Affect Regulation and Purging: An Ecological Momentary Assessment Study in Purging Disorder

Abstract

Research suggests that affect may play an important role in the propensity to purge among women with Purging Disorder (PD). However, prior work has been constrained to cross-sectional or laboratory designs which impact temporal interpretations and ecological validity. This study examined the role of negative affect (NA) and positive affect (PA) in triggering and maintaining purging in PD using ecological momentary assessment. Women with PD (N=24) made multiple daily ratings of affect and behavior for two weeks. Multilevel models examined associations between affect and purging at different levels of analysis, including a novel analytic approach to address the specificity of changes in affect relative to purging behavior by comparing trajectories of change on purge versus non-purge days. For trajectories of affect over time, NA increased before purging and decreased following purging on purge days; however, only the decrease in NA following purging was significantly different from the trajectory of NA on non-purge days. Conversely, PA failed to increase before purging on purge days compared to a matched time-point on non-purge days. These findings suggest unique roles of PA in triggering and NA in maintaining purging in PD and support models in which purging functions to regulate affect. For comparisons of ratings before and after purging, NA increased and PA decreased after purging, highlighting how different analytic strategies produce different findings requiring integration into affect regulation models. These data provide insight into why women with PD purge after consuming normal amounts of food, a crucial first step for developing effective interventions.

Purging disorder (PD) is characterized by recurrent purging (i.e., self-induced vomiting, laxative abuse, diuretic abuse) following consumption of normal or small amounts of food and body image disturbance among women of minimally normal weight (Keel, Haedt, & Edler, 2005). Similar to Anorexia Nervosa (AN) and Bulimia Nervosa (BN), PD involves extreme efforts to control weight coupled with the undue influence of weight or shape on self-evaluation. However, in contrast to AN, purging occurs among individuals who are not underweight, and, in contrast to BN, purging occurs in the absence of objectively large binge episodes. Previous research supports the clinical significance (e.g., Keel et al., 2005; Keel, Wolfe, Gravener, & Jimerson, 2008; Mond et al., 2006; Wade, Bergin, Tiggemann, Bulik, & Fairburn, 2006) and distinctiveness of PD (e.g., Binford & le Grange, 2005; Keel et al., 2005; Keel, Mayer, & Harnden-Fischer, 2001; Keel, Wolfe, Liddle, De Young, & Jimerson, 2007; Pinheiro, Bulik, Sullivan, & Machado, 2008; Sullivan, Bulik, & Kendler, 1998; Wade, 2007; Wade et al., 2006), and suggests that the lifetime prevalence of PD is comparable to that of AN and BN (Favaro, Ferrara, & Santonastaso, 2003; Wade et al., 2006). Reflecting this growing body of literature, the DSM-5 has included PD as a named syndrome among Other Specified Feeding or Eating Disorders (American Psychiatric Association, 2013). The primary rationale for not including PD among formal diagnoses was the absence of data regarding treatment or course of PD to ensure predictive validity and clinical utility of the diagnosis (Brown, Keel, & Striegel, 2012). However, prior to the development of efficacious treatments, more research is needed to understand factors that contribute to and maintain purging behavior in PD. This study sought to test an affect regulation model of purging in PD.

Very little is known about factors which serve to trigger purging in PD. The vast majority of work on purging has been conducted in the context of BN where predominant theoretical models focus on binge eating as the primary trigger for purging behavior. According to cognitive-behavioral theory proposed by Fairburn, compensatory purging develops as a way to minimize weight gain associated with large binge-eating episodes (Fairburn, 2008). As such, this explanation does not explain purging in PD or even purging episodes in AN or BN when they are not preceded by binge eating. Another line of work in BN has focused on changes in affect as an antecedent that triggers behavior and a consequence that maintains behavior to understand both binge eating and purging (Alpers & Tuschen-Caffier, 2001; Corstorphine, Waller, Ohanian, & Baker, 2006; Davis, Freeman, & Solyom, 1985; Elmore & de Castro, 1990; Kaye, Gwirtsman, George, Weiss, & Jimerson, 1986; Powell & Thelen, 1996; Rebert, Stanton, & Schwarz, 1991; Schlundt, Johnson, & Jarrell, 1986; Smyth et al., 2007). Based on evidence that purging in BN is associated with reductions in negative affect (Haedt-Matt & Keel, 2011), changes in affect represent a prime candidate for understanding purging behavior in PD.

Negative affect is a risk factor for general eating pathology (Stice, 2002), and one of the most commonly cited explanations for disordered eating emphasizes the role of regulating emotional distress or negative affect. Previous research suggests that emotional factors are relevant to purging in the absence of binge eating. Cross-sectional studies (Keel et al., 2005; Keel et al., 2008; Wade et al., 2006) support elevated negative affect in PD relative to non-eating disordered controls, and laboratory studies support that food intake triggers significant increases in negative affect in women with PD compared to controls despite no differences in the amount of food consumed (Keel et al., under review; Keel et al., 2007). However, cross-sectional designs cannot establish temporal sequences of changes in affect and changes in purging necessary to establish affect as an antecedent or consequence of purging behavior, and laboratory studies have limited ecological validity. Ecological momentary assessment (EMA; Stone & Shiffman, 1994) involves repeated assessments as participants go about their daily lives and has been increasingly used to assess dynamic changes in affect related to disordered eating (Berg et al., 2013; Engel et al., 2013; Haedt-Matt & Keel, 2011). Recent findings from EMA studies of AN and BN provide important clues regarding affective influences on purging that may extend to PD.

Ecological Momentary Assessment Studies of Purging and Affect in AN and BN

EMA research has found that increased negative affect predicts purging in AN (Engel et al., 2013) and BN (Alpers & Tuschen-Caffier, 2001; Rebert et al., 1991; Schlundt et al., 1986; Smyth et al., 2007). In addition, negative affect increased prior to purging whether or not purging was preceded by objective binge eating in BN (Smyth et al., 2007) or subjective, loss of control eating in AN (Engel et al., 2013). Similarly, EMA studies in AN (Engel et al., 2013) and BN (Smyth et al., 2007) found that positive affect decreased prior to self-induced vomiting/purging. Within AN, this trajectory was not influenced by presence of loss of control eating prior to purging behaviors (Engel et al., 2013). Taken together, previous research points to a potential role of increases in negative affect and decreases in positive affect in triggering purging in PD, which may be used to regulate emotional distress.

In addition to antecedent triggers, evaluating the affective consequences of purging is important for understanding their function, as either reductions in negative affect or increases in positive affect would reinforce purging behavior. Partially supporting an affect regulation model of purging, negative affect decreased from pre- to post-purge in EMA studies of BN (Alpers & Tuschen-Caffier, 2001; Corstorphine et al., 2006; Davis et al., 1985; Elmore & de Castro, 1990; Kaye et al., 1986; Powell & Thelen, 1996). A recent meta-analysis of these studies supports that purging may reduce negative affect induced by binge eating (Haedt-Matt & Keel, 2011). Notably, negative affect levels following purging did not differ significantly from affect before the binge-eating episode which precipitated purging in BN (Alpers & Tuschen-Caffier, 2001; Corstorphine et al., 2006; Powell & Thelen, 1996). In addition, negative affect increased from pre- to post-purge in a recent study of AN (Engel et al., 2013), questioning the ability to generalize findings from one eating disorder to another. Finally, EMA studies support that positive mood increases following purging in AN (Engel et al., 2013) and BN (Smyth et al., 2007). Thus, purging may function to regulate affect through decreases in negative affect or increases in positive affect. However, it is unknown how purging influences affect in PD.

Limitations of Previous Research

EMA studies of affective antecedents and consequences of purging are limited to studies of participants with AN or BN, with the vast majority of studies focusing on BN. Given that previous research supports important psychological and biological differences between BN and PD (Keel, 2007), there may be important differences between factors contributing to purging in BN versus purging in PD. Thus, no previous study has adequately tested the role of affect as an antecedent or consequence of purging in PD as there have been no EMA studies of PD. Comparisons of proximal affect ratings before and after purging for AN (Engel et al., 2013) do not mirror those for BN (Alpers & Tuschen-Caffier, 2001; Corstorphine et al., 2006; Davis et al., 1985; Elmore & de Castro, 1990; Kaye et al., 1986; Powell & Thelen, 1996), underscoring the need for more work to understand the role of negative affect in purging across different syndromes and using different data analytic strategies.

While several EMA studies have examined negative affect in relation to disordered eating behaviors, few EMA studies have examined the role of changes in positive affect (Haedt-Matt & Keel, 2011) and even fewer have examined positive affect as a consequence of purging in eating disorders (Engel et al., 2013; Smyth et al., 2007). Although related, negative and positive affect represent relatively independent constructs that have demonstrated differential patterns of associations with psychopathology (Watson, Clark, & Carey, 1988). Thus, it is important to investigate changes in both negative and positive affect related to purging behavior in PD.

Finally, previous EMA studies have analyzed within-day trajectories of affect before and after a behavior. Based on these trajectories, conclusions have been drawn regarding the role of increases in negative affect in triggering and maintaining binge eating and purging episodes. However, this analytic approach cannot provide information regarding the specificity of changes in affect and the presence/absence of a particular behavior, such as purging. Individuals with eating disorders report a great deal of affective variability (Anestis et al., 2009; Anestis et al., 2010; Benjamin & Wulfert, 2005) and may experience changes in affect over time on days that they do not purge. Supporting this, previous EMA research in BN suggests significant diurnal variation in affect (Smyth et al., 2009). Thus, temporal changes in affect on purge days may be due to extraneous third variables, such as diurnal variation or work/school schedules. Empirically examining whether the trajectory of affect is specific to the presence of purging would provide additional information regarding the role of affect in purging behavior.

The Present Study and Hypotheses

To address each of these limitations, the current study investigated how changes in negative affect and positive affect are related to purging behavior in PD using EMA. The EMA design allowed for the examination of several different aspects of the association between affect and purging. Comparisons were made between days when participants purged versus days when participants did not purge, and we hypothesized greater mean levels of negative affect and lower positive affect on days characterized by purging (Hypothesis 1). We then examined temporal associations between changes in affect prior to and following purging within purge days. We hypothesized that negative affect would increase and positive affect would decrease prior to purging, representing antecedent triggers of this behavior (Hypothesis 2a). Using a novel analytic strategy to determine whether changes in affect were specifically linked to the presence of purging, we compared these trajectories to time-matched trajectories on non-purge days as a within-subject control and expected that posited changes described in Hypothesis 2a would be significantly greater on purge days compared to non-purge days (Hypothesis 2b). As proposed consequences of purging, we hypothesized that negative affect would decrease and positive affect would increase following purging (Hypothesis 3a) and that these trajectories would differ significantly on purge days versus days in which participants did not purge (Hypothesis 3b). Another approach to investigating the consequences of purging behavior was to compare affect before and after purging, and we hypothesized that negative affect would be lower and positive affect would be higher after purging compared to before purging (Hypothesis 3c) providing further support for an affect regulation model of purging in PD.

Method

Participants and Recruitment

Women with PD (N = 24) were recruited from the community in Iowa City, IA and Tallahassee, FL using identical protocols for recruitment, screening, and data collection. Recruitment methods included posters, advertisements, and university mass-emails. If potential participants appeared to be eligible following confidential telephone screens, the caller was invited to participate in an in-person intake assessment that included semi-structured clinical interviews to confirm inclusion/exclusion criteria. Inclusion criteria were: 1) female, 2) age 18–45 years, 3) purging (i.e., self-induced vomiting, laxative abuse, and/or diuretic abuse) at least twice per week for the previous three months, and 4) undue influence of body shape or weight on self-evaluation. Exclusion criteria were: 1) any objectively large binge episodes within the previous 12 months, 2) underweight (i.e., body mass index < 18.5 kg/m2), 3) psychotic disorder, and 4) inability to read English. Criteria for PD are more stringent than research criteria proposed by Keel and Striegel-Moore (2009) to ensure enough instances of purging during data collection to provide detection of reliable temporal patterns.

Of the 32 women invited to complete the intake assessment based on their initial telephone screen, 7 were ineligible due to endorsement of objectively large binge eating episodes during the intake assessment and 25 began daily assessments. Once identified, participants were highly likely to complete study procedures; only one participant withdrew after her first two days of daily assessments (96% retention rate). The remaining 24 participants completed a mean (SD) 16.79 (3.26) days of EMA and are included in data analyses. The average number of days exceeds study requirements because several women continued to participate when unable to schedule their final assessment on day 16.

Participants were predominantly young adult women (mean (SD) age = 23.08 (5.44) years) who were normal weight (mean (SD) body mass index = 21.99 (2.82) kg/m²). Ethnic/racial identification was 87.5% Caucasian, 4.2% Hispanic, 4.2% Black, and 4.2% Asian. Although this was a community sample, 41.7% (n = 10) of participants reported that they were in current psychological treatment, and 66.7% (n = 16) reported a lifetime history of psychological treatment. Three participants (16.7%) reported a lifetime history of AN, and one participant reported a history of AN and BN (8.3%). Overall, 20 participants (83%) had no history of another eating disorder. Participants endorsed a range of comorbid psychopathology, including current (n = 4, 16.7%) and lifetime (n = 16, 66.7%) mood disorders, current (n = 5, 20.8%) and lifetime (n = 10, 41.7%) anxiety disorders, lifetime substance use disorders (n = 10, 41.7%), and lifetime impulse control disorders (n = 2, 8.3%). No participants met criteria for a current substance use or impulse control disorder.

Procedure and Measures

This research study was reviewed and approved by institutional review boards at the University of Iowa and Florida State University. Participants provided written informed consent prior to study participation and were asked to complete the following assessments, including four visits to the research lab: 1) intake assessment including interviews, questionnaires, and training on EMA procedures (study visit 1), 2) daily assessments of purging and affect, 3) intermediate phone and two in-person assessment check-ins (study visits 2 and 3), and 4) final assessment including evaluation of changes in eating disorder symptoms (study visit 4). Participants were offered $50 for the intake assessment, $175 for daily assessments prorated according to degree of response to signals and end-of-day ratings (full payment for completing 85%-100% of random signals), and $25 for the final assessment.

Intake assessment

Height and weight were objectively measured using a digital scale and wall-mounted ruler. In addition, semi-structured clinical interviews of eating and related Axis I disorders were conducted to confirm inclusion/exclusion criteria. Finally, participants received detailed instructions for completing daily assessments on handheld computers and how to deal with problems or questions that might arise.

Eating Disorders Examination (EDE; Fairburn & Cooper, 1993)

This semi-structured clinical interview assesses frequency of disordered eating behaviors and specific features of eating disorders and was used to confirm study eligibility (i.e., inclusion criteria for current PD). The main advantage of the EDE was the inclusion of questions to distinguish objectively large versus subjective binge episodes. Previous research supports the reliability and validity of the EDE (Berg, Peterson, Frazier, & Crow, 2012), and the EDE is considered the “gold standard” in eating disorder assessment (Grilo, 2005). Interrater reliability was examined by randomly selecting five interviews (21% of sample) that were rated by an independent assessor. There was 100% agreement for diagnoses of current PD.

Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I; First, Spitzer, Gibbon, & Williams, 1996)

This structured interview was used to screen for exclusion criteria (i.e., psychotic disorders module). In addition, the SCID-I overview and diagnostic modules were used to characterize the sample for comparisons with previous research. The SCID-I has demonstrated good to excellent test-retest and interrater reliabilities in a large, multi-site study (Zanarini et al., 2000). There was 100% agreement for current and lifetime diagnoses of mood, anxiety, eating, and impulse control disorders, 80% agreement for current substance use disorders, and 60% agreement for lifetime substance use disorders.

Daily assessments

Participants carried handheld computers and completed assessments of purging behavior and affect for 16 days, including two practice days and 14 days of data collection. To minimize concerns about potential reactivity, data collected during the initial two-day practice period were not included in data analyses, consistent with methods used in other EMA studies (Engel et al., 2013; Smyth et al., 2007). In addition, reactivity in the current study was examined by comparing EDE interview assessments of eating disorder attitudes and behaviors before and after EMA monitoring. There were no mean differences between pre- and post-monitoring assessments of eating disorder attitudes or purging frequency and correlations between pre- and post-monitoring assessments were comparable to studies of test-retest reliability over a similar two-week time frame, reducing concerns about reactivity to EMA (Haedt-Matt & Keel, 2013).

This study included three types of assessments (Wheeler & Reis, 1991) implemented in previous studies of AN (Engel et al., 2013) and BN (Smyth et al., 2007). First, participants were signaled at six semi-random times throughout the day (signal-contingent) to complete momentary ratings of affect and to report any purging (self-induced vomiting, laxative abuse, or diuretic abuse) that had not been previously recorded. The time of signals was determined by randomly selecting times within six equally spaced intervals between 8:00am and 11:00pm to ensure representative sampling throughout the day. In addition, participants were instructed to complete momentary ratings of affect as soon after each purging episode as possible (event-contingent). These event-contingent ratings allowed the assessment of immediate consequences of purging that may have been missed by random signals. Finally, participants completed ratings of affect at the end of each day to capture any changes since the last signal (end-of-day).

Purging

Participants reported any purging behaviors as well as how long ago purging occurred. Self-reports of purging behaviors have demonstrated high test-retest reliability (Peterson, Miller, Johnson-Lind, Crow, & Thuras, 2007) and high agreement with interview-based assessments, likely because these behaviors are salient and questions regarding these behaviors are less susceptible to misinterpretation than questions about other eating disorder behaviors (e.g., binge eating; Fairburn & Beglin, 1990; Stein & Corte, 2003). Exploratory analyses also were conducted to examine the concurrent validity of EMA (14 days momentary assessment) and EDE data from the final assessment (retrospectively reported for the same 14 day period). There were significant, positive correlations between EDE and EMA for purging frequency (r = .82, p < .001), and there were no differences between type of assessment (paired t(19) = .83, p > 0.05). Thus, comparisons of momentary EMA and retrospective EDE data in the current sample provide strong support for the concurrent validity of EMA recordings against interview assessments.

Affect

The Positive and Negative Affect Schedule (PANAS; Watson, Clark, & Tellegen, 1988) consists of 20 items that assess positive (e.g., excited, proud, inspired) and negative (e.g., anger, sadness, nervousness) emotions. The PANAS was designed to assess mood over different time periods, including momentary ratings. Thus, this scale has been validated for the assessment of state mood required for momentary assessments in the current study. The PANAS has demonstrated excellent internal consistency and good convergent and discriminant validity (Watson, Clark, & Tellegen, 1988). In the current study, internal consistency (Cronbach’s alpha) was .84 for negative affect and .89 for positive affect.

Intermediate assessment check-ins

Participants returned to the research lab after the first two days of EMA, were given feedback regarding their compliance rates, and any questions or concerns regarding assessment procedures were discussed with a research assistant. Participants returned one week later to upload EMA data mid-way through the assessment phase. Throughout the 14-day assessment period, participants were telephoned at least twice per week to check in with them regarding any problems or concerns.

Statistical Analyses

Preliminary data considerations

Baseline comparisons between sites were made on demographic, eating pathology, and general psychopathology variables, and site was evaluated for use as a covariate in analyses. Prior to analyses, data were examined for normality. Raw scores for negative affect and positive affect were log-transformed to correct for significant positive skew and multiplied by 10 to avoid boundary constraints (Singer & Willett, 2003).

Hypothesis testing

Multilevel model (MLM) analyses were used to test study hypotheses that 1) negative affect would be higher and positive affect would be lower on purge days compared to non-purge days, 2) negative affect would increase and positive affect would decrease prior to purging, and 3) negative affect would decrease and positive affect would increase following purging. MLM is superior to alternative analytical methods, such as repeated measures ANOVA, because of its ability to handle correlated within-person data with unequal variances in unbalanced designs. Data collected using EMA are unlikely to result in the same number of measurements per participant because of differences in purging frequency, differences in participant compliance, and because participants will inevitably miss some random signals due to unavoidable circumstances (e.g., driving) (Schwartz & Stone, 1998). MLM analyses used maximum likelihood estimation methods to include information for participants when data were missing, which is superior to the use of list-wise deletion employed by repeated measures ANOVA.

Applied to EMA data, MLM analyses are ideal to assess momentary ratings (level 1) made by individuals (level 2). The inclusion of random effects for intercept, slope, and their covariation was investigated, and models which specified a random intercept provided the best fit to the data. MLM models further specified a first-order autoregressive covariance structure to model the autocorrelation between within-person random errors because ratings made closer in time were expected to have errors that were more highly correlated compared to ratings made farther apart. Model fit was improved for all analyses using this specification. Full maximum likelihood estimates were used to permit comparisons of model fit. Significance of fixed effects was examined using the t statistic with degrees of freedom equal to N – 1 – number of predictors (Raudenbush & Bryk, 2002) and a significance level of p < .05.

Hypothesis 1: differences between purge and non-purge days

Between-days analyses compared mean levels of affect on purge days versus non-purge days. Each study day was dummy coded to distinguish days when purging occurred (purge day) versus days purging did not (non-purge day). Data were aggregated across within-day assessments so that scores reflect the average values for each participant on each day.

Hypothesis 2a: antecedents of purging on purge days

Within-day analyses examined changes in affect over time in relation to purging behavior. Following recommendations of Singer and Willett (2003), an unconditional means model was first examined to determine if there was significant within-person variance in each dependent variable. Next, unconditional growth models were conducted to examine linear effects of time. Fit indices and the log-likelihood ratio test were used to determine if the addition of linear effects of time improved model fit over the unconditional means model.

Within-person changes in affect prior to purging on purge days were examined using separate general linear mixed models for each dependent variable. The main predictor variable in antecedent analyses was the linear effect of time leading up to purging behavior. Using negative affect as an example, Hypothesis 2a would be supported if there is a significant, positive coefficient for time prior to purging indicating increases in negative affect prior to purging. When multiple purging behaviors were reported on the same day (purging occurred more than once on 51 days [24.4% of purge days]), only ratings made prior to the first purging episode of the day were included in antecedent analyses.¹

Hypothesis 2b: comparisons of antecedent change trajectories between purge and non-purge days

In order to examine whether the trajectory of change differed on purge days compared to non-purge days, an average time of purging was calculated for each participant and momentary ratings on non-purge days were centered relative to average time of purging. For these analyses, predictors included time prior to purging, purge day, and their interaction. Using negative affect as an example, hypothesis 2b would be supported with a significant positive interaction between time prior to purging and purge day, reflecting greater increases in negative affect prior to purging on purge days relative to the trajectory of negative affect prior to average time of purging on non-purge days.

Hypothesis 3a: consequences of purging on purge days

Within-person changes in affect following purging behavior were examined using separate general linear mixed models for each dependent variable. As in antecedent analyses, the predictor variable was the linear effect of time after purging on days in which participants purged. Hypothesis 3a would be supported for negative affect if there is a significant negative slope coefficient for time following purging behavior, indicating decreases in negative affect over time. When multiple purging behaviors were reported on the same day, only ratings made following the last purging episode of the day were included in these analyses.

Hypothesis 3b. comparisons of consequence change trajectories between purge and non-purge days

In order to examine whether the trajectory of change differed on purge days compared to non-purge days, momentary ratings on non-purge days were centered relative to average time of purging and predictors included time after purging, purge day, and their interaction similar to the approach described for Hypothesis 2b. Using negative affect as an example, hypothesis 3b would be supported with a significant negative interaction between time after purging and purge day, reflecting greater decreases in negative affect following purging behavior on purge days relative to the trajectory of negative affect following average time of purging on non-purge days.²

Hypothesis 3c. comparisons of affect before and after purging

General linear mixed models were used to compare affect during the last rating prior to purging and affect during the first rating after purging. Separate models were conducted for each dependent variable. This is consistent with the analytic approach used to compare proximal affect ratings in previous research (Engel et al., 2013; Haedt-Matt & Keel, 2011).

Results

Preliminary Data Analyses

There were 2,445 momentary ratings included in data analyses, including 1,978 signal-contingent ratings, 313 end-of-day ratings, and 154 event-contingent behavior ratings. Compliance with the study protocol was comparable to previous research (Engel et al., 2013; Smyth et al., 2007); 77.9% of all signal and end-of-day ratings were completed. In addition, participants were timely in responding to random signals and completed 71.4% of signal-contingent ratings within 30 minutes. Across participants, there were 194 non-purge days and 209 purge days consisting of 260 purging episodes (two purging episodes were reported on 44 days and three episodes were reported on 7 days). The average time of purging for each participant ranged from 2:15pm to 7:49pm with a mean time of purging of 4:50pm. Within participants, the variability of times of purging ranged from SD = 17 minutes to SD = 7 hours 30 minutes, with an average of 3 hours 20 minutes.³ Participants recruited in Tallahassee, FL reported greater body dissatisfaction, eating disorder-related impairment, and impulsivity, and more frequent purging compared to participants recruited in Iowa City, IA (all ps < .05). Thus, site was included as a covariate in all analyses.

Hypothesis 1: Affective Differences Between Purge and Non-Purge Days

Between-days analyses compared mean levels of affect on purge days versus non-purge days (see Table 1). The estimate for “purge day” reflects the average difference on purge days compared to non-purge days. As hypothesized, participants reported significantly greater negative affect and lower positive affect on purge days compared to days they did not purge.

Table 1.

Between-Days Multilevel Model Analyses

	Negative Affect		Positive Affect
Fixed Effects	Estimate (SE)	t	Estimate (SE)	t
Intercept	11.78 (.30)	39.68***	12.7 (.36)	35.76***
Site	−0.70 (.36)	−1.93	.19 (.47)	.40
Purge Day	.25 (.06)	4.15***	−.19 (.08)	−2.48*
Covariance Parameters	Estimate (SE)	Wald Z	Estimate (SE)	Wald Z
Within-person variance	.61 (.07)	8.57***	.56 (.06)	10.06***
Autocorrelation	.56 (.05)	10.83***	.34 (.06)	5.30***
Between-person intercept	.59 (.21)	2.87**	.94 (.33)	2.85**

^*p < .05,

^**p < .01,

^***p < .001

Note: The estimate for “purge day” reflects the average difference on purge days compared to non-purge days.

Prior to testing Hypotheses 2 and 3, we examined within-day variability for negative affect and positive affect using unconditional means models. Unconditional means models revealed significant within-person variance in both negative affect (variance estimate (SE) = 1.06 (.05), p < .001) and positive affect (variance estimate (SE) = 1.55 (.08), p < .001), indicating significant within-day changes in these variables on days that purging occurred.

Hypotheses 2a and 2b: Affective Antecedents of Purging

Within-day analyses examined trajectories of change prior to purging behavior (Hypothesis 2a). Results from linear growth models are presented in Table 2. The estimate for “hours prior to purging” reflects the rate of change, or slope, of each dependent variable leading up to a purging behavior. As hypothesized, results indicated significant increases in negative affect prior to purging behavior. The linear growth model indicated no significant changes in positive affect.

Table 2.

Within-Days Multilevel Model Analyses: Antecedents of Purging on Purge Days

	Negative Affect			Positive Affect
Fixed Effects	Estimate	SE	t (21)	Estimate	SE	t (16)
Intercept	11.98	.30	40.03***	12.44	.32	38.69***
Site	−.80	.35	−2.24*	.56	.40	1.40
Hours prior to purging	.03	.01	2.68*	.01	.01	1.01
Covariance Parameters	Estimate	SE	Wald Z	Estimate	SE	Wald Z
Within-person variance	.94	.07	12.80***	1.39	.10	13.67***
Autocorrelation	.59	.04	15.73***	.41	.05	8.28***
Between-person intercept	.53	.19	2.79**	.58	.22	2.66**
Model Fit Indices
AIC	1666.76			1696.53
BIC	1693.66			1722.32
−2 Log Likelihood	1654.76			1684.53
Likelihood Ratioa	χ2(2) = 2195.98***			χ2(2) = 2047.05***

^aLikelihood ratio tests reflect improvement in model fit of the linear growth model over the unconditional means model.

^*p < .05,

^**p < .01,

^***p < .001

To test the specificity of these affective trajectories to the occurrence of purging (Hypothesis 2b), additional MLM analyses examined interactions between rates of change prior to time of purging and “purge day” (see Table 3 and Figures 1a and 1b). For negative affect, mean levels were higher on purge days compared to non-purge days, but the trajectory of affect did not differ significantly (see Figure 1a). The trajectory of change in positive affect was significantly different on purge versus non-purge days; positive affect increased over time on non-purge days (estimate (SE) = .14 (.02), p < .001) and failed to change relative to purging on purge days see Figure 1b).

Table 3.

Within-Day Multilevel Model Analyses: Comparisons of Antecedent Growth Trajectories on Purge versus Non-Purge Days

	Negative Affect		Positive Affect
Fixed Effects	Estimate (SE)	t (19)	Estimate (SE)	t (14)
Intercept	11.72 (.29)	40.15***	13.26 (.39)	33.94***
Site	−0.63 (.35)	−1.82	.29 (.49)	.59
Hours prior to purging	.01 (.01)	1.04	.14 (.02)	6.32***
Purge Day	.12 (.12)	1.08	−.56 (.17)	−3.22**
Purge Day * Hours prior to purging	.01 (.02)	.71	−.13 (.03)	−4.73***
Covariance Parameters	Estimate (SE)	Wald Z	Estimate (SE)	Wald Z
Within-person variance	.88 (.05)	18.45***	1.44 (.07)	19.52***
Autocorrelation	.55 (.03)	18.61***	.32 (.04)	8.01***
Between-person intercept	.59 (.18)	3.19**	1.06 (.37)	2.84**

^*p < .05,

^**p < .01,

^***p < .001

Figure 1.

Fitted Linear Trends of Negative Affect (1a.) and Positive Affect (1b.) Prior to Purging.

Hypotheses 3a, 3b, and 3c: Affective Consequences of Purging

Results from linear growth models for analyses of the consequences of purging are presented in Table 4 (Hypothesis 3a). As hypothesized, results indicated significant decreases in negative affect following purging behavior. There were no changes in positive affect over time after purging.

Table 4.

Within-Days Multilevel Model Analyses: Consequences of Purging on Purge Days

	Negative Affect			Positive Affect
Fixed Effects	Estimate	SE	t (21)	Estimate	SE	t (16)
Intercept	12.43	.36	34.40***	11.90	.30	40.39***
Site	−.84	.44	−1.93	.42	.37	1.11
Hours after purging	−.06	.01	−4.62***	−.03	.02	−1.45
Covariance Parameters	Estimate	SE	Wald Z	Estimate	SE	Wald Z
Within-person variance	1.03	.07	13.76***	1.46	.11	12.85***
Autocorrelation	.52	.04	13.00***	.50	.04	11.54***
Between-person intercept	.90	.29	3.09**	.51	.21	2.39*
Model Fit Indices
AIC	1692.47			1590.99
BIC	1718.97			1616.42
−2 Log Likelihood	1680.47			1578.99
Likelihood Ratioa	χ2(2) = 2170.27***			χ2(2) = 2153.29***

^aLikelihood ratio tests reflect improvement in model fit of the linear growth model over the unconditional means model.

^*p < .05,

^**p < .01,

^***p < .001

Within-day MLMs examining comparisons of growth trajectories after purging on purge days versus after average time of purging on non-purge days are presented in Table 5 (and depicted in Figures 2a and 2b; Hypothesis 3b). For these interaction models, results indicated that presence of purging (purge day) influenced the trajectory of change in both negative affect and positive affect. The negative valence of the interaction coefficient for negative affect indicated larger decreases in negative affect over time following purging behavior on purge days relative to average time of purging on non-purge days (see Figure 2a), consistent with hypotheses emerging from an affect regulation model for purging. The positive interaction coefficient for positive affect indicated smaller decreases in positive affect following purging on purge days (see Figure 2b).⁴

Table 5.

Within-Day Multilevel Model Analyses: Comparisons of Consequence Growth Trajectories on Purge versus Non-Purge Days

	Negative Affect		Positive Affect
Fixed Effects	Estimate (SE)	t (19)	Estimate (SE)	t (14)
Intercept	11.62 (.34)	34.85***	13.24 (.34)	39.12***
Site	−.62 (.39)	−1.57	−.12 (.42)	−.30
Hours after purging	.02 (.02)	1.39	−.15 (.02)	−6.26***
Purge Day	.68 (.12)	5.70***	−.98 (.16)	−6.02***
Purge Day * Hours after purging	−.09 (.02)	−4.07***	.12 (.03)	4.06***
Covariance Parameters	Estimate (SE)	Wald Z	Estimate (SE)	Wald Z
Within-person variance	1.01 (.05)	18.86***	1.48 (.08)	17.62***
Autocorrelation	.52 (.03)	18.06***	.47 (.03)	13.88***
Between-person intercept	.78 (.24)	3.26**	.73 (.26)	2.77**

^*p < .05,

^**p < .01,

^***p < .001

Figure 2.

Fitted Linear Trends of Negative Affect (2a.) and Positive Affect (2b.) After Purging.

Comparisons of affect ratings before and after purging (Hypothesis 3c) indicated that negative affect reported during the first rating after purging (M (SD) = 12.75 (1.47)) was significantly higher than negative affect during the last rating before purging (M (SD) = 11.40 (1.36), p < .001). Positive affect reported during the first rating after purging (M (SD) = 12.21 (1.34)) was significantly lower than positive affect during the last rating before purging (M (SD) = 12.75 (1.47), p < .001). However, ratings after purging were made an average of .52 (SD = .92) hours after the behavior, which was significantly more proximal to purging compared to ratings prior to purging (M (SD) = 1.90 (1.61) hours; p < .001).

Discussion

This study sought to further understand psychological factors that contribute to purging in PD, a newly named Other Specified Feeding or Eating Disorder (American Psychiatric Association, 2013), by using EMA to examine associations between affect and purging behavior in PD at different levels of analysis. Consistent with study hypotheses regarding mean level differences, negative affect was higher and positive affect was lower on purge days compared to non-purge days. Further, this study provided robust evidence that dynamic changes in affect over time at the momentary level are related to purging behavior in PD. Trajectory analyses supported the absence of increases in positive affect as an antecedent trigger of purging behavior in PD whereas decreases in negative affect may negatively reinforce and maintain the use of purging as an affect regulatory behavior. However, a different pattern emerged when examining comparisons of single affect ratings before and after purging (see Table 6 for a summary of key findings in relation to analytic approach).

Table 6.

Summary of Key Findings

Comparison	Findings
Between-Day analyses: Purge vs. Non-purge Days	Negative affect: higher on purge days Positive affect: lower on purge days
Within-Day analyses: Affective changes prior to and following purging	Negative affect: increased prior to purging on purge days; decreased following purging on purge days Positive affect: no significant changes on purge days
Within-Day × Between-Day analyses: Affective changes prior to and following purging on purge days vs. non-purge days	Negative affect: trajectory prior to purging on purge days not significantly different from non-purge days; decrease following purging on purge days was significantly different from trajectory on non-purge days Positive affect: increased over time on non-purge days and failed to change relative to purging on purge days; increase following purging on purge days was significantly different from trajectory on non-purge days
Proximal affect ratings before and after purging	Negative affect: higher post-purging Positive affect: lower post-purging

When looking at trajectories of affect leading up to and following purging, we found that negative affect increased prior to purging on purge days, consistent with previous EMA research in AN (Engel et al., 2013) and BN (Alpers & Tuschen-Caffier, 2001; Rebert et al., 1991; Schlundt et al., 1986; Smyth et al., 2007). However, the increase was not significantly different from the trajectory of negative affect on non-purge days, suggesting that these increases may not specifically trigger purging behavior. Instead, positive affect was specifically linked to the presence versus absence of purging behavior. Compared to non-purge days, positive affect failed to increase prior to purging behavior on purge days. Thus, an absence of positive affect may trigger purging behavior compared to the protective effects of increases in positive affect.

Although we did not find evidence that purging was triggered by unique increases in negative affect, decreased negative affect following purging was significantly different from the trajectory of negative affect on non-purge days. This extends previous EMA studies in AN and BN (Alpers & Tuschen-Caffier, 2001; Corstorphine et al., 2006; Engel et al., 2013; Powell & Thelen, 1996) by demonstrating that decreases in negative affect appear to be consequences of purging rather than simple reflections of diurnal changes in affect over the course of the day. In addition, the presence of purging moderated the trajectory of positive affect, with less steep decreases in positive affect following purging on purge days compared to average time of purging on non-purge days. Thus, purging in PD may be maintained primarily through negative reinforcement by reducing negative affect and secondarily by protective effects in altering the trajectory of declines in positive affect that triggered purging episodes.

Findings from trajectory analyses support unique roles of negative and positive affect in triggering and maintaining purging in PD and support an affect regulation model for purging. Although previous EMA research suggests that negative affect and positive affect have inverse relations with purging in AN (Engel et al., 2013) and BN (Smyth et al., 2007), results of this study suggest that negative and positive affect have different functional associations with purging in PD. The absence of positive affect increased vulnerability to purging whereas decreases in negative affect negatively reinforced the behavior.

Comparisons of affect ratings before and after purging produced a different pattern of results, finding that negative affect was higher and positive affect was lower following purging in PD. Results are consistent with previous research in AN (Engel et al., 2013), but inconsistent with hypotheses generated by an affect regulation model of purging and previous research in BN finding that negative affect was lower following purging (Alpers & Tuschen-Caffier, 2001; Corstorphine et al., 2006; Davis et al., 1985; Elmore & de Castro, 1990; Kaye et al., 1986; Powell & Thelen, 1996). One explanation for this discrepancy may be whether pre- and post-purging ratings differ in how proximal each rating was made in relation to purging behavior. In the current study, the first rating after purging was made closer to the purging behavior (~30 minutes after purging) compared to the last rating before purging (~two hours before purging), likely reflecting the fact that event-contingent ratings were better able to capture the immediate consequences of purging whereas signal-contingent ratings were less able to capture the immediate antecedents of purging. The longer duration of time prior to purging behavior influenced our ability to examine proximal changes in affect. Given that negative affect increases over time prior to purging, it is likely that NA continued to increase during the two hours prior to purging behavior and may have been higher immediately before purging compared to immediately after purging. Importantly, due to the gap in the last pre-purge rating and the time of purging, the point in time at which negative affect peaks and begins to decrease remains unknown. It is possible that purging triggers the change in trajectory suggested by trajectory analyses. Thus, although results from trajectory analyses and proximal comparisons appear to be inconsistent, they may be compatible. Alternatively, it is possible that feelings of shame and failure immediately following purging result in heightened negative affect. Future research incorporating alternative assessment designs (e.g., more frequent signals to capture more immediate antecedents) is needed before drawing conclusions about proximal ratings of affect before and after purging.

Taken together, findings support the importance of affect in triggering purging in PD consistent with an affect regulation model. However, findings related to the affective consequences of purging are less clear. This study included data analytic strategies that have been used in previous investigations of temporal associations between affect and purging and were associated with different findings related to the affect regulation model (e.g., Engel et al., 2013; Haedt-Matt & Keel, 2011, Smyth et al, 2007). Proximal ratings of affect immediately surrounding purging may be ideal for examining negative reinforcement models as immediate reductions in NA are likely to be more powerful than delayed reductions in NA in maintaining behavior. However, these comparisons rely on single data points before and after purging and are not informed by the multiple ratings of affect prior to and following purging. Thus, examining trajectories are important for establishing the affective context of purging and other disordered eating behaviors. While inconsistent findings are challenging to interpret, this study highlights the utility of evaluating the affect regulation model using multiple approaches to empirically evaluate the impact of each analytic strategy.

This and future data regarding affect and purging behavior are crucial to understand why women with PD purge after consuming normal or small amounts of food. Findings could have important implications for the development of effective interventions for PD, a syndrome for which no evidence-based treatments currently exist. Current results suggest that increases in positive affect may be protective against the use of purging behavior. Thus, individuals who purge may benefit from treatments incorporating behavioral activation techniques in order to gain greater access to experiences that are likely to augment positive affect (Dimidjian, Barrera, Martell, Muñoz, & Lewinsohn, 2011). In addition, given that purging may function to regulate affect through decreases in negative affect over time following purging, psychological treatment of PD should focus on developing more adaptive affect regulation strategies to cope with intense emotions. Finally, women with PD also may benefit from treatments aimed at improving distress tolerance and impulse-control skills, such as techniques used in Dialectical Behavior Therapy (Linehan, Cochran, & Kehrer, 2001).

As the first EMA study of PD, findings provide important information regarding how affect changes over time in relation to purging behavior. However, the current study is unable to determine why women with PD purge as a form of affect regulation. Previous research suggests that women with PD report greater increases in fullness and stomach discomfort following ingestion of a standardized test meal compared to both non-eating disorder controls and women with BN, suggesting that gastrointestinal distress following food intake may be a unique risk factor for developing purging in PD (Keel et al., 2007). Thus, one possibility is that women who experience negative affect in the context of gastrointestinal discomfort are more likely to purge in an attempt to reduce distress versus engaging in alternative affect regulation behaviors. Future EMA research is needed to examine whether stomach discomfort differentiates those who use purging versus other methods to regulate affect or times when individuals with PD use purging versus. other affect regulation strategies. In addition, future research can incorporate comparisons with other affect regulation behaviors (e.g., alcohol use, cutting) and comparison groups (e.g., Major Depressive Disorder) to further examine why some individuals experience fluctuations in affect and purge versus engaging in other affect regulatory behaviors.

The current study had several notable strengths. First, this study represented the first application of EMA to PD and provided much needed information regarding affective factors that trigger and maintain purging in this syndrome. In addition, this study assessed both negative and positive affect. The influence of positive affect on purging or other disordered eating behaviors has received limited attention in previous research (Haedt-Matt & Keel, 2011). To our knowledge, findings from the current study are the first to identify a potentially unique role for positive affect in triggering purging behavior. Finally, this study applied a novel analytical approach that allowed us to determine whether within-day changes in affect were specifically related to the presence of purging behavior. Previous EMA studies have examined trajectories of affect only within binge and/or purge days (e.g., Engel et al., 2013; Smyth et al., 2007). However, these analyses do not control for changes in posited antecedents and consequences that may be unrelated to the purging behavior, such as work/school schedule or time of day effects. Thus, comparisons of trajectories of change on purge versus non-purge days provided a particularly rigorous test of the affect regulation model.

Several limitations also must be noted. First, while EMA is ideally suited to assess events that precede and follow a behavior within an individual’s natural environment, a key concern for any study utilizing EMA is inability to draw causal inferences from a longitudinal design. A second limitation was the sample size, which limited our ability to adequately examine non-linear trajectories of change and more complex statistical models due to concerns of low statistical power. Thus, MLM analyses examining linear changes may not have fully captured differences in trajectories of affect on purge versus non-purge days. Notably, previous EMA studies of disordered eating published in peer-reviewed journals have had an average sample size of N = 29 participants and ranged from N = 8 to N = 131 participants (Haedt-Matt & Keel, 2011), making the current study comparable in size to several other published studies in the field. Moreover, the current sample size of N = 24 was sufficient for testing hypotheses given the large effect sizes identified from the extant literature (Haedt-Matt & Keel, 2011) and observed in the current study. Third, we did not assess any other affect regulation strategies, such as alcohol use or cutting. Thus, one reason that we did not observe significant differences between antecedent trajectories of negative affect on purge versus non-purge days may be the potential use of alternative affect regulation strategies on non-purge days. Thus, antecedent increases in negative affect may still be relevant to purging behavior. A fourth limitation was the large within-subject variability of time of purging, which may have limited the ability of time-matched trajectories on non-purge days to detect how affective changes differed between purge days and non-purge days. Fifth was the generalizability of our sample, including having to use site as a covariate to control for differences in clinical variables as well as unobserved nuisance variables that may have differed between study locations. Given that participants were recruited from the community, there may be differences between our sample and those seeking treatment that should be examined in future research. However, community-based samples represent the original pool of individuals who may ultimately seek treatment. Finally, criteria for PD in the current study are more stringent than research criteria proposed by Keel and Striegel-Moore (2009) to ensure enough instances of purging during data collection to provide detection of reliable temporal patterns. As a consequence, our participants may have greater severity of illness than those presenting with lower symptom frequencies.

This study highlights the utility of EMA in examining complex temporal relationships among antecedents and consequences of purging in PD. Future EMA research has the unique potential to further improve our understanding of the complex mechanisms maintaining purging in PD and to contribute to the development of effective treatments for this syndrome. It will be important for future studies to include larger samples in order to test non-linear changes in affect, moderating effects of between-subjects factors, such as individual differences in personality, and to gain a deeper understanding of the interaction between trait and state psychological factors in the maintenance of purging in PD as models supported significant variance in trajectories that was not fully accounted for by variables currently examined. Future studies could examine the influence of different facets of negative affect as guilt has been supported as being particularly relevant to purging behavior in BN (Berg et al., 2013). In addition, future research could examine difference facets of positive affect, such as joy and attentiveness, to see if they are differentially related to triggering and maintaining purging. Given that many women with PD report subjective binge episodes (67% of current sample), future research should examine if the presence of subjective binge eating assessed at the momentary level influences associations between changes in affect and purging. Finally, specific targets for intervention should be tested in future treatment trials. If interventions that increase positive affect decrease purging frequency in PD, this would provide compelling causal evidence for the impact of positive affect as a trigger for purging and could lead to large-scale randomized controlled trials to determine the efficacy of these interventions for treating PD. Given the prevalence and clinical severity of PD (Favaro et al., 2003; Keel et al., 2005; Keel et al., 2008; Mond et al., 2006; Wade et al., 2006) and absence of evidence-based treatments, more work in this area has great public health significance.