Habitual starvation and provocative behaviors: Two potential routes
to extreme suicidal behavior in anorexia nervosa
Edward A. Selbya, April R. Smitha, Cynthia M. Bulikb, Marion P. Olmstedc, Laura Thorntonb,
Traci L. McFarlanec, Wade H. Berrettinid, Harry A. Brandte, Steve Crawforde, Manfred M. Fichterf,
Katherine A. Halmig, Georg E. Jacobyh, Craig L. Johnsoni, Ian Jonesj, Allan S. Kaplanc,
James E. Mitchellk, Detlev O. Nutzingerl, Michael Stroberm, Janet Treasuren,
D. Blake Woodsidec, Walter H. Kayeo, Thomas E. Joiner Jr.a,*
aDepartment of Psychology, Florida State University, USA
bUniversity of North Carolina, Chapel Hill, USA
cDepartment of Psychiatry, Toronto General Hospital, University of Toronto, Canada
dCenter for Neurobiology and Behavior, University of Pennsylvania, USA
eDepartment of Psychiatry, Sheppard Pratt Health System, USA
fRoseneck Hospital for Behavioral Medicine, University of Munich, Germany
gWeill Cornell Medical College, Department of Psychiatry, USA
hKlinik am Korso, Center for the Treatment of Eating Disorders, Germany
iLaureate Psychiatric Hospital, USA
jDepartment of Psychological Medicine, University of Birmingham, UK
kNeuropsychiatric Research Institute, USA
lFaculty of Medicine, University of Lübeck, Germany
mDavid Geffen School of Medicine, University of California, Los Angeles, USA
nDepartment of Academic Psychiatry, Kings College London, UK
oDepartment of Psychiatry, University of California, San Diego, USA
a r t i c l e i n f o
Received 28 October 2009
Received in revised form
28 February 2010
Accepted 18 March 2010
a b s t r a c t
Anorexia nervosa (AN) is perhaps the most lethal mental disorder, in part due to starvation-related
health problems, but especially because of high suicide rates. One potential reason for high suicide rates
in AN may be that those affected face pain and provocation on many fronts, which may in turn reduce
their fear of pain and thereby increase risk for death by suicide. The purpose of the following studies was
to explore whether repetitive exposure to painful and destructive behaviors such as vomiting, laxative
use, and non-suicidal self-injury (NSSI) was a mechanism that linked AN-binge-purging (ANBP) subtype,
as opposed to AN-restricting subtype (ANR), to extreme suicidal behavior. Study 1 utilized a sample of
787 individuals diagnosed with one or the other subtype of AN, and structural equation modeling results
supported provocative behaviors as a mechanism linking ANBP to suicidal behavior. A second, unex-
pected mechanism emerged linking ANR to suicidal behavior via restricting. Study 2, which used
a sample of 249 AN patients, replicated these findings, including the second mechanism linking ANR to
suicide attempts. Two potential routes to suicidal behavior in AN appear to have been identified: one
route through repetitive experience with provocative behaviors for ANBP, and a second for exposure to
pain through the starvation of restricting in ANR.
? 2010 Elsevier Ltd. All rights reserved.
Anorexia nervosa (AN), a disorder characterized by an incessant
pursuit of thinness even after reaching an extremely low body
weight, is a deadly disease (Sullivan, 2002), with a standardized
mortality ratio (SMR; the ratio of number of actual deaths to
expected deaths in the population) ranging from 6.2 in recent
studies (Papadopoulos, Ekbom, Brandt, & Ekselius, 2009), to as high
as 17.8. (Norring & Sohlberg, 1993). Contrary to popular belief, the
majority of deaths from AN are by suicide, not from medical
complications that result from prolonged starvation (e.g., Crisp,
Callender, Halek, & Hus, 1992; Fedorowicz et al., 2007; Harris &
Barraclough, 1997). Keel et al. (2003) found that people with AN
were almost 57 times more likely to die by suicide than those in the
* Correspondence to: Thomas E. Joiner, Department of Psychology, Florida State
University, 1107 W. Call Street, Tallahassee, FL 32306-1270, USA. Tel.: þ1 850644
1454; fax: þ1 850 644 7739.
E-mail address: firstname.lastname@example.org (T.E. Joiner Jr.).
Contents lists available at ScienceDirect
Behaviour Research and Therapy
journal homepage: www.elsevier.com/locate/brat
0005-7967/$ e see front matter ? 2010 Elsevier Ltd. All rights reserved.
Behaviour Research and Therapy 48 (2010) 634e645
general population (SMR ¼ 56.90). Remarkably, these rates of death
by suicide are high even as compared to rates for other lethal
mental disorders, such as schizophrenia and major depressive
disorder (Harris & Barraclough, 1997).
The elevated suicide rate evidenced in people with AN begs
a question: What are the mechanisms through which AN confers
such a high risk for suicide? One explanation is that the suicide rates
reported in published studies are inflated due to their reliance on
inpatient samples (which arguably represent more serious cases).
Yet studies have found that not only does a history of hospitalization
for eating disorders appear to have no effect on risk for suicide, but
that eating disordered patients who had been hospitalized actually
appear to have a lower risk for death by suicide than those who have
not been hospitalized (Keel et al., 2003; Papadopoulos et al., 2009).
Another explanation is what some have called the “fragility
hypothesis” (Joiner, Van Orden, Witte, & Rudd, 2009). According to
this hypothesis people with AN die by suicide at an elevated rate
due to the starvation-induced frailty of their bodies. Attempts that
would not be lethal for a normal weight individual might result in
death in people with AN, due to their weakened condition.
However, results from studies that report the methods of suicide
used suggest that the attempts made by people with AN are lethal
enough to kill even the healthiest individuals (Bulik et al., 2008;
Holm-Denoma et al., 2008).
Anorexia nervosa and the acquired capability for suicidal
Joiner, 2005) poses that there are three proximal, jointly necessary,
and jointly sufficient causes which must be present before a person
will die by suicide; these are: 1) feelings of perceived burdensome-
ness, 2) a sense of thwarted belongingness, and 3) an acquired
capability to enact lethal self-injury. Although AN has been found to
influence belongingness and burdensomeness (e.g., De la Rie,
Noordenbos, Donker, & van Furth, 2007; De la Rie, van Furth, De
Koning, Noordenbos, & Donker, 2005; Hillege, Beale, & McMaster,
2006; Simon, Schmidt, & Pilling, 2005; Tiller et al., 1997), the
primary focus of this study is on the third factor, fearlessness about
Death by suicide is inherently fearsome, and often involves
substantial pain. The IPTS suggests that the acquired capability is
or provocative events, ultimately decreasing fear of pain and death,
through habituation. Studies supporting this hypothesis have found
that painful experiences such as car accidents, non-suicidal self-
injury, previous suicide attempts, childhood abuse, and exposure to
combat during wartime have been all linked to later suicidal
behavior (Joiner et al., 2007; Selby et al., 2010; Van Orden et al.,
2008). Other impulsive behaviors, for example skydiving, intrave-
nous drug use, physical fights, and jumping from high places have
also been linked to suicidal behavior (Van Orden et al., 2008).
Importantly, provocative behaviors have also been found to be
correlated with one’s perceived ability to enact lethal self-injury
(r ¼ .29, p < .01; Van Orden et al., 2008). For more detail about the
acquired capability and subsequently death by suicide please see
a review of the IPTS conducted by Van Orden et al. (in press).
Based on the IPTS, then, people with AN should possess high
levels of acquired capability for suicide due to painful and provoca-
tive experiences such as damaging weight control methods, other
associated behaviors such as self-injury, and habitual starvation.
Weight control methods damage tissues, exert pain, and have
adverse health consequences (Baker & Sandle, 1996; Hellstrom,
2007; Sidiropoulos, 2007). Thus, the repetition of the types of
painful behaviors necessary to keep one’s weight drastically low,
accompanied by enduring the painful sequelae of these behaviors, is
likely to inure one to the painful effects of these behaviors, poten-
tially resulting in a fearlessness and stoicism toward the experience
of pain. There is no doubt that the effects of starvation are also
painful; moreover, starvation has deleterious effects on every organ
of the body. Although the acquired capability has not been directly
tested in individuals with AN there is a large body of research which
shows that people with AN have higher pain tolerance, an analogue
of acquired capability, as compared to healthy controls (e.g., Claes,
Vandereycken, & Vertommen, 2006; Lautenbacher, Pauls, Strian,
Pirke, & Krieg, 1991; Raymond et al., 1995).
The idea of fearlessness in the face of pain and death in AN may
seem paradoxical, as previous research has found that AN is asso-
ciated with high levels of harm avoidance (Klump et al., 2004),
a trait that manifests as increased worry, shyness, fearfulness, and
increased behavioral inhibition in order to avoid punishment. Low
harm avoidance, on the other hand, may result in more bold,
confident behavior e and would seem more in line with acquired
capability. Yet harm avoidance and a lack of fear of death may not
be mutually exclusive. For example, some individuals may have no
fearof one thing (e.g. spiders) and be terrified of another (e.g. giving
a speech). Many with AN may avoid situations that are interper-
sonally problematic or that may interfere with personal goals, but
they may also be more tolerant of the discomfort and pain involved
in compensatory behaviors. This same reasoning may extend to
self-injurious and suicidal behaviors.
Differential risk for death by suicide across AN-subtypes?
Importantly, the acquired capability aspect of the IPTS may
explain the high rate of suicide in AN, but it also makes predictions
about which subtype of AN one would expect to be most lethal.
Specifically, the theory would predict that the binge-purge subtype
of AN (ANBP) should confer greater suicide risk because people with
this subtype are likely to engage in even more painful and provoc-
ative behaviors than people with the restricting subtype of AN
(ANR). AlthoughpeoplewithANR and ANBPbothrestrict,thosewith
ANBP will additionally employ painful purging behaviors. Further-
more, individuals with ANBP are more likely to engage in impulsive
behaviors, such as delinquency and substance use (e.g., Wonderlich,
Connolly, & Stice, 2004), which may result in the experience of pain.
Although relatively few studies have been conducted which
closely examine suicidal behavior by AN subtype, those that have
been conducted seem to support this prediction. Some studies have
found that individuals with ANBP had more suicide attempts than
individuals with ANR (Bulik et al., 2008; Favaro & Santonastaso,
1997), diagnostic migration from ANR to ANBP increases the like-
lihood of a suicide attempt (Foulon et al., 2007), and purging
behaviors are highly associated with suicide attempts (Franko &
Keel, 2006; Tozzi et al., 2006). Furthermore, researchers have
found that individuals with ANBP have higher pain tolerance than
individuals with ANR (Pape? zova, Yamamotova, & Uher, 2005).
Despite the findings that ANBP individuals have higher rates of
suicidality, ANR individuals still appear to have elevated levels of
suicide attempts compared to the general population (Franko &
The purpose of the current studies was to determine whether
the acquired capability component of the IPTS provides a useful
mechanism for understanding suicidal behavior in people with AN.
Specifically, we theorized that higher levels of acquired capability
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
would be the mechanism that links those with ANBP to higher
levels of suicidal behavior than those with ANR. Although the other
components of IPTS theory may be relevant to suicide in AN, they
were not examined in these studies. It was hypothesized that the
relation between ANBP and extreme suicidal behavior (i.e. more
frequent and lethal) would be fully mediated by a “Provocative
Behaviors” latent variable, comprised of both eating disordered and
non-eating disordered behaviors. This hypothesis was tested using
structural equation modeling in two large samples of individuals
diagnosed with either ANBP or ANR. As will be seen in both Study 1
and Study 2, support was found for the hypothesized mechanism,
and furthermore a second mechanism linking ANR to suicidal
behavior also emerged, namely, restricting.
Participants consisted of 787 primarily European ancestry,
primarily female (96%) individuals enrolled in a NIH funded Genetics
of Anorexia Nervosa (GAN) Collaborative Study. This was a multi-site
study that took place in various research and clinical settings across
be found in Kaye et al. (2008). Although the number of male partic-
was important to conduct the study with male participants included
because studies indicate that males account forapproximately 10%of
all bulimic and anorexic patients (Carlat, Carmargo, & Herzog,1997),
and this rate may be increasing (Braun,1999). Including males in the
study may also increase the generalizability of the findings.
To be included in this study, all participants had to be over age 16
and have had a lifetime diagnosis of AN according to DSM-IV criteria
(with or without amenorrhea) by age 45. The amenorrhea criterion
was waived due to the lack of reliability of retrospective assessment
inwomen, complications inassessment posed by frequent hormonal
treatment, and that its presence/absence may not characterize AN
individuals in a meaningful way (Gendall et al., 2006; Pinheiro et al.,
2007). The standardized threshold for low weight was defined as
a Body Mass Index (BMI) at or below 18 kg/m2for females and
19.6 kg/m2for males.1Specific criteria for probands included having
a diagnosis of ANat least 3 years before entry into the study and that
they did not engage in regular binge eating as defined by the
frequency and duration set forth by DSM-IV for bulimia nervosa.
as some who engaged in current occasional binge eating, however,
only not at DSM-IV threshold for frequent binge eating.
Lastly, to participate in the study all probands were required to
have at least one first through third degree relative with AN who
was willing to participate in the study. The exceptions to this were
parents and MZ twins, who were not eligible because these rela-
tionships are not genetically informative for linkage analysis. Thus,
all participants in the study had at least one affected relative who
also participated. Exclusion criteria for potential participants
included: a history of severe CNS trauma, psychotic disorders,
developmental disability, or any other medical, neurological, or
substance use disorder that could confound a diagnosis of AN or
interfere with responding during assessment. Those who did not
speak either English or German were also excluded from the study.
Assessment of ED pathology. To assess for a diagnosis of AN, the
Extended Screening Instrument, an expanded modified version of
Module H of the Structured Clinical Interview for Axis I Disorders
(SCID-I; First, Spitzer, Gibbon, & Williams, 1997), was used. Partici-
pants were diagnosed with AN and specified with subtypes of
restricting (ANR), purging (ANP), binging and purging (ANBP), and
AN with a history of bulimia nervosa (ANBN). A total of 357 indi-
viduals (45.4% of the sample) met criteria for ANR, 220 participants
(25% of the sample) were diagnosed with ANP, 116 participants
(14.7% of the sample) met criteria for ANBP, and 94 participants
the eating disorder diagnoses in this study ranged from .93 to 1.0
as currently experiencing symptoms of AN (Kaye et al., 2008).
Provocative behaviors variables. Painful and provocative events
in this study were defined as events or behaviors that would elicit
physical pain and/or fear in most people, such as self-induced
vomiting, severe restriction, or non-suicidal self-injury. Although
many of these behaviors are traditionally thought of as “impulsive
behaviors,” recent research indicates that the relation between
impulsivity and suicide is a function of the pain and fear that can
arise from these impulsive behaviors (Smith et al., 2008; Witte
et al., 2008). Thus, these behaviors may serve as good indicators
of one’s potential habituation to pain and fear. The following vari-
ables were created for use as indicators of a Provocative Behaviors
latent variable created to tap the construct of acquired capability:
Non-eating disordered provocative behaviors (NonEDPB)
The Eatatelife Phenotype (EATATE), Version 2.1, January 19, 2001
(Project EHE, 2001) The EATATE is a retrospective assessment of
childhood perfectionism and rigidity as well as other personality
traits which often predate the onset of an eating disorder. Of
particular relevance to this study, informationwas also collected on
behaviors that could be considered painful and/or provocative:
excessive alcohol consumption, shoplifting or stealing, gambling,
hitting someone or breaking things, provoking fights or arguments,
fire setting, non-suicidal self-injury, overdosing, using street drugs,
excessive spending, and disinhibited or reckless sexual activities.
These behaviors were coded as the patient either engaged (1), or
did not engage (0) in each behavior. The scores for all of these
behaviors were then summed to create a continuous score where
higher scores indicated more experience with these dysregulated
behaviors. This variable demonstrated somewhat low internal
consistency, ranging from a ¼ .57 when either drug use or over-
dosing was included, and a potentially inflated a ¼ .63 when both
are included, due to the nested nature of the two questions.
Eating disordered provocative behaviors (EDPB)
The Structured Interview for Anorexia Nervosa and Bulimic
Syndromes (SIAB; Fichter, Herpetz, Quadflieg, & Herpetz-Dahlmann,
1998) was used to assess many common behaviors found in indi-
viduals with eating disorder psychopathology. The following behav-
iors were assessed: appetite suppressant misuse, enema misuse,
excessive exercise, fasting, Ipecac misuse, laxative misuse, self-
summed to create a continuous variable where higher scores indi-
variable demonstrated adequate internal consistency (a ¼ .63).
Suicidal behavior variables
Previous suicidal behavior was assessed with Section O of the
Diagnostic Interview for Genetics Studies (DIGS; Nurnberger et al.,
1994). This portion of the DIGS contains in-depth questions about
suicidal ideation and behaviors. Each of these variables was used to
1These standardized BMI values correspond to the 5th percentile BMI values of
the National Health and Nutrition Examination Survey epidemiological sample of
females and males, for the average age range (27e29 years) of participants from
previous studies (Hebebrand, Himmelmann, Heseker, Schafer, & Remschmidt, 1996).
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
create a latent variable of Extreme Suicidal Behavior. Information
1) Number of previous suicide attempts.
2) Lethality of most serious attempt, rated by the interviewer as:
0 ¼ no previous suicide attempts,1 ¼ No danger (e.g., held pills
in hand), 2 ¼ minimal (e.g., scratch on wrist), 3 ¼ Mild (e.g.,10
aspirin, mild gastritis), 4 ¼ moderate (e.g., 10 Seconals, briefly
unconscious), 5 ¼ severe (e.g., cut throat), 6 ¼ extreme (e.g.,
respiratory arrest or prolonged coma).
3) Intent to Die during most serious attempt, rated by the inter-
viewer as: 0 ¼ no previous suicide attempts,1 ¼ no or minimal
intent, manipulative gesture; 2 ¼ definite intent, but ambiva-
lent; 3 ¼ serious intent, expected to die.
4) Premeditation prior to most serious attempt, rated by the inter-
viewer as: 0 ¼ no previous suicide attempts,1 ¼ impulsive (less
than 1 h or forethought, used materials immediately at hand),
attempt), 3 ¼ thoroughly premeditated (persistent suicidal
ideation over weeks, months, or longer prior to the attempt).
5) Classification of most serious attempt as a violent attempt: All
individuals were rated on the severity of their attempt method
for their most severe suicide attempt. The attempt was rated
(1) for a violent attempt, as defined as gunshot, stabbing,
hanging, or jumping from a high place. All other suicide
attempts or no attempts were coded as (0). Approximately 2.8%
of the sample was coded as having made a violent attempt.
Assessment of co-occurring psychopathology. Axis I disorders were
diagnosed using the Diagnostic Interview for Genetics Studies
(DIGS; Nurnberger et al., 1994). This diagnostic interview has been
demonstrated to have strong inter-rater and testeretest reliability
(Preisig, Fenton, Matthey, Berney, & Ferrero,1999; Roca et al., 2007).
The inter-rater reliability of Axis I and II diagnoses in the GAN study
ranged from .80 to 1.0 (Kaye et al., 2008). Although the full sample
characteristics of the GAN collaborative study can be found in Kaye
et al. (2008), we have provided information on the most frequent
co-occurring disorders for each AN subtype. For the ANR group the
most common co-occurring lifetime disorders were: major
depressive disorder (64%), obsessiveecompulsive disorder (38%),
social phobia (17%), and alcohol abuse/dependence (14%). For the
ANBP combined group the most common co-occurring lifetime
disorderswere: major depressive
veecompulsive disorder (54%), alcohol abuse/dependence (34%),
social phobia (25%), any drug abuse/dependence (24%), and post-
traumatic stress disorder (15%). Below we provide more detail on
assessments of specific disorders used as covariates in the study.
Lifetime depression. Lifetime depression was used in the analyses
for Study 1 as a covariate in order to demonstrate that the suicidal
behavior of AN individuals was not primarily a function of
depression. In this sample, 538 individuals (68% of the sample) met
diagnostic criteria for a past episode of depression Table 1.
Borderline personality disorder (BPD). All Axis II disorders were
assessed using the Structured Clinical Interview for DSM-IV Person-
ality Disorders (SCID-II; First, Spitzer, Gibbon, Williams, & Benjamin,
1997). Only borderline personality disorder (BPD) was used as
a covariate in the analyses for Study 1, as it is a disorder that also has
elevated suicide rates (Duberstein & Conwell, 1997) and high co-
occurrence with eating disorders (Wonderlich, Swift, Slomik, &
Goodman, 1990). In this sample, 40 individuals (6% of the sample)
met diagnostic criteria for BPD.
Data analytic strategy
The data for this study were analyzed using structural equation
modeling (SEM), with analyses conducted using AMOS 6.0
(Arbuckle & Wothke,1999). The model that was tested is displayed
in Fig.1. This model explores the mediational effects of Provocative
Behaviors (both eating disorder2and non-eating disorder related)
in the relation between AN status (ANR versus ANBP) and Extreme
Suicidal Behavior. Therewere two latent variables in the model that
were created to parse out potential error in measurement and
capture the various facets of each construct better than observed
variables alone can. The first latent variable in the model was
Provocative Behaviors, which consisted of the eating disorder
related provocative behaviors variable (EDPB), and the non-eating
disorder related painful behavior variable (NonEDPB). The second
latent variable was Extreme Suicidal Behavior, which consisted of
the following indicators: number of suicide attempts, lethality of
the most serious attempt, length of premeditation prior tothe most
serious attempt, intent to die during the most serious attempt, and
whether the most serious attempt was classified as violent or not.
Residual variables were placed on the latent variables to
measure error in predicting the latent variables with the indicators.
The following variables were used as covariates in the structural
model, but are not shown in Fig.1 for ease of presentation: age, sex,
lifetime depression, and diagnosis of BPD. Sex and age were used as
covariates because women tend to have more suicide attempts
(McIntosh, 2002), and age is generally correlated with number of
suicide attempts (McIntosh, 2002). All exogenous variables (AN
status and covariates) had direct paths to both latent variables and
correlations with each other. A causal path was indicated between
Provocative Behaviors and Extreme Suicidal Behavior.
In order to evaluate the overall model, the maximum likelihood
chi-square statistic (c2) was used (with non-significance indicating
that the model fit the data perfectly). Due to the chi-square’s
sensitivity to large sample sizes, other fit indices were also used
including the comparative fit index (CFI), the root mean square
error of approximation (RMSEA), and the TuckereLewis index (TLI).
Multiple cutoff criteria were used to determine if the model
adequately fit the data, and consisted of CFI values greater than .95,
RMSEA values of less than .08, and TLI values of .90 or higher (Hu &
Bentler, 1999). To test individual parameter estimates, a cutoff
criterion value for significance was set at p ¼ .05. Due to missing
data for some of the variables (30 participants were missing data at
random; less than 4% of data was missing for the whole sample),
full information maximum likelihood estimation (FIML; Anderson,
1957) was used. FIML provides less biased information than ad-hoc
procedures such as listwise deletion, pairwise deletion, or impu-
tation of means (Little & Rubin, 1987; Schafer, 1997).
The age of participants ranged from 16 to 76 (mean ¼ 29.7,
SD¼ 11.2).ThecurrentBMIsofparticipantsinthisstudy rangedfrom
2Because using eating disordered provocative behaviors as a mediator results in
overlap with the ANBP subtype, which is defined by these behaviors, the overall
model was evaluated without a provocative behaviors latent variable and the non-
eating disordered provocative behaviors variable alone was used as a mediator. This
model resulted in essentially the same fit, path directions and magnitudes as the
model presented in the results. This finding indicates that the overlap of the eating
disorder provocative behaviors and ANBP does not account for the significant
findings of the model.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
and of those 3% were classified as having made a recent, violent
the study indicated that there were significant group differences on
most variables between the ANR group and the other subtypes, but
there were no significant differences between ANP, ANBP, and ANBN
were separated into two groups: AN restricting (ANR) and AN binge
participants.3For all analyses individuals with ANR were coded (1),
and individuals with ANBP were coded (2).
Measurement model. Before examining a structural equation
model, Kline (2005) recommends that the measurement model be
examined in order to detect potential problems with the latent
variables to be used in the structural model. Accordingly,
a confirmatory factor analysis was conducted with the two latent
variables and consisted only of the two latent variables displayed
in Fig. 1. The measurement model met criteria for adequate fit
(c2¼ 83.33, df ¼ 13, p < .001, CFI ¼ .98, TLI ¼ .96, RMSEA ¼ .08). All
factor loadings of the manifest indicators were significant
(p < .001) and were between .33 and .98. The correlation between
Provocative Behaviors and Extreme Suicidal Behavior was signif-
icant (r ¼.58, p < .001) e a result, incidentally, consistent with the
IPTS. Thus, an exploration of the measurement model indicated
that the latent variables provided adequate fit to the data as two
correlated factors, and all indicators significantly contributed to
their corresponding latent variable.
Structural model. The hypothesized model met criteria for adequate
fit (c2¼ 209.00,df¼ 38, p< .001,CFI ¼ .96,TLI¼ .92,RMSEA¼ .076).4
When interpreting the results, it is important to note that ANR was
path from AN status to another variable indicates a stronger relation
to ANBP while a negative path from AN status indicates a stronger
relation to ANR.
The model results indicated that our hypothesized mediational
relation between ANBP and Extreme Suicidal Behavior through
Provocative Behaviors was supported, with a significant, positive
path from AN status to Pain Behaviors (p < .001, b ¼ .73), and
a significant, positive path from Pain Behaviors to Extreme Suicidal
Behavior (p < .001, b ¼ .78).5These findings suggest that there was
a positive association between ANBP and Provocative Behaviors,
and that Provocative Behaviors had a positive association with
Extreme Suicidal Behavior.
The path from AN status to Extreme Suicidal Behavior remained
p < .05), indicating that a partial mediation effect was supported,
rather than the predicted fully mediational effect. To confirm the
original valence and magnitude of the relation between AN and
Extreme Suicidal Behavior the analysis was rerun without the
the path was both significant and positive (b ¼ .20, p < .05); the
positive relation indicates that ANBP and suicidality are related, such
Study 1 means and standard deviations for, and intercorrelations between, all measures.
123456789 1011 12
1. AN Status
4. Suicide Attempts
8. Violent Attempt
10. BPD Diagnosis
Note. N ¼ 787, * ¼ p < .05, ** ¼ p < .001. AN Status ¼ Anorexia Nervosa (Restricting ¼ 1, Binging/Purging ¼ 2); AN Mean ¼ % with ANR; EDPB ¼ Eating Disordered Painful
Behaviors; NonEDPB ¼ Non-Eating Disordered Painful Behaviors; Violent Attempt Mean ¼ % with a violent attempt; Lifetime depression Mean ¼ % with depression;
BPD Diagnosis ¼ Borderline Personality Disorder, Mean ¼ % with BPD; Sex Mean ¼ % female.
3The same analyses presented in the Results section were analyzed
comparing the ANR group to those formally diagnosed ANBP instead of the
combined ANBP group, and the results were essentially the same as the anal-
yses presented, which combine the ANP, ANBP, and ANBN groups into one
4Because these data were intended to study the genetics of AN, all participants
were related to one or more other participants in the study. This resulted in
a potential problem with non-independence of observations. In order to ensure that
this issue of non-independence was not influencing the results of this study, all
analyses were run a second time using a modified dataset that included only one,
randomly-selected member from each family. This ensured that all participants in
this second dataset were independent from each other. The model fit and results
remained essentially the same (N ¼ 382; c2¼ 145.32, df ¼ 38, p < .001, CFI ¼ .96,
TLI ¼ .91, RMSEA ¼ .086); there were no changes in significant findings for either
regression paths or factor loadings, indicating that the pattern of findings is unlikely
to be influenced by data non-independence.
5It is possible that the indicators for the painful and provocative behaviors latent
variable served primarily as markers of disorder severity of AN, rather than as
indicators of level of acquired capability. If this were true, then severity of disorder
may play a larger role in suicidal behavior than our hypothesis of acquired capa-
bility. In order to test this potential alternative, we evaluated the exact same model
tested in Study 1 with the exception that the mediator latent variable had different
indicators. Instead of indicators of painful and provocative behaviors, we used
lowest lifetime BMI as one indicator, and highest lifetime BMI as second indicator.
Research has found these indices to be reliable markers of illness severity; specif-
ically, the lower an individual’s BMI the less likely that individual is to recover from
anorexia and the greater the probability of death due to the condition (Hebebrand
et al.,1997; Howard, Evans, Quintero-Howard, Bowers, & Andersen,1999). Similarly,
those with more severe conditions are also likely to have a lower overall highest
lifetime BMI, potentially indicating longer experience with the condition and less
likelihood of full recovery. This alternative model fit the data well (c2¼ 4475.53,
df ¼ 56, CFI ¼ .98, RMSEA ¼ .046) and both indicators significantly loaded onto the
latent mediator variable. Importantly, however, this severity latent variable did not
have a significant path to the extreme suicidal behavior latent variable (b ¼ .02,
p ¼ .72), and the direct path from AN status to suicidal behavior remained signif-
icant with a positive magnitude. The failure of the severity variable to mediate the
relationship between AN status and suicidal behavior suggests that the findings of
Study 1 are not due to overall severity of AN. A similar alternative analysis was
conducted for Study 2, with similar, consistent results.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
their ANR counterparts.
Interestingly, as noted above, when the Provocative Behaviors
latent variable was introduced into the model, the relation between
AN and Extreme Suicidal Behavior flipped from being significant
and positive (b ¼ .20; indicating a stronger relation between ANBP
and suicidality) to being significant and negative (b ¼ ?.40; indi-
cating a stronger association between ANR and suicidality). Given
the finding that ANBP individuals tend to engage in more serious
suicidal behavior than ANR individuals (Bulik et al., 2008), this
finding was unexpected. This reverse in signs indicates that indi-
viduals with ANBP may have a stronger relation with suicidality
through painful and provocative behaviors, but when these
behaviors are accounted for, a unique relation between ANR and
suicidality emerges such that individuals with ANR endorse more
Extreme Suicidal Behavior than those with ANBP.
on Extreme Suicidal Behavior was b ¼ .57, indicating a significant
impact of Provocative Behaviors on the relations between AN status
and Extreme Suicidal behavior, the PRODCLIN program was used.
This program was developed by MacKinnon, Fritz, Williams, and
Lockwood (2007), and tests meditational effects without some of
theproblemsinherent inothermethodsof testing formediation(e.g.
inflated rates of Type I error, see MacKinnon, Lockwood, Hoffman,
West, & Sheets, 2002). In addition, the logic for this method is well
suited to testing for mediation in structural equation modeling
(Bollen, 1987). PRODCLIN examines the product of the unstandard-
coefficients (ab/sab) and a confidence interval is generated, with
a statistically significant mediation effect indicated by the absence of
zero in the confidence interval. The unstandardized path coefficients
and standard errors of the path coefficients for the indirect effect of
and produced a 95% confidence interval of .72e2.54. Because zero
was not included in the confidence interval, we concluded that the
relation between ANBP and Extreme Suicidal Behavior was signifi-
cantly mediated by Provocative Behaviors.
Additional information regarding the model that was of interest,
though not central to our main aims, was obtained. Both lifetime
depression (b ¼ .20, p < .001) and BPD (b ¼ .19, p < .001) had
significant paths to Provocative Behaviors. The direct paths from
both depression and BPD to Extreme Suicidal Behavior were not
significant, which suggests that there may be a full mediation effect
of painful behaviors on the relation between these disorders and
suicidal behavior (a result that is corroborative of the IPTS and of
expanded upon further). The other covariates (sex and age) did not
significantly predict either Provocative Behaviors or Extreme
Suicidal Behavior. There were significant correlations between the
following: AN status and depression (r ¼ .18, p < .001) such that
individuals with ANBP had a higher prevalence of depression; AN
statusandBPD(r ¼.10, p<.01) suchthatindividualswithANBPhad
a higher prevalence of BPD; depression and BPD (r ¼ .15, p < .001)
such that individuals diagnosed with BPD had higher rates of life-
time depression than those without BPD diagnoses; AN status and
sex (r ¼ .13, p < .001) such that men in the sample tended to have
ANBP slightly more than ANR; and AN status and age (r ¼ .08,
p < .05) such that the ANBP group was somewhat older.
Due to the unexpected switch of the relation between AN status
and Extreme Suicidal Behavior after controlling for Provocative
Behaviors, we generated a post-hoc hypothesis that restricting itself
could cause a great deal of pain and increase risk for Extreme
Suicidal Behavior, and we reanalyzed our model without fasting
included in the EDPB variable. Removing fasting from the EDPB
indicator would allow for more variance to be allocated from the
Provocative Behaviors mediator to the direct path from AN to
Extreme Suicidal Behavior. Thus, if fasting was driving the associa-
tion between ANR and Extreme Suicidal Behavior, removing fasting
from the mediator should have increased the strength of the direct
path from AN to Extreme Suicidal Behavior. After rerunning the
original model with fasting removed from the EDPB variable, the
pathfromANtoExtreme SuicidalBehaviorchanged fromb¼ ?.40to
b ¼ ?.55, indicating that the strength of the relation between ANR
and Extreme Suicidal Behavior increased, providing some support
forour post-hoc hypothesis thatfasting may be a second mechanism
linking AN to suicidal behavior.
Study 1 discussion
The results of Study 1 support our hypothesis that the relation
between ANBP and Extreme Suicidal Behavior is mediated in part by
behaviors such as laxative use and NSSI. This supports the IPTS
notion that the ability to engage in suicidal behavior may develop
through repetitive experience with these behaviors, causing habit-
uation to pain and fear. A novel finding of Study 1 was that when
controlling for Provocative Behaviors, the association between ANR
and Extreme Suicidal Behavior became significant, suggesting that
there may be a specific relation between ANR and Extreme Suicidal
Behavior separate from that of the mechanism linking ANBP to
Extreme Suicidal Behavior.
What might explain this unpredicted effect? Because most
painful and fear inducing behaviors employed by individuals with
AN were accounted for by the Provocative Behaviors latent variable,
there was only one apparent difference between the two diagnostic
subtypes that might be involved in the association between ANR
and Extreme Suicidal Behavior: the degree of restricting engaged
in, which for ANR individuals involves extremely limited food
intake or complete refusal of food (APA,1994). Extensive restriction
is more characteristic of people with ANR and is perhaps qualita-
tively different from the restriction seen in people with ANBP, as
people with ANBP may occasionally alleviate their hunger through
binge eating episodes. Accordingly, we believe that extreme
restricting may be a second route to development of acquired
capability in AN.
Removing the fasting item from the EDPB indicator of the
Provocative Behaviors mediator increased the magnitude of the
* *6 4 . * * 37 .
Fig. 1. Study 1 model with Anorexia Nervosa (AN) status predicting Provocative
Behaviors and Extreme Suicidal Behavior. N ¼ 787. AN Status was coded as ANR
(restricting) ¼ 1, ANBP (binging/purging) ¼ 2; * ¼ p < .05; ** ¼ p < .01; / indicates the
path before and after including Provocative Behaviors in the model. EDPB ¼ Eating
Disordered Provocative Behaviors; NonEDPB ¼ Non-Eating Disordered Provocative
Behaviors. Lifetime depression, borderline personality disorder, age, and sex were used
as covariates in the analysis, but are not displayed for clarity of presentation.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
relation between ANR and Extreme Suicidal Behaviors. Allowing
the fasting item to contribute to the path between ANR and Extreme
Suicidal Behavior, rather than to the Provocative Behaviors mediator,
may explain this increase in relation strength. This finding suggests
that fasting/restriction may be a painful behavior influencing the
relation between ANR and suicidal behavior. It may be thatextensive
experience with the pain induced by restriction increases the
acquired capability for suicide in individuals with ANR, and that this
represents a second mechanism through which AN is linked to
There was an important limitation of note with Study 1. The
Provocative Behaviors variables were created from Likert scale
questions regarding how much one engaged in a particular behavior
(e.g. for vomiting the response options were: never, rarely less than
twice a week), sometimes (at least twice a week), frequently (up to
once a day), very frequently (several times a day). For this study,
these responses were then dichotomized so that the participant was
scored a 0 if they never engaged in the behavior and 1 one for all
other responses.Thismethod of assessment gauges whetherone has
used a behavior, which is relevant to our hypothesized mechanism
linking ANBP to suicidality, but it does not account for frequency of
engagement in a behavior. For example, one might use only one of
the behaviors list (e.g., vomiting), but engage in that behavior
frequently, which may have a similar influence on suicidal behavior
as using a broad array of these behaviors.
Despite finding two possibly different routes to suicidal
behavior in AN, the detection of the relation between ANR and
suicidal behavior was not originally predicted. Furthermore, Study
1 assessed the number of types of painful behaviors that might be
used in AN, but it did not account for the frequency of those
behaviors. In order to explore both mechanisms a second time, and
to include frequency of provocative behaviors in the model, we
analyzed a similar model in a second sample of ANR and ANBP
participants. We again predicted a positive, indirect effect between
ANBP and suicidal behavior, which flowed through a Provocative
Behaviors mediational variable. We also made a riskier prediction:
in the presence of the Provocative Behaviors mediator the relation
between AN and suicidal behavior would be negative, indicating
that ANR also has a distinct relationwith suicidal behavior, through
persistent restriction. Study 2 also aimed to test the hypotheses,
without the limitation noted from Study, 1 by including the
frequency of engaging in painful behaviors in the model Table 2.
Participants. Participants were 249 current patients drawn from an
outpatient and day hospital eating disorders treatment setting at
Toronto General Hospital in Toronto, Canada.6All patients were
assessed at intake, and all provided full, informed, and written
consent for research participation. All patients in this study were
diagnosed with current AN, with 106 (43%) specified with ANR and
143 specified with ANBP. The age of patients ranged from 16 to 68,
with an average age of 26.30 (SD ¼ 8.50). The sample was 98%
female and consisted of almost entirely Caucasian individuals. The
intake BMIs of patients ranged from 11.8 to 18.5 kg/m2, with an
average BMI of 16.75 kg/m2(SD ¼ 1.33). Patients were not given full
Axis I diagnostic interviews, given the eating disorder specialty of
the treatment center, so information on diagnostic co-occurrence
for this sample was not available.
Anorexia diagnoses. All patients were assessed with a modified
version of the Eating Disorder Examination (EDE; Fairburn &
Cooper, 1993) to diagnose AN as well as other eating disorders.
Only patients with AN were included in this study, and all met
criteria for AN according to DSM-IV-TR criteria (except for the
amenorrhea criterion, which was waived). Diagnoses of AN were
further specified according to specific subtypes, ANR and ANBP. The
ANBP group consisted of individuals who purge only as well as
those who binge and purge.
Provocative behaviors. A latent variable similar to that in Study 1
was created using three indicators: vomiting frequency, laxative
abuse, and NSSI. Information on vomiting frequency was assessed
by asking each participant to indicate how many times he or she
engaged in vomiting behavior each month. Information on general
frequency of laxative abuse was obtained by having each partici-
pant rate his or her frequency of laxative abuse with a four-point
Likert scale (1 ¼ never, 4 ¼ often). Frequency of NSSI was assessed
by asking all patients how many times they had engaged in self-
injury, without suicidal intent, in their lives.
Lifetime suicide attempts. All patients were asked how many times
in their lives they had attempted suicide. Information on the
severity of the attempt, the preparation for the attempt, or the
amount of desire to die during the attempt was not collected in this
study. The number of lifetime suicide attempts was used as the
outcome variable in the structural model.
Substance use. A Substance Use latent variable was created as
a covariate in the structural model. All patients were asked to rate
how frequently they used each of the following substances during
their periods of most heavy use: barbiturates, amphetamines,
cocaine, and hallucinogens. Each drug category was rated on a six-
point Likert scale (0 ¼ never, 5 ¼ very frequently). Information on
alcohol use was also collected. For use as an indicator in the
Substance Use latent variable, the number of drinks per day during
their heaviest period of alcohol use was included in order to
represent problematic alcohol use.
Data analytic strategy
Structural equation modeling, using AMOS 6.0, was again
employed in an attempt to replicate the results from Study 1. For the
Study 2 model, a similar Provocative Behaviors latent variable was
created using the following indicators: frequency of vomiting in the
last month, general frequency of laxative abuse, and frequency of
NSSI. Although this model did not provide an extensive list of
painful/provocative behaviors as in Study 1, it nonetheless contained
important behaviors that may contribute to acquired capability and,
importantly, the frequency of those behaviors.
A latent variable of Substance Use was also created to use as
a covariate in the Study 2 model. Although Study 1 tested a number
of important covariates in the relation between AN and suicidal
behavior, substance use was not included. Given the strong relation
between substance abuse and suicidal behavior (Maseret al., 2002),
substance abuse was another important covariate to rule out in the
suicidal behavior of individuals with AN. The Substance Use latent
6There was some concern about patient overlap between the samples of Study 1
and Study 2, as both involve AN individuals from the Toronto area. Unfortunately
records linking participants’ identification to the sample for Study 2 were destroyed
at the end of the study, making identification impossible. In order to control for any
participants potentially included in both samples, the analyses from Study 1 were
reanalyzed without any of the participants (N ¼ 82) from the Toronto site included.
The results were essentially unchanged, and the model maintained similar fit
statistics (c2¼ 207.31, df ¼ 38, p < .001, CFI ¼ .96, TLI ¼ .94, RMSEA ¼ .079). Thus,
even if there is some overlap between the two samples, the results remain
consistent even without those participants included.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
variable was created using the following indicators: period of
heaviest hallucinogen use, heaviest cocaine use, heaviest barbitu-
rate use, heaviest amphetamine use, and number of drinks per day
during heaviest period of drinking. A causal path was drawn
between AN status and Substance Use because we hypothesized
that an AN diagnosis, particularly ANBP, would likely influence
substance use behavior. We also predicted Substance Use would
contribute to Provocative Behaviors as these behaviors may result
in habituation to pain as well (i.e., intravenous use).
In a similar manner as Study 1, AN individuals for Study 2 were
coded with (1) if specified with ANR and (2) if specified with ANBP.
We hypothesized that ANBP would have a positive relation with
number of suicide attempts through the mechanism of the
Provocative Behaviors latent variable. Thus, we expected to find
a significant, positive indirect effect of the AN status variable on
suicidal behavior. Furthermore, we had the a-priori hypothesis for
Study 2 that ANR would have a relation with number of suicide
attempts after controlling for pain behaviors because severe
restricting is another way that one might develop the acquired
capability for suicidal behavior. Thus, we expected to see the
positive relation between AN status and increased suicidality
change to a significant negative relation between the AN variable
and suicide attempts when controlling for Provocative Behaviors.
Age and sex were included as covariates in the model, and residual
predictors were included on the latent variables to model error.
Establishing excellent model fit was not the goal of Study 2;
instead the purpose was to replicate the two mechanisms for
suicidal behavior in AN that were identified in Study 1. Thus, less
stringent model fit criteria were used to evaluate model fit, with
a CFI > .90 and an RMSEA < .10 indicating adequate fit. Mediational
effects were again explored using the PRODCLIN program. Due to
data missing at random for 22 people in the sample (9% of total
sample), FIML was again used to estimate those missing data points
for the model.
Approximately 30% of this sample reported at least one lifetime
suicide attempt. The average number of suicide attempts in this
sample was .81 (SD ¼ 1.63). No significant difference between the
two groups was found regarding number of suicide attempts
(F (1, 248) ¼ 1.22, b ¼ .07, p ¼ .27). Both groups appeared to have
similar statistics on suicide attempts, although the ANBP group
(M ¼ .91, SD ¼ 1.66) trended toward more suicide attempts than the
ANR group (M ¼ .68, SD ¼ 1.60).
The results of the Study 2 model are displayed in Fig. 2. This
model provided adequate fit to the data (c2¼ 87.12, df ¼ 44,
p < .001, CFI ¼ .91, RMSEA ¼ .06).7All latent variable indicators
significantly loaded onto their respective latent variables. The path
from AN status to Provocative Behaviors was positive and signifi-
cant (b ¼ .82, p < .001), indicating an association between ANBP
and Provocative Behaviors. The path from Provocative Behaviors to
number of suicide attempts was also significant and positive
(b ¼ 2.1, p < .001), suggesting that the more patients engaged in
purging, laxative abuse, and self-injury the more suicide attempts
they were likely to have. Yet, also as predicted based on the results
of Study 1, the direct path from AN type to suicide attempts was
negative and significant (b ¼ ?1.70, p < .001), after controlling for
There was a positive indirect effect from AN type to number of
suicide attempts (b ¼ 1.77), as was hypothesized. This positive
indirect effect indicated that the relation between ANBPand number
of suicide attempts flowed through the Provocative Behaviors latent
variable. The mediational effects of Provocative Behaviors on the
relation between ANBP and suicide attempts was explored with
PRODCLIN, described in more detail in Study 1. A confidence interval
of 2.82e8.99 was obtained, and because the range did not include
zero, a significant mediation effect was indicated.
Although not central to our aims, there was additional infor-
mation of importance regarding the model. The significant positive
path fromAN status toSubstance Abuse (b ¼ .15, p < .001) indicated
that the ANBP group tended to engage in more substance use than
the ANR group. There was a significant positive relation between
the Substance Use latent variable and the Provocative Behaviors
latent variable (b ¼ .25, p < .01), supporting the hypothesis that
substance use also contributes to increasing acquired capability.
The path from Substance Use to suicide attempts was not signifi-
cant. Age and sex did not significantly predict either Provocative
Behaviors or Substance Abuse, nor did they predict suicide
Study 2 means and standard deviations for, and intercorrelations between, all measures.
123456789 1011 12
1. AN Status
3. Laxative Use
5. Suicide Attempts
6. Hallucinogen Use
7. Cocaine Use
8. Barbiturate Use
9. Alcohol Use
10. Amphetam. Use
Note: N ¼ 249. * ¼ p < .05; ** ¼ p < .01; ^ ¼ percent of sample diagnosed with ANR; þ ¼ percent of sample that is female; AN Status ¼ anorexia nervosa restricting subtype
coded as 1 and anorexia nervosa binge-purging subtype coded as 2; NSSI ¼ non-suicidal self-injury; Amphetam. Use ¼ Amphetamine Use.
7When this model was analyzed a negative error variance was obtained for the
residual indicator for suicide attempts. This was likely a function of the residual
being near zero, a problem which is minor enough that AMOS was still able to
conduct the analysis. The recommended way to remedy a negative error variance is
to remove the residual indicator and reanalyze the model (Dillon, Kumar, & Mulani,
1987). After running a second analysis where the residual predictor for suicide
attempts was removed, the model was empirically identified (i.e. there were no
negative error variances) and model fit was the same. Following this second
analysis, the empirical identification of the model remains intact and the fit indices
generated by the model remain valid.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
attempts. The correlations of both age and sex with AN status were
Although various studies have documented the relation between
AN and suicidal behavior, few studies have explored potential
mechanisms in this relation. Inspired by the IPTS (Joiner, 2005), we
tested the hypothesis that employment of behaviors that result in
provocation and/or pain would mediate the relation between ANBP
and extreme suicidal behavior. The results of both Study 1 and Study
2 support the mediational effects of painful/fear-inducing behaviors
on the association between ANBP and extreme suicidal behavior.
The second major finding of these studies was that, only after
accounting for pain-inducing behaviors, there appears to be
a substantive association between ANR and extreme suicidal
behavior. Individuals with ANBP tend to endorse more suicidal
behaviors than their ANR counterparts, perhaps due to the painful
and provocative behaviors they engage in. But, after accounting for
these painful behaviors, ANR appears to have a stronger association
with suicidal behaviors. In Study 1 we hypothesized that the rela-
tion between ANR and suicidal behavior emerges because the act of
starving oneself is extremely painful, and restriction is the only
other painful behavior diagnostically specific to ANR that remains
after accounting for the other painful behaviors found in AN.
Furthermore, those with ANR engage in a type of restriction that is
different in severity from those with ANBP, which may indepen-
dently increase acquired capability.
path between ANR and Extreme Suicidal Behavior increased when
the fasting item was removed from the Provocative Behaviors
mediation latent variable, allowing for the variance accounted for by
Although our interpretation of this finding in Study 1 was post-hoc,
the a-priori hypothesis of this effect and replication of this finding in
Study 2 suggest that severe and unrelenting restriction may be
another important behavior that influences suicide risk in AN. It is
also important to note that the suicidal behavior latent variable in
Study 1 consisted of indicators concerning number of attempts, the
use of a violent attempt method, lethality, intent to die, and
premeditation for the most serious attempt. This suggests that many
individuals with AN reported experience with extreme forms of
suicidality, as opposed to a solitary, tentative attempt. Thus, the
simplya “cry forhelp,” assome families orclinicians may mistakenly
think. Importantly, both studies included key covariates such as
depression, BPD diagnosis, and substance abuse, suggesting that
these findings are unlikely a result of co-occurring psychopathology.
Two routes to suicidal behavior in AN
Support was found for two independent routes to increased
suicidal behavior in AN: one through provocative behaviors, such as
vomiting and laxative abuse, the second route through unwavering
restriction. The connection between both subtypes of AN, a dimin-
ished fear of death, and serious suicidality can be seen in a recent
series of case studies on death by suicide and AN (Holm-Denoma
et al., 2008). Among these cases, the authors identified individuals
with ANR and ANBP who died bysuicide and these individuals used
particularly lethal methods that would have killed healthy indi-
viduals, rather than less lethal methods that might have only killed
of trains (two cases), ingested a household cleaning product con-
taining hydrochloric acid, severely overdosed, and died by hanging.
On the other hand, individuals with ANBP died by jumping in front
of a train, hanging, severe overdosing, and fire-induced carbon
monoxide poisoning. The take away point from the Holm-Denoma
et al. (2008) study, interpreted in context of the findings of this
study, is that most of the individuals with both ANR and ANBP who
died by suicide did so through violent methods with a high proba-
bility of death. The ability to enact lethal self-injury in these cases
may have developed by overcoming a fear of death through famil-
iarity with the pain of restricting and compensatory methods.
Strengths and limitations
These studies have a number of important strengths that speak
to the robustness and generalizability of our findings and conclu-
sions. Importantly, we found the same pattern of results across two
large, clinically-impaired, yet distinct, groups of individuals with
AN e one treatment seeking and the other not. This includes the
replication of the positive indirect effect between ANBP and
suicidal behavior, through painful behaviors, as well as the repli-
cation of the association between ANR and suicidal behavior when
including provocative behaviors as a mediator/covariate in the
model. A second strength was that both studies used different
indices of provocative behaviors and suicidality, yet the results
remained essentially the same. Finally, both studies included
rigorous covariates, all of which were important because stronger
conclusions can be made for the specificity of the relation between
AN and suicidal behaviors, with evidence that suicidality in AN was
not an artifact of depression, BPD, or substance abuse.
Although this study makes important contributions to the
literature on AN and suicidal behavior, there are a few limitations
that should be considered when interpreting the findings. The most
important limitation of this study was that both samples were
cross-sectional, a problem that can lead to difficulty identifying the
flow of a meditational model. Yet, given the preliminary stages of
understanding suicidal behavior in AN, there is still much that can
be learned fromthis study. Another important limitation is that this
study was about suicide attempts, rather than death by suicide.
Even though the findings of this study suggest that engaging in
painful behaviors (bothrelated and unrelated to eating disorders) is
associated with increased severity of suicidal behavior, these
findings may not necessarily generalize to AN individuals who
ANR versus ANBP
Number of Suicide
Fig. 2. Study 2 Model. N ¼ 249 (Anorexia e Restricting [ANR] ¼ 106, Anorexia-binge-
Purging [ANBP] ¼ 143). ANR is coded as 1, while ANBP is coded as 2; * ¼ p < .05;
** ¼ p < .01; / indicates the path before and after including Provocative Behaviors in
the model. Model includes sex and age as covariates, which are not displayed for
clarified presentation. NSSI ¼ non-suicidal self-injury.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
eventually die by suicide. Yet, given the use of a latent variable of
severity of suicidal behavior (more attempts, longer premeditation,
greater intent, and greater lethality) in Study 1, it is a reasonable
possibility that individuals expressing extreme suicidal behavior at
the time of this study may have higher risk for death by suicide in
Another limitation is that the ANR group and ANBP group in
Study 2 did not significantly differ on number of suicide attempts
(though the means were in the expected direction). Although
inconsistent with our results from Study 1, other studies have also
found similar rates of suicidal behavior between ANR and ANBP
individuals (e.g., Eddy et al., 2001). In general suicidality appears to
be higher in individuals with ANBP, but it is clearly high in both
ANBP and ANR; and thus, it is not surprising that occasionally the
groups are found not to differ on this behavior. Furthermore,
because the sample used in Study 2 was a treatment seeking
sample, patients may have had more severe impairment, including
suicidal behaviors, and thus the ANRgroupin Study 2 may have had
more severe suicidality than the population based subsample in
Study 1. Final limitations include the cross-sectional design, and
both samples consisted of primarily female, Caucasian participants.
Although this is an important group to study with regard to eating
disorders and suicidal behavior, the findings may not generalize to
other ethnic groups or males with these disorders.
It is also important to note that there may be alternative factors
influencing the link between AN and suicidal behavior, making the
mechanisms proposed in this study somewhat speculative. For
be influenced through impaired serotonin function, which is known
to increase depression and suicidality (e.g., Purselle & Nemeroff,
2003). If this were the case, we would expect to find more impair-
ment in serotonin function inpeople with ANR versus ANBP. Though
some differences in serotonin function have been found in recovered
ANR versus ANBP individuals, it is unclear if these differences exist
during the active phase of the illness (Kaye, 2008). Moreover, Favaro
et al.(2004) foundsimilarlylowlevelsof cholesterolinbothANRand
ANBP patients as compared to controls. It is believed that low
the serotonin transporter (Engelberg, 1992). Thus, these findings
imply that differences in suicidality between ANR and ANBP are
unlikely due solely to impaired serotonin functioning.
Yet, other alternative mechanisms may account for the relation
between AN and suicide. For example, intense psychological pain,
termed psychache by Shneidman (1996), could be more prevalent
in individuals with AN, and psychache could then be further
aggravated by painful and provocative behaviors such that the pain
of living is worse than that of dying. The IPTS theory of suicide
would suggest that this psychache is a result of feelings of bur-
densomeness and belongingness, however, and that acquired
capability is still needed to actually make a serious suicide attempt.
More research on psychache and burdensomeness/belongingness
is needed to address this alternative explanation. Similarly,
research on behaviors such as NSSI and purging suggests they may
be used to cope with negative affect due to physical distractions
such as pain (Selby, Anestis, & Joiner, 2008), yet many patients who
engage in seemingly painful behaviors report an analgesic state
during these behaviors, suggesting that these behaviors might not
actually instill pain (Lieb, Zanarini, Schmahl, Linehan, & Bohus,
2004). This could be evidence against the development of
acquired capability through these behaviors. Yet, along the lines of
the IPTS framework, not feeling pain during these behaviors could
be the result of opponent-processes where, through repetition,
habituation to pain occurs and feelings of relief from emotional
distress take its place. Essentially, the analgesia reported during
these behaviors could be evidence of habituation and increased
acquired capability. More research is needed to fully address this
point, however. Future research on the relation between AN and
suicide may benefit from more strict assessment of feelings of
fearlessness about pain and death, perhaps with the use of the
Acquired Capability for Suicide Scale (ACSS; see Van Orden et al.,
2008). Replication of the findings from the current studies with
a more precise measurement of acquired capability may provide
more support for the speculative theoretical underpinnings of the
current study. Future studies should also attempt to test all three
components of the IPTS in AN todetermine if all components of this
theory are relevant to those with AN, and future studies should
examine if the behaviors examined in this study actually contribute
to the development of acquired capability.
The results of this study provide two potential routes to
increased suicidal behavior in AN. The first route, applying
primarily to individuals with ANBP, may be through the use of
painful compensatory behaviors as well as other self-damaging
behaviors, such as NSSI. The second route, which may be somewhat
more relevant to the ANR group (but relevant to some degree for
both groups), may be through the constant pain associated with
obstinate dietary restriction. Future research should continue to
explore the proposed mechanisms as well as identify other mech-
anisms that may increase suicide risk in AN. AN is undoubtedly
a painful disorder to live withdso painful, in fact, that for these
afflicted individuals the pain involved with death may no longer
serve to deter suicidal desire.
This study was funded, in part, by National Institute of Mental
E. Joiner) and byan NIMH grant F31MH083382 to A. R. Smith (under
the sponsorship of T. E. Joiner), as well as the following National
Institutes of Health Grants (MH066122, MH066117, MH066145,
MH066288, MH066146). The content of this paper is solely the
responsibility of the authors and does not necessarily represent
the official views of the National Institute of Mental Health or the
National Institutes of Health. The authors thank the study managers
and clinical interviewers of the Genetics of Anorexia Nervosa
Collaboration for their efforts in participant screening and clinical
assessments. Genotyping services for the Genetics of Anorexia
Nervosa Collaboration were provided by the Center for Inherited
Disease Research (CIDR). CIDR is fully funded through a federal
contract from the National Institutes of Health toThe Johns Hopkins
receives support from the Franklin Mint Chair in Eating Disorders.
The authors are indebted to the participating families for their
contribution of time and effort in support of this study.
American Psychiatric Association. (1994). Diagnostic and statistical manual of mental
disorders (4th ed.). Washington, D.C.: Author.
Anderson, T. W. (1957). Maximum likelihood estimates for a multivariate normal
distributions when some observations are missing. Journal of American Statis-
tical Association, 52, 200e203.
Arubckle, J. L., & Wothke, W. (1999). AMOS 4.0. Chicago, IL: SPSS Inc.
Baker, E. H., & Sandle, G. I. (1996). Complications of laxative abuse. Annual Review of
Medicine, 47, 127e134.
Bollen, K. A. (1987). Total, direct, and indirect effects in structural equation models.
Sociological Methodology, 17, 37e69.
Braun, D. L. (1999). More males seek treatment for eating disorders. International
Journal of Eating Disorders, 25, 415e424.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
Bulik, C. M., Thornton, L., Pinheiro, A. P., Plotnicov, K., KIump, K. L., Brandt, H., et al.
Carlat, D. J., Carmargo, C. A., & Herzog, D. B. (1997). Eating disorders in males:
a report of 135 patients. American Journal of Psychiatry, 15, 1127e1132.
Claes, L., Vandereycken, W., & Vertommen, H. (2006). Pain experience related to
self-injury in eating disorder patients. Eating Behaviors, 7, 204e213.
Crisp, A. H., Callender, J. S., Halek, C., & Hus, L. K. (1992). Long-term mortality in
anorexia nervosa. A 20-year follow-up of the St George’s and Aberdeen cohorts.
The British Journal of Psychiatry, 161, 104e107.
De la Rie, S. M., van Furth, E. F., De Koning, A., Noordenbos, G., & Donker, M. C.
(2005). The quality of life of family caregivers of eating disorder patients. Eating
Disorders, 13, 345e351.
De la Rie, S. M., Noordenbos, G., Donker, M., & van Furth, E. F. (2007). The patient’s
view on quality of life and eating disorders. International Journal of Eating
Disorders, 40, 13e20.
Dillon, W. R., Kumar, A., & Mulani, N. (1987). Offending estimates in covariance
structure analysis: comments on the causes of and solutions to Heywood cases.
Psychological Bulletin, 101(1), 126e135.
Duberstein, P. R., & Conwell, Y. (1997). Personality disorders and completed suicide:
a methodological and conceptual review. Clinical Psychology: Science and Prac-
tice, 4, 359e376.
Engelberg, H. (1992). Low serum cholesterol and suicide. Lancet, 339, 727e729.
Eddy, K. T., Keel, P. K., Dorer, D. J., Delinsky, S. S., Franko, D. L., & Herzog, D. B. (2001).
Longitudinal comparison of anorexia nervosa subtypes. International Journal of
Eating Disorders, 31, 191e201.
Fairburn, C. G., & Cooper, Z. (1993). The eating disorder examination. In
C. G. Fairburn, & G. T. Wilson (Eds.), Binge eating: Nature, assessment and
treatment (12th ed.). (pp. 317e360) New York: Guilford Press.
Favaro, A., Caregaro, L., Di Pascoli, L., Brambilla, F., & Santonastaso, P. (2004). Total
serum cholesterol and suicidality in anorexia nervosa. Psychosomatic Medicine,
Favaro, A., & Santonastaso, P. (1997). Suicidality in eating disorders: clinical and
psychological correlates. Acta Psychiatrica Scandinavica, 95, 508e514.
Fichter, M., Herpetz, S., Quadflieg, N., & Herpetz-Dahlmann, B. (1998). Structured
interview for anorexic and bulimic disorders for DSM-IV and ICD-10: updated
(third) revision. International Journal of Eating Disorders, 24, 227e249.
First, M. B., Spitzer, R. L., Gibbon, M., & Williams, J. B. W. (1997). Structured clinical
interview for DSM-IV axis I disorders. Research Version, Non-Patient Edition
(SCID-I/NP). New York: Biometrics Research Department, New York State
First, M. B., Spitzer, R. L., Gibbon, M., Williams, J. B. W., & Benjamin, L. (1997).
Structured Clinical Interview for DSM-IV Personality Disorders (SCID-II). Wash-
ington, D.C.: American Psychiatric Publishing.
Foulon, C., Guelfi, J. D., Kimpan, A., Ades, J., Romo, L., Houdeyer, K., et al. (2007).
Switching to the bingeing/purging subtype of anorexia nervosa is frequently
associated with suicidal attempts. European Psychiatry, 22, 513e519.
Franko, D. L., & Keel, P. K. (2006). Suicidality in eating disorders: occurrence,
correlates,and clinical implications.
Gendall, K., Joyce, P., Carter, F., McIntosh, V., Jordan, J., & Bulik, C. (2006). The
psychobiology and diagnostic significance of amenorrhea in patients with
anorexia nervosa. Fertility and Sterility, 85, 1531e1535.
Harris, E. C., & Barraclough, B. (1997). Suicide as an outcome for mental disorders:
a meta-analysis. The British Journal of Psychiatry, 170, 205e228.
Hebebrand, J., Himmelmann, G. W., Herzog, W., Herpertz-Dahlmann, B. M.,
Steinhausen, H. C., Amstein, M., et al. (1997). Prediction of low body weight at
long-term follow-up in acute anorexia nervosa by low body weight at referral.
American Journal of Psychiatry, 154, 566e569.
Hebebrand, J., Himmelmann, G. W., Heseker, H., Schafer, H., & Remschmidt, H.
(1996). Use of percentiles for the body mass index in anorexia nervosa: diag-
nostic, epidemiological, and therapeutic considerations. International Journal of
Eating Disorders, 19, 359e369.
Hellstrom, I. (2007). Oral complications in anorexia nervosa. European Journal of
Oral Sciences, 85, 71e86.
Hillege, S., Beale, B., & McMaster, R. (2006). Impact of eating disorders on family life:
individual parents’ stories. Journal of Clinical Nursing, 15, 1016e1022.
Holm-Denoma, J. M., Witte, T. K., Gordon, K. H., Herzog, D. B., Franko, D. L.,
Fitcher, M., et al. (2008). Deaths by suicide among individuals with anorexia as
arbiters between competing explanations of the anorexiaesuicide link. Journal
of Affective Disorders, 107, 231e236.
Howard, W. T., Evans, K. K., Quintero-Howard, C. V., Bowers, W. A., & Andersen, A. E.
(1999). Predictors of success or failure of transition to day hospital treatment
for inpatients with anorexia nervosa. American Journal of Psychiatry, 156,
Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure
analysis: conventional criteria versus new alternatives. Structural Equation
Modeling, 6(1), 1e55.
Joiner, T. E., Jr., Sachs-Ericsson, N. J., Wingate, L. R., Brown, J., Anestis, M. D., &
Selby, E. A. (2007). Childhood physical and sexual abuse and lifetime number of
suicide attempts: A persistent and theoretically important relationship.
Behaviour Research and Therapy, 45(3), 539e547.
Joiner, T. E., Jr., Van Orden, K. A., Witte, T. K., & Rudd, M. D. (2009). The interpersonal
theory of suicide: Guidance for working with suicidal clients. Washington, DC:
American Psychological Association.
Kaye, W. (2008). Neurobiology of anorexia and bulimia nervosa. Physiology and
Behavior, 94, 121e135.
Kaye, W. H., Plotnicov, K., Thornton, L., Devlin, B., Fichter, M., Treasure, J., et al.
(2008). The Geneticsof Anorexia Nervosa (GAN)
methods and sample description. International Journal of Eating Disorders, 41,
Keel, P. K., Dorer, D. J., Eddy, K. T., Franko, D., Charatan, D. A., et al. (2003).
Predictors of mortality in eating disorders. Archives of General Psychiatry, 60,
Klump, K. L., Strober, M., Bulik, C. M., Thornton, L., Johnson, C., Devlin, B., et al.
(2004). Personality characteristics of women before and after recovery from an
eating disorder. Psychological Medicine, 34, 1407e1418.
Lautenbacher, S., Pauls, A., Strian, F., Pirke, K., & Krieg, J. (1991). Pain sensitivity
in anorexia nervosa and bulimia nervosa. Biological Psychiatry, 29(11),
Lieb, K., Zanarini, M. C., Schmahl, C., Linehan, M., & Bohus, M. (2004). Borderline
personality disorder. Lancet, 364, 453e461.
Little, R. J. A., & Rubin, D. B. (1987). Statistical analysis with missing data. New York:
McIntosh, J. L. (2002). U.S.A. suicide statistics for the year 1999: Overheads and
a presentation guide. Washington, D.C.: American Association of Suicidology.
MacKinnon, D. P., Fritz, M. S., Williams, J., & Lockwood, C. M. (2007). Distribution of
the product confidence limits for the indirect effect: program PRODCLIN.
Behavior Research Methods, 39(3), 384e389.
MacKinnon, D. P., Lockwood, C. M., Hoffman, J. M., West, S. G., & Sheets, V. (2002).
A comparison of methods to test mediation and other intervening variable
effects. Psychological Methods, 7(1), 83e104.
Maser, J. D., Akiskal, H. S., Schettler, P., Scheftner, W., Mueller, T., Endicott, J., et al.
(2002). Can temperament identify affectively ill patients who engage in lethal
or near-lethal suicidal behavior? A 14-year prospective study. Suicide & Life-
Threatening Behavior, 32, 10e32.
Norring, C. E., & Sohlberg, S. S. (1993). Outcome, recovery, relapse and mortality
across six years in patients with clinical eating disorders. Acta Psychiatrica
Scandinavica, 87, 437e444.
Nurnberger, J. I., Blehar, M. C., Kaufmann, C. A., York-Cooler, C., Simpson, S. G.,
Karakavy-Owens, M. J., et al. (1994). Role of serotonininthe pathophysiology
of depression: focus on the serotonin transporter. Clinical Chemistry, 40,
Papadopoulos, F. C., Ekbom, A., Brandt, L., & Ekselius, L. (2009). Excess mortality,
causes of death and prognostic factors in anorexia nervosa. British Journal of
Psychiatry, 194(1), 7e10.
Pape? zova, H., Yamamotova, A., & Uher, R. (2005). Elevated pain threshold in eating
disorders: physiological and psychological factors. Journal of Psychiatric
Research, 39, 431e438.
Pinheiro, A., Thornton, L., Plotnicov, K., Tozzi, T., Klump, K., Berrettini, W., et al.
(2007). Patterns of menstrual disturbance in eating disorders. International
Journal of Eating Disorders, 40(5), 424e434.
Preisig, M., Fenton, B. T., Matthey, M. L., Berney, A., & Ferrero, F. (1999). Diagnostic
Interview for Genetic Studies (DIGS): inter-rater and testeretest reliability of
the French version. European Archives of Psychiatry and Clinical Neuroscience,
Project EHE. (2001). Eatatelife phenotype interview. Version 2.1.
Purselle, D. C., & Nemeroff, C. B. (2003). Serotonin transporter: A potential substrate
in the biology of suicide. Neuropsychopharmacology, 28(4), 613e619.
Raymond, N. C., de Zwaan, M., Faris, P. L., Nugent, S. M., Ackard, D. M., Crosby, R. D.,
et al. (1995). Pain thresholds in obese binge-eating disorder subjects. Biological
Psychiatry, 37, 202e204.
Roca, M., Martin-Santos, R., Raiz, J., Obiols, J., Serrano, M. J., Torrens, M., et al.
(2007). Diagnostic Interview for Genetic Studies (DIGS): inter-rater and
testeretest reliability and validity in a Spanish population. European Psychi-
atry, 22, 44e48.
Schafer, J. L. (1997). Analysis of incomplete multivariate data. London: Chapman & Hall.
Selby, E. A., Anestis, M. D., Bender, T. W., Ribeiro, J. D., Nock, M. K., Rudd, M. D., et al.
(2010). Overcoming the fear of lethal injury: evaluating suicidal behavior in the
military through the lens of the interpersonal-psychological theory of suicide.
Clinical Psychology Review, 30, 298e307.
Selby, E. A., Anestis, M. D., & Joiner, T. E., Jr. (2008). Understanding the relationship
between emotional and behavioral dysregulation: emotional cascades. Behav-
iour Research and Therapy, 46, 593e611.
Sidiropoulos, M. (2007). Anorexia nervosa: the physiological consequences of
starvation and the need for primary prevention efforts. McGill Journal of
Medicine, 10, 20e25.
Simon, J., Schmidt, U., & Pilling, S. (2005). The health service use and cost of eating
disorders. Psychological Medicine, 35, 1543e1551.
Shneidman, E. S. (1996). The suicidal mind. Oxford: Oxford University Press.
Smith, A. R., Witte, T. K., Teale, N. E., King, S. L., Bender, T. W., & Joiner, T. E. (2008).
Revisiting impulsivity in suicide: implications for civil liability of third parties.
Behavioral Sciences and the Law, 26(6), 779e797.
Sullivan, P. F. (2002). Course and outcome of anorexia nervosa and bulimia nervosa.
In C. G. Fairburn, & K. D. Brownell (Eds.), Eating disorders and obesity: A
comprehensive handbook (2nd ed.). (pp. 226e232) New York: The Guilford
Tiller, J. M., Gaynor, S., Schmidt, U., Troop, N., Power, M., & Treasure, J. L. (1997).
Social support in patients with anorexia nervosa and bulimia nervosa. Inter-
national Journal of Eating Disorders, 21, 31e38.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645
Tozzi, F., Thornton, L. M., Mitchell, J., Fichter, M. M., Klump, K. L., Lilenfeld, L. R., et al., Download full-text
, Collaborative Group. (2006). Features associated with laxative abuse in indi-
viduals with eating disorders. Psychosomatic Medicine, 68, 470e477.
Van Orden, K. A., Witte, T. K., Cukrowicz, K. C., Braithwaite, S., Selby, E. A., & Joiner, T.
E., Jr. The interpersonal theory of suicide. Psychological Review, in press.
Van Orden, K. A., Witte, T. K., Gordon, K. H., Bender, T. W., & Joiner, T. E., Jr. (2008).
Suicidal desire and the capability for suicide: tests of the interpersonal-
psychological theory of suicidal behavior among adults. Journal of Consulting
and Clinical Psychology, 76, 72e83.
Witte, T. K., Merrill, K. A., Stellrect, N. E., Bernert, R. A., Hollar, D. L.,
Schatschneider, C., et al. (2008). “Impulsive” youth suicide attempters are not
necessarily all that impulsive. Journal of Affective Disorders, 107, 107e116.
Wonderlich, S. A., Connolly, K. M., & Stice, E. (2004). Impulsivity as a risk factor for
eating disorder behavior: assessment implications with adolescents. Interna-
tional Journal of Eating Disorders, 36, 172e182.
Wonderlich, S. A., Swift, W. J., Slomik, H. B., & Goodman, S. (1990). DSM-III-R
personality disorders in eating disorder subtypes. International Journal of Eating
Disorders, 9, 607e616.
E.A. Selby et al. / Behaviour Research and Therapy 48 (2010) 634e645