Trait Neuroticism is Associated with how Often People Switch Between Emotion Regulation Strategies Used to Manage Negative Emotions in Daily Life

Most emotion regulation researchers have focused on how people down-regulate negative (or unpleasant) affect. However, how people respond to positive (or pleasant) emotional experiences also has implications for mental health and well-being (Raes et al., 2011). Indeed, regulating positive emotions through many emotion regulation strategies was more strongly associated with overall mood than using many strategies to regulate negative emotions (Heiy & Cheavens, 2014). Given that both negative and positive emotion-focused emotion regulation are associated with psychopathology (Carl et al., 2013), and yet are unique systems (Diener & Emmons, 1984) associated with unique motives (Tamir et al., 2020), our analyses separately examine the relations between neuroticism with strategy switching in negative emotion-focused regulation and neuroticism with strategy switching in positive emotion-focused regulation. Here we use negative emotion-focused regulation and positive emotion-focused regulation to refer to the valence of the target emotion being regulated without value judgment on the specific strategy being used.

All hypotheses were preregistered on the Open Science Framework (OSF, https://​osf.​io/​d3tyn/​) and an overview of all hypotheses is provided in Table 2.

Ninety-two undergraduate students at a large Midwestern university participated in the present study for class credit. Sample size for this secondary data analysis was constrained by the sample size of the parent data collection; however, given our available sample size and assuming 80% power, an alpha of 0.05, and one predictor, we are positioned to detect small-to-medium effect sizes (f² = 0.09; Cohen, 1998). Study participation was open to undergraduate students enrolled in an introductory psychology course. No exclusion criteria were used. Thirty-two participants were invited to participate due to elevated scores on the NEO Personality Inventory-Revised (NEO-PI-R-N; Costa & McCrae, 1992) which pertained to scores at or above 117 for females and at or above 107 for males. These cut scores reflect “High” and “Very High” reports of neuroticism on the measure. The remaining 60 participants were invited regardless of their trait neuroticism score. We used this recruitment approach to obtain a normally distributed sample with respect to trait neuroticism to better understand how emotion regulation functions along the full range of this dimension of personality. Neuroticism’s distribution in the present sample evidenced signs of normality: minimal skewness (-0.31), kurtosis close to 3 (2.16), and the mean and median values were approximately equal (105.74 and 109.5, respectively). There were no other differences in recruitment eligibility across participants and all participants could opt not to participate. Participants ranged in age from 18 to 31 years old (M = 19.73, SD = 2.25) and were predominantly female (54%) and White (81%).

After providing informed consent to participate in this ethics board-approved study (protocol 2008B0320), participants completed a range of emotion regulation and psychopathology questionnaires at baseline. They were then trained on how to answer the ecological momentary assessment (EMA) survey using a study-provided personal digital assistant (PDA), also known as a handheld personal computer. The PDA-administered surveys asked participants to report on their experiences of emotion, emotion regulation attempts, and current mood three times a day for 10 days. Surveys occurred randomly within four-hour intervals throughout anticipated waking hours; assessments typically occurred around 1pm, 5pm, and 9pm and the PDA alerted participants whenever a survey was available to be answered. Participants submitted 1966 surveys out of a possible 2760 surveys (71.23% compliance).

Data were collected in 2009 and were leveraged for the present analysis due to their structure being ideally suited for the newly developed stability metric (i.e., binary time series data with high dimensionality; Daniel et al., 2022). All hypotheses and plans for analyses were preregistered on OSF prior to viewing the data (https://​osf.​io/​d3tyn/​ ).¹ Deviations to preregistered plans, and justifications for those deviations, are detailed below. Analysis scripts are openly provided on OSF (https://​osf.​io/​d3tyn/​) and data are available upon reasonable request to AUTHOR. Data are not stored in an online repository because participants from this 2009 data collection did not consent to having their de-identified data shared publicly. We report how we determined our sample size, all data exclusions, and all relevant measures.

We chose to measure strategy switching with stability rather than conceptually related transition metrics (e.g., Markov models, multidimensional recurrence quantification analysis, autocorrelation/inertia) because stability was specifically designed to analyze high-dimensional multivariate binary time series data. Stability can handle upwards of 30 variables whereas alternative methods tend to return unidentifiable parameters when transition states between more than four variables are investigated (see Daniel et al., 2022 for greater discussion; Tian & Anderson, 2000; Wallot, 2019).

Further, stability has been demonstrated through simulation to be robust to time interval misspecification (Daniel et al., 2023). Specifically, stability is unbiased and has approximately 95% coverage when it is applied to data that are collected with between-person random sampling (relative to fixed sampling), making it a trustworthy metric when applied to data with these unequal time spacing characteristics. This quality is important for the present application because these data were collected with a between-person random sampling schedule, so the amount of time between survey observations varies both within and between participants.

First, we calculated stability within the 20 negative emotion-focused emotion regulation strategies according to the steps outlined in Daniel et al. (2022). Specifically, we used the ‘buildTransArray’ function in the TransitionMetrics package (Daniel & Moulder, 2020) to build a sliding series of transition matrices per participant with 20-by-20 dimensions. We set the window size (W) to nine and used a windowing lag of one.² We set W to nine (a pre-registered decision) so that a minimum of three days’ worth of surveys—with each survey having been administered three times daily—would be included in each transition matrix. We made this decision to increase the likelihood that participants would be reporting emotion regulation attempts in response to multiple, distinct events that took place over multiple days, as opposed to switching their strategies throughout the regulation of a single stressor on a given day, which would likely be a qualitatively different experience. We elected not to set W higher than nine to retain as many participants as possible, as increasing W would increase the number of surveys needed by each participant to be included in analysis. Further, this decision is in line with guidelines offered by Daniel et al. (2022), which suggest that W should be at minimum five. We applied this procedure using all surveys submitted by each participant, including surveys where emotions were denied, which were scored as 0 for all strategies. Given multiple imputation may be less effective for handling missingness in non-linear daily life data (due to dependencies in the results based on the specific imputation model used; Boker et al., 2018), we only use complete cases. As such, the number of transition matrices built per participant varied based on how many total observations they submitted throughout the study. Participants who submitted fewer than nine surveys (n = 3) were excluded from analyses given insufficient observations upon which to build a minimum of one complete transition matrix.

To illustrate how the transition matrices are built from the data, imagine participant i rated whether they had used each of four different emotion regulation strategies (ER₁, ER₂, ER₃, ER₄) at seven surveys (T₁, T₂, T₃, T₄, T₅, T_6, T₇). With these data—which are given on the left side of Fig. 1—and assuming a window of six and a lag of one, two transition matrices are constructed (X_i1, X_i2).

To construct X_i1, a 4 × 4 matrix with all cells set to zero is created. The example data show that ER₁ was endorsed at the first survey (T₁) and no strategy was endorsed at the second survey (T₂). In the present study, no strategies are endorsed on surveys when participants deny experiencing any positive or negative emotion, respectively, since the previous survey. Because this method defines a transition as a switch or a repeat between two endorsed variables in back-to-back surveys (Daniel et al., 2022), this means that a pairwise transition did not occur between the first two time points, so all cells in transition X_i1 remain zero. At the next survey (T₃), ER₁ was endorsed, but because no strategy was endorsed at T₂, no pairwise transition is observed between T₂ and T₃ and all cells remain zero. At the next survey (T₄), ER₁ was endorsed again, indicating a pairwise transition from ER₁ to ER₁ between T₃ and T₄. To reflect this pairwise transition in the transition matrix, we add one to the (1,1) cell of X_i1. At the next survey (T₅), ER₂ was endorsed, indicating a pairwise transition from ER₁ to ER₂ between T₄ and T₅. Thus, we add one to the (2,1) cell of X_i1. At the next survey (T₆), ER₂ and ER₄ were endorsed, indicating two pairwise transitions between T₅ and T₆: one from ER₂ to ER₂ and one from ER₂ to ER₄. To reflect these transitions, we add one to the (2,2) cell of X_i1 and add one to the (4,2) element of X_i1. At this point, all pairwise transitions between the four emotion regulation variables across the first six time points are exactly accounted for in X_i1 (see Fig. 1).

To construct X_i2, we would start with a second 4 × 4 matrix with all cells set to zero. Because we specified a lag of one, the window of observations being read into X_i2 would be shifted down the time series by one compared to what was read into X_i1. As such, the transitions between T₁ and T₂ would not be captured by X_i2. The transitions between T₂ and T₃, T₃ and T₄, T₄ and T₅, and T₅ and T₆, however, would be reflected in X_i2 just like in X_i1. Moreover, because the window of observations was shifted down by one, there would be one new survey to read into X_i2 (i.e., the transition between T₆ and T₇). At this final survey, ER₂ was endorsed, indicating that two pairwise transitions occurred between T₆ and T₇: one from ER₂ to ER₂ and one from ER₄ to ER₂. To reflect these transitions, we add one to the (2,2) cell of X_i2 (such that the (2,2) cell now equals two) and add one to the (2,4) element of X_i2. With all transition matrices populated, the stability equation, which is described below, would then be applied to each completed transition matrix.

After constructing all transition matrices in the present data, we used the ‘transStats’ function in the TransitionMetrics package to calculate stability for each transition matrix. Notably, stability leverages one important characteristic of the transition matrix to measure the extent to which the multivariate system transitions from one variable of interest to the same variable of interest relative to all observed pairwise transitions. If the same variable is endorsed in two back-to-back surveys, then a cell along the on-diagonal of the matrix is incremented to reflect this “repeat.” Conversely, if one variable is endorsed prior to a different variable being endorsed at the next survey, then a cell along the off-diagonal of the matrix is incremented to reflect this “switch.” As such, stability is given by

Where \(tr\left({\mathbf{X}}_{ij}\right)\) is the sum of the elements along the on-diagonal of a given 20-by-20 transition matrix X_ij (i.e., the number of repeats); \(\sum \sum {\mathbf{X}}_{ij}\) is the sum of all elements of X_ij (the number of both repeats and switches). As a ratio, stability is therefore bounded between 0 and 1 where values closer to 1 indicate a tendency to repeat the same variable between successive surveys and values closer to 0 indicate a tendency to switch between at least two different variables between successive surveys. Phrased differently, since \(tr\left({\mathbf{X}}_{ij}\right)\) is the number of times the person did not switch strategies and \(\sum \sum {\mathbf{X}}_{ij}\) is the number of times that person either switched or did not switch strategies, then stability (\({St}_{ij}\)) can be thought of as the proportion of times that a person did not switch strategies.

Finally, we averaged all stability scores at the person level to arrive at one stability score per participant. This average stability in negative-emotion regulation strategies was used in subsequent analyses. We then repeated these steps for the 20 positive emotion-focused emotion regulation strategies to arrive at a measure of average stability in positive emotion-focused regulation strategies.

We took identical steps as above, although the dimensions of the transition matrices varied depending on how many strategies composed each emotion regulation strategy class. For example, stability within the adaptive engagement strategies was calculated using a 9 × 9 transition matrix, whereas stability within the behavioral strategies was calculated using a 4 × 4 transition matrix.

Stability measures the extent of different types of transitions (e.g., repeat strategy vs. change to a different strategy) that occur between successive survey responses. By definition, a transition can only be observed when at least one strategy is endorsed at each of two successive surveys. If no strategy-to-strategy transitions are observed within a given transition matrix, then that transition matrix is assigned a “noUse” value because the denominator in its calculation would be zero, which results in an undefined term. “noUse” values could be observed in one of two ways: (1) each of the nine surveys contributing to a given transition matrix denied experiencing an emotion such that the entire segmented data frame was all 0 or (2) surveys that denied experiencing an emotion were inserted between active regulation attempts such that no two active regulation attempts were ever observed between back-to-back surveys within the nine-survey segment. If at least one strategy-to-strategy transition is observed within a transition matrix, then stability returns a meaningful numeric result. In line with our preregistration, we removed all “noUse” transition matrices prior to calculating each participant’s average stability value given that an average could not otherwise be computed.

We decided to retain all surveys and treat lack of emotion endorsement as 0 for all strategies, rather than either remove all surveys where no strategy was endorsed or add a binary timeseries that gives 1 if an emotion was denied or 0 if an emotion was endorsed, to aid interpretability. For example, although removing all surveys where an emotion was denied would have prevented the “noUse” case from occurring (because all rows of data would have had at least one 1), it also would have meant that we would be implying transitions in the timeseries that did not directly occur by removing meaningful rows of data. Similarly, adding a new variable that gives 1 during all surveys when an emotion was denied and therefore no strategies were endorsed would have equated repetition of no strategy with repetition of a strategy. Both alternative options seemed especially problematic for investigations into class-specific switching when lack of relevant strategy endorsements was likely to be higher (e.g., a person who never reported using an aversive cognitive perseveration strategy would be scored as equally stable in their use of this strategy class as someone who repeatedly used the same aversive cognitive perseveration strategy). For additional details regarding how the stability calculation manages lack of emotion regulation strategy use, see the online supplement.

We conducted all analyses in R version 4.1.3 (R Core Team, 2022) using the “rq” function in quantreg package version 5.88 (Koenker et al., 2021). We regressed continuous neuroticism on centered average ER stability (linear effect). We then added centered average ER stability squared (quadratic effect) into the model and conducted model comparisons to test whether the quadratic effect should be retained in the final model. We interpreted the best-fit model.

Our pre-registration outlined a plan to use regression on the mean to conduct these analyses; however, once viewing the data, we observed that the residuals from these models did not conform to assumptions of linearity, normality, and homoscedasticity due to the presence of outliers. To better model these data, we deviated from our pre-registration in favor of using a robust method: quantile regression on the median.³ In total, we conducted eight sets of hierarchical quantile regression models, all predicting neuroticism but using the following emotion regulation stability predictors: (1) overall negative-focused; (2) adaptive engagement negative-focused; (3) aversive cognitive perseveration negative-focused; (4) disengagement negative-focused; (5) overall positive-focused; (6) enhancement positive-focused; (7) disengagement positive-focused; (8) behavioral positive-focused.

To investigate how the relationship between neuroticism and strategy switching changes as a function of W, we conducted sensitivity analyses that systematically varied W but otherwise used identical procedures. Specifically, W was iteratively set to: {6, 12, 15, 18, 21, 24, 27}. Given our design of three surveys per day, these Ws resulted in stability values calculated across a minimum of 2, 4, 5, 6, 7, 8, and 9 study days, respectively. Sensitivity analyses were only conducted at the overall level, resulting in 14 additional sets of analyses (7 for overall negative-focused strategy switching and 7 for overall positive-focused strategy switching).