Abstract
To date, a substantial body of research exists suggesting an association between indicators of intelligence and various deleterious outcomes, including externalizing and internalizing behavioral problems. Much of this research, however, has focused on samples drawn from the general population, thus it remains less clear how (and if) intelligence relates to problem behaviors in samples of highly at-risk individuals. The current study seeks to contribute to this knowledge base by examining the associations between intelligence and internalizing, externalizing, and total scores on the Child Behavioral Checklist in a sample of approximately 2,500 highly disadvantaged and at-risk respondents. While the two measures of intelligence performed differently, there emerged some association between overall lower IQ and higher total behavioral problem scores. There was some evidence that low IQ also predicted higher internalizing scores, but this relationship varied greatly by measure and model. Results, limitations, and implications of the current study are discussed.
Introduction
The study of human intelligence over the last century has continued to provide meaningful insights about a range of important developmental and social outcomes (Kline, 2013; Ritchie, 2015). Indeed, what seems rather clear at this point, is that variation in intelligence predicts variation across key social outcomes across all phases of the life course, beginning in childhood and spanning into adulthood (Gottfredson, 2004; Plomin & Deary, 2015; Ritchie, 2015). In what constitutes one of the more notable tests on the topic, Moffitt and colleagues (2011), using a sample of respondents tracked from birth until the third decade of life, uncovered evidence that along with self-control, intelligence was consistently associated with indicators of health and economic success across years of the lifespan.
Other studies have uncovered very similar patterns of effects, revealing positive associations between intelligence and educational attainment, accrual of wealth, increased selfregulation, upward social mobility, selection of friends and romantic partners (i.e., assortative paring), and even longer life expectancy, across multiple independent samples (Arden et al., 2015; Beaver et al., 2016; Calvin et al., 2017; Gottfredson, 2004; Boutwell et al., 2017; Meldrum et al., 2017; Plomin and Deary, 2015). Not only have higher levels of intelligence been linked to positive social outcomes, the inverse also seems to be true, in that lower levels of intelligence predict various adverse life events. Beaver et al (2016), for instance, using a nationally representative sample of over 15,000 participants, found that those in the bottom 25% of IQ scores were almost twice as likely to be victimized as those in the top 25%. Lower levels of intelligence are also associated with an increased risk of self-reported criminal justice processing (being arrested and incarcerated) in adulthood (Beaver et al., 2013).
Finally, Calvin and colleagues (2017) recently analyzed a large sample of Scottish participants, and uncovered an association between childhood intelligence and various causes of mortality across several decades of the life course. In particular, children who scored higher on childhood measures of intelligence, were less likely to die from a variety of adulthood conditions such as coronary heart disease and smoking related cancer (see Calvin et al., 2017 for additional detail; see also, Arden et al., 2015). Though far from exhaustive, this body of evidence clearly suggests that variation in measures of intelligence represents a robust correlate for a wide swath of phenotypes (see also Aarons, James, Monn, Raghavan, Wells, & Leslie, 2010; Ritchie, 2015; Vaughn, Shook, & McMillen, 2008).
Remaining Questions to Ask About Intelligence
Despite the consistent pattern of findings briefly outlined above, there remains several interesting gaps in the literature that need to be examined. In particular, less effort to date has been aimed at examining whether, and to what extent, intelligence predicts adverse or antisocial outcomes in samples of highly at-risk participants. Children exposed to abuse, trauma, neglect, and maltreatment on the part of their caregivers, for instance, are at risk for a host of maladaptive outcomes psychological and behavioral problems among them (Flouri, Midouhas, & Joshi, 2015; Jonson-Reid & Drake, 2017). Given the broad nexus of risk factors these children are often exposed to (in general, see Jonson-Reid & Drake, 2017), it seems important to further explore whether indicators of intelligence might explain variation for behavioral and psychological problems, in a population broadly exposed to a range of risk factors.
Additionally, most of the studies outlined above examined openly manifested forms of antisocial behavior (i.e., violence, arrests, etc.), as opposed to using more clinically relevant instruments such as scales designed to assess externalizing psychopathologies early in the life course. Additionally, even less work (relatively speaking), has specifically focused on the emotional, or “internalizing” form that certain psychopathologies can assume. And to be sure, very little evidence exists concerning the association between intelligence and both externalizing and internalizing problems in a sample of extremely disadvantaged respondents—such as children exposed to abuse and neglect early in life. Below we discuss some of the literature that has attempted to address this topic, then transition to outlining the goals of the current study.
Intelligence, Adverse Behavior, & At-Risk Populations
As a brief reminder, psychopathology—from a clinical standpoint—can be broadly subsumed in two main categories: externalizing and internalizing problems (Caspi et al., 2014; Kotov et al., 2017), with a smaller third category sometimes reported for thought disorder or psychotic experiences (Wright et al., 2013). These hierarchical clusters have been replicated cross-culturally (Kessler et al., 2011) and across the life course, manifesting in both children/adolescents (Achenbach, 1966; Lahey et al, 2011) and adults (Krueger & Markon, 2006). Generally speaking, externalizing problems encompass overt displays of aggression and impulsivity (del Giudice, 2016). Internalizing problems, on the other hand, involve difficulties with anxiety, depression, and other less overt forms of psychopathology. Researchers examining the origins and nature of psychopathology have long recognized the tendency for externalizing and internalizing problems to be comorbid with one another (Caspi et al., 2014; Lilienfeld, 2003), perhaps representing the manifestations of a singular latent construct (Caspi et al., 2014). Put another way, individuals with internalizing (or externalizing) problems, are at higher odds of experiencing some form of the other class of psychopathology.
At the same time, extant research has examined the possible risk factors, both phenotypic and genetic, which might account for variation in these psychopathologies, and also perhaps shed light on the co-occurrence of internalizing and externalizing problems (Del Giudice, 2016). To date, intelligence seems to have a clear association with externalizing problems (Guay, Ouimet, & Proulx, 2005; Hinshaw, 1992; Menting, Hirschi & Hindelang, 1977; Menting, Van Lier & Koot, 2011; see also Aarons, James, Monn, Raghavan, Wells, & Leslie, 2010; Vaughn, Shook, & McMillen, 2008). In a review of the longitudinal research on IQ, school achievement, and externalizing behavior problems, Hinshaw (1992) concluded that low IQ was a strong predictor of increased behavioral problems.
Yet, to reiterate, it remains unclear whether similar findings would emerge in highly at-risk samples of children —children, for example, who are members of families that have been actively investigated by Child Protective Services (CPS) for alleged abuse or neglect—or for measures of internalizing disorders. One recent study has helped to shed some insight on to this topic. Using a sample of socioeconomically disadvantaged families in the United Kingdom, Flouri, Midouhas, & Joshi (2015) found that intelligence measured between ages 3 and 7 was associated with both internalizing and externalizing disorders. Interestingly, however, the tendency for intelligence measures to moderate the relationship between a host of environmental threats (socioeconomic disadvantage, adverse life events, and neighborhood poverty) and changes in child externalizing behavioral problems over time seemed more mixed. Nonetheless, less intelligent children, irrespective of environmental threat, experienced more internalizing and externalizing problems over time compared to more intelligent children.
Additionally, another recent study, in this case using a U.S. based group of subjects, examined the relationship between measures of intelligence and behavioral problems in a longitudinal sample of maltreated children (Harpur, Polek, & Harmelen, 2015). The authors found that both higher spatial and verbal intelligence measured at age 6 was associated with lower levels of anxiety and depression at age 14. Thus, it may indeed be the case that intelligence is inversely related to externalizing and internalizing behaviors across childhood and approaching adolescence. At this point, however, more evidence is required before any strong inferences can be drawn.
The Current Study
To further examine the association between indicators of intelligence and problem behaviors, the current study makes use of a unique dataset of highly at-risk American children.
In particular, we analyze a sample that is nationally representative of children who have been investigated by CPS based on allegations of abuse and neglect on the part of their caregivers within the United States. The sample represents children who may be considered the most-imperiled members of the population in that they are predominately minority respondents residing in lower socioeconomic strata. To date, it remains less clear whether variation in measures of intelligence function to predict behavioral problems in this segment of the population. Although this population may have on average more externalizing and internalizing behavioral problems, we hypothesize an inverse relationship between intelligence and both externalizing and internalizing behaviors: as intelligence scores increase, problem behaviors and symptoms should decrease.
METHODS
Sample
For the current analyses, we employed data drawn from the second cohort of the National Survey of Child and Adolescent Well-Being (NSCAW II). NSCAW II, has been discussed in detail in a variety of other studies and as a result, we restrict our description to an abbreviated discussion. The study designers utilized a two-stage stratified sampling design. The first step was to select nine sampling strata consisting of the eight states with the largest child welfare caseloads and the remainder of the US (see Dolan, Smith, Casanueva, & Ringeisen, 2011). The primary sampling units (PSUs), were then selected within each of the nine strata. The same numbers of families were then sampled within each of the 83 selected PSUs. For the NSCAW II sample, cases from CPS investigations that were closed between February 2008 and April 2009 nationwide (n = 5873) were included. The final sample of children was representative of the national population of children birth to 17 years of age in families being investigated for allegations of maltreatment (Dowd et al., 2011).
An important aspect of the sample is that it included both substantiated and unsubstantiated investigations. Moreover, NSCAW II also contained cases that received family preservation services, as well as those who did not receive services in the wake of an investigation. Finally, the sample also included families who had their children removed to foster care, following CPS investigations. Face-to-face interviews with children and current caregiver by trained NSCAW practitioners were completed on average 4 months after the close of the investigation, and again one year (approximately) following the baseline interview (Waves 1 and 2 of NSCAW). Due to age restrictions on our standardized measures, we restricted the data to children over the age of 3 at baseline interview (n=2591).
Measures
Child behavior
The caregiver reported Child Behavioral Checklist (CBCL; Achenbach & Ruffle, 2000) was used to assess behavioral problems in the current study. The CBCL measures internalizing and externalizing behaviors as well as other domains of behavior, including social, thought, and attention problems. For internalizing behaviors, three subscales were combined and averaged: somatic complaints, anxiety and depression, and withdrawn. For externalizing problems, two subscales were combined and averaged: rule breaking and aggressive behaviors. A total score was also derived using all of the above subscales with an additional 33 items subsumed under “other problems” domain. The CBCL gives a standardized score with a mean of 50 and standard deviation of 10, with higher scores indicating greater behavioral problems.
Indicators of Intelligence
We utilize two indicators of intelligence in the current analysis. First, the Kaufman Brief Intelligence Test (KBIT; Kaufman & Kaufman, 1990) test was used, which assesses both verbal intelligence (word knowledge and verbal concept formation) and nonverbal intelligence. Additionally, we utilized a normalized sum of standard scores for vocabulary and matrices scores at points in the statistical modeling. The KBIT was administered to children directly by an NSCAW interviewer. Scores ranged from 40 to 142, with a mean of 100, a standard deviation of 15 (in the general population). Second, we examined scores on the Woodcock Johnson III Tests of Cognitive Abilities (WJ-III). Each scale in the measure—Letter Identification, Passage Comprehension and Applied Problems—has national norms for each age and an average score of 100 and standard deviation of 15 in the general population.
An important note to mention is that children over the age of 11 (i.e., 12 and over) were not administered the Passage Comprehension tests in NSCAW. Thus, when creating the total WJ-III score for the current study, a child’s average score of the available scale based on age was used. In other words, for the creation of the total WJ-III composite measure, we summed and averaged the scores on available subscales. If the child was 12 years of age and older, the other two subscales were included. If a child was 11 years old or under, all three subscales are used. Finally, we created a summated and averaged total intelligence score based on the total K-BIT and the total WJ-III. To clarify further, every participant possessed a measure of total intelligence, however, children over the age of 11 did not have the Passage Comprehension subscale included in their total intelligence score (given that they were not administered this particular subscale by NSCAW staff).
Child demographics
The child’s age, race/ethnicity, and sex were assessed by structured interview with primary caregiver at time of NSCAW interview.
Family Poverty
Poverty was measured by calculating the family’s income-to-needs ratio, which was estimated by dividing family income by its corresponding poverty threshold in 2009. The poverty threshold varies by family size and is based on the money necessary for the minimally accepted amounts of food, with 1.00 representing the overall poverty threshold (Bishaw & Iceland, 2003). Caregivers reported both family income and household size. This measure was divided into four categories: at or below 50% of the poverty line, between 51% and 100% of the poverty line, between 101% of the poverty line and 200%, and above 200% of the poverty line.
Analytic Approach
All analyses were performed using STATA Statistical Software Release 13 (Stata, 2013). Due to NSCAW’s complex sampling design, special STATA survey commands were applied to obtain unbiased estimates of population parameters (NSCAW Research Group, 2002). All percentages were weighted for sample probabilities; therefore, percentages reported in tables represent national estimates. To accommodate certain aspects of the survey design, χ2 statistics were converted to F-statistics with noninteger degrees of freedom using a second-order Rao and Scott correction (StataCorp, 2003). All intelligence and behavioral measures were normally distributed, making ordinary least squares (OLS) regression appropriate. Three separate regressions were used to examine the relationship between intelligence and future behavior. The first regression modeled only behavior at wave 2 by intelligence at wave 1 and the second model added important covariates such as child age, race, sex, and family poverty. In the third model, we added behavior scores at wave 1 as a type of lag variable, thereby providing the ability to test change in behavioral scores across waves as a function of intelligence while still controlling for covariates from baseline (Cohen, Cohen, West, Aiken, 2003). Unstandardized coefficients are presented in tables. Missing data were limited to 40 cases, or 1.5 percent of the sample.
Results
Table 1 reports a description of the sample and also shows that the average behavioral scores were slightly higher than the general population of children, which indicated greater behavioral problems. Likewise, average intelligence was lower than the general population of children, with the sample verbal scores in the K-BIT and applied and passage comprehension scores close to one standard deviation below the population mean. Average child age was 10 years, gender was equally distributed, and less than half of the sample was non-Hispanic white. Over half (54.7%) of children were living in families that were under the federal poverty line. Twenty-four percent of CPS investigations were found credible (i.e. substantiated), and 60% of children remained in-home without CPS services, 24% remained in-home with CPS services, and 16% were removed from care and placed with an out-of-home caregiver (results not reported in tables).
In Table 2, beginning with the K-BIT, an interesting pattern of results emerged (in order to simplify the presentation of results, covariates are excluded from the tables. However, full tables are available in supplementary files). Both the matrices and total scores were related to externalizing problems in Model 1—as scores on these items increased, scores on the externalizing scale declined—however, none of the K-BIT scores (matrices, verbal, or total) predicted externalizing problems in Model 2 after correcting for demographic and socioeconomic factors. Interestingly, Model 3 revealed that all K-BIT measures predicted changes in externalizing behavior scores over time.
Next, we examined the WJ-III. Model 1 revealed that the applied problems, letter identification, and total scale scores were significantly associated with externalizing problems. No such effect emerged for the passage compression subscale. Similar to the results using the KBIT, none of the intelligence items predicted externalizing problems in Model 2 for the WJ-III once we controlled for covariates. However, for Model 3, all of the WJ-III measures (except passage comp) predicted changes in externalizing problems measured at Wave 2. Our final set of analyses for the externalizing measure examined a summed score of intelligence across both the K-BIT and the WJ-III. In this case, the intelligence item was significant in all three models, predicting behavioral problems in the bivariate and multivariate models, as well as predicting change in model 3.
Table 3 presents the analyses using the internalizing behavioral problems scale as the dependent variable. The K-BIT measures of intelligence appeared to be unrelated to the internalizing problems in either Models 1 or 2. In Model 3, however, each of the K-BIT items were significantly related to changes in internalizing problem scores even when controlling for covariates: as intelligence increased, internalizing problems decreased. Moving next to the WJ-III measures, both the letter identification and total scale score items were significantly (and inversely) related to internalizing problems. These effects remained in Model 3, suggesting that the letter identification and total scores were associated with changes in internalizing scores over time and when controlling for confounders. Finally, we examined the global intelligence score and found that intelligence was significantly associated with changes in internalizing behavior over time, but not when controlling for confounders (see Model 3).
Table 4 presents the total CBCL scores as the dependent variable. The K-BIT scores of intelligence were not significantly related to total behavioral problems in Models 1 or 2. However, Model 3 suggested that increased intelligence as measured by the K-BIT was associated with a decrease in total behavioral problems. Moving next to the WJ-III, higher scores in the applied problems, letter identification, and total intelligence scales were significantly associated with decreased behavioral problems in Model 1. Model 3 revealed an identical pattern of findings for these three measures over time. The final row of Table 4 examined the relationship between the composite intelligence measure and the total behavioral problems scale. Similar to the overall relationship between intelligence and externalizing behaviors in Table 2, increases in total intelligence were related to a decrease in total behavioral problems in all Models.
Finally, and owing to possible concerns about spuriousness, some sensitivity analysis regarding a determination of substantiation (of abuse/neglect) and placement following the investigation was conducted by entering these variables into the analyses as additional covariates1. In almost every model, the introduction of these variables did not substantively alter the relationship between intelligence and behavior. However, the relationship between intelligence as measured by the KBIT matrices score and internalizing behavioral problem (top row of Table 3) was rendered insignificant.
Discussion
The association between internalizing and externalizing problems has been well known to, and widely discussed by psychopathology researchers for some time (Caspi et al., 2014; Del Giudice, 2016). Increasingly, scholars have attempted to understand the source of the comorbidity in order to better gauge whether certain risk factors are common across both types of disorders. At the same time, intelligence researchers have reported well-replicated associations between indicators of general intelligence and various deleterious outcomes including overt antisocial behavior, violence, aggression, low impulse control, as well as a host of economic, educational, and other social outcomes (see citations above). In other words, there exists evidence for a link between intelligence and externalizing behaviors but less so for internalizing problems. Our study was intended to examine internalizing behaviors alongside externalizing, as well as test if the relationship between intelligence and difficult behaviors held in a highly at-risk and disadvantaged sample of American participants.
Intelligence, as measured by both the K-BIT and WJ-III, in the present study of at-risk youth seemed to be more consistently associated with externalizing and total behavioral problems rather than internalizing problems. Interestingly, the K-BIT did not seem to predict externalizing problems as robustly as the WJ-III, which it should be noted, generally (though not entirely) assess variation in achievement and crystalized intelligence (see Ritchie, 2015 for details on the distinction). A less consistent association emerged for internalizing problems and the indicators of intelligence, though the WJ-III did evince some evidence of correlation with the internalizing items of the CBCL. Finally, the total CBCL was associated with lower scores on the composite intelligence measure in the sample. Thus, the results did deviate from expectation in some instances. It does seem worth noting, though, that some general picture emerged of an association between intelligence and externalizing problems, along with a much less consistent relationship for internalizing problems, and an overall association for behavioral problems (broadly defined) and intelligence. The smaller effect of intelligence on internalizing compared to externalizing behaviors, it should also be mentioned, is consistent with previous associations found in more advantaged populations (Flouri, Midouhas, & Joshi, 2015).
Findings pertaining to the WJ-III warrant some additional comment, as they may suggest a role for crystalized intelligence in general, and reading performance in particular, in the origins of externalizing problem behavior in at-risk children. This observation is in line with prior research finding an association between verbal ability and delinquency, in part because increased verbal ability might permit greater frustration tolerance, and solving interpersonal conflict via communication as opposed to violent outbursts (Bellair, McNulty, and Piquero, 2016; Moffitt, Lynam, and Silva, 1994). From a more pragmatic perspective, this finding (very tentatively) suggests the possibility that reading intervention for struggling readers in this at-risk population may represent one possible avenue for reducing externalizing behavior (Vacca, 2008). Such a recommendation should be tempered, however, as the results of the K-BIT verbal test (as well as the WJ-III in various models) were somewhat at odds with this, generally only emerging as significant in models examining changes in behavior over time. Any concrete policy recommendations at this juncture would be decidedly premature.
Prior to concluding, there are a number of limitations and considerations that are important to mention. First, and more in the vein of being a consideration, the sample represents an at-risk portion of American children, and the findings will not generalize to the broader population. Nonetheless, because there has been relatively little effort to examine the phenotypic association between indicators of intelligence and behavioral outcomes in this population (compared to general population samples), this represents a desirable quality of the current sample. Second, the assessment of externalizing and internalizing problem behaviors is based solely on caregiver reports and is therefore subject to the usual threats of social desirability bias and over and under reporting. Another consideration involves the tendency for the statistically significant association of intelligence and the outcome variables to dissipate entirely in Model 2 when other key control variables were introduced. While this may be in part due to diminished power and small increases in variation attributable to these covariates, the actual effect size was not substantially attenuated across models. Moreover, and as is often the case in associational research, issues such as collider bias and residual confounding cannot be ruled out.
Finally, the primary independent and dependent variables in the study are moderately to highly heritable constructs (Plomin & Deary, 2015), however, the nature of our data did not permit us to model the effects of heritability directly. This is an important issue, because as prior research has noted (Barnes, Boutwell, Gibson, Beaver & Wright, 2014), even a moderate genetic correlation left unaccounted for may fully confound a phenotypic correlation. Thus, it remains to be seen whether the phenotypic association between intelligence and externalizing/internalizing behaviors observed herein can withstand correction for genetic confounding (for review of genetic influences on intelligence and psychopathology, see Smoller et al., 2018).
In conclusion, we observed a set of findings in an at-risk US sample that seem—at times—in line with the limited prior research on this topic, though there were some divergences too. Higher intelligence was, in various models, associated with lower instances of child externalizing behavioral problems. This same relationship was also observed with internalizing behaviors - including feelings of anxiety and depression - but in substantially fewer of the models. Ultimately, our paper offers additional insight into the association (and perhaps lack thereof, in some cases) between indicators of intelligence and problem behaviors. To the extent that our findings replicate in other samples, using other measures of intelligence and psychopathological outcomes remains an open and important question for future researchers to address.
Footnotes
↵1 We would like to thank our anonymous reviewers for mentioning this point, and suggesting the additional analyses.