Childhood Adversity, Monoamine Oxidase A Genotype, and Risk for ConductDisorder FREE

There is a long history of research on risk factors for conduct disorder,^1- 3 in part because of thedeleterious effects of aggressive, antisocial, and criminal acts on individualsand society at large. Evidence for aggregate genetic effects has been reportedby many twin, family, and adoption studies, based on the pattern of resemblancebetween different classes of relatives, but reliable evidence for specificgenetic effects is limited. Recently, an interaction between a functionalpolymorphism in the promoter of the monoamine oxidase A (MAO-A) gene and childhoodmaltreatment was reported to be associated with a significantly increasedrisk for antisocial behavior in boys from the Dunedin Multidisciplinary Healthand Development Study.⁴ There was an independenteffect of maltreatment but no independent effect of low MAO-A activity onrisk for antisocial behavior. Low MAO-A activity was therefore a detectablerisk factor only in the presence of an adverse childhood environment, andthe effect was not trivial. Individuals with both the low-activity MAO-A genotypeand maltreatment accounted for 12% of the birth cohort but 44% of the cohort'sviolent convictions. A genotype-environment interaction may therefore accountfor a significant portion of individual variation in risk for antisocial behavior.This finding is a rare example of a measured gene and a measured environmentjointly affecting human behavior. Such effects have been documented in experimentalorganisms⁵ and described in theoretical treatises,⁶ but they have rarely been explored and replicatedin studies of human behavior.

In this study, we tested for the presence of an interaction betweenlow MAO-A activity and exposure to childhood adversity that increases riskfor conduct disorder in a community-based sample of boys from the VirginiaTwin Study for Adolescent Behavioral Development (VTSABD).⁷ Wedid not survey maltreatment as defined by Caspi et al⁴ (maternalrejection, repeated loss of a primary caregiver, harsh discipline, physicalabuse, sexual abuse), but instead we defined adversity by exposure to 3 knownrisk factors for conduct disorder: interparental violence, parental neglect,and inconsistent discipline.^3,8- 9 Wehypothesized that the findings of Caspi and colleagues indicate a robust effecton risk for conduct disorder of low MAO-A activity contingent on exposureto a variety of childhood adversities.

Subjects comprised 514 white male twins from the community-based, longitudinalVirginia Twin Study for Adolescent Behavioral Development. The recruitmentand assessment of the 1412 twin families from which this subsample deriveshave been described in detail elsewhere.^10- 11 Maletwins were included in the present study if data were available on their historyof conduct disorder, exposure to childhood adversity, and MAO-A genotype.The VTSABD assessed subjects on up to 4 occasions, depending on their age(<18 years) and their willingness to participate in continued assessments.Recent history (past 3 months) of conduct disorder was surveyed at times 1,2, 3, and 4, and a history of exposure to environmental adversities and DNAwere collected at times 3 and 4. Eligible male subjects were aged 8 to 17years (mean ± SD, 12.23 ± 2.81 years) at entry into the study.

Twins and their parents were personally interviewed at home by trainedfield workers. Twins were randomly assigned to different interviewers, andeach co-twin was interviewed simultaneously in separate areas of the familyhome. Interviewers held a master's degree in social work or an equivalentprofessional degree, or they had extensive experience in psychiatric interviewing.Interviewer training consisted of 3 weeks of residential instruction, andall field interviews were audiotaped and reviewed by senior monitors. Regularmeetings were conducted to avoid drift in use of the interview or other aspectsof the protocol. A residential meeting was held annually to ensure that theassessment protocol was being implemented in a standardized manner.¹⁰ Informed consent was obtained in writing from parents,and assent was obtained from the child prior to the personal interview.

The past 3 months' history of conduct disorder was assessed followingthe DSM-III-R criteria using the Child and AdolescentPsychiatric Assessment (CAPA)–Child and Parent Version.¹² Conductdisorder was diagnosed using a symptom or-rule; that is, a symptom was ratedas present if endorsed by any informant. We used child, maternal, and paternalCAPA data.

The VTSABD did not survey the variables that were used to constructthe maltreatment index used by Caspi et al.⁴ Anadversity index was therefore constructed with 3 known correlates of conductdisorder: parental neglect, exposure to interparental violence, and inconsistentparental discipline. All 3 adversities were assessed at wave 3 of the VTSABDwith the same assessment method (personal interview) and time frame of survey(ever).

Three items rated by parents were used to define parental neglect: (1)Did anyone ever say you weren't looking after the children properly? (2) Hasanyone ever thought that 1 of the children became ill because the child wasn'tlooked after properly or because the home wasn't clean enough? (3) Was thereever a time when 1 of the children was very ill, but, at the time, you didn'tthink the child needed to see a doctor, or was there ever a time when a doctorsaid you should have brought the child in earlier?

Two items rated by child subjects were used to define exposure to interparentalviolence: (4) Have your parents ever pushed or shoved each other during anargument? (5) When your parents fight, do they (or have they) ever hit eachother? Two items rated by child subjects were used to define inconsistentparental discipline: Some parents can be strict one day, and the next dayit seems as though they don't care whether you broke a rule or not. Is itlike that with either of your parents? This was rated for (6) mother and (7)father.

A maternal history of 7 antisocial personality symptoms were surveyedby personal interview with mothers. These symptoms were (1) inability to sustainconsistent work behavior; (2) failure to conform to laws or social norms;(3) irritable, aggressive, or involved in fighting or assaults; (4) failureto honor financial obligations; (5) impulsivity, moving from place to place;(6) recklessness regarding own or others' safety; and (7) no relationshipthat lasted at least 1 year and was monogamous.

Cytology brushes were used to obtain a sample of buccal cells from thetwins for DNA analysis. DNA was isolated using the InstaGene Matrix kit (Bio-RadLaboratories, Hercules, Calif) protocol for cell lysis product absorption.Each sample was diluted to a working concentration of 5 to 20 ng/µL.We used primer sequences described previously,¹³ MAOAPT1 (5′-ACAGCCTGACCGTGGAGAAG-3′), 5′-labeled with the FAM-6fluorophore and MAO APB1 (5′-GAACGGACGCTCCATTCGGA-3′). We amplifiedpolymerase chain reaction products in 96-well microtitre plates in 20-mL volumecontaining 50 to 200 ng of genomic DNA, 1X GeneAmp PCR Gold buffer (AppliedBiosystems, Foster City, Calif), 1.5 mM magnesium chloride, 10 pmols eachforward and reverse primer, 0.3 mM each 2′-deoxynucleoside 5′-triphosphate,and 1.5 U HotMaster Taq DNA Polymerase (Eppendorf, Hamburg, Germany). Cyclingreactions were performed on a PTC-225 DNA engine (MJ Research Inc, Waltham,Mass) with 3 minutes initial denaturation at 95°C, followed by 35 cyclesof 95°C for 3 minutes, 62°C for 1 minute, 72°C for 1.5 minutes,with a final extension at 72°C for 8 minutes. We analyzed products usingan SCE-9610 capillary sequencer (SpectruMedix, State College, Pa) with ROX-labeledGS-500 (Applied Biosystems) as size standard, and we determined allele sizesusing Genospectrum v2.6 DNA fragment analysis software (SpectruMedix).

Male twins who were included in the current study (N = 514) were comparedwith male twins who were not included because they had become too old forthe study or were otherwise lost to follow-up at time 3/4 (N = 823). We madethis comparison to assess the representativeness of the subsample of the currentstudy. Variation in the prevalence of conduct disorder at time 1 by participationstatus was evaluated using a 2-tailed χ² test. Variation inthe level of maternal antisocial personality symptoms, age at time 1, andordinally scaled census tract variables were evaluated using a 2-tailed Wilcoxonrank sum test. Logistic regression was used to estimate risk for the clinicaldiagnosis of conduct disorder in association with MAO-A activity, exposureto childhood adversity, and the interaction between MAO-A activity and childhoodadversity. Modeling conduct disorder symptom counts within a linear regressionframework may be more powerful statistically, but the results obtained withlogistic regression are likely to be more robust. An interaction identifiedby linear regression may reflect heteroscedastisity, that is, an unequal mean-variancerelationship over the range of scale scores and an artifact of the scale ofmeasurement. We therefore sought to avoid misinterpretation of scalar effectsby modeling the clinical diagnosis of conduct disorder. Logistic regressionwas performed in PROC GENMOD (SAS version 6.12; SAS Institute Inc, Cary, NC)to adjust for the correlation between co-twins.¹⁴ One-tailedprobability levels are reported for the interaction term estimated in theregression model because this is a replication study and the direction ofeffects can be specified. A 2-tailed Fisher exact test was used to comparethe prevalence of conduct disorder among subjects with low vs high MAO activityin association with no, probable, or definite exposure to childhood adversity.

Male twins who participated in the current study were, as expected,younger at entry into the study (mean ± SD age, 10.4 ± 1.6 years)than nonparticipants (mean ± SD age, 13.04 ± 2.51 years, P<.001), and participants had a lower prevalence ofconduct disorder at time 1 (3.89%) than nonparticipants (8.45%, P< .001). Mothers of participants had fewer antisocial personalitysymptoms (mean ± SD, 0.67 ± 0.87) at time 1 than mothers ofnonparticipants (mean ± SD, 0.95 ± 1, P<.001).Participants did not differ significantly from nonparticipants on the followingcensus-based indicators of regional socioeconomic status: median family income(P = .31), rural vs urban residence (P = .69), or proportion of college-educated adults (P = .55). Participants were more likely, however, to reside in areaswith lower levels of male unemployment (3.03%) than nonparticipants (3.75%, P = .02). Participants and their mothers were thereforeless symptomatic at entry into the study than twins who became too old forthe study or were otherwise lost to follow-up. Our subsample is thereforenot representative of the total sample at time 1, but any bias would likelyserve to attenuate rather than overstate the magnitude of effects observedin this study.

Among participating male twins, the prevalence of conduct disorder acrosstime 1-4 was 11.48%. The prevalence of any low-activity MAO-A allele was 29.38%.Broken down by the number of repeats at the MAO-A promoter polymorphism (andby activity type), the frequency of each allele was 2 repeat, 0.39% (low);3 repeat, 28.79% (low); 3.5 repeat, 2.33% (high); 4 repeat, 68.29% (high);and 5 repeat, 0.19% (low). The prevalence of exposure to any interparentalviolence was 2.53%; any parental neglect, 12.65%; and inconsistent maternalor paternal discipline, 17.32%.

As expected, all 3 adversities were significantly associated with conductdisorder (univariate odds ratio [OR]: interparental violence, OR, 3.6, P = .04; parental neglect, OR, 2.46, P = .008; inconsistent parental discipline, OR, 2.15, P = .01), even after controlling for their intercorrelations withina multiple regression framework (multivariate OR: interparental violence,OR, 3.38, P = .05; parental neglect, OR, 2.55, P = .006; inconsistent discipline, OR, 2.05, P = .03). Parental neglect, interparental violence, and inconsistentdiscipline were therefore largely independent correlates of risk for conductdisorder.

Childhood adversity was coded as a categorical variable to replicatethe study by Caspi et al.⁴ In that study, thepresence of any 1 adversity was coded as probable exposure, and the presenceof any 2 adversities was coded as definite exposure. The prevalence of probableadversity (25.49%, 131/514) or definite adversity (2 adversities, 3.5%, 18/514)in the VTSABD was somewhat lower than the prevalence of probable maltreatment(28%) or definite maltreatment (8%) in the Dunedin sample. This likely reflectsthe much larger number of items (n = 21) used to assess maltreatment in theDunedin study than the number of items (n = 7) used to assess adversity inthe VTSABD. Given the low prevalence of the definite exposure category, adversitywas also coded as an ordinal variable by counting the positive responses tothe 7 items used to assess parental neglect, interparental violence, and inconsistentdiscipline to maximize statistical power in the data analyses described below.

When we modeled exposure to childhood adversity as an ordinal variable,there was a significant main effect of childhood adversity but not MAO-A onrisk for conduct disorder (Table 1).Low MAO-A activity increased risk for conduct disorder only in the presenceof an increasingly adverse childhood environment. After controlling for theinteraction between low MAO-A activity and childhood adversity and the maineffect of adversity, low MAO-A activity was associated with a lower risk ofconduct disorder.

The prevalence of conduct disorder was plotted by level of adversityamong boys with low vs high MAO-A activity to interpret the basis for theinteraction (Figure 1). The rawdata used to create this graph are given in Table 2. Most of our power to detect an interaction derives fromthe extremes of the distribution (Fisher exact test). Because of our samplesize, we had low power to detect a statistically significant difference inthe prevalence of conduct disorder in association with MAO-A genotype in childrenexposed to multiple adversities.

A statistical interaction such as we observed may result from a genotype-environmentinteraction or, under certain circumstances, from a genotype-environment correlation.This distinction is important because a genotype-environment interaction reflectsgenetically mediated sensitivity to environmental influences and/or environmentallymediated effect of genotype, whereby genes and environment together affectthe phenotype. A genotype-environment correlation, in contrast, reflects anonrandom distribution of environments among different genotypes. In our specificcase, this may be due to either a direct influence of the child's genotypeon experienced adversity (an evocative genotype-environment correlation) oran indirect influence of the child's genotype on experienced adversity viacorrelated parental characteristics (a passive genotype-environment correlation).

If adverse parental treatment is elicited by a child with low MAO-Aactivity, this could create an evocative genotype-environment correlation.In a test for an evocative genotype-environment correlation, low MAO-A activitydid not predict level of exposure to childhood adversity (linear regressionanalysis, β = .08, P = .38). The child's owngenotype therefore had no discernible impact on their exposure to adversityin this study.

Children derive both genotype and familial environment from their parents.A passive genotype-environment correlation is present if parental characteristicsassociated with the child's exposure to familial adversity are correlatedwith the child's genotype. If a parent with a particular genotype is morelikely to create a particular family environment, this will create a passivegenotype-environment correlation in the child. We therefore tested if theobserved interaction is attributable to a passive genotype-environment correlation.We have chosen to adjust for maternal symptoms of antisocial personality inour regression model because males inherit their X-linked MAO-A allele fromtheir mother and because antisocial personality is associated with poor parentingand thus potentially with exposure to the adversities assessed in this study.

We found a significant correlation between the child's exposure to adversitiesand maternal antisocial personality symptoms (Spearman rank correlation coefficient, r = 0.24, P<.001); this wasconsistent with our expectation that antisocial personality is associatedwith poor parenting. Adjustment for the main effects of the child's MAO-Agenotype, the child's level of exposure to adversity, and maternal antisocialpersonality symptoms in our regression model did not, however, attenuate themagnitude or the statistical significance of the association between conductdisorder and the interaction between MAO-A activity and adversity (Table 1).

A passive genotype-environment correlation therefore does not accountfor the observed association. Our findings in both tests for genotype-environmentcorrelation are therefore consistent with a random distribution of low-activitygenotypes among children exposed to different levels of childhood adversity.

When we modelled childhood adversity following Caspi and colleagues,there was a significant main effect of probable (OR, 2.76; 95% confidenceinterval [CI], 1.55-4.93; P<.001) or definiteexposure to childhood adversity (OR, 4.35; 95% CI, 1.4-13.5; P = .01) but not a significant main effect of MAO-A (OR, 0.63; 95%CI, 0.33-1.22; P = .18) on risk for conduct disorder(Table 2). There was also a marginallysignificant association between risk for conduct disorder and the interactionbetween low MAO-A activity and definite (OR, 7.56; 95% CI, 0.59-95.95; P = .059) but not probable exposure to childhood adversity(OR, 1.32; 95% CI, 0.32-5.41; P = .34) after controllingfor the main effect of definite (OR, 2.67; 95% CI, 0.69-10.23; P = .15) or probable exposure to childhood adversity (OR, 2.62; 95%CI, 1.35-5.05; P = .004) and low MAO-A activity (OR,0.48; 95% CI, 0.17-1.3; P = .15). In a test for apassive genotype-environment correlation, there was still a marginally significantassociation between risk for conduct disorder and the interaction betweenlow MAO-A activity and definite (OR, 5.84; 95% CI, 0.44-77.97; P = .09) but not probable exposure to childhood adversity (OR, 1.36;95% CI, 0.32-5.72; P = .33) after controlling formaternal symptoms of antisocial personality disorder (OR, 1.18; 95% CI, 0.94-1.5; P = .14) and the main effect of definite (OR, 2.22; 95%CI, 0.54-9.07; P = .26) or probable exposure to childhoodadversity (OR, 2.45; 95 % CI, 1.22-4.93; P = .01)and low MAO-A activity (OR, 0.5; 95% CI, 0.18-1.35; P =.17). In a test for an evocative genotype-environment correlation, low MAO-Aactivity did not predict level of exposure to childhood adversity handledas a categorical variable (χ²₂= 0.69, P = .7).

This study replicates the report by Caspi et al⁴ ofa genotype-environment interaction that predicts individual variation in riskfor antisocial behavior. We detect an interaction between MAO-A and exposureto a different set of childhood adversities in those surveyed by Caspi andcolleagues. This suggests a relatively robust effect of MAO-A in combinationwith exposure to environmental adversity on risk for conduct disorder. Bothsets of findings are also consistent with reports from adoption studies describingan increased risk for antisocial behavior in boys in association with an interactionbetween aggregate genetic effects and exposure to adversities within the adoptivefamily.¹

A genotype-environment interaction may indicate a variety of underlyingprocesses,^15- 17 andit is therefore important to distinguish the basis for the observed interaction.An interaction against a background of significant main effects may resultin a fanning out of differences among groups. This would occur, for example,when there is an independent effect of a high-risk genotype and a high-riskenvironment. If the presence of both factors increases risk in a nonadditivefashion, this will be observed as an interaction. In this case, the high-riskenvironment potentiates the effects of an already high-risk genotype, andvice versa. An interaction without a significant main effect of genotype resultsin a crossing-over or reordering of risk among groups contingent on environmentalexposure. In this case, risk associated with a specific genotype differs qualitativelyin association with different environments. Variation in exposure to environmentalrisks reorders genotypic effects. Our data are consistent with the latterexplanation, but a larger study is required to definitively address this possibility.

In our study, there was no independent effect of low MAO-A activityon risk for conduct disorder in a model that included only the main (additive)effects of MAO-A and exposure to childhood adversity. In fact, in that model,the high-activity MAO-A genotype was associated with an insignificantly higherrisk for conduct disorder than the low-activity MAO-A genotype. This effectwas reversed, however, when the joint effect of genotype and exposure to thehigh-risk environment was considered. The low-activity genotype was associatedwith an increasing risk for conduct disorder only in association with exposureto increasingly severe environmental adversity. After we controlled for thisinteraction, the low-activity MAO-A genotype was associated with a significantlylower risk for conduct disorder. This is an important finding because it suggeststhat specific genotypes may be associated with increasing or decreasing risksfor psychiatric disorder contingent on environmental exposures. Moreover,relevant genetic factors may not be detected at all unless the target sampleis stratified by salient environmental risk factors.

If genotype-environment interactions are generalized phenomena, thereare several implications for psychiatric genetic studies. The effects of genotypemay be small (or nonexistent) in gene-phenotype association studies if variabilityin exposure to environmental adversity is not taken into account. The absenceof an independent effect of MAO-A on risk for conduct disorder in both theDunedin and the Virginia studies underscores this point. It will thereforebe important to collect detailed environmental histories as well as DNA onsubjects participating in projects designed to identify genetic risk factorsfor psychiatric disorder. When samples differ in the prevalence of exposureto salient environmental risk factors, failures to replicate may remain therule. Epidemiological samples are therefore likely to play an increasinglyimportant role in the detection of genes for psychiatric disorder becausethey routinely sample a diversity of environments in unselected subjects froma representative range of families. Identification of a reordering of genotypiceffects contingent on environmental exposure also requires subjects with abroad range of genetic and environmental risk and protective factors, includingsubjects who lack the environmental risk factors that may moderate genotypiceffects.

A reordering of genotypic effects contingent on environmental exposurehas been described in domesticated animal and plant species,^17- 18 andsuch effects may reflect selection for a variable response to environmentalfactors.¹⁷ This seems quite plausible for behavioraloutcomes and may explain why specific environmental risk factors for juveniledisorders have been consistently replicated across studies,¹⁹ whereasthe identification of specific genetic risk factors has been disappointinglyelusive. In our own study, low MAO-A activity should not be considered a high-riskgenotype per se, but it is perhaps poorly suited to tolerating specific formsof environmental adversity.

Many previous genetically informative studies have focused on estimatingthe main effects of genes on risk for psychiatric disorders independent ofenvironmental exposures. This paradigm assumes a direct path from genes todisorders with no interaction between genetic and environmental risk factors.In the case of measured genetic effects, this approach has yielded very disappointingresults despite significant investment.^20- 21 Mostcandidate gene findings have failed consistent replication, and even thosethat have been verified account for only a very small fraction of variationin risk.²² If genes primarily exert effectson risk for psychiatric disorders contingent on environmental exposures, thismay explain the very limited success of studies directed at detecting themain effects of genes via traditional linkage or association studies. Giventhe importance of epigenetics in human brain development²³ andthe complexity of human behavioral repertoires, it should not be surprisingif complex genetic effects on psychiatric phenotypes are more readily discernedwhen relevant environmental contexts are taken into account. The implicationsfor twin and family studies are not trivial. If a genotype-environment interactionis present but not modeled explicitly, it will be confounded with estimatesof the main effects of genes or the environment in twins reared together.^24- 25 An interaction between genetic andenvironmental risk factors unshared by relatives will be confounded with estimatesof the environment and will be identified as a main effect of the environmentunique to individuals. An interaction between genes and environmental factorsshared by relatives will be confounded with estimates of both genetic effectsand common environmental effects, and will be identified as a main effectof both genes and the common environment.

Genotype-environment interactions have been reported to be importantfor the development of depression as well as antisocial behaviors in studiesestimating aggregate genetic effects,^26- 29 andCaspi et al³⁰ have recently reported that afunctional polymorphism in the promoter region of the serotonin transportergene moderates the influence of stressful life events on depression. Therehas traditionally been a strong emphasis on the likely importance of the independenteffects of genes and environments on risk for complex disease in humans,³¹ but the findings of Caspi et al,⁴ whichare replicated here, suggest that analysis of the joint effects of measuredgenes and environments may indicate 1 important way forward for psychiatricgenetics.

The results presented here should be interpreted in light of the followinglimitations and caveats. First, the male twins who are the subject of thecurrent report are not representative of all male twins in the VTSABD. Atentry into the VTSABD, participants in the current study had a significantlylower prevalence of conduct disorder, and had mothers with significantly fewersymptoms of antisocial personality disorder, than twins who did not participatein the current study. However, any associated bias would likely serve to attenuaterather than overstate the magnitude of effects reported here. Second, theVTSABD did not survey the variables that were used to construct the maltreatmentindex used by Caspi et al,⁴ and we thereforecannot estimate the association between their maltreatment index and our ownadversity index. It is likely, however, that these 2 measures are significantlycorrelated, and they may indicate an overlapping set of environmental risks.Third, our findings are consistent with a reordering of genotypic effectscontingent on environmental exposure, but a larger study is needed to confirmthis possibility. Fourth, population stratification was not formally evaluatedby a genomic control analysis, but we have partly allowed for the effectsof stratification by controlling for the effects of maternal antisocial personalityin our test for a passive genotype-environment correlation. Fifth, previoustwin studies have largely focused on estimating the main effects of autosomalgenes on risk for conduct disorder. However, MAO-A is an X-linked gene, andan increased risk for conduct disorder in association with an interactionbetween MAO-A and exposure to an adverse childhood environment suggests thatthere are sex differences in the heritable transmission of risk to male andfemale offspring. It will therefore be important to determine if the patternof familial resemblance in extended (twin) family data is consistent withthe presence of an X-linked gene with a large effect and if such effects arediscernible only when samples are stratified by salient environmental exposures.Sixth, analysis of other psychiatric outcomes and a broader range of environmentaladversities are required to determine the specificity of the effects on psychiatricdisorder of MAO-A contingent on variation in exposure to salient environmentalrisk factors.

Correspondence: Debra Foley, PhD, Department of Human Genetics, VirginiaCommonwealth University, PO Box 980003, Richmond, VA 23298-0003 (dfoley@hsc.vcu.edu).

Submitted for publication November 28, 2003; final revision receivedFebruary 19, 2004; accepted February 26, 2004.

This work was supported by grants MH-45268 (Dr Eaves) and MH-57761 (DrFoley) from the National Institute of Mental Health, Bethesda, Md; the CarmanTrust for Scientific Research, Richmond, Va (Dr Silberg); and by a MacArthurJunior Faculty Award (Dr Silberg). We acknowledge the contribution of theVirginia Twin Study for Adolescent Behavioral Development, now part of theMid-Atlantic Twin Registry, Richmond, to ascertainment of subjects for thisstudy. The Mid-Atlantic Twin Registry, directed by Linda Corey, PhD, has receivedsupport from the National Institutes of Health, Bethesda; the Carman Trustfor Scientific Research, Richmond; the John T. and Katherine D. MacArthurFoundation, Chicago, Ill; the W. M. Keck Foundation, Los Angeles, Calif; theJohn Templeton Foundation, Radnor, Pa; and The Robert Wood Johnson Foundation,Princeton, NJ.

We thank Adrian Angold, MD, for his helpful comments on an earlier versionof the manuscript.