Assessing Eating Disorders in Adults with Type 1 Diabetes: An Item Response Theory Analysis of the Diabetes Eating Problems Survey-Revised (DEPS-R)

Abstract

The 16-item Diabetes Eating Problems Survey-Revised (DEPS-R) is one of the few instruments assessing disordered eating behaviors (DEB) in type 1 diabetes (T1D). Although a DEPS-R total score of ≥20 is oft used to identify a probable eating disorder, previous work has shown that some items may reflect T1D management struggles, rather than DEB, with other work indicating that different response profiles (e.g., overeating vs. diabulimia) correspond to varying DEB risk. To identify DEPS-R items of greatest utility in detecting DEB in T1D, we used a form of Item Response Theory (IRT) to evaluate items variation in difficulty and discrimination, as well as scale precision across the latent DEB continuum. Participants were 329 adults with T1D who completed the DEPS-R when joining a participant registry or enrolling in a study on eating and weight concerns in T1D. We found a number of DEPS-R items (e.g., “I skip meals or snacks”) that demonstrated low-difficulty and low-to-moderate-discrimination. These items were often endorsed by even the lowest-severity participants, yet provided relatively limited information for differentiating the DEB construct. Other items with high-difficulty and moderate-discrimination (e.g., “I make myself vomit”) provided relatively modest information even at the highest pathology levels, as they were unlikely to be endorsed even by the most severe cases. Most high-discrimination items provided the most diagnostic information at highest pathology levels (e.g., “I try to eat to the point of spilling ketones in my urine”). Results suggest revising the DEPS-R scale, as several items provided limited contribution to DEB detection.

Introduction

Disordered eating behaviors (DEB) are fairly prevalent in persons with type 1 diabetes (T1D), posing significant risks to physical and psychological health (Dean et al., 2024). These behaviors, which include dysregulated eating (e.g., binge eating) and maladaptive weight control strategies (e.g., intentional restriction of insulin), are associated with elevated HbA1c (Merwin et al., 2014; Ripoli et al., 2022), higher 2-hour postprandial blood glucose (Moskovich et al., 2019), and more time spent above target glycemic range during continuous glucose monitoring (Eisenberg Colman et al., 2018). DEBs are also associated with more hospitalizations, more episodes of diabetic ketoacidosis and 3 times greater risk of early and severe diabetes-related complications and 6 times greater risk of premature death (Goebel-Fabbri, 2009; Marks et al., 2024). Thus, the ability to detect these problems and coordinate care is essential for improving clinical outcomes (Shapiro, 2022).

The Diabetes Eating Problems Survey-Revised (DEPS-R) is one of the most commonly used instruments for screening DEB in persons with T1D (Markowitz et al., 2010; Wisting et al., 2013b). This 16-item measure assesses general eating disorder attitudes and behaviors, and diabetes-specific concerns, such as insulin restriction. The DEPS-R has demonstrated internal consistency and construct validity (e.g., Markowitz et al., 2010; Wisting et al., 2019). The DEPS-R clinical cutoff of ≥20 accurately identifies 81.8% of people as having an eating disorder diagnosis or not as validated by the gold standard structured eating disorder diagnostic interview, although individuals engaging exclusively in either binge eating or purging behaviors are less likely to be detected, with a 40% false-negative rate (Moskovich et al., 2025). Despite its limitations, the adoption of the DEPS-R in both research and clinical settings underscores its utility (Atik-Altınok et al., 2023; Hummadi et al., 2023; Karastogiannidou et al., 2021; Lv et al., 2021; Ryman et al., 2019).

There are several limitations of the DEPS-R that have been identified and warrant further investigation. Although several studies demonstrated that the DEPS-R may measure a heterogeneous construct with distinct factors (e.g., Atik-Altınok et al., 2023; Gagnon et al., 2017; Lok et al., 2023; Park et al., 2025; Pinna et al., 2017; Wisting et al., 2019), the nearly universal approach to scoring the DEPS-R in clinical practice is to sum the 16 items as reflective of a unidimensional construct, and use a clinical cutoff score of DEPS-R scores ≥20 to identify persons with clinically significant DEB or a probable eating disorder. This practice is based on studies linking scores ≥20 to poorer glycemic outcomes (Markowitz et al., 2010; Cezar et al., 2022). However, the cutoff has not undergone rigorous psychometric evaluation, raising concerns about its ability to distinguish individuals with diagnosable eating disorders from those whose elevated scores reflect broader challenges in diabetes self-management (Moskovich et al., 2025; Embaye et al., 2024).

Indeed, some DEPS-R items may reflect common struggles with diabetes management rather than clinical eating disorders (Goebel-Fabbri et al., 2008; Wisting et al., 2013a). This includes items assessing weight concerns (e.g., “Losing weight is an important goal for me”) or the perceived difficulty of controlling both weight and diabetes (e.g., “I feel it is difficult to lose weight and control my diabetes at the same time”), which could be highly relevant to daily diabetes management without indicating maladaptive behavior. Conversely, items designed to capture severe problems, such as intentionally maintaining hyperglycemia to influence weight, may lack sensitivity for detecting less overt or emerging forms of DEB (Goebel-Fabbri et al., 2008). Emerging evidence also highlights the importance of considering distinct DEB phenotypes in T1D populations, rather than a single total DEPS-R score. For example, a latent class analysis of the DEPS-R identified 4 distinct profiles across DEPS-R items (Bulimia, Binge Eating Disorder, Overeating, Low Pathology), with different patterns of item endorsement and glycemic outcomes (Merwin et al., 2024). These limitations suggest the need for a more nuanced understanding of the DEPS-R's item performance and how best it may be used or improved on for use in research or clinical practice.

The current study used a two-parameter logistic item factor analysis, a version of Item Response Theory (IRT) appropriate for ordered categorical variables, to evaluate the psychometric properties of the DEPS-R. IRT offers a sophisticated framework for analyzing item- and test-level performance. Unlike the summary DEPS-R score derived from Classical Test Theory, IRT does not assume identical discrimination and difficulty values at the item level and does not assume stable scale-level values for reliability (e.g., Cronbach’s alpha) and measurement error. IRT explicitly estimates item difficulty and discrimination to assess how well each item identifies individuals with varying symptom severity and provides standard errors of measurement values that vary on the ability continuum, and shows how item and scale suitability vary across the continuum of the latent trait (Embretson & Reise, 2000).

The study aims to identify DEPS-R items with greater utility in screening and detecting eating disorders. This includes examining which DEPS-R items may effectively distinguish DEB from general diabetes-related struggles or concerns and identifying items with limited value for identifying DEB. Given the most common practice of treating the DEPS-R as a unidimensional construct, and in line with our goal to evaluate the psychometric properties of the DEPS-R as it is typically used in clinical practice, we focused our psychometric testing on the single-factor DEPS-R. For the interested reader, we also supplemented our results with IRT parameters for the 3-factor model as presented in the original DEPS-R proposal among adult T1D patients (Wisting et al., 2013b). By evaluating the DEPS-R, this study seeks to enhance its specificity and clinical utility as a screening tool for DEB in persons with T1D, or to inform the development of new measures. Measures that quickly and effectively detect DEB would be particularly useful in time-sensitive clinical settings such as primary care or endocrinology clinics. Ultimately, this work aims to improve diabetes care by enabling more precise DEB detection, reducing clinical or diagnostic ambiguities, and facilitating targeted, phenotype-specific interventions to improve outcomes for persons with T1D.

Method

Participants were 329 adults with T1D who either completed the DEPS-R as part of joining an online research participant registry for persons with T1D (n=236), or as part of a pilot study and larger investigation (n=93) on eating and weight concerns in T1D. Some individuals both joined the registry and participated in the later pilot study. For these participants, only their most recent DEPS-R scores were included to avoid double counting any one participant. The studies were each approved by the Duke University Medical Center’s Institutional Review Board and participants provided informed consent.

Measures

Diabetes Eating Problems Survey – Revised (DEPS-R)

The DEPS-R assesses disordered eating behaviors in individuals with diabetes. The DEPS-R consists of 16 items rated on a six-point Likert scale (0 = Never to 5 = Always; Table 1). Higher scores indicate more severe disordered eating symptoms, with a score of ≥20 commonly used as a threshold for clinically significant symptoms or a probable eating disorder (Markowitz et al., 2010). Previous studies have validated the DEPS-R for use in identifying diabetes-specific disordered eating behaviors, including insulin restriction and binge eating, with established internal consistency and construct validity (Markowitz et al., 2010).

Table 1.

Diabetes Eating Problems Survey-Revised (DEPS-R) Item Number and Content

Item Number	Item Content
1	Losing weight is an important goal for me.
2	I skip meals and/or snacks.
3	Other people have told me that my eating is out of control.
4	When I overeat, I don’t take enough insulin to cover the food.
5	I eat more when I am alone than when I am with others.
6	I feel that it’s difficult to lose weight and control my diabetes at the same time.
7	I avoid checking my blood sugar when I feel like it is out of range.
8	I make myself vomit.
9	I try to keep my blood sugar high so that I will lose weight.
10	I try to eat to the point of spilling ketones in my urine.
11	I feel fat when I take all of my insulin.
12	Other people tell me to take better care of my diabetes.
13	After I overeat, I skip my next insulin dose.
14	I feel that my eating is out of control.
15	I alternate between eating very little and eating huge amounts.
16	I would rather be thin than have good control of my diabetes.

Analytic Strategy

We first confirmed sufficient unidimensionality and local independence of the scale. We then fit a single-factor 2-parameter logistic (2PL) item factor analytic (IFA) model for polytomous items (Samejima, 2016), a form of IRT. We used weighted least squares mean and variance adjusted estimator (WLSMV), theta parameterization (item residual variance fixed to 1), and a probit link function (Asparouhov & Muthén, 2020). Participants’ trait-level theta (θ; the level on the underlying DEB trait) was estimated from the 16 DEPS-R observed indicators. The θ mean and variance were set to 0 and 1, respectively, to freely estimate factor loadings and thresholds for each of the 16 indicators. Following previous work on model fit indices (Hu & Bentler, 1999; MacCallum, Browne, & Sugawara, 1996), we used the following cutoffs for acceptable model fit: ≥.95 for Comparative Fit Index (CFI); ≥.95 for Non-Normed Fit Index (TLI); ≤.08 for root-mean-square error of approximation (RMSEA); and ≤0.08 for standardized root-mean-square residual (SRMR). Local misfit was assessed by inspecting residual covariances among item pairs using their respective modification indexes.

To assess the diagnostic utility of each item we present figures with traditional item information curves, and represent difficulty and discrimination values in a graphical form following recent IRT work on other scales (e.g., Hsu, Hoffman & Thomas, 2023). The functions representing the test information curve and standard error of measurement were used to assess scale reliability across the pathology continuum.

Factor loadings and transformed thresholds correspond to IRT item discrimination ($a_{\mathrm{i}}$) and difficulty ($b_{ik}$) parameters. Item-specific discrimination, $a_i$, is proportional to the item characteristic curve slope at $\theta =b_{ik}$, and indicates how well an item distinguishes between individuals at a pathology level that corresponds to the difficulty of response option $k$ for item $i$. Item-specific difficulties, $b_{i1}-b_{i5}$, correspond to five item-specific thresholds at response options Rarely, Sometimes, Often, Usually and Always. The resulting difficulty parameters $b_{ik}$ thus represent the latent level of disordered eating pathology at which the probability of endorsing $k$ or greater response option of item $i$ reaches 50%. We did not fit $b_{i3}$ for DEPS-R 10, as no participant endorsed Often eating to the point of spilling ketones.

We used previously published (Asparouhov & Muthén, 2020; Paek et al., 2018) formulas to convert IFA-derived parameters and their standard errors to their IRT counterparts, using the Mplus model constraint commands. IRT discrimination for item $i$ is given by $a_i={\lambda }_i\sqrt{\psi }$ ¹, where ${\lambda }_i$ is the item’s unstandardized IFA factor loading, and $\psi$ is the factor variance. IRT difficulty parameter for response option $k$ of item $i$ is determined by the following equation: $b_{ik}=\frac{{\tau }_{ik}-{\lambda }_i\alpha }{{\tau }_i\sqrt{}\psi }$ ², where ${\tau }_{ik}$ is the item’s unstandardized threshold at response option $k$, ${\lambda }_i$ is the item’s unstandardized factor loading, $\alpha$ is the factor mean, and $\psi$ is the factor variance.

Baker (2001) suggests interpreting discrimination values as very low (<0.25); low (0.25 to 0.64), moderate (0.65 to 1.34), high (1.35 to 1.69), and very high (≥1.7). IRT modeling best practices do not specify an ideal cut off for item difficulty, rather, the preferred item difficulty is determined based on the goal of the test. For example, items should vary across the broad range of theta when the interest is to identify participants across the full continuum of the latent trait, and items should focus on the narrower high range of the trait when the goal is to identify high-trait respondents (e.g., high ability scholarship applicants, patients at risk of a clinical diagnosis). Given our interest in supporting research and clinical practice on individuals at risk of DEB, we classified items with a difficulty for $\underset{¯}{b_{i1}}$ below 0 (corresponding to a ≥50% change of endorsement of ≥Rarely among participants below the mean of the trait where θ=0) as low difficulty. We classified items with $b_{i1}$ difficulty between 0 and 1 as medium difficulty, and items with $b_{i1}>1$ as high difficulty.

Data Quality

The analysis involved existing data that were previously described (an online research registry: Merwin et al., 2014; an investigation on eating and weight concerns in T1D: Merwin et al., 2015). To ensure quality and integrity, data were regularly reviewed throughout the study, and the dataset was examined for completeness, uniqueness (e.g., potential duplicate records and, in the case of duplicate records, use the most recent response set in analyses), and accuracy (e.g., all data values were within the acceptable ranges). All data were entered electronically by participants, thus not necessitating the use of double-data entry.

Transparency and Openness

All data, analysis code, and research materials are available upon request from the first author. Data were analyzed using Mplus version 9 (Muthén & Muthén, 1998-2025). This study’s design and its analyses were not preregistered.

Results

Sample characteristics

The sample was predominantly White (88.45%), at least college-educated (72.34%) women (72.04%), with a mean age of 43.27 years (SD=13.53; range: 18-76). Twenty-five percent of the sample reported they had never been married, whereas the majority were either currently (63.53%) or previously (i.e., separated, divorced or widowed; 11.25%) married. The mean self-reported age of T1D diagnosis was 20.45 years (SD=13.23) and mean self-reported most recent HbA1c was 7.55% (SD=1.47). DEPS-R scores ranged from 0-76 with a mean of 16.61 (SD=12.10). Twenty-eight percent of the sample had a DEPS-R score ≥20, suggesting the presence of clinically significant DEB. Response option frequencies are displayed in Supplemental Table 1.

Dimensionality and Local Independence

Unidimensional Maximum Likelihood Model.

We first fit a single-factor confirmatory factor model with maximum likelihood (ML) estimation, and confirmed that the DEPS-R is not unidimensional (CFI=0.748; TLI=0.709; RMSEA=0.136 (90% CI [0.127, 0.145]); SRMR=0.080), as has been previously shown (e.g., Wisting et al., 2019). Modifying the model to use ML with robust standard errors (MLR; which provides model fit and SE adjustments for multivariate non-normality) did not result in acceptable model fit (CFI=0.752; TLI=0.714; RMSEA=0.104 (90% CI [0.095, 0.114]); SRMR=0.080). This indicates that a model for continuous responses is not appropriate, as the categorical DEPS-R data would not be well represented by a single latent trait with continuous indicators.

Unidimensional WLSMV Model.

We then fit a single-factor 2PL IFA model using the WLSMV estimator. Although there were fit improvements, the polytomous IFA model still did not reach acceptable fit (CFI=0.934; TLI=0.924; RMSEA=0.095 (90% CI [0.085, 0.105]); SRMR=0.073). We then examined evidence for local dependence, a condition when item responses exhibit residual correlations beyond those explained by the IRT model parameters (i.e., beyond the θ, $a_i$ and $b_{ik}$). The largest residual correlation was between DEPS-R 1 (“I skip meals and/or snacks”) and DEPS-R 6 (“I feel that it’s difficult to lose weight and control my diabetes at the same time”) (standardized expected parameter change of 0.848).

WLSMV Model Adjusted for Local Dependence.

Guided by previous research on covariance adjustment for the DEPS-R, we freed the residual correlation between DEPS-R 1 and 6. Results from two previously published independent samples demonstrated that the 1- and 3-factor model does not sufficiently account for the association between DEPS-R 1 and 6 (Lok et al., 2023; Wisting et al., 2019), indicating that the residual correlation is not an artifact of our data’s vagaries, but a consistent content-level link between the experiences of skipping meals/snacks and the struggle of losing weight while controlling diabetes, which is not explained by the latent factor(s) of the DEPS-R. Following the freed residual correlation, fit indexes demonstrated acceptable model fit (CFI=0.968; TLI=0.963; RMSEA=0.066 (90% CI [0.056, 0.077]); SRMR=0.058), so we chose to proceed with this model.

Multidimensional 3-Factor Model.

We did not pursue a multidimensional IRT approach using the 3-factor model, due to a lack of established subscales corresponding to the distinct factors of the DEPS-R scale, and due to the nearly-predominant use of the total DEPS-R score in clinical practice. However, given previous research supporting the 3-factor model (e.g., Atik-Altinok et al., 2023; Lok et al., 2023; Pinna et al., 2017; Wisting et al., 2019), we briefly describe the 3-factor model fit, and provide interested readers with item- and scale-level results in Supplemental Table 2 and Supplemental Figure. We modified the unidimensional WLSMV model to include the following three correlated factors: maladaptive eating behaviors (DEPS-R 2, 3, 4, 5, 7, 12, 13, 14, 15), preoccupation with thinness/weight (DEPS-R 1, 6, 11, 16), and maintaining high glucose to lose weight (DEPS-R 8, 9, 10). The 3-factor model demonstrated poor fit (CFI=0.949; TLI=0.939; RMSEA=0.085 (90% CI [0.075, 0.095]); SRMR=0.066), and the 3-factor model with the freed residual correlation between DEPS-R 1 and 6 reached acceptable fit (CFI=0.971; TLI=0.965; RMSEA=0.064 (90% CI [0.054, 0.075]); SRMR=0.053).

IRT Results

Supplemental Table 3 shows item-level difficulty and discrimination parameters, while Figure 1 graphically demonstrates the range of these parameters. To correspond to our Analytic Strategy, we arranged Figure 1 estimates in ascending discrimination, within three difficulty groups (i.e., “low,” “medium,” and “high”). Given that θ had a mean of zero and a variance of one, the observed values of theta exhibited a mean of 0.037 (variance=0.861, SD=0.928), ranging from −2.068 to 3.403. On average, the DEPS-R items demonstrated high discrimination (mean $a_i=1.248$, ranging from 0.393 to 3.030) and low difficulty for the lowest threshold (mean $b_{i1}=-0.012$, ranging from −2.945 to 2.130).

Figure 1. Item Response Theory parameters for the Diabetes Eating Problems Survey-Revised.

Note. DEPS-R = Diabetes Eating Problems Survey–Revised. (A) Item Discrimination Parameters. (B) Item Difficulty Location Parameters. Solid line = moderate discrimination; dotted line = high discrimination; circle = difficulty for ≥rarely; triangle = difficulty for ≥sometimes; square = difficulty for ≥often; cross mark = difficulty for ≥usually; diamond = difficulty for always.

Item difficulty and discrimination are independent variables (in addition to the participant-specific θ trait score) in the 2PL IRT model. Together, discrimination and difficulty parameters inform the shape of the logistic response probability, and both must be considered in concert to determine its discrimination power of the item for a given difficulty (Embretson & Reise, 2000). For example, the moderately-discriminating DEPS-R 6 (“I feel that it’s difficult to lose weight and control my diabetes at the same time”) has $b_{i1}$ difficulty of −1.431. Given the observed θ mean and SD of approximately 0 and 1, respectively, a difficulty level of $b_{i1}=-1.431$ indicates that the ≥Rarely response option of this item has the optimum discrimination for participants who are at approximately 1.5 standard deviation below the mean on the DEB trait. Indeed, although DEPS-R 6 discriminates moderately well, most of the information comes from the low end of the trait, with 81.3% of all participants in the sample endorsing DEPS-R 6 at least Rarely. Even the highest response option (Always, $b_{i5}=1.306$) has maximum discrimination among participants at ~1.3 SDs above the average DEB. Such an item provides little to inform a clinician interested in identifying a high-risk participant, but may be valuable for detecting very mild or early DEB manifestations that could be used in targeted prevention. In contrast, DEPS-R 14 (“I feel that my eating is out of control”) provides high discrimination; with 56.8% of the sample endorsing ≥Rarely, with a corresponding difficulty for $b_{i1}$ of −0.068 (almost exactly at the mean). This item may be very valuable in detecting mild disordered eating and increased risk for maladaptive weight control, but might not discriminate individuals in greatest need of clinical intervention.

IRT Results Overview.

Figure 1 demonstrates that 1) higher difficulty items also tend to be more discriminating (panel A), and 2) the breadth of coverage across theta (i.e., difficulty) becomes more narrow as the items become more difficult (panel B). For many items, particularly those of moderate-to-high difficulty, not much meaningful information is provided by the distinction between response options Often and Usually (and perhaps between Usually and Always), as the markers for those response options are close or overlapping for many items. In contrast, the theta coverage from Rarely to Sometimes, and from Sometimes to Often is relatively wide, indicating that these distinctions may be of use in clinical practice. Below, we first present the difficulty results and then integrate how discrimination contextualizes the items’ difficulty parameters.

Low Difficulty Items.

Eight items – DEPS-R 1, 2, 4, 5, 6, 7, 12, and 14 – demonstrated $b_{i1}<0$ (corresponding to difficulties for ≥Rarely below the mean of the trait where θ=0), and therefore were considered low difficulty. DEPS-R 2 (“I skip meals or snacks”) was the easiest item in the scale, with 50% likelihood of endorsing ≥Rarely occurring among participants with an estimated theta of −2.945. For context, the lowest observed theta score (calculated by exporting factor scores) in our sample was almost 1SD above this score (i.e., −2.068; approximate total DEPS-R score=0). Although DEPS-R 2’s lowest response option is unlikely to be of clinical use, as it provides maximum information among participants with an impossibly-low DEPS-R score, its breadth of θ coverage across response options is impressive (e.g., $b_{i5}=5.248$). Our findings on its wide breadth of coverage are tempered by DEPS-R 2’s low r² value (Supplemental Table 3), indicating that although the overall model is acceptable in fit across 16 items, it is explaining the variance in DEPS-R 2 relatively poorly.

Low Difficulty and Low Discrimination.

DEPS-R 1³ (“Losing weight is an important goal for me”) and DEPS-R 2 (“I skip meals or snacks”) exhibited the lowest discriminations and did not meet the threshold of moderate discrimination (≥0.65). As is the case for all low discrimination items, individual differences on the latent trait have relatively little impact on response probabilities of DEPS-R 1 and 2. These items’ relatively low discriminations indicate they have limited utility for differentiating even modest disordered eating pathology.

Low Difficulty and Moderate Discrimination.

Five items reached the threshold of moderate discrimination, but were below the threshold indicating high (≥1.35) discrimination. These items consisted of DEPS-R 4 (“When I overeat, I don’t take enough insulin to cover my food”), DEPS-R 5 (“I eat more when I am alone than when I am with others”), DEPS-R 6⁴ (“I feel that it’s difficult to lose weight and control my diabetes at the same time”), DEPS-R 7 (“I avoid checking my blood sugar when I feel like it is out of range”), and DEPS-R 12 (“Other people tell me to take better care of my diabetes”). These five items showed their (albeit modest) highest discrimination at θ levels below the mean, suggesting their limited use for identifying significant pathology.

Low Difficulty and High Discrimination.

DEPS-R 14 (“I feel my eating is out of control”) was the only highly discriminating low difficulty item. In general, trait level differences have a strong impact on response probabilities among high discrimination items. DEPS-R 14 ≥Rarely difficulty level was below the θ mean (−0.068), hence, DEPS-R 14 may be appropriate for identifying mild/early disordered eating pathology.

Medium Difficulty Items.

Items with moderate difficulty included the DEPS-R 3 (“Other people have told me my eating is out of control”), DEPS-R 11 (“I feel fat when I take all my insulin”), DEPS-R 15 (“I alternate between eating very little and eating huge amounts”), and DEPS-R 16 (“I would rather be thin than have good control of my diabetes”), showing ≥Rarely difficulty within one SD above theta mean (ranging from 0.037 to 0.853).

Medium Difficulty and Moderate Discrimination.

All four medium difficulty items demonstrated moderate discrimination, with $a_i$ ranging from 0.920 to 1.344. Although moderate in discrimination, these three items may still be of some interest for those identifying mild-to-moderate disordered eating pathology.

High Difficulty Items.

DEPS-R items 8, 9, 10 and 13 exhibited the highest difficulty levels. Items DEPS-R 8 (“I make myself vomit”) and DEPS-R 10 (“I try to eat to the point of spilling ketones in my urine”) were the most difficult items, endorsed primarily by individuals with the most severe disordered eating pathology.

High Difficulty and Moderate Discrimination.

One high-difficulty item – DEPS-R 8 (“I make myself vomit”) – had moderate discrimination of 1.092. This item demonstrated the highest ≥Rarely difficulty across all 16 items ($b_{i1}=2.130$), and was endorsed at least rarely by only ~5% of the sample. In our sample, the closest observed theta score (calculated by exporting factor scores) was 2.139; this participant’s corresponding raw total score was 63, suggesting that only the highest scoring participants have a reasonable chance of endorsing DEPS-R 8 even Rarely. When compared to the three highly discriminating items described in the next section, DEPS-R 8 appears to provide modest utility for discriminating high severity disordered eating behaviors in T1D, and may index DEB symptomatology that that is not universally experienced by high-severity DEB in T1D. Indeed, vomiting is a severe DEB endorsed by only a subset of individuals with EDs in general and may be even less frequently endorsed by individuals with T1D for whom vomiting has increased complications for diabetes management.

High Difficulty and High Discrimination.

Finally, the following three items coupled high discrimination with high difficulty: DEPS-R 9 (“I try to keep my blood sugar high so that I will lose weight”), DEPS-R 10 (“I try to eat to the point of spilling ketones in my urine”), and DEPS-R 13 (“After I overeat, I skip my next insulin dose”). These three items exceeded the threshold of high discrimination (≥1.35) and reached very high (≥1.7) discrimination. These items are particularly useful for identifying those with more severe symptoms. Figure 2 illustrates each item’s Item Information Curve (IIC), which is a square of the item’s discrimination, and is inversely proportional to the item’s measurement error. IICs demonstrate (1) the level of information provided by the item across trait level, (2) the inverse of the item measurement error across trait level, and (3) the range where the item provides the most precise measurement (i.e., highest information and lowest measurement error). Higher peaks at a particular location on the theta indicate greater precision in identifying individuals at specific levels of the latent trait. Items with wider peaks contribute information across a broader range, while those with narrower, sharper peaks are more discriminative within specific severity levels. This visualization highlights the variability in item performance, which is critical for refining the DEPS-R for enhanced detection in persons with diabetes.

Figure 2. Item Information Curves (IICs) for the Diabetes Eating Problems Survey-Revised (DEPS-R).

Figure 2 confirms that information contained in items varied by trait level. For example, items DEPS-R 9 (“I try to keep my blood sugar high so that I will lose weight”), DEPS-R 10 (“I try to eat to the point of spilling ketones in my urine”), and DEPS-R 13 (“After I overeat, I skip my next insulin dose”) provided the highest amounts of information, and were most diagnostic at high pathology levels, with pronounced peaks at the far end of the eating disorder continuum, indicating strong utility for identifying severe DEB. Figure 2 confirms the numerical results, showing that there was one highly discriminating item of low difficulty, DEPS-R 14 (“I feel my eating is out of control”). This item provided relatively moderate information, and its highest performance was at low-to-moderate pathology levels.

Some items offered minimal information across the continuum, as indicated by their low peaks. This included items DEPS-R 1 (“Losing weight is an important goal for me”), DEPS-R 2 (“I skip meals or snacks”), DEPS-R 3 (“Other people have told me my eating is out of control”), DEPS-R 4 (“I eat more when I am alone than when I am with others”), DEPS-R 6 (“I feel that it’s difficult to lose weight and control my diabetes at the same time”), DEPS-R 7 (“I avoid checking my blood sugar when I feel like it is out of range”), and DEPS-R 12 (“Other people tell me to take better care of my diabetes”).

Test Information

Figure 3 displays the Test Information Curve (TIC; panel A) and the corresponding Standard Error of Measurement (SEM; panel B) curve for the DEPS-R scale. In IRT, the TIC shows test information at a given ability level and is comprised of the sum of individual IICs. The TIC is inversely proportional to the function of the scale’s SEM. The range with relatively high TIC and low SEM values demonstrates the latent trait range where the test performs best (i.e., the scale shows a relatively high ability to distinguish among respondents who vary on trait level DEB).

Figure 3. Functions illustrating test information and measurement error for the Diabetes Eating Problems Survey-Revised (DEPS-R).

Notes. A: Test Information Curve (TIC). B: Standard Error of Measurement (SEM; dashed line); Test Information Curve on a 0-1 reliability metric (solid line); Traditional cutoff for good reliability ≥0.80 (dotted line)

The TIC can also be calculated on a 0-1 reliability metric as follows: $\frac{\text{information}}{\text{information}+1}$ (McDonald, 1999). The TIC on the reliability metric is displayed on Panel B of Figure 3 (solid line), and can be compared to the conventional cutoff for good reliability of ≥0.80 (dotted line). Panel B shows that the scale performs relatively well (i.e., ≥0.80 reliability) among participants ranging from approximately −1.5 to +4 on the latent trait. In Panels A and B, the solid lines peak and the dashed line troughs at theta 2.5, illustrating the powerful impact of DEPS-R 9, 10 and 13’s high discriminations at this latent level.

Thus, among participants scoring between −1.5 to +4.0 on the latent trait, greater than 64% (i.e., 0.8²) of the variance in their DEPS-R scores is due to their DEB level. In contrast, greater than 36% (i.e., 1 – 0.8²) of the variance in DEPS-R scores is due to measurement error for participants outside of the −1.5 to +4.0 range. Test information and reliability were thus relatively low, and SEM was relatively high for individuals scoring low on the latent trait. There is also significant variability in score precision within the relatively reliable −1.5 to +4.0 range. For example, a participant at 1.5 standard deviations above the mean has an estimated SEM of 0.18, whereas a participant at −0.5 SD below the mean has an estimated SEM of 0.36. This shows that even within the well-performing range of the scale, some respondents (θ=−0.5) are measured with lower precision, as their standard errors are more than double those scoring in the moderately-high (θ=1.5) range.

Discussion

This study used IRT to evaluate the psychometric properties of the DEPS-R, identifying strengths and weaknesses in its ability to detect disordered eating behaviors (DEB) in adults with T1D. By applying IRT, the study provides valuable insights into the functioning of individual items and the measure’s overall utility. The findings revealed significant variability in how well different items discriminated between individuals with and without clinically significant DEB, highlighting the nuanced nature of DEB detection in this population. These findings align with prior research, which has emphasized the challenges associated with detecting eating disorders in persons with diabetes (Wisting et al., 2013a). Notably, even diabetes-specific measures, such as the DEPS-R, may capture attitudes and behaviors that are more normative in the context of diabetes (e.g., difficulty losing weight), or general population (e.g., skipping meals or snacks, weight loss as a goal) rather than identifying clinically significant eating disorder symptomatology.

Item-Specific Utility

Findings from the current study suggest that the DEPS-R items have widely varying utility. Some items demonstrate exceptionally high difficulty that preclude endorsement even by the highest DEB participants, while others have low discrimination at the low end of the of the DEB continuum, thereby minimally contributing to detecting an eating disorder. For example, we saw that the most severe individuals often did not achieve DEPS-R 8 endorsement, and the lowest DEB respondents often still endorsed DEPS-R 1 and DEPS-R 2. Other low difficulty items (DEPS-R 4, 5, 6, 7, 12) only had up to moderate discrimination and low item information. All of these items were endorsed Rarely or more by well over 1/2 of participants, demonstrating frequent endorsement among respondents with estimated pathology well below one standard deviation below the mean. For example, items DEPS-R 5 ("I eat more when I am alone than when I am with others") and DEPS-R 12 ("Other people tell me to take better care of my diabetes") may reflect typical diabetes-related challenges, rather than clinically significant DEB. This discrepancy further emphasizes the need to revise the DEPS-R to avoid over pathologizing normative experiences.

We note that the two lowest difficulty items, DEPS-R 1 (“Losing weight is an important goal to me”) and DEPS-R 2 (“I skip meals or snacks”) demonstrated Rarely difficulties of −2.945 and −2.604, respectively. Such low thresholds demonstrate that even individuals at 2 and 3 standard deviations below the mean of the latent construct are likely to endorse this item. Even though many IRT assessments of clinical measures suggest that a low number of items tapping into the low end of a continuum demonstrates the need to develop additional low-difficulty items, we suggest this depends on one’s assessment goal. For instance, if the goal is to efficiently identify those in a clinically significant range of DEB, the scale should focus administration time on those items that are high difficulty. Removing or replacing low difficulty items could help reduce false positives while ensuring the measure remains sensitive to clinically significant DEB.

Figure 1 demonstrated that item difficulties tightly correspond to item discriminations, such that the items with the least utility in identifying DEB (i.e., low discrimination) are low in difficulty. That said, we did identify one low difficulty item with high discrimination, DEPS-R 14 (“I feel my eating is out of control”). Given its lowest threshold difficulty of approximately 0, DEPS-R 14 may be appropriate for identifying mild/early disordered eating pathology. There were no other highly discriminating items at low or medium difficulty. Of the medium difficulty items, DEPS-R 11 (“I feel fat when I take all my insulin”) was the most discriminating question. Together with DEPS-R 14, DEPS-R 11 may be useful in differentiating clinically significant symptoms from more normative experiences.

Some items were very good at detecting eating disorders. However, most of these items provided information only at very high levels of pathology (DEPS-R 8, 9, 10, and 13). For example, approximately 5% of participants endorsed engaging in DEPS-R 8 (“I make myself vomit”) or DEPS-R 10 (“I try to eat to the point of spilling ketones into my urine”) at least rarely. Only a small subset of individuals was likely to reach the threshold of DEB severe enough to endorse DEPS-R 8 and 10, likely underestimating the score of some high-risk participants. DEPS-R 8 (vomiting) shows more modest discrimination and may index symptomatology not universally experienced even by high-severity DEB in T1D. One explanation for this is that vomiting has increased complications for diabetes management, so severe DEBs in T1D may manifest differently than in the general population.

Relative to the relatively infrequently endorsed DEPS-R 8 and 10, DEPS-R items 9 (“I try to keep my blood sugar high so that I will lose weight”) and 13 (“After I overeat, I skip my next insulin dose”) demonstrated a moderately high difficulty of ~1.3 standard deviations above the mean. DEPS-R 9 and 13 items were more frequently endorsed (by ~10% of the sample, compared to about 5%) suggesting they are the most effective items for identifying significant symptomatology without being unduly restrictive.

Of note, DEPS-R 9 (“I try to keep my blood sugar high so that I will lose weight”) showed remarkably high discrimination, which is not uncommon in studies of clinical constructs, and may indicate its measurement of a conceptually narrow latent construct (for review, see Reise & Waller, 2009). Such high discrimination may also indicate a high correlation between other high-difficulty items that index a conceptually narrow but clinically significant behavior (i.e., insulin omission as measured by items DEPS-R 10 and 13).

Unfortunately, the DEPS-R appears to have specific limitations when assessing insulin restriction for weight control. For many items, particularly those of moderate-to-high difficulty, the distinction between Often and Usually (and perhaps between Usually and Always) did not offer much meaningful information, as the markers for those response options were overlapping for many items (Figure 1). Moreover, many items were either too difficult and only discriminate people with severe problems (e.g., DEPS-R 10 “I try to eat to the point of spilling ketones into my urine” and DEPS-R 9 “I keep blood sugar high so that I will lose weight”) or too easy (DEPS-R 4 “When I overeat, I don’t take enough insulin to cover my food”). The best-performing item with regard to insulin omission was an attitudinal item (DEPS-R 11 “I feel fat when I take all my insulin”), rather than a behavioral item.

Purposeful insulin restriction is a critical aspect of DEB in people with T1D that is not adequately assessed by the DEPS-R. The mSCOFF (Zuijdwijk et al., 2014), which is a common screening tool for EDs that has been modified for T1D with the inclusion of an insulin item, is also suboptimal. The item assesses insulin restriction (“I take less insulin than I should”), but not specifically for weight control purposes and uses non-person-centered language. We propose a novel item for testing that we identified in focus groups with persons with T1D-ED: "I take less insulin than I need because of concerns about my weight."

DEPS-R Scale Performance

In addition to revealing differences in the utility of specific items, the current study also highlights limitations in the use of total DEPS-R scores. Our results align with previously published evidence suggesting it is not unidimensional. Although the 16 items formed a reasonably unidimensional scale after fitting a residual correlation between DEPS-R 1 and 6, the lack of support for the truly unidimensional covariance-unadjusted model suggest that low- (e.g., DEPS-R 4 "When I overeat, I don’t take enough insulin to cover the food") and high-difficulty (e.g., DEPS-R 9 "I try to keep my blood sugar high so that I will lose weight") items cannot be aligned along a single continuum, as they measure two correlated but distinct DEB domains. Future research should consider our supplementary 3-factor multidimensional IRT results (Supplemental Table 2), and further examine the potential subscales and their corresponding high-risk cutoffs of the DEPS-R.

However, previous factor analytic approaches to identify the DEPS-R factor structure have not yet yielded subscale consensus. For example, studies differ in the number of factors identified by the 16 DEPS-R items (e.g., compare Park et al., 2025 4-factor solution vs. Wisting et al., 2019 3-factor solution), and the solutions with the identical number of factors can use different items to identify each factor (e.g., compare Embaye et al., 2024 vs. Wisting et al., 2019). The oft-selected three-factor solution frequently requires fitting numerous residual correlations to achieve acceptable fit (e.g., 9 residual correlations, Wisting et al., 2019). This complex factor structure of the DEPS-R contrasts with conclusions in most published literature, which refer practitioners and researchers to use the sum score or the dichotomized cutpoint based on the DEPS-R that is assumed to reflect a single-factor construct. Alternatively, DEPS-R scale revision may include dropping low performing items to determine whether this improves DEPS-R factor structure and clinical utility.

Previous factor analytic research on the DEPS has used full information maximum likelihood approaches (e.g., Wisting 2019, ML discrepancy method) to estimate the response distribution. The ML approach inherently assumes continuous responses, yet our model fit results suggest that ML estimators are inappropriate for the inherently categorical DEPS data, as even the MLR estimator did not provide acceptable model fit. ML-based estimators, which recreate the Pearson correlation matrix, even with rescaling robust to non-normality (i.e., MLR estimator), cannot overcome the inherent incongruity between the ordered-categorical response options of the DEPS-R and the traditional ML-based approach that assume continuous responses. The use of inappropriate ML-based estimators likely resulted in the need for an exceedingly high number of residual covariance adjustments in previous research (e.g., Wisting et al., 2019: 4 for the unidimensional and 9 for the 3-factor model; Karastogiannidou et al., 2021: 3 for the unidimensional and the 3-factor models; Atik-Altinok et al., 2023: 9 for the 3-factor model). We achieved acceptable fit with a single residual correlation, and recommend that future DEPS-R measurement researchers use an estimator appropriate for ordered-categorical variables.

Finally, the TIC and SEM functions demonstrated that the DEPS-R is not equally informative across the continuum of the latent DEB. Rather, scale precision varies across the latent DEB trait, with the strongest scale performance among participants exhibiting middle to middle-high theta values, and good performance (i.e., ≥0.80 reliability) among participants ranging from −1.5 to +4 on the latent trait. Even within the “good” reliability range, we showed significant variation in precision, such that some participants (e.g., those at −0.5 SD relative to mean theta) have standard errors more than double of their counterparts at +1.5 SD relative to mean theta. This finding has relevance for studies analyzing change in DEPS-R (e.g., intervention studies), as change scores may have different precision for individuals at different baseline levels on the underlying trait, and may lead to a mischaracterization of changes due to interventions and/or development.

Limitations and Constraints on Generality

The findings from the current study should be considered in the context of its limitations. The sample was predominantly White, college-educated, and female, which may limit the generalizability of the findings to more diverse populations, particularly in terms of race, ethnicity, and gender. Although these demographics align with existing trends noted broadly in eating disorder research (Egbert et al., 2022), they are not necessarily indicative of the overall population of individuals with eating disorders (Halbeisen et al., 2022). Future research should aim to increase the diversity of eating disorder research samples including within the T1D space by, for instance, looking beyond traditional healthcare systems to recruit non-treatment-seeking samples. Such measures can improve sample diversity and generalizability.

Additionally, the reliance on self-reported data introduces potential biases, as participants may underreport or misrepresent sensitive behaviors like insulin restriction or purging. These self-report biases are a known limitation in eating disorder research (Norwood et al., 2016; Vitousek et al., 1991). Finally, our decision to fit a residual covariance between DEPS-R 1 and 6 to improve model fit may have underestimated their discrimination parameters, as their discrimination parameters now represent the impact of θ over above the residual covariance. Although this adjustment was necessary to achieve reliable results for the majority of the items, researchers should interpret our DEPS-R 1 and 6 discrimination results with caution, and seek to replicate them.

Future Directions

The findings also point to important directions for future research. This might include refining the DEPS-R (such as dropping lower performing items, like DEPS-R 1, 2, and 8, and adding the proposed insulin restriction item), and testing it with diverse populations, including racial and ethnic minorities, critical to ensuring its applicability and relevance. Longitudinal studies should explore the predictive validity of the DEPS-R, examining the relationship between revised item performance and long-term outcomes, such as glycemic management, and physical and psychological well-being. Additionally, incorporating the proposed insulin-related item into larger studies will help address a significant gap in the current measure and improve its utility in detecting diabetes-specific DEB. Finally, the development of a shorter screener, using items that are most likely to discriminate individuals in need of additional support (with moderate to high symptom severity), could enhance the tool’s usability in clinical settings, particularly in primary care and endocrinology clinics, where time constraints limit the ability to administer long surveys.

In conclusion, this study underscores the importance of refining the DEPS-R to improve its accuracy and efficiency as a screening tool for DEB in persons with T1D. By addressing gaps in item performance, incorporating a novel insulin-related item, and tailoring the scale to meet diverse clinical needs, this work will contribute to advancing diabetes care and improving both metabolic and psychological outcomes for people living with T1D. As diabetes care continues to evolve, ensuring the accurate identification and treatment of DEB is essential for improving the quality of life and health outcomes for individuals affected by both T1D and eating disorders.

Public Significance Statement.

Disordered eating in type 1 diabetes (T1D) is a significant problem, contributing to diabetes medical complications and premature death. Practitioners need effective and efficient ways to detect these problems so that individuals receive the care that they need. The current study examined individual item- and scale-level usefulness of the Diabetes Eating Problems Survey – Revised (DEPS-R) for this purpose. Findings may improve screening of disordered eating in adults with T1D.

Data Availability Statement:

Data are available upon reasonable request.