Associations of the serotonin transporter promoter polymorphism with aggressivity, attention deficit, and conduct disorder in an adoptee population☆☆☆★

Sample selection 

The overall design and methodology of this adoption study, including the subgroup of individuals from whom these swabs were gathered have been described in detail previously.23, 24 Briefly, this study recruited 197 adoptees and their adoptive parents from four different adoption agencies starting in 1989. Adoptees were separated at birth from their biologic parents and raised by nonrelatives. Adoptees and their adoptive parents were interviewed initially (1990 to 1992) and 181 adoptees were reinterviewed approximately 5 years later (1995 to 1997). At the time of second interview, all subjects were adults between 23 and 53 years of age. At the second interview buccal swabs were solicited. Swabs were obtained from 154 of these reinterviewed subjects, but usable DNA could be extracted from only 98 of these swabs for this study. This number of subjects for 5HTTLPR behavior analyses was further reduced to 87 because of missing information from adoptive parents who declined to be interviewed. All protocols in this study were approved by the Institutional Review Board of the University of Iowa College of Medicine.

Biological parent diagnosis was established as previously described.25 Briefly, biological parent names were compared with the records of state mental health and correctional facilities. Apparent matches were corroborated by comparison of available family and demographic records. Coded hospital and correctional records were reviewed independently by three psychiatrists and diagnoses were assigned using DSM-III criteria.26 Disagreements were settled by conference and those individuals about whom a consensus could not be reached were excluded. The κ for interrater reliability ranged from 0.67 to 0.79.

Development of outcome measures 

At the time of the first interview, adoptive parents (usually the mother) answered questions about specific symptoms of preschool (age 2 to 6 years) and grade school (age 6 to 12 years) attention deficit disorder and conduct disorder (including aggressive behavior). Nearly all of these questions were rated on a four-point Likert scale ranging from 0 (absent) to 3 (strongly present). Occasionally items were rated on other categorical scales (e.g., absent or present). These were recoded so that their scaling approximated the above four-point scale (e.g., absent = 0, present = 2). Information on preschool and grade school externalizing behaviors were available for 175 of the adoptees (six adoptive parents were not interviewed). For the factor analyses described next, we used all 175 reinterviewed subjects rather than just those who with a known 5HTTLPR genotype.

Development of behavior scales 

We derived the ADHD subscales using categorical factor analysis based on a latent probit model.27, 28 We applied unweighted least squares fitting to the estimated tetrachoric correlation matrix using M-plus software.28 We obtained nonorthogonal (i.e., correlated) solutions via Promax rotation.29 Our sample size was rather low compared to what is desirable to assure reasonable estimate stability for this categorical data approach. We therefore compared the results for consistency with those obtained from traditional principle component and principal factor analyses of the same data. (These continuous data methods have their own limitations when analyzing categorical, highly skewed distributions. We view comparison of results obtained from each of these methods as the most satisfactory approach to categorical data factor analysis with fewer than several hundred observations.) Our final results were qualitatively consistent across all three approaches in the sense that they would have led to nearly identical factor scores, which are described below.

We chose the number of factors primarily on the basis of the number of eigenvalues greater than 1 in the decomposition of the implied tetrachoric correlation matrix, although we also paid heed to the slope of eigenvalue decrease and interpretability of the solutions. We encountered no serious contradictions when comparing these criteria. A small number of items were deleted from the final models due to low correlations with any postulated factor or high correlations with multiple factors indicating nonspecificity.

Finally, we defined the “working” behavior scores in a simplified two-step manner. In the first step, we simply added the original Likert scores (0 to 3 or the equivalent) of the items assigned to each factor. (We assigned each item to the factor that it loaded on most highly.) Such simplified scores typically correlate quite highly with analytically estimated factor scores29 and, in addition to their simplicity, have the advantage of robustness when the analysis is based on relatively small sample sizes.29, 30 In an additional step, we collapsed certain contiguous values of each scale—especially at the upper extreme—so that the observed probability distributions contained at least several observations at each possible value. This was to improve the stability of our estimates of 5HTTLPR association via ordinal logistic regression (see below). The resultant behavior scores were reduced to ordinality rather than retaining a true approximate scaling; however, the possible undue influence of outliers with extreme behavioral scores was minimized and an alternative implicit scaling of the underlying behavior was recovered as the model was fit. The items comprising each scale, along with the descriptive scale names, are listed in Table 1. Note that these scales follow commonly accepted DSM concepts quite closely, including subtypes of behaviors that are identified under current ADHD diagnostic criteria.31 We have also listed in Table 1 the scales developed prior to this study that used items from both the adoptee interview as well as the adoptive parent interview. Adoptee adolescent aggressivity was defined by a scale described by Loney et al32 and modified by an additional five items (described in Table 1). Conduct disorder in adolescence was defined by the sum of behavior items which were adopted from Robins.34

Table 1. Defining Items for Preadolescent Behavior Factors and Adolescent and Grade School Aggressivity and Conduct Disorder

	Pre-Adolescent Behavior Factors: Defining Items
	Preschool Factors (ages 2-6)
	Conduct Problems	Impulsiveness	Attention Deficit Problems	Hyperactivity Problems
	Physically attacks other children	Doesn't stick to play	Fails to finish tasks	Trouble waiting turns
	Defies adults	Refuses to wait turn	Appears not to listen	Runs and climbs excessively
	Difficult to discipline	Grabs things from others	Easily distractable	Trouble sitting quietly
		Interrupts others	Trouble concentrating	On the go-like motordriven
		Loud and overexcitable

	Grade School Factors (ages 6-12)
	Conduct Problems	Attention Deficit Problems	Hyperactivity Problems
	Talk back	Fails to finish tasks	Impulsive
	Profanity	Doesn't listen	Shifts activity excessively
	Lies frequently	Easily distracted	Disrupts class
	Trouble with teacher	Trouble maintaining attention	Trouble sitting quietly
	Quarrelsome	Trouble organizing work
	Temper tantrums	Disorganized with possessions
	Loud/abusive	Trouble remaining on task
		Changes activity without finishing
		Trouble studying for tests
		Trouble organizing responsibilities

	Previously Defined and Published Adolescent and Grade School Aggressivity and Conduct Disorder: Defining Items
	Preadolescent Aggressivity	Adolescent Aggressivity	Adolescent Conduct Disorder
	Distructive (vandalism)	Distructive (vandalism)	Truant
	Temper tantrums	Temper tantrums	Expelled from school
	Bully/cruelty	Bully/cruelty	Run away from home
	Defiant	Defiant	Lies
	Fights	Fights	Early sex
	Lies	Lies	Early substance use/abuse
	Quarrelsome	Quarrelsome	Steals
	Rebellious	Rebellious	Vandalism
	Steals	Steals	Poor school grades
	Swears	Swears	School problems fights
	Teases	Teases
	Won't mind	Won't mind
	Insolent/sassy	Insolent/sassy
	Sets fires	Sets fires
	Physically attacks adults	Physically attacks adults
	Verbally abuses adults	Verbally abuses adults
	Threatens others	Threatens others

Our final behavior scores had mostly moderate intercorrelations, both among behaviors related to the same childhood period and between behaviors related to different periods. These correlations are listed in Table 2.

Table 2. Pearson Correlations Between Measures of Externalizing Behaviors

		PS Conduct	PS Impulse	PS AD	PS Hyper	GS Conduct	GS AD	GS Hyper-Impulse	Preadol Aggress	Adol Aggress	Adol Conduct
	Preschool conduct	1.00	.58	.47	.42	.48	.38	.47	.53	.49	.37
	Preschool impulsivity	.58	1.00	.64	.52	.54	.52	.69	.59	.41	.41
	Preschool attention-deficit	.47	.64	1.00	.54	.50	.66	.69	.53	.44	.42
	Preschool hyperactivity	.42	.52	.54	1.00	.47	.42	.64	.52	.40	.32
	Grade school conduct	.48	.54	.50	.47	1.00	.51	.58	.89	.61	.53
	Grade school attention-deficit	.38	.52	.66	.42	.51	1.00	.59	.48	.38	.48
	Grade school hyperactivity-impulsivity	.47	.69	.69	.64	.58	.59	1.00	.60	.53	.50
	Preadolescent aggressivity	.53	.59	.53	.52	.89	.48	.60	1.00	.60	.56
	Adolescent aggressivity	.49	.41	.44	.40	.61	.38	.53	.60	1.00	.60
	Adolescent conduct disorder	.37	.41	.42	.32	.53	.48	.50	.56	.60	1.00

NOTE. Correlations are based on 175 subjects, not all of whom had serotonin genotypes available for further analysis. Abbreviations: GS, grade school; Preadol, preadolescent; Adol, adolescent, AD, attention-deficit; hyper, hyperactivity; aggress, aggressivity.

Analysis of 5HTTLPR associations 

We conducted regression analyses of the relationship between 5HTTLPR and various measures of developmental behavior problems. Other previously known or suspected predictors (see below) were included in the analyses in order to control for potential confounding. Due to the small number and skewness of ordered categories for most of the outcome behaviors, we elected to use proportional odds ordinal logistic regression35, 36 rather than conventional linear regression. While logistic regression coefficients are typically interpreted as the natural logarithm of a constant odds ratio,36 we have elected to report results from the perspective of analyzing underlying, unobserved, latent continuous variables. Under this interpretation, the logistic regression coefficients are interpretable as estimates of the corresponding regression coefficients that would be obtained if linear regression could be performed using the underlying continuous variables as the outcomes.35 Note that logarithms are not involved in this interpretation. With appropriate rescaling, these regression coefficients can be interpreted as the standardized effect size associated with one unit change in the corresponding predictor variable. The standardized effect size is given in terms of standard deviations of the model's residual error.

As reviewed in the introduction, we have previously observed in these data that antisocial biological parents (ASP bio), adverse adoptive home environment, and gender are significant predictors of conduct or aggressive behaviors.33, 37, 38 In order to avoid potential confounding and work with a consistent model, we automatically included all of these terms as additional predictors in the present models. We also included alcoholic biological (ALC bio) background because of evidence in earlier studies for significant association with substance abuse, conduct disorder, and ADHD.37, 38, 39 For some outcomes, an interaction between ASP bio and adverse adoptive home environment has also been observed.14 In the present models, this interaction was included if the estimated P value for its association with the behavioral outcome was less than .15.

Having established the background model containing the variables just noted, we proceeded to add 5HTTLPR genotype as an additional predictor. Due to the paucity of short homozygotes (see Table 3 for allele frequencies), we could not analyze this genotype separately. As an alternative, we collapsed the genotypes of short-short (S-S) and short-long (S-L) into a single category and compared this combined genotype to the long-long (L-L) combination. This amounts to an assumption that the short allele operates under classical Mendelian dominance, an interpretation with some support in the literature.4, 5, 6

Table 3. 5-HTTLPR Genotype by Sex of Adoptee

		Genotype
	Sex	SS	SL	LL	Total
	Female	10	22	27	59
	Male	5	13	21	39
	Total	15	35	48	98

Abbreviations: SS, homozygous for short allele; SL, heterozygous; LL, homozygous for long allele.

Our first set of regression models, which are not presented in any detail, failed to show significant simple 5HTTLPR associations (main effects) with any of the 10 behavioral outcomes. Observing this, we moved on to models that also investigated potential interactions between 5HTTLPR and previously established risk factors. We found no evidence for interactions between 5HTTLPR and the adverse adoptive home environment scale. However, as detailed in the Results section, we did find evidence of multiple interactions between 5HTTLPR and both biological parent status (ASP bio and ALC bio) and adoptee gender. To standardize the structure of our predictive models and control for potential confounding among these terms, our final models included all three 5HTTLPR interactions—with adoptee gender, and with ALC or ASP bio parents—regardless of their individual significance in predicting a particular outcome.

The problem of multiple comparisons 

Given the scope of our final design—a study with 10 related but distinct behavioral outcomes and three 5HTTLPR interactions to check for each outcome—we felt it essential to address the issue of multiple comparisons. Within a particular model, we use the 4-df Wald tests for the joint significance of 5HTTLPR and its interactions as our primary tests for 5HTTLPR association. However, in deciding which exploratory results to report, we adopted a rather liberal standard; P values as high as .17 for this overall test will be noted below. One-degree-of-freedom tests for specific interactions were then reported if their P values were below .10.

The issue of multiple comparisons across multiple models was addressed by using a simulation technique that compared the number of our results that met the above 4-df criteria to the approximate number that would be expected by chance. We first obtained the Pearson correlation matrix of the residuals that resulted from fitting each behavioral outcome to a logistic regression model containing only our background predictors (i.e., the non-5HTTLPR terms). We then generated random multivariate normal data sets with the same number of observations as our actual data and the same correlation structure observed in the above residuals. We did this repeatedly (10,000 times), and on each replication, fit conventional linear regression models to the 10 simulated normally distributed outcomes using the 5HTTLPR term and its three interactions as the predictor variables. This method allowed us to simulate the approximate expected joint distribution of test statistics under the null hypothesis of no 5HTTLPR associations with any of our outcomes. We estimated the probability under the null hypothesis of observing n positive 4-df tests vis-à-vis an arbitrary P value criterion by tabulating the proportion of simulated chi-square statistics that exceeded the appropriate critical value. (The method is approximate because of the switch—for the sake of practical simulation—to normal data and linear regression. We used chi-square approximations of the usual linear regression Fstatistic to more closely duplicate the actual statistics used in logistic regression.)

Molecular genetic protocol 

Buccal swabs were obtained using Cytotech brushes (Medical Packaging Corp, Camarillo, TX) according to the manufacturer's suggestions. These swabs were assigned a study code, stripped of all other subject identifiers, then stored in a refrigerator at 4°F until used. DNA from these swabs was prepared using a QIAmp DNA minikit (Qiagen, Inc, Valencia, CA).

Genotyping of the 5HTTLPR locus was carried out using the primers F- GGCGTTGCCGCTCTGAATGC and R-GAGGGACTGAGCTGGACAACCAC as given by Persico et al.48 Vent polymerase was used according to the manufacturer's suggestion (New England Biolabs, Beverly, MA) and 100 μmol/L 7-deaza guanosine triphosphate (GTP) (Boehringer Mannheim, Indianapolis, IN) was added to aid amplification through this GC-rich region. Cycling parameters were as follows: 98°C × 2 minutes, then 45 cycles of 96°C × 15 seconds, 68°C × 15 seconds, and 72°C × 45 seconds, with a 7-minute final extension at 72°C. Approximately 3 μL of each of the above polymerase chain reaction (PCR) products were denatured, then loaded on a standard 6% polyacrylamide sequencing gel and electrophoresed for 2 to 3 hours. The gels were exposed to standard x-ray film and the visualized PCR products sized by comparison to an internal sequencing ladder. Gels were read independently and blindly with respect to subject behavioral outcome or genetic background by two of the authors (R.P. and H.S.).

General comments 

This study of adopted persons has found evidence that 5HTTLPR variants may play a significant role in the manifestation of externalizing behaviors (such as attention deficit hyperactivity [ADH], aggressivity, and conduct disorder) when interacted with additional factors: (1) adoptee gender, (2) alcoholism in the biologic parent, and (3) antisocial personality in the biologic parent. The interactions shown by 5HTTLPR with ALC and ASP bio parental diathesis do not allow easy generalization, except for the classic interpretation of interaction: the difference in outcome between SS or SL and LL depends on the presence or absence of the genetic diathesis for alcoholism or antisocial personality. The reader can see this variation in differences presented in Fig 1 by imagining, for example, a line drawn between the mean effect size for alcohol LL and alcohol S and contrasting this upwardly sloping line with the downwardly sloping line between the two mean effect sizes for no alcohol LL and no alcohol S. Similar results can be obtained with the antisocial interaction.

The analyses of behaviors in Table 4 further suggest that these three types of interactions are separate and that the direction and magnitude of each effect appears to be similar across varied externalizing behaviors. For example, in females the S allele produces the lowest effect (see Table 5). In subjects with biologic parent alcoholism the short allele is consistently associated with higher aggressivity and other externalizing behaviors (see Table 7), whereas with ASP bio parent, the short allele correlates with low aggressivity (see Table 6 as well as interaction figures on top row of Fig 1). Failure to find a main 5HTTLPR effect in our preliminary analyses can be explained by the presence of these interactions and the distribution of the interacting variables in our data.

Multiple comparisons 

The problem of multiple comparisons in our analyses requires some detailed comment. Since the outcomes of interest have substantial overlap, a conventional multiple comparison correction such as Bonferroni's is unnecessarily conservative. (Consider the situation in which an experimenter unwittingly retests the same outcome multiple times, calling it by a different name each time so that multiple tests appear to have been performed.) As described in the Methods, we performed more accurate corrections based on computer simulation of the observed correlation of residuals.

Typically, multiple comparison corrections are constructed to answer the following question: If there are no real relationships in the underlying population, what is the chance that I would have found anyof the nominally significant observed results to have P values at least as low those actually observed? While this is obviously an important question, it is not the only one that can be asked. A second question of similar interest is this: If there are truly no findings of interest in the underlying population, what is the chance that I would have found allof the nominally significant results to have P values at least as low those observed? We have answered both of these questions via simulation.

We first note that, among the models with “interesting” interaction results, the highest P value for the overall (4 df) serotonin test was .17 (Table 4). Under the null hypothesis, our simulations suggested that at least one P value ≤ .17 would have occurred 81% of the time. However, we have eight of 10 overall serotonin tests with P values that are at least this low. The same simulations suggest that eight or more tests with P values ≤ .17 would occur only 0.5% of the time under the null hypothesis of no real interactions (i.e., P ~ .005). Since the P value is well below the usual level (5%) of significance we can reject the null hypothesis and assert that it is unlikely that none of the interaction results are real. Which interactions are spurious can only be determined by repeating the study on an independent sample.

Statistical estimation of effect sizes 

The use of ordinal logistic regression rather than conventional regression was motivated primarily by problems that arise when studying behaviors associated skewed distributions of relatively crude measurements. Other than this wrinkle, our analyses were based on well-established epidemiological methods for controlling confounding variability in an observational experiment.36

The magnitude of effect sizes shown in Table 6 through 8 requires comment. The mean effect sizes are in the general range of 0.5 to 1.0 and should be considered of substantive significance. These are mean effects. It may be that an individual has to be 1.5 to 2 SD beyond the norm to be clinically abnormal, so these effect sizes do not imply that the average person with the relevant combination of risk factors will be clinically deviant. However, these effect sizes suggest that their risk for clinical deviance is substantially elevated.

Important confounds 

As mentioned in the introduction, the rationale for our focus upon 5HTTLPR variants was the extensive literature from neuropsychiatry implicating serotonin in a variety of externalizing behaviors. However, genetic association studies of 5HTTLPR have been equivocal with some showing correlations with externalizing behaviors such as suicide attempts, neuroticism, severe habitual aggressive behavior, and alcohol dependence.41, 42, 43, 44, 45, 46 The “positive” studies point to the involvement of the short variant of the 5HTTLPR gene in the correlation with deviant behavior. In contrast, other studies of independent samples have failed to show a correlation of 5HTTLPR variants with similar types of behavior.12, 47, 48, 49 In some cases, “positive” studies may be due to population stratification: the presence of genetically different populations in the control or comparison groups. This possibility was specifically discussed in one study contrasting different ethnic groups. Gelernter et al. found race × promotor system genotype interaction that they interpreted as evidence of population stratification.12 Our results are not likely due to population stratification for the following reasons: (1) comparison adoptees in the analysis came from the same adoption agencies as the adoptees with ASP or ALC bio parents; (2) there were no 5HTTLPR allele frequency differences among the four adoption agencies providing subjects; and (3) in adoptees there were no 5HTTLPR allele correlations with ASP bio or ALC bio background. The last point is especially important since it demonstrated that breaking down the sample on the basis of biologic parent psychopathology did not result in two different subsamples with different allele frequencies. Thus the data answer the question of whether population stratification is the explanation for findings in this study—it is not.

Gender × 5HTTLPR interaction: Supporting evidence 

The interaction found between gender and 5HTTLPR deserves further comment. Gender differences in externalizing behaviors are nothing new: boys are generally more aggressive and hyperactive for example. However, in this study the interaction with gender is significant in spite of the correction in the regression model for gender as well as for the two other interactions. The mechanism of interaction is not apparent and could range from molecular genetic reasons to different environmental expectation for behaviors in the two genders. There are some examples in the literature of gender × genotype interaction in humans: one involving 5HTTLPR and the second, the DRD4 receptor. One study by Gelernter at al.12 tried to replicate previously reported association between 5HTTLPR and the harm avoidance scale of the tri-dimensional personality questionnaire. They failed to find a main effect of 5HTTLPR but did report a significant gender × 5HTTLPR interaction instead (P = .04). Males with a short allele had higher harm avoidance score than those with LL; just the opposite was found with females where LL was associated with higher mean scored for harm avoidance and S with lower scores. The second study by Vandenbergh et al.50 reported that long forms of the dopamine receptor (DRD4) gene manifesting VNTR were significantly more prevalent in male substance abusers than in females. The interactions described in the Gelernter and Vandenburgh studies also suggest that gender × genotype interactions as found in this study might be successfully detected in general population samples.

Defining the role of 5HTTLPR in externalizing behaviors 

Among recent association studies there is a trend to look at more narrowly defined behaviors. The intent is to achieve a more homogeneous subgroup and possibly better show a genetic correlate (because of more homogeneous etiology), e.g., one study that correlated 5HTTLPR variants with alcoholics who also had a history of bothimpulsive and violent behaviors found that the short allele frequency was higher among such alcoholics compared to alcoholics without the violence.51 Another example of narrowing the behavior to achieve a more homogeneous group of subjects is a study that showed the short 5HTTLPR allele conferred susceptibility to subjects whose behavior profile was characterized by both high novelty seeking and low harm avoidance.52 Todd et al. have used latent-class analysis in a twin study of ADHD to find ADHD subtypes that appear to be independently transmitted in families, and they suggest that their classes “may be more appropriate targets for molecular genetic studies”.53 Another group of investigators have further narrowed behavior using an ethological approach: by adding a developmental aspect to “externalizing” behavior such as excessive aggressivity in lower primates (monkeys).54 Variations in the 5HTTLPR gene have been found to affect development depending on early social-rearing history. Male monkeys with a short allele (LS) were more likely to have lower cerebrospinal fluid 5-hydroxyindoleacetic acid (5-HIAA) concentration than males with the LL allele when both were reared by peers as infants. However, LS and LL monkeys raised with their mothers did not show this difference in cerebrospinal fluid 5-HIAA. This involvement of 5HTTLPR with behavior depends upon rearing circumstances and is evidence of specific genetic-environmental interaction.55

A developmental perspective of behavior makes biologic as well as social-behavioral sense in explaining and understanding externalizing behaviors that show considerable variability between individuals,56, 57 as well as over time (e.g., childhood ADHD predicts adult antisocial personality in individuals who are aggressive as children19). Thus more recent behavioral genetic association studies are attempting to look at more specific (and narrowed) behaviors at different developmental periods. In the ethologic approach (see above) the developmental aspect is further studied by manipulating earlier environmental experience and noting its effect on future behavior. The current study is a further example of delineating and focusing on specific behaviors at different developmental periods. The results of focusing on more narrowly defined behaviors suggest that several important kinds of externalizing childhood and adolescent behaviors may be influenced by the 5HTTLPR allele variation through interaction with other genetic influences.

The information in the adoption paradigm has provided a rational etiologic perspective for defining homogeneous subgroups of subjects with regard to other background sources of genetic variability, e.g., those with a genetic diathesis for a specific behavioral disturbance known to have a substantive genetic etiology (such as personality deviance and substance abuse). The regression analysis also expedites analysis of genetic effects by including environmental factors (such as disturbed adoptive parents) that also influence behavior. In connection with this novel application of the adoption design to genetic association studies more discussion about epistasis is indicated because of our finding of evidence for gene-gene interactions.

Epistasis is a concept from classical Mendelian genetics. It was a term originally used to describe specific gene-gene interactions occurring between non-allelic (and nonlinked genes) which affected a phenotypic trait.58 Here, we use a closely related phenomenon, gene-biological family history interaction to describe the 5HTTLPR interaction with ALC biol parent and ASP bio parent diatheses. We imply that epistasis may be occurring because the offspring of parents with alcoholism or antisocial personality are very likely to inherit genes that predispose themselves to these behaviors and which in turn lead to a variety of deviant behaviors during early development such as ADHD. The term gene-family history interaction used here refers to the interaction of the 5HTTLPR with a group of genes of unknown number since we are dealing with psychiatric conditions with complex polygenic etiologies that are presumably inherited from the biological parents.

Before molecular analysis of genes was available, epistasis was usually inferred by examining the deviations of phenotypes from carefully selected animal or plant crosses.59 Gene-biological family history interactions in human behavior were very difficult to study for the following reasons: (1) controlled crosses were and are not practical (or can be immoral for that matter); (2) complex behavioral traits are difficult to characterize unlike “simple” physical traits such as hair or eye color; and (3) analysis of gene-biological family history interactions through examining human family trees requires large numbers of cases. However, the adoption paradigm allows a more direct analysis for these interactions. The regression methodology developed here can be used to identify candidate genes (or other genomic polymorphisms, e.g., 5HTTLPR). Additional genes found to be important in behavior can in turn be tested directly for epistasis by constructing regression models containing both the new gene along with 5HTTLPR or with other genes identified through regression models as well as interaction terms between the candidate genes. Whether this regression approach to identifying specific genes will prove fruitful can only be tested by further repetition with independent samples and with other candidate genes within the present sample.

Although the present report deals with an adopted group of subjects, a similar regression approach seeking gene-biological family history interaction might be fruitful in association designs with probands from twin and family studies where behavioral genetic diatheses have been defined by psychiatric conditions in blood relatives. If gene-environment interaction plays a relatively minor role in outcome, such studies could be effective in detecting gene-biological family history interaction.

Biologic significance of epistasis 

Many human illnesses, such as diabetes or depression, are recognized now as “complex” genetic illnesses, thought to be caused by multiple genes (often of small individual effect) acting together (as well as interacting with the environment). In studying such illnesses, the assumption often made when model-building is that the effect of multiple genes can be assumed to be additive. Interaction is often ignored in analysis.60 However, our results raise the possibility that non-linearity and nonadditivity may occur. This study has looked at only one gene in the serotonin system and has found evidence of interaction between its variant and two other groups of behaviorally important diatheses (antisocial personality and alcoholism). This suggests that epistasis might be much more common. It would then appear to be imperative to survey other serotonin receptors as well as other neurotransmitter systems (e.g., dopamine and glutamate and acetylcholine receptors and transmitters that are involved in important behaviors such as reward systems, conditioning, memory, etc.). If interactions as described here are frequent then we can expect much more variability in behavior from fewer genes than have been postulated in strictly linear polygenic models of such outcomes as height or psychiatric conditions.60