Skip to main content

Clustering individuals’ temporal patterns of affective states, hunger, and food craving by latent class vector-autoregression



Eating plays an important role in mental and physical health and is influenced by affective (e.g., emotions, stress) and appetitive (i.e., food craving, hunger) states, among others. Yet, substantial temporal variability and marked individual differences in these relationships have been reported. Exploratory data analytical approaches that account for variability between and within individuals might benefit respective theory development and subsequent confirmatory studies.


Across 2 weeks, 115 individuals (83% female) reported on momentary affective states, hunger, and food craving six times a day. Based on these ecological momentary assessment (EMA) data we investigated whether latent class vector-autoregression (LCVAR) can identify different clusters of participants based on similarities in their temporal associations between these states.


LCVAR allocated participants into three distinct clusters. Within clusters, we found both positive and negative associations between affective states and hunger/food craving, which further varied temporally across lags. Associations between hunger/food craving and subsequent affective states were more pronounced than vice versa. Clusters differed on eating-related traits such as stress-eating and food craving as well as on EMA completion rates.


LCVAR provides novel opportunities to analyse time-series data in affective science and eating behaviour research and uncovers that traditional models of affect-eating relationships might be overly simplistic. Temporal associations differ between subgroups of individuals with specific links to eating-related traits. Moreover, even within subgroups, differences in associations across time and specific affective states can be observed. To account for this high degree of variability, future research and theories should consider individual differences in direction and time lag of associations between affective states and eating behaviour, daytime and specific affective states. In addition to that, methodological implications for EMA research are discussed.


Over the last decades, eating behaviours have become an intensely researched topic, as they are associated with multiple physical health outcomes − e.g., cardiovascular diseases [1] and diabetes [2] − as well as mental health outcomes − e.g., eating disorders [3] and depression [4]. Thus, eating behaviours impact not only individuals’ quality of life but also health care systems [5, 6]. Although eating behaviours are strongly affected by homeostatic processes such as hunger or energy deficits, they can also be expressions of habits, lifestyle [7], and self-regulation [8]. Compared to traditional models that focus on independent homeostatic and hedonic processes, recent models (e.g., regarding appetite control) include ‘cross-talk’ of homeostatic, cognitive and reward mechanisms [9]. Therefore, eating can be viewed as a result of interactions between physiological, affective, and psychological variables. Central non-homeostatic variables influencing eating behaviours include food cravings [10], positive and negative emotions [11] as well as perceived stress [12] / stress coping [13, 14], among others. Given both the importance and the complexity of eating behaviours, several approaches within psychological research were developed to assess the nature and direction of associations between eating behaviour and affective states (hereinafter used as an umbrella term for emotions as well as experiencing and coping with stress [15]). In the following, we will briefly review these approaches and highlight perspectives on variability in associations between eating behaviour and affective states.

From the main effect model to the individual difference model of affect-eating relationships

For several decades, effects of stress on food intake have been documented in both rodent and human studies: While some studies observed an increase in food intake after a stressor, others described the opposite, i.e., a decrease in food intake in response to a stressor [16]. Such effects have been explained in terms of ‘main effect models’, meaning that, depending on the theory, stress either increases or decreases food intake. To illustrate, different models have tried to explain these effects psychologically (e.g., eating more to alleviate negative feelings [17]) and physiologically (e.g., acute stress leading to a fight-or-flight based reduction of appetite, thus, decreased eating vs. chronic stress leading to increased eating driven by the hypothalamic-pituitary-adrenal axis [16]). Yet, human individuals can differ notably in the direction and strength of such stress-eating relationships, which led to the individual difference model [18]: Some individuals increase their eating behaviour in reaction to stress or negative affect, while others decrease it or show no effect on eating, resulting in pronounced between-subject variability that can be (partially) explained by individual characteristics. To assess this variability, trait-level questionnaires were developed to capture individual differences, such as the Dutch Eating Behavior Questionnaire (DEBQ [19]) tapping into a range of negative emotional states and measuring to what degree these states are reported to increase subsequent craving or eating.

However, as different affective states are related to different physiological responses, their effects on eating behaviour likely differ as well. Thus, further differentiation on the level of affective states became necessary: First, stress has been differentiated from various negative emotions (e.g., anxiety or anger) in recently developed psychometric scales: The Salzburg Stress Eating Scale (SSES [20]) and the Salzburg Emotional Eating Scale (SEES [21]) revealed that stress eating appears to be a distinct concept compared to emotional eating. Second, within emotions, negative and positive emotions need to be distinguished from each other and even within negative emotions, highly arousing emotions such as anger or anxiety showed distinct associations with eating in comparison to low arousal emotions such as sadness [21]. Thus, as the field of stress-eating and emotional eating matured, the number of possible affect-eating relationships increased and with that theories and approaches became more complex to account for this variability [17, 22].

Individual and temporal differences: ecological momentary assessment studies

While trait questionnaires revealed strong between-subject variability and marked differentiation on the level of affective states, they do not account for momentary and contextual influences and circadian patterns that characterise naturalistic eating behaviours. Instead of relying on one-time, trait questionnaires inquiring about ‘general’ or ‘typical’ affect-eating patterns [17, 23], ecological momentary assessment (EMA) can survey affective states and eating behaviour multiple times a day in natural settings. Typically smartphone-based, EMA gathers ‘real-time’ and ‘real-world’ data through repeated short questions across the day, thereby limiting retrospective memory biases [24]. Importantly, multiple assessments per day allow a lagging of variables: relating one variable at measurement point tx (e.g., negative mood at noon) to another variable at a subsequent measurement point tx + 1 (e.g., food craving at 4 pm) thus uncovering temporal and potentially causal effects of one variable on another. Studies using EMA in addition to questionnaire data have for instance shown that relationships between stress and eating behaviour can differ based on the time point and aggregation level (i.e., associations lagged within a day vs. overall day level associations after aggregation [25]).

Accounting for variability within and between individuals in EMA data: multilevel models and network models

Several analytical approaches to the variability inherent in EMA data have been taken. Theory-driven analytical approaches using multilevel modelling (MLM [26]) relate momentary (within-subject, Level 1) variables to each other, while the variables and the direction of influence/lagging (e.g., affect influences eating instead of eating influencing affect) are predetermined by the researcher and thus also lead to a limited perspective on associations assumed. This statistical technique also allows for the simultaneous modelling of trait level questionnaires (between-subject, Level 2), e.g., stress and eating behaviour are associated within participants with high scores on the SSES [25]. Exploratory approaches, by contrast, do not constrain the choice of variables that vary within individuals: Vector-autoregressive (VAR) models allow calculating associations of variables among each other based on their previous values and by considering the respective autoregression of Xt on Xt + 1 (each variable predicts itself on a future time point) [27]. Besides contemporaneous and between-subject analyses, multilevel vector-autoregression (mlVAR [28]) can fit a temporal network model based on EMA (time-series) data. Thus, directed (i.e., lagged) and unique (i.e., partial correlative) effects of Xt on Yt + 1 and Yt on Zt + 1 (etc) can be obtained. The temporal network structure is estimated by considering all variables of interest and their unique lagged associations among each other [27]. A network analysis study by our group found that higher anticipated stress coping preceded higher goal-congruent eating [14]. Novel findings revealed through network models illustrate how exploratory analytical approaches can influence theory building.

Exploring similar patterns: latent class vector-autoregressive modelling

Although such a temporal network structure is rather comprehensive, it lacks the consideration of between-subject differences in the identified within-subject temporal relationships. MLM, by contrast, requires a theory-driven choice for Level 2 (between-subject) data as potential moderators of associations (e.g., trait questionnaires). Until recently, it was not possible to determine between-subject variability in within-subject associations in a data driven manner. The development of latent class vector-autoregression (LCVAR) closes this gap by exploring latent classes of individuals (i.e., distinct subgroups within a sample) based on similarities in lagged Level 1 associations, thus based on similar temporal dynamics [29]. Furthermore, LCVAR provides opportunities to explore within-subject variability: Several lags can be considered, for instance, stress at tx might be associated with other variables not only at tx + 1 but also at tx + 2 or even at tx + 3, while the strength and direction of these relationships might differ between lags. This important complexity has already been noted in EMA research [25] but has not been addressed comprehensively so far. In the following, we will briefly describe LCVAR, its advantages and its potential for affective eating behaviour research.

Latent class vector-autoregressive models

LCVAR models can be characterised as multilevel models based on repeated measures (i.e., occasions) at Level 1 that are nested within individuals at Level 2. In LCVAR, time-series Level 1 data provide covariate coefficients (concerning trends and exogenous variables, such as time) and a lagged VAR model (concerning dynamics). Using VAR coefficients, dynamic processes can be investigated as effects of the endogenous variables at tx on themselves at tx + 1, resulting in autoregressive and cross-regressive coefficients. Time-series data are handled by maximum likelihood estimation which allows modelling intensive and unequally long time-series data. LCVAR utilises the Expectation Maximization (EM) algorithm [30] to estimate probabilistic latent class membership and class-wise time-series parameters. Thus, at Level 2, LCVAR models latent classes (i.e., clusters) with fixed coefficients for each cluster. The method utilises an adaptive model estimation which can result in different numbers of lags for each cluster. Setting an information criterion such as the Hannan-Quinn (HQC [31]) or the Akaike information criterion (AIC) enables a model selection of different cluster solutions (i.e., solutions with different numbers of clusters and different numbers of lags). Lastly, by considering interpretability and estimation convergence, a final cluster solution is chosen. Such a solution contains several distinct clusters with (potentially different) lag orders. Each individual is assigned to one cluster only. For detailed information on LCVAR, please also see the original paper [29].

Aims of the present study

The goal of our study was to apply the recently developed exploratory method of latent class vector-autoregressive modelling [29] to naturalistic affective and eating behaviour research. We aimed to explore between-subject differences in the dynamics of affective states, hunger, and food craving (i.e., urges to eat specific foods [32]). Such temporal associations might be similar across some individuals and could thus form clusters of dynamic patterns. Given the exploratory approach, hypotheses are tentative only. Overall, we expect high variability regarding a) specific associations (coefficients) between specific affective and eating-related variables (e.g., anger might relate differently to subsequent craving than sadness does, or craving might increase subsequent anger), b) regarding temporal resolution of relationships (based on different lags), and c) regarding differences between clusters (positive associations in one cluster, negative associations in another) as suggested by the respective questionnaire-based research reviewed above. Lastly, we aimed to characterise clusters using visualisations of daytime courses of all variables, cross-sectional questionnaire data and EMA completion rates.



The data analysed in the present study were taken from a mid-study subsample of individuals who took part in a randomised controlled trial (RCT) that was preregistered at the German register of clinical studies (DRKS, DRKS00017493). The RCT was based on a three-week EMA study during which the active group received eating behaviour related short messages. Data of both groups (i.e., active and control group) were included in the present analyses. For the exploratory LCVAR, data are based on 14 days of EMA that constituted the ‘main phase’ (which served as the ‘treatment’-phase for the active group, respectively as the control phase) of the superordinate study. Individuals were recruited for the RCT through study announcements via e-mail at several universities in Austria and Germany, Facebook groups, Twitter, and by word of mouth. Similar to previous study inclusion criteria [14], agreeing to one of the following two questions enabled individuals to take part in the study: 1) ‘Do you currently pay attention to your nutrition to maintain or reduce your body weight’? and 2) ‘Do you currently cut down on your food intake to maintain or reduce your body weight’? Thus, generally, diet-interested individuals were recruited. Furthermore, it was required to own an Android smartphone to install the app for the EMA. In the beginning, participants reported demographic variables and answered several trait questionnaires (see Trait questionnaires). After that, participants were instructed on how to use our customised smartphone app ‘SmartEater’ that was installed manually on participants’ Android smartphones and that served as the assessment tool for the EMA. At the beginning of the study, participants received written and oral information on the study and signed an informed consent form according to the ethics committee at the University of Salzburg, Austria that granted ethical approval (EK-GZ: 37/2018). After the EMA period, participants answered questions regarding subjective reactivity: 1) “How much did the assessment itself change your eating behaviour?” (0 = negatively, 50 = unchanged, 100 = positively) and 2) “Did you eat differently than you would normally have in some situations because of the app answers?” (0 = not at all, 50 = partly, 100 = very much). Both questions were answered on a scale from 0 to 100 in steps of ten. At the end of the study, participants received visualised feedback of their EMA data and financial compensation or course credits for their study participation. Please note that other parts of the dataset have previously been analysed and published [33].

EMA measures

EMA data used in this study were gathered by signal-contingent sampling: Six times a day (9:00 a.m., 11:30 a.m., 2:00 p.m., 4:30 p.m., 7:00 p.m., 9:30 p.m.) participants received app notifications (‘beeps’) on their smartphones. Each time, they were asked to rate their momentary affects, stress (coping), hunger and food craving on a horizontal, continuous rating slider from 0 (= not at all) to 100 (= very much). For a list of variables measured but not used in the present study, please see the ‘List of additional variables’ in the supplementary materials. Items used in this study have also been used in previous studies [14]: The affect items were partially taken from the Positive and Negative Affect Schedule (PANAS) with additional items added that feature different thresholds or levels of arousal. Positive affect items included ‘cheerful’, ‘enthusiastic’, ‘relaxed’, ‘calm’ and ‘active’, whereas negative affect items included ‘irritated’, ‘worried’, ‘depressed’, ‘bored’ and ‘nervous / stressed’. Stress coping was also assessed similar to previous research [14]: Participants reported their present stress coping efficacy (‘Do you feel that you are on top of things?’) and their anticipated stress coping efficacy (‘Do you feel that you can cope with all upcoming things that you will have to do?’). The questions were worded in the style of the Perceived Stress Scale (PSS [34]) and were adapted to a state level. Furthermore, hunger and food cravings were also assessed as in previous studies [35]: ‘How strong is your hunger right now?’ (hunger) and ‘How strong is your desire to eat certain foods right now?’ (food cravings). Participants were instructed that ‘certain foods’ refer to specifically tasty foods. For reasons of safety (e.g., while driving) and practicability, participants had up to 1 h to answer the questions of each EMA beep. Thereafter, the respective EMA questionnaire disappeared and was registered as missing values.

Trait questionnaires

Before the EMA period, participants answered several online questionnaires that were presented via the platform LimeSurvey [36]. The following trait-questionnaires were used for the present study and aimed to capture between-subject variability: the three subscales (emotional eating, restrained eating and external eating) of the Dutch Eating Behavior Questionnaire (DEBQ [19]) that queries different influences on eating, the Food Cravings Questionnaire Trait reduced (FCQ-T-r [37]) which assesses trait-level experiences of food craving, the Salzburg Stress Eating Scale (SSES [20]), and the four subscales (anger, anxiety, happiness, sadness) of the Salzburg Emotional Eating Scale (SEES [21]). The SSES and SEES assess how individuals perceive stress, respectively specific emotions to influence the amount of food they consume.

Latent class vector-autoregressive modelling

All analyses were run in R (version 4.0.3) using RStudio (version 1.3.1093). We analysed inter-individual patterns of associations among the fourteen variables described above (i.e., five positive affect items, five negative affect items, two stress coping items, one hunger item and one food craving item). For the analyses, we included participants with high adherence/compliance, i.e., completing at least 60% of all EMA questionnaires (i.e., at least 50 out of 84 EMA questionnaires across the 14 days). Having at least 50 observations per person is recommended for running an LCVAR [29]. Missing values were imputed using ‘Multivariate Imputation by Chained Equations’ (MICE), as implemented in the mice package [38], with 500 iterations of a single imputation. All functions necessary for the LCVAR were derived from GitHub [39]. Since the present study is the first to apply LCVAR to eating behaviour research, our theoretical assumption of cluster numbers is based on the literature reviewed above, which allows, among others, two hypothetical (general) distinctions: Firstly, individuals might be grouped together based on a ‘dichotomous’ main effect model (i.e., observable associations vs. no observable associations between affective states and eating behaviour). Secondly, individuals might also be clustered into three groups based on similar ‘individual differences’ (e.g., direction of coefficients) regarding these associations (i.e., no observable associations vs. mostly positive associations vs. mostly negative associations between affective states and eating behaviour). Previous studies have reported similar subgroup distinctions regarding stress and the amount of food consumed [20, 40]. Given these theoretical assumptions, the exploratory nature of our study, and the present (limited) sample size we thus set the minimum number of clusters to two and the maximum number of clusters to three. Regarding cluster memberships, we decided to require each cluster to have at least two members. Moreover, we set the lowest lag required to one and the highest potential lag to three which considers the assessment frequency given and provides a plausible timeframe for temporal associations of affective states and eating behaviour. As previously suggested [29], we used the extended Hannan-Quinn information criterion (HQC) for LCVAR as it enables finding a model solution that ideally combines lags and memberships of clusters. Furthermore, we oriented our LCVAR settings towards the ones of the original article describing the analysis [29]. We set the number of pseudo-random starts (i.e., randomly selecting k individuals as cluster centres) of the EM algorithm to 15 for fitting each potential model, while rational starts (i.e., k-means partitioning based on individuals’ VAR and coefficients of covariates) of the algorithm were also used. When posterior probabilities of cluster memberships were reset, every element of Sigma was increased by 10. After every start of the EM algorithm, 50 EM-iterations were allowed before this EM-start had to terminate. Lastly, to determine the convergence of an EM-start, the convergence criterion of the log-likelihood was set to 1e-07. In summary, our present LCVAR combined each lag order, i.e., one, two or three lag(s) for both numbers of clusters, i.e., two or three clusters. Each of these combinations was based on its own statistical model. The start of the EM algorithm was based on 15 pseudo-random initialisations, k-means based rational initialisations and previous solutions. After a maximum of 50 EM-iterations or convergence of the likelihood, each EM-start led to a solution, eventually resulting in multiple solutions out of which the ideal solution for two, respectively, three clusters (across all lag combinations and EM-starts) was chosen based on interpretability (i.e., theoretically plausible differentiation between patterns of lagged associations) and the HQC. Resting upon our sampling plan and frequency, we additionally used the time of day at which participants answered EMA questionnaires as the exogenous variable in the analyses. The exogenous variable allows plotting the mean values of each endogenous variable (i.e., each item) at all six time points (‘beeps’) for all clusters separately which allows insights into daytime courses of the items. For more details regarding potential parameters, settings and optional arguments of LCVAR, please see the original paper [29]. To support the understanding of the meaning of the present LCVAR results, we additionally added a simulated LCVAR to the supplementary materials (see 'Simulated random data LCVAR') for which we used the same settings as described above but randomised data, as suggested by a reviewer of this paper.


Participants’ descriptive statistics

The initial sample of the present study consisted of 136 participants. Twenty-one participants were excluded from the analyses because they completed less than 60% of all EMA questionnaires. Thus, 115 individuals (i.e., a total number of 9660 Level 1 occasions) were included in the LCVAR. Because of missing descriptive data for one participant, descriptive statistics and additional analyses are based on 114 participants, of whom 95 were female (83.3%) and 19 were male (16.7%). The mean age was 23.2 years (SD = 3.81 years, range = 18–40 years). The group allocation of individuals was rather balanced with 56 participants out of the treatment group (49.12%) and 58 participants out of the control group (50.88%). The mean BMI was 23.32 kg/m2 (SD = 3.72 kg/m2, range = 17.8–37.65 kg/m2). Participants reported a low to moderate influence of the assessment itself (M = 35.4, SD = 23.4, N = 112) and of their app answers (M = 30.8, SD = 26.5, N = 112) on their eating behaviour.

Latent class vector-autoregression

The LCVAR resulted in two cluster solutions. Solution 1 contained two clusters with three lags each. Solution 2 contained three clusters, also with three lags each. Given the slightly smaller HQC of Solution 2 (HQC2 = 77.72) compared to Solution 1 (HQC1 = 77.86), we decided to further investigate Solution 2 containing three clusters. Moreover, as it may be reasonable to obtain clusters that differ regarding the ‘roles’ of hunger and food craving (i.e., negative/positive, respectively, small/large coefficients) in the respective clusters, we chose the solution containing three clusters displayed as heatmaps in Fig. 1 for all further analyses. Additionally, this solution appeared to capture variability both within and between clusters. Eight EM-iterations were repeated before the forced end of the algorithm. Cluster membership proportions were distributed as follows: Cluster 1: 46.1%, Cluster 2: 31.3%, Cluster 3: 22.6%. Since we used the time of day at which participants answered EMA questionnaires as the exogenous variable in our LCVAR, we plotted the mean values of each endogenous variable (item) at all six daily time points for all clusters separately (see Fig. 2).

Fig. 1
figure 1

Heatmaps of the selected cluster solution. Note. VAR = vector-autoregressive, _x = x lags. Vertical lines were added to visualise the beginning of a new lag. Words are in bold for visualisation purposes only. For full item meanings please see ‘EMA measures’ in the Methods section

Fig. 2
figure 2

Mean values of all fourteen endogenous variables for all three clusters. Note. X-axis: daytime, y-axis = mean values of the respective clusters. Range of all item scales: 0 to 100. Enthus. = Enthusiastic, Relax. = Relaxed, P.coping = Present stress coping, A.coping = Anticipated stress coping, N/stress. = Nervous / stressed, Depress. = Depressed. Colours are based on colour palettes from the wesanderson package [41]

In the following, we will focus specifically on variability concerning hunger and craving-related associations (see the first two rows as well as the two columns before the vertical ‘lag lines’ in the cluster heatmaps in Fig. 1). The LCVAR resulted in three subgroups in which the direction and strength of coefficients appear to differ both across lags within clusters and – expectedly - between clusters. The ‘daytime courses’ in Fig. 2 display Cluster 2 as the most fluctuating of the clusters regarding the endogenous variables, with the lowest mean positive and highest mean negative affects in general. Moreover, Cluster 2 also showed the strongest increase in cravings, with high values especially in the afternoon. By contrast, Cluster 3 showed the strongest but also the most differentiated associations between craving and hunger on the one hand and subsequent affective states on the other hand: unexpectedly, the associations of affective variables with subsequent hunger and food craving were very small (‘emotional eating’, see last two columns in each lag quadrant of the heatmap). By contrast, the reverse direction of association emerged in Cluster 3 (‘emotional about eating’, first two rows of the heatmap). Food craving and hunger were associated with subsequent affective states in different ways: Food craving showed negative coefficients for subsequently feeling calm (βlag_1 = −.154, βlag_2 = −.156) and anticipated stress coping (βlag_2 = −.157) while also positive associations with subsequent positive emotions were observed, e.g., cheerful (βlag_1 = .049) and active (βlag_1 = .100). Also in Cluster 3, hunger showed negative coefficients with subsequently feeling depressed (βlag_1 = −.140, βlag_2 = −.149): experiencing more hunger preceded feeling less depressed. For all associations between affective states, hunger, and food craving please also see Fig. 1, respectively Fig. 2 (for daytime courses) as well as the ‘Coefficients tables’ in the supplementary materials.

Characterising clusters using questionnaires and completion rates

After obtaining the algorithm-based cluster solutions, we aimed to characterise the clusters by analysing potential Level 2 (i.e., between-subject data) differences between clusters by means of univariate ANOVAs. Table 1 displays the results of the ANOVAs including the above-mentioned questionnaires (see Trait questionnaires in the Methods section) and completion rates. Post hoc contrasts revealed specific significant differences between clusters regarding cross-sectional data: Cluster 3 had the significantly lowest mean FCQ-T-r score (mean = 37.58) compared to Cluster 1 (Tukey multiple comparisons of means (Tukey HSD), ∆(3-1) = − 7.52, p = .031) and Cluster 2 (Tukey HSD, ∆(3-2) = − 11.57, p = .001). Moreover, Cluster 3 also had the significantly highest mean compliance (mean = 0.92) compared to Cluster 1 (Tukey HSD, ∆(3-1) = 0.08, p = .004) and Cluster 2 (Tukey HSD, ∆(3-2) = 0.10, p = .001). In addition, Cluster 3 had a significantly lower mean SSES score (mean = 2.84) compared to Cluster 2 (Tukey HSD, ∆(3-2) = − 0.52, p = .022), however it did not differ significantly from Cluster 1 (Tukey HSD, ∆(3-1) = − 0.29, p = .246). Lastly, Cluster 2 showed the lowest mean SEES happiness score (mean = 2.74) compared to Cluster 1 (Tukey HSD, ∆(2-1) = − 0.28, p = .038) and Cluster 3 (Tukey HSD, ∆(3-2) = 0.32, p = .049). Please also see Fig. 3 that displays raincloud plots illustrating the mean, confidence intervals, and information on distributions of the four significant differences in cross-sectional data (i.e., compliance, SSES, SEES happiness, FCQ-T-r) between clusters. In addition, we conducted a Pearson’s Chi-squared test regarding group allocation in the RCT. Cluster memberships did not differ between the active and the control group, X2 (2, N = 114) = 4.14, p = .126. Moreover, since only a few men participated in the present study, we conducted Fisher’s exact test instead of Pearson’s Chi-squared test regarding gender and cluster memberships. The test revealed a significant difference (p = .019, Fisher’s exact test). In Cluster 1, there were 41 females (77.4%) and 12 males (22.6%), while in Cluster 2, there were 34 females (97.1%) and only one male (2.9%), whereas in Cluster 3, there were 20 females (76.9%) and six males (23.1%).

Table 1 Differences in trait characteristics between clusters
Fig. 3
figure 3

Raincloud plots of cross-sectional data of the three clusters. Note. N = 114. Only variables that revealed significant differences between clusters are displayed (see also Table 1 for all ANOVA results). Plots include cluster specific means, 95% confidence intervals, information on distributions as well as added random jitter. Y-axis = clusters, x-axis = respective score. CI = confidence interval, EMA = ecological momentary assessment, FCQ-T-r = Food Cravings Questionnaire Trait reduced, SSES = Salzburg Stress Eating Scale, SEES Happiness = Salzburg Emotional Eating Scale (Happiness Subscale). Colours are based on colour palettes from the nord package [42]. For more information on raincloud plots, please see the original paper [43]


The present study applied the novel analytical approach of latent class vector-autoregression (LCVAR) to affective eating behaviour research. We aimed to explore inter-individual patterns of lagged associations between momentary measures of affect, stress (coping), hunger, and food craving. We analysed 14 days of EMA data of 115 diet-interested individuals. The LCVAR resulted in three clusters with three lags each. Lastly, we characterised the clusters by means of trait-level data and completion rates of EMA questionnaires. Overall, the LCVAR results highlight the substantial variability in the relationships between affective states and hunger/food craving. Results support the assumption that temporal patterns of associations between affective and eating-related states might be similar across some individuals who could be clustered to form subgroups within a population. This appears plausible given the concept of ‘individual differences’ regarding the influence of affective states on eating behaviours [17, 18].

Characterisation of clusters

In the following, we will characterise the three clusters using the LCVAR coefficients (Fig. 1), the daytime courses (Fig. 2) and the results of the ANOVAs and Tukey HSD, (see also Table 1 and Fig. 3). We will discuss the clusters in reverse order beginning with Cluster 3. Firstly, Cluster 3, the ‘smallest’ cluster containing 22.6% of participants, showed the most pronounced coefficients regarding associations between affective states and hunger/craving (see the first two rows in the heatmap in Fig. 1). Higher food cravings in Cluster 3 preceded feeling less calm (at one and two lags) and lower anticipated stress coping. By contrast, higher hunger was followed by feeling less depressed. Thus, in this cluster, hunger and craving appear to have specific and partially different associations with subsequent mood. Previous research has already differentiated effects of hunger and craving in that hunger might positively influence goal-congruent eating, while craving might have slightly negative influences on goal-congruent eating [14]. This is in line with the literature drawing further distinctions between the two [44]: e.g., hunger was negatively related to snack food consumption when taking craving intensity into account [45]. Based on the ANOVAs, Cluster 3 might be considered hosting ‘healthy eaters’: Participants in Cluster 3 had the lowest mean FCQ-T-r score of all clusters (thus most likely included individuals with rather low trait food craving) and descriptively showed the lowest mean state cravings at all six intra-day measurement time points (Fig. 2). In addition, Fig. 3 illustrates that Cluster 3 showed balanced happiness eating (in this case: mostly eating as usual when feeling happy) and balanced stress eating (in this case: mostly eating slightly less to eating as usual when stressed). Importantly, contrary to what we had expected in general, we mainly observed associations between hunger/food craving and subsequent affective states (‘emotional about eating’), i.e., the opposite direction as traditionally assessed via the above-mentioned trait questionnaires (‘emotional eating’ [46], please also see the General discussion and methodological implications below). Lastly, from a methodological point of view, results might be influenced by a high compliance: The subsequent ANOVAs revealed that Cluster 3 had the significantly highest mean completion rate of EMA questionnaires of all three clusters.

Secondly, Cluster 2 (containing 31.3% of all participants) could be described as the ‘affectively troubled’ cluster, based on rather pronounced associations between affect and eating behaviour, which might not be revealed at first glance by the coefficient heatmap (Fig. 1). However, Cluster 2 showed the highest mean negative affect scores and the lowest mean positive affect scores of all clusters in the ‘daytime courses’ (Fig. 2). Interestingly, results suggest a simultaneous increase in food craving and negative affect as well as a decrease in positive affect (which were most pronounced in Cluster 2).Footnote 1 Furthermore, the ANOVAs and the post hoc contrasts indicate that Cluster 2 might be the subgroup with the lowest happy eating (SEES happiness) and the highest stress-eating (SSES) traits (see also Fig. 3) which might correspond to the above-mentioned daytime courses. Additionally, although not differing significantly from Cluster 1, Cluster 2 showed a relatively high mean value of trait craving (FCQ-T-r [47]), especially compared to Cluster 3. Interestingly, the ratio of males in Cluster 2 was notably smaller than in the other two clusters. Considering the above-mentioned results of happy eating and stress-eating in Cluster 2, results correspond to previous findings that women might show higher stress-eating than men (thus the high value in this cluster), while men might show higher happy eating than women (thus the low mean value in this cluster) [20, 21].

Lastly, Cluster 1 is characterised by small coefficients that do not suggest strong lagged associations between these two eating-related variables and affective states (see first two rows as well as the two columns before the lag lines). The results of the ANOVAs and the post hoc contrasts underline the assumption of Cluster 1 being the ‘average’ subgroup within our sample, as the main scores of trait craving (FCQ-T-r), stress-eating (SSES), happy eating (SEES happiness) and the main completion rate of answered EMA questionnaires are mostly between those of Cluster 2 and 3.

General discussion and methodological implications

Most notably, associations of eating behaviours (hunger and craving) with subsequent affective states were stronger than the reverse direction of association. Although most trait questionnaires and most research on emotional eating focus on emotion-induced eating behaviour [11, 48, 49], the reverse direction, i.e., affective changes in response to eating or food stimuli, is equally plausible and has also been documented [17, 50, 51]. The present results thus suggest that future research should routinely consider both directions of association between affect and appetitive behaviours. With the reverse direction, the question of actual food intake is pressing, which however was not included in the present study, as only momentary measures (instead of retrospective ones) were of interest. Yet, it is reasonable to assume that hunger and craving peak around lunch and dinner time and thus might peak in food intake which has been previously suggested by correlational data [44]. Interestingly, craving was negatively associated with subsequently feeling relaxed in Cluster 3, suggestive of potential non-rewarding food intake experiences, yet it was positively associated in Cluster 1 and 2. These between-cluster differences suggest that not all individuals respond to cravings (and potentially food intake) in a similar manner: For some individuals, (yielding to) food cravings might increase positive affect, for instance because of reward effects [52]. In this context, diet-interested individuals (as in the present study) might also crave healthy foods [45] that may be diet-goal-congruent. In contrast, other individuals might respond to cravings with increased negative affect or decreased positive affect, either because cravings might be perceived as potential threats to diet / weight goals [10] or because of actual subsequent consumption of high calory (i.e., often diet-incongruent) foods [45]. Thus, differences in individual coefficients can be based on different subjective experiences, making multiple ‘concurrent’ explanations possible. Besides these specific differences between clusters, craving also showed similarities across clusters: The daytime courses (Fig. 2) permit insight into diurnal trends of hunger and craving. All three clusters showed similar trends with higher values of hunger around midmorning and late afternoon as well as higher values of cravings in the afternoon. Such general trends (‘M-curve’) are in accordance with the results of previous studies concerning hunger and craving [44] as well as latent classes of eating patterns [53].

Importantly, we observed variability both between and within clusters: In Cluster 3, both the direction and the strength of coefficients can be observed to change most prominently across lags. For instance, the association between hunger and feeling depressed was notable for one and two lags but decreased with three lags. Additionally, the relationship between momentary craving and stress coping appeared to be slightly negative for one lag but positive for two lags. The variability within clusters underlines the relevance of a well-considered temporal resolution when studying relationships between affective states and eating behaviour: Assessment frequency and length of intervals between (thus the length of lags) might influence the direction and strength of results obtained later. Therefore, it appears reasonable to sample with a high frequency to best capture potential lagged associations and at the same time keep questionnaires short to reduce participant burden [54]. Lastly, a notable finding is that up to three lags of associations contributed to the best fit of the cluster solution (and even to the second-best fitting model). This indicates that the correlations between momentary data can ‘spread’ (systematically) over several hours – at least within certain subgroups or individuals. Such insights may indicate regularity and potentially even predictability in the oscillation of appetitive-affective experience in (different subgroups of) individuals which might relate to predictive modelling or machine learning approaches as often used in the context of just-in-time adaptive interventions [55].

Limitations and future research

We utilised LCVAR to analyse time-series data aiming to identify potential subgroups. However, the existence of subgroups (i.e., clusters) within our data is an assumption of our model. Moreover, our description and characterisation of the clusters were mainly based on eating-related variables while also all associations among affective states contributed to cluster separation. For instance, lagged associations between different affective states might have been equally or even more influential for cluster membership allocation. Our clusters might thus reflect between-subject differences in affective dynamics just as well as they do reflect affective-appetitive dynamics. Therefore, we aimed to characterise the clusters using eating-related trait questionnaires in subsequent ANOVAs, suggesting that cluster separation might be related to differences in affective eating behaviours. Clearly, future studies should characterise clusters also with affective (e.g., emotion regulation, stress coping) trait questionnaires. Furthermore, we explicitly focussed on diet-interested individuals since hunger and food craving can play important roles in dieting. However, results of this mostly young and female sample (which is not representative for the general population) have a limited generalisability. Moreover, we did not directly analyse actual eating behaviour (e.g., retrospectively reported amounts of consumed foods), but related appetitive states (hunger and craving). Therefore, we did not take into account whether participants consumed foods after reporting high hunger or cravings and which eating behaviours actually took place between two measurements. Although this might provide important information, especially in hindsight regarding the surprisingly reverse direction of effects observed, we had decided to focus on momentary measures of states that were assessed simultaneously. From an analytical point of view, at the time of this publication, LCVAR does not yet support nested structures (e.g., measurements within a day). Thus, each day’s last measurement was used to predict the next day’s first value. Therefore, coefficients in the present study might underestimate associations, as the ‘predictive power’ of such overnight associations might be lower than intra-day ones. However, we refrained from interpolating values to add them ‘between two days’, because it does not appear reasonable to estimate hunger and craving (but also affect) values for night hours. Moreover, we set the threshold to at least 50 repeated measures per person to ensure good performance of the LCVAR [29], though a higher number of participants and even more measurement time points might be beneficial in future studies. Lastly, because of the exploratory nature of the present study, the results need to be replicated in order to draw concrete conceptual implications.


LCVAR appears to be a promising approach to investigate between-subject differences as well as variability in affect and stress-related eating behaviours. LCVAR enables researchers to visually explore temporal patterns in the clusters obtained: Specific associations between hunger/craving on the one hand and specific affective states on the other hand can be compared across lags and clusters. Thus, LCVAR makes specific, data-driven exploration of temporally lagged relationships possible and visible, both within and between clusters. In the present study, associations between hunger/craving and affective states differed across variables as well as across lags within and between clusters. Temporal associations between hunger/craving and subsequent affective states were more pronounced than vice versa. The present study revealed empirically how highly variable affect-eating-relationships are. We exemplified the potential of LCVAR with a set of affective and eating-related variables, however, the present paper is not meant as a ‘how-to-guide’ and other (subjective or objective) variables may of course be of great interest for health or clinical research. Different analytical approaches might be combined to fully exploit the potential of such time-series data analyses, both in a data-driven as well as theory-based way. Future applications to other populations, e.g., individuals with eating disorders, might reveal clearer ‘patterns’ of relationships between affective states and eating behaviour [56] and may also help to modify existing theories and develop novel ones.

Availability of data and materials

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.


  1. This finding is not visible in the heatmaps, as contemporaneous associations are not included, however, based on the daytime courses corresponding patterns might be plausible.


  1. Shan Z, Li Y, Baden MY, Bhupathiraju SN, Wang DD, Sun Q, et al. Association between healthy eating patterns and risk of cardiovascular disease. JAMA Intern Med. 2020;180(8):1090–100.

    Article  PubMed  Google Scholar 

  2. Basu S, Yoffe P, Hills N, Lustig RH. The relationship of sugar to population-level diabetes prevalence: an econometric analysis of repeated cross-sectional data. PLoS One. 2013;8(2):e57873.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  3. Walsh BT. The importance of eating behavior in eating disorders. Physiol Behav. 2011;104(4):525–9.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  4. Lassale C, Batty GD, Baghdadli A, Jacka F, Sánchez-Villegas A, Kivimäki M, et al. Healthy dietary indices and risk of depressive outcomes: a systematic review and meta-analysis of observational studies. Mol Psychiatry. 2019;24(7):965–86.

    Article  PubMed  Google Scholar 

  5. Ágh T, Kovács G, Supina D, Pawaskar M, Herman BK, Vokó Z, et al. A systematic review of the health-related quality of life and economic burdens of anorexia nervosa, bulimia nervosa, and binge eating disorder. Eat Weight Disord. 2016;21(3):353–64.

    Article  PubMed  PubMed Central  Google Scholar 

  6. van Hoeken D, Hoek HW. Review of the burden of eating disorders: mortality, disability, costs, quality of life, and family burden. Curr Opin Psychiatr. 2020;33(6):521–7.

    Article  Google Scholar 

  7. Chapman CD, Benedict C, Brooks SJ, Schiöth HB. Lifestyle determinants of the drive to eat: a meta-analysis. Am J Clin Nutr. 2012;96(3):492–7.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  8. Dohle S, Diel K, Hofmann W. Executive functions and the self-regulation of eating behavior: a review. Appetite. 2018;124:4–9.

    Article  PubMed  Google Scholar 

  9. Higgs S, Spetter MS, Thomas JM, Rotshtein P, Lee M, Hallschmid M, et al. Interactions between metabolic, reward and cognitive processes in appetite control: implications for novel weight management therapies. J Psychopharmacol. 2017;31(11):1460–74.

    Article  PubMed  PubMed Central  Google Scholar 

  10. Boswell RG, Kober H. Food cue reactivity and craving predict eating and weight gain: a meta-analytic review. Obes Rev. 2016;17(2):159–77.

    Article  PubMed  Google Scholar 

  11. Evers C, Dingemans A, Junghans AF, Boevé A. Feeling bad or feeling good, does emotion affect your consumption of food? A meta-analysis of the experimental evidence. Neurosci Biobehav Rev. 2018;92:195–208.

    Article  PubMed  Google Scholar 

  12. O'Connor DB, Jones F, Conner M, McMillan B, Ferguson E. Effects of daily hassles and eating style on eating behavior. Health Psychol. 2008;27(1, Suppl):S20–31.

    Article  PubMed  Google Scholar 

  13. Spoor STP, Bekker MHJ, van Strien T, van Heck GL. Relations between negative affect, coping, and emotional eating. Appetite. 2007;48(3):368–76.

    Article  PubMed  Google Scholar 

  14. Pannicke B, Kaiser T, Reichenberger J, Blechert J. Networks of stress, affect and eating behaviour: anticipated stress coping predicts goal-congruent eating in young adults. Int J Behav Nutr Phys Act. 2021;18(1):9.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Gross JJ. Emotion regulation: current status and future prospects. Psychol Inq. 2015;26(1):1–26.

    Article  Google Scholar 

  16. Torres SJ, Nowson CA. Relationship between stress, eating behavior, and obesity. Nutrition. 2007;23:887–94.

    Article  PubMed  Google Scholar 

  17. Macht M. How emotions affect eating: a five-way model. Appetite. 2008;50:1–11.

    Article  PubMed  Google Scholar 

  18. Greeno CG, Wing RR. Stress-induced eating. Psychol Bull. 1994;115:444–64.

    Article  CAS  PubMed  Google Scholar 

  19. van Strien T, Frijters JE, Bergers G, Defares PB. The Dutch eating behavior questionnaire (DEBQ) for assessment of restrained, emotional, and external eating behavior. Int J Eat Disord. 1986;5(2):295–315.

    Article  Google Scholar 

  20. Meule A, Reichenberger J, Blechert J. Development and preliminary validation of the Salzburg stress eating scale. Appetite. 2018;120:442–8.

    Article  PubMed  Google Scholar 

  21. Meule A, Reichenberger J, Blechert J. Development and preliminary validation of the Salzburg emotional eating scale. Front Psychol. 2018;9:88.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Reichenberger J, Schnepper R, Arend A-K, Blechert J. Emotional eating in healthy individuals and patients with an eating disorder: evidence from psychometric, experimental and naturalistic studies. Proc Nutr Soc. 2020;79:290–9.

    Article  PubMed  PubMed Central  Google Scholar 

  23. van Strien T, Herman CP, Anschutz DJ, Engels RCME, de Weerth C. Moderation of distress-induced eating by emotional eating scores. Appetite. 2012;58(1):277–84.

    Article  PubMed  Google Scholar 

  24. Shiffman S, Stone AA, Hufford MR. Ecological momentary assessment. Annu Rev Clin Psychol. 2008;4:1–32.

    Article  PubMed  Google Scholar 

  25. Reichenberger J, Pannicke B, Arend A-K, Petrowski K, Blechert J. Does stress eat away at you or make you eat? EMA measures of stress predict day to day food craving and food intake as a function of trait stress-eating. Psychol Health. 2021;36(2):129–47.

    Article  PubMed  Google Scholar 

  26. Raudenbush SW, Bryk AS. Hierarchical linear models: applications and data analysis methods. 2nd ed. Thousand Oaks: Sage; 2002.

    Google Scholar 

  27. Epskamp S, Waldorp LJ, Mõttus R, Borsboom D. The Gaussian graphical model in cross-sectional and time-series data. Multivar Behav Res. 2018;53(4):453–80.

    Article  Google Scholar 

  28. Epskamp S, Deserno MK, Bringmann LF. mlVAR: multi-level vector autoregression. R package version 042. 2019.

    Google Scholar 

  29. Ernst AF, Albers CJ, Jeronimus BF, Timmerman ME. Inter-individual differences in multivariate time-series. Eur J Psychol Assess. 2020;36:482–91.

    Article  Google Scholar 

  30. Dempster AP, Laird NM, Rubin DB. Maximum likelihood from incomplete data via the EM algorithm. J R Stat Soc Ser B Methodol. 1977;39(1):1–22.

    Google Scholar 

  31. Quinn BG. Order determination for a multivariate autoregression. J R Stat Soc Ser B Methodol. 1980;42(2):182–5.

    Google Scholar 

  32. Hill AJ. The psychology of food craving. Proc Nutr Soc. 2007;66:277–85.

    Article  PubMed  Google Scholar 

  33. Kaiser T, Butter B, Arzt S, Pannicke B, Reichenberger J, Ginzinger S, et al. Time-lagged prediction of food craving with qualitative distinct predictor types: an application of BISCWIT. Front Digital Health. 2021;3:694233.

    Article  Google Scholar 

  34. Cohen S, Kamarck T, Mermelstein R. A global measure of perceived stress. J Health Soc Behav. 1983;24:385–96.

    Article  CAS  PubMed  Google Scholar 

  35. Reichenberger J, Kuppens P, Liedlgruber M, Wilhelm FH, Tiefengrabner M, Ginzinger S, et al. No haste, more taste: an EMA study of the effects of stress, negative and positive emotions on eating behavior. Biol Psychol. 2018;131:54–62.

    Article  PubMed  Google Scholar 

  36. LimeSurvey GmbH. LimeSurvey: An Open Source survey tool: LimeSurvey GmbH, Hamburg, Germany; [Available from:

  37. Meule A, Hermann T, Kübler A. A short version of the food cravings questionnaire-trait: the FCQ-T-reduced. Front Psychol. 2014;5:190.

    PubMed  PubMed Central  Google Scholar 

  38. van Buuren S, Groothuis-Oudshoorn K. Mice: multivariate imputation by chained equations in R. J Stat Softw. 2010;1-68.

  39. Ernst AF. LCVAR / Functions [Available from:

  40. Gibson EL. Emotional influences on food choice: sensory, physiological and psychological pathways. Physiol Behav. 2006;89:53–61.

    Article  PubMed  CAS  Google Scholar 

  41. Ram K, Wickham H. wesanderson: A Wes Anderson palette generator. R package version 0.3. 6 2018 [Available from:

    Google Scholar 

  42. Kaupp J. nord: Arctic Ice Studio's Nord and Group of Seven Inspired Colour Palettes for ‘ggplot2’. R package version 1.0.0 2019 [Available from:

    Google Scholar 

  43. Allen M, Poggiali D, Whitaker K, Marshall TR, Kievit RA. Raincloud plots: a multi-platform tool for robust data visualization. Wellcome Open Res. 2019;4:63.

    Article  PubMed  PubMed Central  Google Scholar 

  44. Reichenberger J, Richard A, Smyth JM, Fischer D, Pollatos O, Blechert J. It’s craving time: time of day effects on momentary hunger and food craving in daily life. Nutrition. 2018;55-56:15–20.

    Article  PubMed  Google Scholar 

  45. Richard A, Meule A, Reichenberger J, Blechert J. Food cravings in everyday life: an EMA study on snack-related thoughts, cravings, and consumption. Appetite. 2017;113:215–23.

    Article  PubMed  Google Scholar 

  46. Bongers P, Jansen A. Emotional eating is not what you think it is and emotional eating scales do not measure what you think they measure. Front Psychol. 2016;7(1932):1–11.

    Google Scholar 

  47. Meule A. Food cravings in food addiction: exploring a potential cut-off value of the food cravings questionnaire-trait-reduced. Eat Weight Disord. 2018;23(1):39–43.

    Article  PubMed  Google Scholar 

  48. Cardi V, Leppanen J, Treasure J. The effects of negative and positive mood induction on eating behaviour: a meta-analysis of laboratory studies in the healthy population and eating and weight disorders. Neurosci Biobehav Rev. 2015;57:299–309.

    Article  PubMed  Google Scholar 

  49. Devonport TJ, Nicholls W, Fullerton C. A systematic review of the association between emotions and eating behaviour in normal and overweight adult populations. J Health Psychol. 2019;24(1):3–24.

    Article  PubMed  Google Scholar 

  50. Macht M, Dettmer D. Everyday mood and emotions after eating a chocolate bar or an apple. Appetite. 2006;46(3):332–6.

    Article  PubMed  Google Scholar 

  51. Desmet PM, Schifferstein HN. Sources of positive and negative emotions in food experience. Appetite. 2008;50(2–3):290–301.

    Article  PubMed  Google Scholar 

  52. Alonso-Alonso M, Woods SC, Pelchat M, Grigson PS, Stice E, Farooqi S, et al. Food reward system: current perspectives and future research needs. Nutr Rev. 2015;73(5):296–307.

    Article  PubMed  PubMed Central  Google Scholar 

  53. Leech RM, Worsley A, Timperio A, McNaughton SA. Temporal eating patterns: a latent class analysis approach. Int J Behav Nutr Phys Act. 2017;14(1):3.

    Article  PubMed  PubMed Central  Google Scholar 

  54. Eisele G, Vachon H, Lafit G, Kuppens P, Houben M, Myin-Germeys I, et al. The effects of sampling frequency and questionnaire length on perceived burden, compliance, and careless responding in experience sampling data in a student population. Assessment. 2020;29(2):1–16.

    Google Scholar 

  55. Nahum-Shani I, Smith SN, Spring BJ, Collins LM, Witkiewitz K, Tewari A, et al. Just-in-time adaptive interventions (JITAIs) in Mobile health: key components and design principles for ongoing health behavior support. Ann Behav Med. 2018;52(6):446–62.

    Article  PubMed  Google Scholar 

  56. Reichenberger J, Schnepper R, Arend AK, Richard A, Voderholzer U, Naab S, et al. Emotional eating across different eating disorders and the role of body mass, restriction, and binge eating. Int J Eat Disord. 2021;54(5):773–84.

    Article  PubMed  PubMed Central  Google Scholar 

Download references


We would like to thank Anja F. Ernst for making available the basis for the heatmap and daytime course plots. Moreover, we would like to thank our former students Elena, Sandra, and Vivien for supporting the recruitment and data collection.


This work was supported by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (ERCStG-2014 639445 NewEat) and the Austrian Science Fund (FWF) [KLI 762-B]. BP and JR were supported by the ERC. Publishing open access is supported by the FWF. The funding bodies were not involved in the study design, the assessment and analyses of data, the interpretation of results, and writing the manuscript.

Author information

Authors and Affiliations



BP designed the study, collected, analysed, and interpreted the data and wrote the manuscript. JB designed and supervised the study, interpreted the data and contributed to the manuscript. JR designed the study, collected and interpreted the data and contributed to the manuscript. TK analysed and interpreted the data and contributed to the manuscript. The authors read and approved the final manuscript.

Corresponding author

Correspondence to Björn Pannicke.

Ethics declarations

Ethics approval and consent to participate

Participants consented to participate by signing an informed consent form according to the relevant ethics committee at the University of Salzburg, Austria, which also granted ethical approval for the present study (ethics approval number: EK-GZ: 37/2018). Participants also received detailed oral information on the study.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1.

Coefficients tables (Cluster 1, Cluster 2, Cluster 3).

Additional file 2.

List of additional variables.

Additional file 3.

Simulated random data LCVAR.

Additional file 4.

Information regarding the present empirical study.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and Permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Pannicke, B., Blechert, J., Reichenberger, J. et al. Clustering individuals’ temporal patterns of affective states, hunger, and food craving by latent class vector-autoregression. Int J Behav Nutr Phys Act 19, 57 (2022).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: