores have the inverse.

Total behavioural and emotional problems
We also combined behavioural and emotional problems as a sum score in order to provide a point of reference for analyses of the differentiation scores described above.Total levels of emotional and

Symptoms of mental health conditions at 8 years
We used the 13-item Short Mood and Feelings Questionnaire (SMFQ; Angold et al., 1995) to assess symptoms of depression, a 5-item version of the Screen for Child Anxiety Related Disorders (SCARED; Birmaher et al., 1997) to assess symptoms of anxiety, and a 34-item version of the Rating Scale for Disruptive Behaviour Disorders (RS-DBD; Silva et al., 2005) to assess symptoms of hyperactivity and inattention (subscales related to attention-deficit hyperactivity disorder [ADHD], 18 items), conduct disorder (CD, 8 items), and oppositional defiant disorder (ODD, 8 items).The measures were treated as continuous, and scores were standardised to have a mean of 0 and standard deviation of 1. Information about psychometric properties of all symptom scales at age 8 are in Appendix S1 and Table S2.

Diagnoses of mental health conditions
We linked to KUHR and NPR to obtain diagnoses of mental health conditions from medical records (depression, anxiety, ADHD, and disruptive behaviour disorders [DBD], combining CD and ODD; see Appendix S2 for information about diagnostic codes).Individuals were classified as a "case" if they had received a relevant diagnosis at least twice in either primary or secondary health care, or once in each, between their eighth birthday and July 2021, the end of the available follow-up.

Early life exposures
We included measures of parental income and education, and both mother-and father-reported adverse life events, mental distress, relationship problems, smoking and at-risk drinking, and maternal alcohol problems and postnatal depression as predictors.A full overview of variables, including information about variable processing, is in Table S3.All deviations from the preregistered list of early life exposures, and reasons for these deviations are described in Appendix S3.

Covariates
We included the child's sex (as registered at birth in MBRN) and parity as time-invariant covariates and the child's age at questionnaire return as a time-varying covariate.Mother ID was used for clustering in all models.

Statistical analyses
All modelling was carried out in R version 4.1.2,including analyses requiring Mplus (Muthén & Muthén, 1998), through the MplusAutomation package version 1.0.0 (Hallquist & Wiley, 2018).The phenotools package version 0.2.8 was used to process questionnaire and registry data (https://github.com/psychgen/phenotools). We used the False Discovery Rate (FDR) to preserve the Type 1 error rates at 5% in all models (Benjamini & Hochberg, 1995).We used a maximum likelihood estimator which computes robust standard errors and a scaled test statistic.

Modelling differentiation as a developmental process
Using latent growth modelling (LGM), we parameterised the development of behavioural-emotional differentiation across three time points (when the children were 1.5, 3, and 5 years of age).Individual differences in the extent and rate of differentiation of behavioural and emotional problems were captured by latent slope and intercept factors (see Figure S1 for an illustration of this model).The intercept was set at 5 years, as we were primarily interested in the endpoint of the children's differentiation process.We also ran equivalent LGMs of total childhood behavioural and emotional problems.In all models we accounted for dependencies in the data (due to participating siblings) by clustering within mothers.

Validation of differentiation
To validate our operationalisation of differentiation, we ran LGMs where both the slope and intercept, and each growth factor individually, were allowed to influence 8-year symptoms of depression, anxiety, hyperactivity, inattention, CD, and ODD.Similarly, we explored associations between differentiation and individuals' likelihood of having a recorded diagnosis of depression, anxiety, ADHD, and DBD, using extracted factor scores as predictors of diagnostic outcomes in a logistic regression framework.To assess the extent of bias arising from non-random non-participation in MoBa, we conducted sensitivity analyses using inverse probability of participation weighting based on aggregated statistics on parity and cohabitation status from Statistics Norway (see Supporting Information Appendixes for further details).

Investigating predictors of differentiation
We entered the 16 early life exposures into the LGMs, initially allowing them to influence both slope and intercept factors to derive estimates of the observational exposure-outcome associations in the full sample.We then tested sub-models in which all exposures influenced the growth process via, alternately, only the slope or intercept factors.Finally, we restricted to the sibling sub-sample (23,945 full siblings) and ran multilevel LGMs, where variance in the exposure variables is decomposed into variance that is specific to a given mother for a given child (level 1) and variance that is specific to a given mother but shared across siblings (level 2).Effects of the exposures on the growth factor(s) were estimated in two versions of this model: (1) unadjusted for familial confounding (equivalent to the observational exposure-outcome associations run in the full sample) and ( 2) adjusted for nuclear family-level confounding.Whilst only controlling for 50% of genetic effects at the child level, this design completely adjusts associations for family-wide factors.This includes most plausible "third variable" confounders of associations between early environmental exposures and child outcomes (i.e., 100% maternal genetic and stable environmental factors influencing predictors).
To account for differences between the sibling sub-sample and the overall sample, due to within-sample selection bias (parents of MoBa siblings, by definition, participate on more than one occasion), we included inverse probability weights based on a logistic regression of sibling-singleton status on all other study variables in all sibling-only models (see Supporting Appendix S5 for full details).
All analyses were again repeated for the total problems outcomes to contextualise the differentiation results.

Inference criteria
Likelihood ratio tests were used to select the best-fitting LGMs.Effect sizes and confidence intervals were used to draw inferences about associations between exposures and outcomes in the full sample analyses.The relative attenuation of exposure-outcome associations in the adjusted versus unadjusted multilevel models was used to guide inferences about the likely role of confounding in observational exposure-outcome associations.
The following fit statistics were reported for each model: the comparative fit index (CFI), Tucker-Lewis index (TLI), standardised root mean square residual (SRMR), and root mean square error of approximation (RMSEA).By convention, an SRMR below 0.1 and an RMSEA of less than 0.05 implies a good fit, as well as CFI and TLI values over 0.95 (Kline, 2011).We also reported the Akaike Information Criterion and Bayesian Information Criterion (BIC), for which lower values are preferred when comparing models.For nested models, −2LL based likelihood ratio tests were used to select the better fitting model.
For non-nested models, the one with the lower BIC was preferred.

Inclusion criteria and sample size
We included all MoBa children with available CBCL data on at least one measurement occasion (see Table S4 for response numbers at each wave, and Appendix S4 for details on handling of outliers).The overall sample size was 78,982 (38,544 female), including 23,945 full siblings.Note that a more limited sample of 31,854 children had CBCL information on all waves, but we used a full information maximum likelihood approach, which includes all available data and yields unbiased parameter estimates assuming data is missing at random or missing completely at random (Enders & Bandalos, 2001).

Availability of data and analytic code
The MoBa data are not publicly available as the consent given by the Data preparation and analysis code for all elements of the project is publicly available on Github: https://github.com/psychgen/childhood-differentiation.

RESULTS
Descriptive statistics for all symptom measures are presented in Table S4.There was evidence of a low level of selective attrition based on the CBCL subscales (see Appendix S6).As expected, the differentiation and total scores were empirically independent (r range : −0.01-0.01).

Modelling differentiation as a developmental process
A linear growth model with both slope and intercept factors provided excellent fit to the data using the differentiation scores of behavioural and emotional problems from age 1.5-5 years (CFI = 0.99, TLI = 0.98, RMSEA = 0.01, SRMR = 0.01; see Table S5 for model comparisons).The correlation between the slope and the intercept was positive (r = 0.64 [0.62, 0.66]).There was evidence of sex differences where males showed greater extent of behavioural relative to emotional problems at age 5 (β intercept = 0.09 [0.08, 0.10]) but similar rates of differentiation over time compared to females (β slope = 0.01 [−0.01, 0.02]).Later birth order was associated with greater extent (β intercept = 0.05 [0.04, 0.06]) and rate (β slope = 0.06 [0.04, 0.07]) of differentiation toward behavioural problems.For total problems, a growth model with both slope and intercept factors also provided excellent fit (CFI = 0.99, TLI = 0.98, RMSEA = 0.02, SRMR = 0.01; see also Table S5).

Validation of differentiation
We tested the associations between differentiation and symptom domains at age 8 (see Figure 2).The best-fitting models included effects from both intercept and slope factors to the 8-year symptom outcomes (i.e., dropping effects led to significant decrement of fit, see Table S6).We observed associations between the slope and/or intercept of differentiation with all outcomes (Figure 2).Overall, differentiation towards behavioural relative to emotional problems in early childhood predicted higher symptoms of behavioural conditions in middle childhood (i.e., hyperactivity, inattention, CD, and ODD), and differentiation towards emotional problems predicted higher anxiety.For total problems, effects from both slope and intercept F I G U R E 2 Results from the validation analysis showing that differentiation in early childhood predicts 8-year symptoms over and above total levels of behavioural and emotional problems.Intercept and slope of differentiation and total problems predicting symptoms of mental health conditions at age 8; ANX, anxiety; CD, conduct disorder; DEP, depression; HYP, hyperactivity; INAT, inattention; ODD, oppositional defiant disorder; R 2 , R squared.factors provided best fit (see Table S6), and associations with 8-year symptoms were generally stronger (see Figure 2).
We then estimated the likelihood of having received relevant diagnoses based on the rate and extent of differentiation during early childhood (see Figure 3).In these analyses, the largest associations were seen for the rate (OR slope = 3. 07 [2.21, 4.27]

Investigating predictors of differentiation
The best-fitting model for differentiation included effects of early life exposures on both slope and intercept factors (see Table S7).In the full sample observational LGM, parental education predicted the extent (β intercept = −0.05[−0.07, −0.04]) and rate (β slope = −0.05[−0.07, −0.03]) of differentiation toward emotional problems.In addition, maternal prenatal and concurrent distress were predictive of the extent and/or rate of differentiation toward emotional problems after FDR correction (see Figure 4).The following predictors of the extent and/or rate of differentiation toward behavioural problems remained after FDR correction in the observational LGM: maternal prenatal smoking and parental relationship problems.In addition, maternal at-risk drinking predicted a higher extent of behavioural problems at age 5 but similar rate of change over time.
Parental education and maternal concurrent distress were the most important predictors of total problems (see Figure S2).
We then ran multilevel models in the sub-sample of siblings, comparing the point estimates for exposure-outcome associations before and after adjusting for confounding by unobserved familial risk (see Figure 5).Maternal at-risk drinking predicted relatively higher behavioural than emotional problems consistently strongly in both unadjusted and adjusted models.Maternal postnatal depression also predicted relatively higher emotional than behavioural problems consistently strongly in both models, whereas the associations for parental education and relationship problems were attenuated.Effects were broadly attenuated for total problems, with parental education and maternal concurrent distress being attenuated to near zero (see Figure S3).Maternal adverse life events remained associated with higher total behavioural and emotional problems, whereas parental income seemed to emerge as a protective factor after adjustment.
To help contextualise the findings relating to differentiation, we also conducted sensitivity analyses with the CBCL behavioural and emotional problems subscales included as separate, correlated outcomes (results for behavioural problems are shown in Tables S9 and   S10, and emotional problems in Table S11 and S12).This sensitivity analysis showed that maternal at-risk drinking was associated with both higher behavioural problems and lower emotional problems, such that the relative distance between them made maternal drinking an important predictor of differentiation.Conversely, the adjusted effect for postnatal depression on differentiation appeared to be primarily driven by postnatal depression being associated with fewer behavioural problems.

DISCUSSION
In summary, we explored the differentiation of behavioural and emotional problems across childhood using data from a large population-based birth cohort.Supporting the validity of our approach, early-life differentiation was associated with symptoms of mental health conditions in middle childhood and clinical manifestations of those conditions later in childhood and adolescence.We identified potentially modifiable parental factors associated with children's differentiation toward behavioural problems (e.g., maternal at-risk drinking) and emotional problems (e.g., maternal prenatal distress).In contrast to observational links between differentiation and parental education or relationship problems, these associations appeared not to be inflated by unobserved familial confounding.
There has recently been much emphasis on the stability and cooccurrence of mental health conditions in childhood, which can be captured by a general "p" factor (Allegrini et al., 2020;Lahey et al., 2015;Murray et al., 2016).However, it remains the case that specific components of liability to psychopathology not captured by this general factor are consequential (Caspi et al., 2014).Here, taking a statistically distinct but theoretically consistent approach to extracting this specific signal, we find evidence of systematic signal in the process of differentiation of behavioural and emotional problems in early childhood, independently of co-occurring (i.e., total) problems.It is important to note that the total level of behavioural and emotional problems was a considerably stronger predictor of symptoms in middle childhood than differentiation between these domains.This is consistent with the power of the "p" factor in predicting F I G U R E 3 Validation analysis showing associations between differentiation and total problems in early childhood and odds of receiving diagnoses later in childhood and adolescence.Estimates were derived based on extracted factor scores, predicting diagnoses of mental health conditions (after age 8) in logistic regressions; ANX, anxiety; DEP, depression; ADHD, attentiondeficit hyperactivity disorder; DBD, disruptive behaviour disorders, including oppositional defiant disorder and conduct disorder.future outcomes, which is well established (Caspi et al., 2014;Lahey et al., 2015).Nonetheless, the proportion of the variance in middle childhood symptoms accounted for by differentiation was also substantial.Modelling differentiation may provide additional information about the aetiology of specific patterns of emerging mental health problems.Separating general and specific aspects of mental health outcomes as we do here, or by means of a p-factor model with specific, orthogonal sub-factors (e.g., Chen et al., 2022) may enhance the informativeness and discriminatory power of any identified risk factors.
Having established that differentiation operationalised as the difference between behavioural and emotional problem scores has predictive validity for later outcomes, we estimated associations with a range of early life exposures.Observational associations were found for maternal at-risk drinking and parental relationship problems predicting differentiation towards behavioural problems, and maternal prenatal distress, concurrent distress, and parental education predicting differentiation towards emotional problems.However, when adjusting for unobserved familial confounding using a sibling comparison design, most associations were attenuated.First, it is worth noting that random measurement error could be sufficient to attenuate the within-sibship estimates in comparison to unadjusted estimates, even if there was no confounding (Frisell et al., 2012).
Nonetheless, the relative consistency of effect estimates for maternal at-risk drinking and prenatal distress across observational, unadjusted (sibling only), and adjusted (sibling only) analyses suggests a robustness to these results.Some associations switched sign or seemingly emerged after adjustment, which could indicate that familial confounding may, in some cases, mask rather than inflate effect estimates.

F I G U R E 4
Observational results of early life exposures on the extent and rate of differentiation in childhood in the full sample.Standardised betas (+behavioural and -emotional) for the observational associations of predictors with the intercept (β I ) and slope (β S ) of differentiation; the upper left quadrant represents converging behavioural problems; the lower left quadrant represents differentiation toward emotional problems; the upper right quadrant represents differentiation toward behavioural problems; the lower right quadrant represents converging emotional problems; predictors that are labelled in the plot were significant after False Discovery Rate (FDR) correction, and predictors that are transparent did not; lines denote 95% confidence intervals; B, behavioural; E, emotional.
F I G U R E 5 Full sample, sibling unadjusted and adjusted associations between early life exposures and differentiation in childhood.Standardised betas (−emotional and +behavioural) for observational (N = 78,982), as well as sibling adjusted and unadjusted (N = 23,945 full siblings) associations between predictors and the intercept (β I ) and slope (β S ) of differentiation; only False Discovery Rate (FDR) significant predictors from the observational analysis are shown here; note that the sibling adjusted results are imprecise due to the weighting procedure used to adjust for differences between siblings and singletons, meaning that attention should mainly be paid to the attenuation (or not) of the point estimates.

DIFFERENTIATION OF BEHAVIOURAL AND EMOTIONAL PROBLEMS IN CHILDHOOD
For instance, paternal prenatal distress seemed to emerge as a predictor of differentiation toward behavioural problems.However, we refrain from -and caution against -interpreting these associations as such.This is partly because they could be unreliable, or biased (to a larger extent than the unadjusted estimate) by confounders not shared by siblings (Frisell et al., 2012).Moreover, despite our attempts to adjust for this, they could be affected by selection bias due to being a sibling participating in MoBa.Due to these alternative possible explanations, we restrict our interpretations to effects that are consistent across the observational, unadjusted, and sibling adjusted analyses, in line with a triangulation approach (Lawlor et al., 2016).
For some early life exposures (e.g., maternal concurrent distress), results were inconsistent across the observational analyses in the whole sample and the unadjusted analyses in the sibling sub-sample.
In this sub-sample, some selection effects remained after weighting (see Appendix S7 and Figure S6).Selection bias related to repeated participation (i.e., being a sibling vs. a singleton) may be the reason behind these seemingly conflicting results.Also, the associations between early life exposures and the rate of differentiation were of much smaller magnitude in the weighted sibling sub-sample.This could perhaps suggest that the slope factor picked up effects in the whole sample that were in part due to selective attrition, which were attenuated in the sibling-only analyses.
The attenuation of the effects of parental education and maternal concurrent distress after accounting for familial confounding is also notable.Parental education is virtually invariant between siblings, which most likely accounts for this.Results for parental education should be considered tentative and warranting further investigation (Costello et al., 2003;Torvik et al., 2020).Similarly, for other risk factors that rarely vary within sibling pairs (i.e., maternal smoking during pregnancy; see within and between cluster variation for each risk factor in Table S8) the adjusted results rely on fewer observations and should not be considered conclusive.Other designs, such as the children-of-siblings design (Kuja-Halkola et al., 2014) are likely superior when estimating the effect of exposures that seldom vary within families on offspring outcomes.The lack of consistent effects of maternal concurrent distress in our study seemingly conflicts with prior reports (i.e., Gjerde et al., 2017).This could be explained by methodological differences between our and previous studies.

Limitations
There are notable strengths to this study, including the large sample size, the detailed preregistration, and the ability to adjust for confounding by unobserved familial risk.Nevertheless, we acknowledge some limitations of our approach.First, we use a differentiation score and draw a contrast with approaches that combine emotional and behavioural problems into a single, total score.However, it is equally common that these subdomains are investigated independently (as per the sensitivity analyses reported in Figures S4 and S5) -in which case, the information contained is exactly equivalent to a model with both a differentiation and total score.Therefore, it should be noted as a clarification (if not a limitation per se), that using a differentiation score is a way to re-frame existing information, rather than a way to access novel information from a measure.For our research questions, this re-framing is pragmatic and helpful; in other cases, use of the variables as separate subdomains may be more appropriate.Second, as with all difference scores, measurement unreliability means that our differentiation scores include noise in addition to signal (Edwards, 1994; see also reliability estimates in Table S1).Here, latent growth factors capture signal that is shared across waves without measurement error, which helps with mitigating this issue.Moreover, our validation results indicate that the differentiation measure does capture meaningful signal.Third, since mothers reported on both the predictors and outcomes, shared method variance might inflate some of the observed associations (Podsakoff et al., 2003).This limitation is inherent to MoBa and similar cohorts, although recent work suggests that any resulting bias may be limited (Olino et al., 2021).Here, any time-invariant maternal rating bias would to some extent be adjusted for in the multilevel SEM models.
In addition we have data on predictors from the fathers, which would be less impacted by shared method variance.Fourth, self-report of certain risk factors (e.g., smoking during pregnancy) is limited to the extent that parents are willing to report on such sensitive topics.
Fifth, differentiation could result from unreliability of symptom measures and clinician-assigned diagnoses.However, this would again be adjusted for in the multilevel SEM.Finally, we found some evidence of selective attrition in our sample, although this was of limited scope to bias the results.Overall, MoBa is not fully representative of the general population due to non-participation bias.
Previous studies have found that the youngest women, those living alone, smokers, women with previous stillbirths, and women with more than two previous births, are under-represented in MoBa (Nilsen et al., 2009).However, sensitivity analyses using inverse probability of participation weighting to account for potential biases showed some, but overall limited, differences between weighted and unweighted results (see Figures S4 and S5 for the validation models, and Tables S13-S20 for the predictor models).

CONCLUSION
We showed that differentiation of behavioural and emotional problems in early childhood was associated with symptoms of mental health conditions in middle childhood, and clinical manifestations of these conditions later in childhood and adolescence.These effects were independent of the total extent of individuals' behavioural and emotional problems in early childhood.Further, we identified specific predictors of differentiation remaining after adjustment for confounding by unobserved familial risk.Identifying modifiable factors associated with differentiation of behavioural and emotional problems in early life may help efforts to detect and eventually prevent impairment from mental health conditions.acquisition; Methodology; Software; Supervision; Validation; Visualization; Writing -original draft; Writing -review & editing.

Sample
We used data from a population-based sample of children from the NorwegianMother, Father, and Child Cohort Study (MoBa;Magnus et al., 2016Magnus et al., , 2006) ) conducted by the Norwegian Institute of Public Health.Participants were recruited from all over Norway from 1999 to 2008.The women consented to participation in 40.6% of the pregnancies.The cohort now includes 114.500 children, 95.200 mothers and 75.200 fathers.The analyses were based on version 12 of the quality-assured data files released for research in January 2019.We also used data from the Medical Birth Registry of Norway (MBRN), a national health registry containing information about all births in Norway.In addition, we obtained data on diagnoses of specific mental health conditions from the Norwegian Patient Registry (NPR) and "Kontroll og utbetaling av helserefusjoner" (KUHR).
behavioural problems were computed by adding up standardised scores of the two dimensions in the overall sample (behavioural problems + emotional problems = total score).By definition, these scores are completely orthogonal to the differentiation scores.

F
Operationalisation and properties of CBCL differentiation/total scores.Panel A shows how the differentiation score is constructed based on individual scores on the behavioural and emotional subscales of the CBCL in 100 randomly selected individuals from a simulated dataset (N = 1000; r beh_emo = 0.5, M emo = 5, M beh = 6, SD emo/beh = 2.5); B and C show that since the computed scores are orthogonal to one another, it is possible for the differentiation score (in B) to be associated with an outcome which the total score (in C) is not; CBCL, Child Behaviour Checklist.DIFFERENTIATION OF BEHAVIOURAL AND EMOTIONAL PROBLEMS IN CHILDHOOD participants does not open for storage of data on an individual level in repositories or journals.Researchers who want access to data sets for replication should submit an application to datatilgang(at)fhi.no.Access to datasets requires approval from The Regional Committee for Medical and Health Research Ethics in Norway and an agreement with MoBa.