Aggregate statistics such as means, variances, and correlations are commonly used to summarize sample-level observations. Such statistics are undeniably useful when attempting to describe observations in the aggregate (i.e., for the group) but are not without controversy or limitations. It is well known, for instance, that the value of a sample mean can be greatly influenced by a small minority of extreme scores which can also produce a skewed distribution of data. Further, aggregate-level statistics are fraught with similar imprecision when examining associations between variables. Anscombe [16] for example demonstrated how data can produce identical bivariate correlations, despite having radically different patterns (i.e., linear; linear with an outlier; nonlinear with an extreme outlier; and curvilinear). Considered alone, group-level correlation statistics can miss important pattern differences like these in the data.

Perhaps more importantly, aggregate statistics may have limited applicability to individual outcomes. For instance, Speelman and McGann [17] created four repeated-measures data sets with mean difference scores equal in magnitude, differing only slightly in their variances across data sets. Aggregate statistical analyses and accompanying significance tests of these data revealed comparable hypothesis-consistent group differences across the four samples. However, inspection of individual difference scores in each sample revealed remarkable variability, with many standing in direct opposition to the hypothesized effect. Similarly, re-analysis of data from previously published studies have shown how group-level conclusions based on aggregate statistics may not match conclusions based on individual patterns of responses [18][19].

With respect to mediation modeling specifically, Brown and Grice [20] and Grice etal. [21] found that aggregate analyses may reveal partial or full mediation effects and explain statistically significant proportions of variance in the outcome variable, but a majority of individuals in the sample may exhibit patterns of responses inconsistent with the model. Thus, although a particular mediation model or analysis may account for significant variance for participantsasagroup, it may not accurately reflect individual behaviors or patterns of behaviors for participantsinthegroup. Stated differently, significant results based on traditional aggregate data analysis inform us how participants on average perform in a certain way but cannot inform us of the proportion of participants in a sample for whom a given model actually fits. In practice, when using aggregate-based analysis we lose the individual outcomes but often mistakenly interpret findings as if they represent the majority of individuals in the group under study (i.e., group-to-persongeneralizabilityproblem; [19]).

Given these limitations, there is increased interest in person-centered methods of data conceptualization and analysis that permit researchers and practitioners to overcome the group-to-person generalizability problem [17]-[19][22]. These methods are not rooted in traditional aggregate statistics but are instead based on the raw observations (i.e., data) themselves and analyzed in such a way as to draw explanatory inferences about individuals rather than groups. Because these methods require researchers to examine patterns of individual responses, they overcome the pitfalls and limitations of aggregate statistics like those discussed above.

More specifically, with person-centered methods like the various pattern analysis techniques in Grice’s [23] Observation Oriented Modeling (OOM) software, individuals’ responses in the data are examined, and those who fit the predicted theoretical pattern are identified and then tallied. Grice converts this tally to a percentage and refers to it as the PercentCorrectClassifications (PCC) index, whereas Speelman and McGann [17] refer to it simply as pervasiveness. Further, distribution-free randomization tests can be used to compute probability statistics (c-values), which can aid in the interpretation of results from person-centered analyses. Low c-values (e.g., <0.10) generated from randomization tests indicate that the observed patterns in the data can be plausibly explained by the theoretical model. High c-values indicate that the most plausible explanation for the results is physical chance [24]. Unlike traditional p-values used in null hypothesis significance testing, these tests are not subject to restrictive and unrealistic assumptions such as homogeneous population variances, normal population distributions, and random sampling.

In the present study we used person-centered methods to re-analyze data from two pediatric IBD samples. Following the procedures of Brown and Grice [20] and Grice etal. [21] we attempted to trace individual youths through the illness stigma → thwarted belongingness → depressive symptoms mediation model reported by Gamwell etal. [1] and Roberts etal. [2]. Specifically, we analyzed each sample separately using the CrossedOrderingsPatternAnalysis in the OOM software [25] to determine the number of youths who could be correctly classified by the proposed mediation model. In effect, youths with relatively high scores on the illness stigma and thwarted belongingness measures were expected to have corresponding high scores on the depressive symptoms measure, and youths with relatively low scores on the illness stigma and thwarted belongingness measures were expected to have corresponding low scores on the depressive symptoms measure.

Participants were 178 youth aged 10 - 18 with a confirmed diagnosis of IBD from independent data sets previously reported in Gamwell etal. ([1]; N = 80) and Roberts etal. ([2]; N = 75). Also, subsequent to the publication of Roberts etal., 25 additional participants were recruited into the study protocol and were included in the present study. Two participants (one from each sample) were excluded due to missing data, resulting in N’s for each sample of 79 and 99, respectively. Participants from both studies were recruited during outpatient appointments at a pediatric gastroenterology clinic within a large teaching hospital in the Southwest United States. Separate protocols for the original studies were conducted in accordance with American Psychological Association ethical guidelines and were approved by the hospital Institutional Review Board.

Data collection for the Gamwell etal. sample took place from 2015-2017. The majority of participants were female (41 female, 38 male) and self-identified as white (71%). Participants were diagnosed with either Crohn’s disease (N = 50; 63%) or ulcerative colitis (N = 29; 37%) and had an average illness duration of 2.9 years. The vast majority of participants (N = 67; 85%) were rated as having quiescent disease, 8 (10%) as mild, 3 (4%) as moderate, and 1 (1%) as severe. Sixteen participants (20%) were prescribed steroid medication.

Data collection for the Roberts etal. sample took place between 2018-2020. The majority of participants were male (43 female, 56 male) and self-identified as white (78%). Participants were diagnosed with either Crohn’s disease (N = 50; 51%) or ulcerative colitis (N = 49; 49%) and had an average illness duration of 2.0 years. The majority of participants (N = 58; 59%) were rated as having quiescent disease, 23 (23%) as mild, 15 (15%) as moderate, and 3 (3%) as severe. Twenty-two participants (22%) were prescribed steroid medication.

The PhysiciansGlobalAssessment (PGA) was completed by the pediatric gastroenterologist to provide an estimate of IBD disease activity. The PGA is based on physician clinical impressions of disease factors (e.g., abdominal pain/tenderness, stool characteristics, etc.). It yields an index of clinical disease activity, with higher scores indicating greater IBD activity (i.e., 1 = quiescent, 2 = mild, 3 = moderate, and 4 = severe).

The StigmaScale-Child (SS-C) is an 8-item self-report measure of youth experiences of stigma related to IBD. The SS-C is adapted from the Child Stigma Scale [26] to include wording related specifically to IBD (e.g., “How often do you feel people may not like you if they know you have IBD?”). Items on the SS-C are measured on a 5-point Likert scale (1 = “never” to 5 = “very often”). A total score, ranging from 8 to 40, is calculated by summing the items with higher scores indicating greater perceived illness stigma.

The InterpersonalNeedsQuestionnaire-ThwartedBelongingnessSubscale (INQ-TB; [27]) is a 9-item self-report measure that assesses feelings of not belongingness to a social group or connecting with others (e.g., “I often feel like an outsider at school or with my friends”) INQ-TB items are rated on a 7-point Likert scale (1 = “not at all true for me” to 7 = “very true for me”). Total scores, ranging from 9 to 63, are calculated by summing the items; higher scores indicate increased perceived thwarted belongingness.

The Children’sDepressionInventory-2^ndEdition (CDI-2; [28]) is a youth self-report measure that assesses the presence of depressive symptoms over the past two weeks. Each of the 28 items is scored on a Likert Scale of 0 to 2 (e.g., 0 = “I am sad once in a while”, 1 = “I am sad many times”, 2 = “I am sad all the time”), with higher sum scores (ranging from 0 to 56) indicating more severe depressive symptoms.

For each sample, preliminary analyses were conducted to obtain descriptive statistics for illness stigma (SS-C), thwarted belongingness (INQ-TB), and depressive symptoms (CDI-2), and to describe how these variables were dichotomized. Although mediation models are ideally tested using the original scaling of the variables, the wide and varied ranges of scale values and markedly skewed distributions in the present data made this approach impractical. Following Brown and Grice [20] and Grice etal. [21], each variable was consequently dichotomized using a median split to create high and low groups.

The primary analyses were conducted in two steps using a CrossedOrderingsPatternAnalysis in the latest version of the Observation Oriented Modeling (OOM; [25]) software. Pairwise relationships between the three variables were first examined. The proposed mediation model (illness stigma → thwarted belongingness → depressive symptoms) was then tested. The principal statistic of interest was the Percent Correct Classifications (PCC) index, which indicates the percentage of children correctly classified by each analysis. A PCC value of 50% was furthermore considered the baseline level of classification, as a clinically meaningful result should correctly classify more than half of the children in the sample. A randomization test with 5000 iterations was also performed for each analysis to generate a chance value (c-value). A low c-value (e.g., ≤0.10) indicates that the observed pattern summarized by the PCC is unlikely to have arisen by chance alone, thereby supporting a causal inference [24].

3.1.1. Preliminary Analyses

Descriptive statistics for the present sample revealed positively skewed distributions for the illness stigma (SS-C; z_skew = 3.01, M = 17.66, Mdn = 16.00, SD = 7.65), thwarted belongingness (INQ-TB; z_skew = 3.12, M = 18.42, Mdn = 15.00, SD = 8.54), and depressive symptom measures (CDI-2; z_skew = 3.10, M = 8.71, Mdn = 7.00, SD = 6.63). The three variables were dichotomized using their respective median values. The SS-C scores ranging from 8 to 15 were classified as “low stigma” and those with scores from 16 to 40 were classified as “high stigma”. The INQ-TB scores from 9 to 15 were classified as “low thwarted belongingness” and those ranging from 16 to 63 were classified as “high thwarted belongingness”. The depressive symptom scores were reordered such that scores from 0 to 7 on the CDI-2 were classified as “low depression” and those from 8 to 56 were classified as “high depression”.

3.1.2. Primary Analyses

Pairwisepatternanalyses. Initial pairwise pattern analyses were conducted to examine the relation between stigma (SS-C) and depressive symptoms (CDI-2), and between thwarted belongingness (INQ-TB) and CDI-2 depressive symptoms. Results revealed that 58 of 79 youth in the sample were correctly classified by corresponding high-low scores on the SS-C and CDI-2 measures (PCC = 73.42%). Pairwise pattern analysis also indicated that 57 youth in the sample were accurately classified by corresponding high-low scores on the INQ-TB and CDI-2 measures (PCC = 72.15%). Randomization tests yielded c-values of <0.001 for both analyses, suggesting that the PCC’s for the SS-C → CDI-2 and INQ-TB → CDI-2 patterns of observations were not likely the result of physical chance.

Mediationpatternanalyses. Pattern analysis for Segment 1 of the mediation model (SS-C → INQ-TB) revealed that 54 of the 79 youth responded in a way consistent with predictions (i.e., corresponding high-low scores on the SS-C and INQ-TB), resulting in a Percent Correct Classifications (PCC) index of 68.35% (c-value = 0.002). Of the 54 youth accurately classified in Segment 1, 45 were correctly classified in Segment 2 of the mediation model (INQ-TB → CDI-2; PCC = 83.33%, c-value ≤ 0.001). Thus, 45 of 79 youths in the sample could be accurately traced through the proposed mediation model (PCC = 56.96%, c-value < 0.001), and this pattern of observations was not likely the result of physical chance. As can be seen in Figure 1, results for the full SS-C→ INQ-TB → CDI-2 model show that 21 youth were correctly classified as having relatively high scores on all three variables, whereas 24 youth were correctly classified as having relatively low scores on all three variables.

Figure 1.Sample 1 pattern analysis of the stigma→thwarted belongingness→depressive symptoms model.

With person-centered analyses it is often instructive to examine particular individuals at the top and bottom ends of the distribution of the outcome variable in a mediation model. Six youth (8%) scored ≥ 20 on the CDI-2 depressive symptom inventory. Of these, only two failed to be classified correctly. One youth reported relatively low stigma (SS-C = 11) despite reporting the highest level of depressive symptoms (CDI-2 = 27) in the sample, while another youth reported relatively low thwarted belongingness (INQ-TB = 21) despite reporting relatively high depressive symptoms (CDI-2 = 12). By contrast, six youth (8%) scored zero on the CDI-2 scale. Of these, only one was classified incorrectly due to reporting a relatively high level of thwarted belongingness (INQ-TB = 22). Seventy-five percent of the youth at the extreme ends of the depression scale distribution could therefore be successfully traced through the mediation model.

3.2.1. Preliminary Analyses

Descriptive statistics revealed positively skewed distributions for the illness stigma (SS-C, z_skew = 3.25, M = 15.42, Mdn = 14.00, SD = 6.12), thwarted belongingness (INQ-TB, z_skew = 3.92, M = 20.44, Mdn = 18.00, SD = 10.62), and depressive symptom measures (CDI-2, z_skew = 5.37, M = 8.81, Mdn = 7.00, SD = 6.97). The three variables were again dichotomized using their respective median values. The SS-C was reordered such that “low stigma” was associated with scores ranging from 8 to 13, and those with scores from 14 to 40 were classified as “high stigma.” The INQ-TB was reordered such that scores from 9 to 17 were classified as “low thwarted belongingness” and those with scores ranging from 18 to 63 were classified as “high thwarted belongingness”. The ordering for CDI-2 remained the same as in Sample 1 (low depression = 0 to 7, high depression = 8 to 56).

3.2.2. Primary Analyses

Pairwisepatternanalyses. Pairwise pattern analysis of the relation between stigma (SS-C) and depressive symptoms (CDI-2) revealed that 62 of 99 youth in the sample were correctly classified by corresponding high-low scores on the SS-C and CDI-2 measures (PCC = 62.63%). Pattern analysis also indicated that 77 of 99 youth were correctly classified by corresponding high-low scores on the thwarted belongingness (INQ-TB) and depressive symptoms (CDI-2) measures (PCC = 77.78%). Randomization tests yielded low c-values for the SS-C → CDI-2 and INQ-TB → CDI-2 analyses indicating the PCCs for both patterns of observations were not likely the result of physical chance (0.01 and <0.001, respectively).

Mediationpatternanalyses. Segment 1 analysis of the mediation model revealed that a little more than half of participant responses (60 out of 99) were consistent with the SS-C → INQ-TB prediction (PCC = 60.61, c-value = 0.03). Of the 60 youth who were accurately classified in Segment 1, 50 were correctly classified by corresponding high-low scores in Segment 2 of the pattern analysis (INQ-TB → CDI-2; PCC = 83.33%, c-value < 0.001). Thus, 50 of the total 99 participants were correctly classified in the SS-C → INQ-TB → CDI-2 mediation model (PCC = 50.51%). The c-value (<0.001) suggests that the pattern of observations for the total model was not likely the result of physical chance. As shown in Figure 2, 24 youth were correctly classified as having high scores on all three variables, whereas 26 youth were correctly classified as having low scores on all three variables.

With respect to examining particular youth with the highest and lowest scores on the outcome variable, eight youth (8%) in this sample scored ≥20 on the CDI-2 depressive symptom inventory. Of these, only one failed to be classified correctly. One youth reported a relatively low stigma score (SS-C = 13) despite reporting a high level of depressive symptoms (CDI-2 = 22). By contrast, six youth (6%) scored zero on the CDI-2. Of these youth, two were classified incorrectly due to reporting relatively high levels of stigma (SS-C values of 15 and 17). Thus, seventy-nine percent of these fourteen youth at the extreme ends of the depression scale distribution could be successfully traced through the mediation model.

Figure 2. Sample 2 pattern analysis of the stigma→thwarted belongingness→depressive symptoms model.