Application of a Latent Class Analysis to Empirically Define EatingDisorder Phenotypes FREE

The term anorexia nervosa (AN) was first usedto describe a self-starvation syndrome predominantly affecting adolescentgirls in the latter half of the 19th century.¹ Approximately100 years later, Russell² introduced the term bulimia nervosa (BN) to describe women who exhibited recurrentbouts of binge eating and self-induced vomiting. Although most of these womenhad histories of AN, their binge-purge pattern was sustained at normal weight,leading Russell to conclude that BN represented an "ominous variant of anorexianervosa." Anorexia nervosa and BN now represent the formally recognized syndromesin the category of eating disorders in the DSM-IV, TextRevision (DSM-IV-TR)³ and International Classification of Diseases and Related Health Disorders, 10thRevision (ICD-10).⁴

Although data on psychological correlates^5- 7 andthe course of illness^8- 11 supportcurrent nosologic schemes, other data suggest considerable overlap among AN,BN, and eating disorders not otherwise specified (EDNOS).^12- 15 Inas many as 50% of individuals with AN, BN develops, and among individualswith BN, approximately 30% report histories of AN.¹⁶ Furthermore,the DSM-IV-TR and ICD-10 definitionsare not isomorphic regarding diagnosis of concurrent AN and BN. Consideringthe longitudinal instability of the distinction between disorders and sharedcharacteristics, it is not surprising that family studies suggest substantialcross-transmission of AN, BN, and EDNOS.^13- 14 Givenevidence that eating disorders are heritable,^17- 18 itremains unclear whether familial cross-transmission reflects the existenceof a broad eating disorder phenotype with shared genetic predispositions¹⁹ or limitations in the systems we currently use todistinguish among eating disorders.²⁰ Moreover,the identification of genetic susceptibility loci for illnesses with complexinheritance requires the identification of valid and reliable phenotypes.²¹

Because diagnostic criteria influence how we recognize, research, andtreat eating disorders, it is important to ensure their empirical validity.That is, beyond clinical experience in seeing patients who present with certainsyndromal patterns, is there evidence that certain symptoms coaggregate atabove-chance levels to form distinct disorders of eating? A recent investigationexamined this question using latent class analysis (LCA).²² Briefly,LCA posits that a heterogeneous group can be reduced to several homogeneoussubgroups through evaluating and then minimizing associations among responsesacross multiple variables. For example, within a group of healthy controlsubjects and patients with BN and in whom patient status was unobserved, LCAwould detect an association between the likelihood of endorsing binge eatingand the likelihood of vomiting. By creating 2 latent classes (individualsdenying vs individuals endorsing both symptoms), the association between thesevariables would be eliminated within classes, and the presence of controlsvs patients would be revealed. Thus, LCA is capable of determining the numberand composition of unobserved latent classes that produce observed data. Buliket al²² detected 6 latent classes, of which3 represented clinically significant eating disorders resembling AN, BN, andbinge-eating disorder. Limitations of this investigation included a low baserate of eating disorders in the population-based twin sample and restrictionof eating disorder symptoms included in the LCA to DSM-IV-TR diagnostic criteria. In addition, the interview used "skip" rulesbased on DSM-IV-TR definitions to reduce the overalllength of assessment. For example, individuals who did not binge were notasked about purging.²² Thus, the potentialutility of symptoms not included in the DSM-IV-TR wasnot evaluated, and evaluation of symptoms was shaped by current conceptualizationsof eating disorder typology.

The present study is an empirical investigation of eating disorder phenotypesbased on data from 2 international, multisite studies of genetic factors inthe pathogenesis of AN and BN. Although inclusion as a proband required a DSM-IV-TR eating disorder, inclusion as an affected relativedid not require conformation with current eating disorder diagnostic criteria.This allows assessment of how symptoms coaggregate within individuals whennot required by definition.

Because of the size and scope of the studies from which data are drawn,separate reports describe the methods in detail.^19,23 Relevantmethodological information is reviewed herein.

Participants (N = 1179) were recruited as members of an affected relativepair (ARP). An ARP consisted of 1 proband and at least 1 affected biologicalrelative. For the first study (AN-ARP),²³ probandswere required to have a lifetime diagnosis of DSM-IV-TR AN (excluding the amenorrhea criterion), with AN onset before 25 yearsof age and at least 3 years before participation. For the second study (BN-ARP),¹⁹ probands were required to have a lifetime diagnosisof DSM-IV-TR BN, self-induced vomiting as the primarymethod of purging, and minimum frequency/duration of binge eating and vomitingof twice per week for 6 months. For both studies, affected relatives wererequired to meet criteria for a clinically significant eating disorder. Clinicalsignificance was defined by present distress, disability, or increased riskfor suffering. However, the disorder did not need to conform to DSM-IV-TR criteria for AN or BN. Thus, clinical presentation amongaffected relatives could range from DSM-IV-TR criteriafor AN, BN, or EDNOS characterized as (1) partial AN, (2) recurrent bingeeating and inappropriate compensatory behavior below the minimum frequencyfor a DSM-IV-TR BN diagnosis, or (3) recurrent useof extreme measures to control weight (eg, self-induced vomiting, laxativeabuse, fasting, excessive exercise) in the absence of binge-eating episodes.Binge-eating disorder could not be the sole eating disorder diagnosis foraffected relatives. Table 1 presentslifetime DSM-IV-TR eating disorder diagnoses forthe study group. Most participants (n = 1135 [96.3%]) were female. The mean(SD) age of participants was 28.2 (9.7) years.

Participants were recruited by 11 different centers selected for experiencein assessment of eating disorders and geographic distribution. Centers (principalinvestigator[s]) included the University of Pittsburgh, Pittsburgh, Pa (W.H.K.);Cornell University, White Plains, NY (K.A.H.); University of California–LosAngeles (M.S.); University of Toronto, Toronto, Ontario (A.S.K. and D.B.W.);Roseneck Hospital for Behavioural Medicine, University of Munich, Prien, Germany(M.F. and N.Q.); University of London, London, England (J.T.); Universityof Pisa, Pisa, Italy (A.R., M.M., and G.C.); University of North Dakota, Fargo(J.E.M.); University of Minnesota, Minneapolis (S.J.C.); Harvard University,Cambridge, Mass (P.K.K.); and University of Pennsylvania, Philadelphia (W.H.B.).Each center obtained institutional review board approval from its own humansubjects committee. All participants completed written informed consent beforeparticipation.

All measures used in the AN-ARP and BN-ARP studies have demonstratedhigh reliability (interrater or internal) and reasonable convergent and discriminantvalidity.^19,23

The Structured Interview for Anorexic and Bulimic Disorders²⁴ assesses the presence of eating disorder symptomsand characteristics when each was at its worst and also when symptoms co-occurredto assess lifetime prevalence of DSM-IV-TR eatingdisorders. Lifetime symptoms/diagnoses were examined owing to the longitudinalinstability of current symptoms/diagnoses.

The Yale-Brown Obsessive Compulsive Scale²⁵ assessesthe presence and severity of obsessions and compulsions found among individualswith obsessive-compulsive disorder. The mean score for patients with obsessive-compulsivedisorder is 25.1.²⁶

Among the measures conducted in the BN-ARP study alone, we collapsedthe Structured Clinical Interview for Axis I DSM-IV Disorders²⁷ diagnoses into mood disorders (bipolar I, bipolarII, major depressive, or dysthymia), anxiety disorders (specific phobia, socialphobia, panic, agoraphobia, or obsessive-compulsive, posttraumatic stress,or generalized anxiety disorder), and substance use disorders (abuse or dependencefor alcohol, sedatives, cannabis, stimulants, opioids, cocaine, or hallucinogens)for analyses. With the Structured Clinical Interview for DSM-IV Axis II Personality Disorders,²⁸ diagnoseswere collapsed into clusters B (borderline, antisocial, narcissistic, or histrionic)and C (obsessive-compulsive, avoidant, or dependent) personality disordersfor analyses. Cluster A personality disorders were not assessed owing to theirlow prevalence among individuals with eating disorders and time constraints.

Across the 2 studies and 11 centers, there were 38 clinical interviewers.All had clinical experience in eating disorder treatment or research. Highestdegrees ranged from doctorates (17 interviewers [45%]) to masters' (15 [39%])and bachelors' (6 [16%]) degrees in psychology. Each clinical interviewercompleted a training program that involved (1) viewing trained raters performingassessments (minimum of 3), (2) scoring 3 videotaped interviews with an acceptedlevel of accuracy, and (3) taping practice interviews reviewed by the headof assessment until the criterion was met (mean of 3). Subsequent to thistraining, every 10th interview was audiotaped for review by the head of assessmentto prevent interviewer drift. The κ reliabilities for DSM-IV-TR diagnoses are reported in the tables.

The Temperament and Character Inventory²⁹ assesses7 dimensions of temperament (novelty seeking, harm avoidance, reward dependence,and persistence) and character (self-directedness, cooperativeness, and self-transcendence).

The Multidimensional Perfectionism Scale³⁰ evaluatesoverall perfectionism along with 6 specific dimensions of perfectionism (concernover mistakes, high personal standards, high perceived parental expectations,high perceived parental criticism, doubt about quality of performance, andneed for organization, order, and precision). For this self-report measurealone, participants were asked to make ratings for when their eating disorderwas most severe.

The State-Trait Anxiety Inventory³¹ assessesanxiety "at this moment" (state) and general levels of anxiety (trait).

The following self-report assessments were conducted in the BN-ARP studyalone: the Beck Depression Inventory,³² whichmeasures levels of depression, and the Revised NEO Personality Inventory,³³ which evaluates 5 personality factors (neuroticism,extraversion, openness to experience, agreeableness, and conscientiousness).

We used LCA³⁴ to determine the numberand composition of groups in which participants aggregated on the basis oftheir eating disorder symptoms and characteristics. We used log-linear andevent history analysis with missing data using the EM algorithm³⁵ toconduct LCA and formulas presented by McCutcheon³⁴ toassign participants to their most likely latent class.

We included the DSM-IV-TR diagnostic criteriaand features commonly associated with but not required for a diagnosis ofan eating disorder in the LCA from the Structured Interview of Anorexic andBulimic Disorders (Table 2). Amenorrheawas not included because of missing data for AN-ARP participants. Data concerningfrequency/duration of symptoms were not included to prevent violation of theconditional independence assumption within LCA. Finally, data concerning undueinfluence of shape and weight on self-evaluation, fear of becoming fat, andthe importance of being thin were not included because each was endorsed bymore than 98% of participants (reflecting study inclusion criteria).

An LCA was conducted with the affected relatives alone (n = 632) toreduce the influence of proband inclusion criteria on the resulting solution,with probands alone (n = 547) to examine latent classes underlying cases inwhich DSM-IV-TR criteria were met, and with the fullsample (N = 1179). The following 4 criteria were used to evaluate the bestsolution: (1) a nonsignificant χ² statistic, suggesting nosignificant difference between predicted and observed data distributions;(2) a significant improvement in fit from the previous solution (the differencebetween the log-likelihood statistics and degrees of freedom was evaluatedalong a χ² distribution); (3) insignificant improvement infit between the solution and subsequent solution; and (4) stability of thelatent class solution across multiple runs to avoid the problem of local maxima.All 3 LCAs produced the same solution characterized by similar symptom profilessupporting the reliability of the solution. Results of the full LCA are presented,and validation analyses were conducted with the full data set.

Validation analyses served the following 2 purposes: examination ofcorrelates that may reflect "more fundamental abnormality" supporting clinicalvalidity of categories³⁶ and description oflatent classes. For example, temperament is a posited risk factor for theemergence of distinct eating disorders, whereas highest body mass index (BMI;calculated as self-reported weight in kilograms divided by the square of heightin meters) was included for descriptive purposes. Measures were selected onthe bases of prior empirical support for their associations with eating disordersubtypes as well as expert opinion among coinvestigators of clinical relevance.

Validation analyses were conducted using generalized estimating equations^37- 39 to control for inclusionof biological relatives in the sample. For continuous dependent variables(such as Multidimensional Perfectionism Scale scores), we calculated the correlationwithin pairs of female probands and their sisters and then entered this estimatedassociation within an exchangeable correlation matrix as a conservative correctionfor similarity in response patterns within families. For nominal dependentvariables (such as mood disorder diagnosis), only the independent correlationmatrix is possible. Application of the exchangeable vs independent correlationmatrix for analyses of continuous dependent variables did not influence theresults in terms of effect sizes or statistical significance. Therefore, wedo not believe that this variation in controlling for related data would haveany impact on interpretation of results.

As already stated, data came from 2 studies and across 11 centers. Differencesin validation variables due to center were found within latent class nestedwithin study; however, differences due to study were not found within latentclass nested within center. Therefore, only center was included as an additionalclustering variable in analyses. Type 3 tests (score [χ²] statistics)were used for testing the significance of latent class for each variable.Means and frequencies adjusted for cluster relationships (family and center)were generated by means of generalized estimating equations, and post hoccontrasts using score statistics were conducted on these adjusted values toexamine differences between latent classes on the dependent variables. Becauseof the large sample size relative to the number of cases within families orcenters, the adjusted and unadjusted values are quite similar, and unadjustedvalues are reported for descriptive purposes.

A Bonferroni-corrected α level of .0016 was used to evaluate associationsbetween latent class and validation variables (representing 32 independentanalyses), whereas an α level of .0083 was used to evaluate post hoccontrast analyses (representing 6 independent contrasts for each overall effect).Statistical analyses were conducted using the GENMOD procedure of SAS 8.0.2.⁴⁰

After fitting a single-class model, the addition of further classesimproved the fit of the model up to a 4-class solution reliably (model χ²_{167 862} = 117 185 [P>.99];improvement of fit χ²₁₆ = 349 [P<.001]). Of the total sample, 369 (31.3%) were members of latentclass 1 (LC1); 565 (47.9%), latent class 2 (LC2); 46 (3.9%), latent class3 (LC3); and 199 (16.9%), latent class 4 (LC4). Biological relatives weresignificantly likely to share a latent class (χ²₃ =17.96 [P<.001]), supporting the role of familialfactors in the etiology of these empirically derived phenotypes. Table 2 presents the distribution of theeating disorder symptoms/features across the 4 latent classes.

Distributions of symptoms across latent classes suggested an associationbetween our latent classes and DSM-IV-TR diagnosticcategories, supported by regression analysis with generalized estimating equationcorrections (χ²₃ = 190.58 [P<.001]).Post hoc contrast analyses suggested that this was largely attributable tothe unique association between LC1 and women with lifetime histories of AN(Figure 1). The distribution of DSM-IV-TR diagnoses differed significantly between LC1and the other latent classes, but no other significant differences existedamong latent classes. The primary sources of disagreement between latent classesand diagnostic categories were the relatively even distribution of individualswith EDNOS across latent classes and the overrepresentation of participantswith lifetime histories of BN in LC2 (Figure1). In addition to low weight, compulsive rituals around eatingand the use of multiple methods of purging further differentiated LC2 fromLC4 (Table 2). Approximately 81%of individuals in LC2 and LC4 engaged in self-induced vomiting; however, 64.1%of individuals in LC2, compared with 24.6% of individuals in LC4, used laxatives,diuretics, or appetite suppressants to control their weight. Indeed, LC2 wasthe only group in which most members used multiple methods for purging. AlthoughLC1 and LC3 resembled restricting AN, LC3 was distinguished by a relativeabsence of eating and body-related preoccupations and compulsions. Thus, wedeveloped the following clinical descriptions for the 4 latent classes: forLC1, restricting AN (RAN); for LC2, AN and BN (ABN) with multiple methodsof purging; for LC3, RAN without eating and body-related obsessive-compulsivefeatures; and for LC4, BN with self-induced vomiting.

Table 3 presents demographicdata for the latent classes. Although men represented 3.7% of all participants,they were underrepresented in LC2 and overrepresented in LC3. The BMIs werelowest in LC1 and highest in LC4. Current and highest BMIs were also lowerin LC3 compared with LC2, suggesting that binge eating is associated withincreased weight. In contrast to similarities in BMI, LC2 reported a youngerage of onset compared with LC4, and LC3 reported the oldest age of onset.

Table 4 presents validationanalyses of the latent classes on personality and attitudinal measures. Latentclass 1 reported the highest need for organization, order, and precision andalso reported high obsessions and compulsions. In addition, LC1 was distinguishedfrom other classes by low novelty seeking, high persistence (tendency to maintainbehavior that is no longer rewarded), and high conscientiousness. Overall,LC2 reported the greatest level of distress, with high scores for perfectionism,obsessions and compulsions, anxiety, harm avoidance, neuroticism, and depression.In addition, LC2 was distinguished by the lowest levels of self-directedness(ie, the tendency not to view the self as autonomous). In contrast, LC3 wasassociated with the lowest scores on perfectionism and obsessions and compulsions,confirming the relative absence of obsessive-compulsive features in this group.Finally, LC4 did not have the highest or lowest scores on any measure. Formany comparisons, LC3 and LC4 reported similar scores that were less severethat those reported by LC1 and LC2.

Table 5 presents validationanalyses of the latent classes on lifetime history of Axis I and II disorders.Compared with the other classes, LC2 reported the highest lifetime prevalenceof mood, substance use, and cluster B personality disorders.

We found 4 latent classes among participants with clinically significanteating disorders. The primary distinctions that arose across classes confirmedsome of the criteria used by the DSM-IV-TR to differentiateeating disorder diagnoses and introduced other criteria by which to make distinctions.In support of the DSM-IV-TR criteria, this studyfound substantial empirical support for differentiating a restricting subtypeof AN. No individuals with a lifetime history of BN were found in LC1 or LC3.In addition, our analyses supported distinguishing among women who binge andpurge on the basis of low weight, reflecting the diagnostic hierarchy presentedin the DSM-IV-TR but not the ICD-10. However, our findings suggested that low weight alone was inadequatefor differentiating these groups. Instead, the presence of multiple methodsof purging appeared to identify a form of eating disorder that was associatedwith particularly high scores on measures of distress and high comorbidity.Similarly, our analyses suggested that a further subdivision may be meritedamong individuals with restricting AN on the basis of obsessive-compulsivefeatures. Finally, our study did not support a distinction between individualswith lifetime histories of EDNOS and those with lifetime histories of AN orBN.

Our LCA supported 2 groups that could be broadly characterized as havingRAN. The larger group (LC1) was characterized by greater perfectionism, obsessions,compulsions, rigidity, conscientiousness, lower levels of novelty seeking,and higher levels of harm avoidance, consistent with early clinical descriptionsof patients with AN.⁴¹ Comorbid obsessive-compulsivedisorder cannot account for this clinical presentation because a substantialproportion of LC1 had no lifetime history of anxiety disorders. In contrast,the smaller LC3 presented with the lowest levels of disturbance across validatinganalyses, and the combination of higher levels of novelty seeking and lowerlevels of harm avoidance provided good separation from LC1 on temperamentalfactors. The older age of onset for LC3 could reflect an increased premorbidperiod during which normal personality may develop. In the AN outcome literature,obsessive-compulsive personality features have been associated with worseoutcome,⁴² suggesting that these may serveas a marker for the presence of a particularly chronic form of RAN.

If obsessive-compulsive features in RAN serve as a potential markerof chronicity, then the presence of multiple methods of purging and low weightamong women who binge serve as markers of severity. Across 32 validation analyses,LC2 reported the most severe psychological disturbance in 21, and they wereunique in presenting with a combination of higher levels of novelty seekingand harm avoidance. The variables on which they did not report greatest severityincluded lowest, highest, and current BMI and the need for organization, order,and precision. Also, compared with the RAN groups (LC1 and LC3), LC2 reportedlower levels of self-directedness, persistence, and conscientiousness, potentiallyrevealing the causes and consequences of binge-purge symptoms in this low-weightsample. Individuals in whom ABN with multiple methods of purging developsmay be ill-suited to maintain a lowered body weight. This could result inhigher BMIs relative to their counterparts with RAN and the emergence of bingeeating after sustained food restriction. Furthermore, binge eating is likelyto contribute to weight gain, as purging is largely ineffective in eliminatingcalories.⁴³ Women with ABN with multiple methodsof purging appear to lack the extreme levels of rigidity and control reportedby their counterparts with RAN, and this may make them temperamentally lesssuited to sustained food refusal. The combination of difficulties in maintaininga low weight and high levels of neuroticism and depression may make theseindividuals more likely to use multiple purging methods to control their weight.Use of multiple purging methods could also contribute to feelings of distressand shame. These individuals were also most likely to report doubts abouttheir own performance and high parental criticism, suggesting an acute awarenessof discrepancies between their highly perfectionistic goals and their abilityto meet these standards. Moreover, our results support distinctions betweenthe RAN and AN binge-purge subtypes within the DSM-IV-TR and suggest further means for distinguishing the AN binge-purge subtypefrom BN.

Although LC4 was not the most or the least severe group across analyses,they were distinct from the first 2 latent classes on a number of measures,particularly those evaluating perfectionism, obsessive-compulsive features,anxiety, harm avoidance, and persistence. Like LC2, LC4 was less likely toattain or maintain a low body weight. However, this may represent a less salientgoal for these individuals, as they were less likely to report extreme caloricrestriction, qualitative restrictions in what they ate, or urges to eat forbiddenfoods. Similarly, LC4 reported less perfectionistic strivings for extremegoals along with less perfectionism and less rigidity. Thus, their reduceddistress compared with LC2 may partially reflect a diminished discrepancybetween goals and achievements in BN with vomiting relative to ABN with multiplemethods of purging.

We did not produce any qualitatively distinct EDNOS categories. Therelatively even distribution across latent classes of individuals with lifetimediagnoses of EDNOS who never met full criteria for AN or BN suggests thatspecific criteria currently used to distinguish AN and BN from EDNOS werenot empirically supported. In some of these individuals, full-threshold eatingdisorders may develop,⁴⁴ particularly giventhe presence of AN or BN or both in their biological relatives. However, ourresults are consistent with those of other studies.^6- 7

Although our results could be interpreted as demonstrating 4 categoriesof eating disorders, LCA tends to identify spurious classes when an underlyingcontinuum of severity causes associations among observed variables withineach true latent class.⁴⁵ For example, it ispossible that both RAN groups (LC1 and LC3) belong to 1 true latent classbut represent different levels of severity within this class. This tendency,as well as differences in study group ascertainment, could contribute to thediscrepancy in the number of classes found between this study and that byBulik et al.²² Moreover, LCA does not precludedimensional associations among classes. Questions concerning the presenceof discrete taxa vs continua underlying eating disorder diagnoses have beenexamined using taxometric analyses.^46- 47 Resultsof such studies support the categorical distinctiveness of RAN vs binge-eating/purgingAN⁴⁶ and AN vs BN.⁴⁷ Thus,from different analytic approaches, there is excellent support for differentiatingcategories of restricting vs bulimic eating disorders. However, past taxometricanalyses have relied on DSM definitions of eatingdisorders. Future application of taxometric analyses to the phenotypes presentedherein can address the validity of 4 eating disorder categories.

This study had several strengths, including the careful assessment oflifetime eating disorders and use of highly reliable and valid assessmenttools. This study also has one of the largest samples of individuals withclinically significant eating disorders. This enabled us to produce a richcharacterization of the empirically derived phenotypes without encounteringproblems with power when controlling for multiple comparisons. By using conservativecorrections for multiple comparisons and similarities in responses due tofamily membership and center, we increased the robustness of our findingsand the likelihood that they will be replicated in future investigations.

This study may not have been sensitive to identification of eating disorderphenotypes that differ markedly from those in the DSM-IV-TR owing to restrictions on proband phenotypes and the familiality oflatent classes. Indeed, the failure to detect a distinct BN-nonpurging subtypemay reflect proband inclusion criteria and heritability of vomiting.^48- 49 As a consequence, our results mayhave limited relevance in developing a typology for eating disorders in children.Selection of affected relatives suitable for linkage analyses may limit generalizabilityof findings for defining eating disorders in individuals with no family historyof eating disorders. We did not include variables concerning frequency orduration of behaviors or co-occurrence of symptoms owing to the requirementfor conditional independence. Although these are important factors in makingeating disorder diagnoses, the original and present studies were not designedto determine the threshold between eating disorders and normal eating. Comparisonsof latent classes on personality or attitudinal measures should be interpretedas comparisons among groups with clinically significant eating disorders.Thus, for example, lower perfectionism in the BN with vomiting group (LC4)does not reflect that absence of perfectionism in this group. Finally, useof lifetime diagnoses involved retrospective recall that may introduce biases.

This study supports low weight and multiple purging methods as importantindicators of distinct phenotypes among individuals who binge and purge. Furthermore,our findings suggest that obsessive-compulsive features may distinguish 2groups within the category of RAN. Indeed, using the drive for thinness andobsessionality as covariates in the AN-ARP study revealed several regionsof suggestive linkage.⁵⁰ Evidence that thelatent classes coaggregated within families increases their relevance in thesearch for genetic loci with moderate to large effects on disorder expression.²¹ Phenotypes with longitudinal stability and specificbiological and psychological risk factors are needed to identify susceptibilitygenes for eating disorders.⁵¹ Future investigationscan evaluate the predictive and etiologic validity of these phenotypes withlongitudinal and genetic designs. Longitudinal studies may reveal distinctivecourse, treatment response, and outcome across phenotypes, and genetic studiescan examine unique fundamental abnormalities contributing to the emergenceof these phenotypes.³⁶

Corresponding authors and reprints: Pamela K. Keel, PhD, Departmentof Psychology, University of Iowa, Iowa City, IA 52242 (e-mail: pamela-keel@uiowa.edu); Walter H. Kaye, MD, Department of Psychiatry, University of PittsburghMedical Center, Suite 600, Iroquois Bldg, Pittsburgh, PA 15213 (e-mail:kayewh@msx.upmc.edu).

Submitted for publication February 24, 2003; final revision receivedJuly 3, 2003; accepted July 10, 2003.

This study was supported by the Price Foundation, Geneva, Switzerland(Dr Kaye), for the clinical collection of subjects and genotyping, and contributionto the support of data analysis. Data analyses also were supported by grantsR01-MH61836 and R01-MH63758 PK from the National Institutes of Health, Bethesda,Md.

We thank the staff of the Price Foundation Collaborative Group for theirefforts in participant screening and clinical assessments, with particularthanks to Katherine Plotnicov, PhD (head of assessment), and Christine Pollice,MPH (project coordinator). We also thank the participating families for theircontribution of time and effort in support of this study.