Body Mass is Linked With a Broad Range of Personality Nuances, but Especially Those With Behavioral Content: A Multi-Sample Exploration

Various personality domains and facets correlate with body mass index (BMI), but recent studies suggest that using narrower personality traits—nuances—could contribute to a more detailed understanding of personality–body weight associations. We used three large datasets with different inventories to describe nuances’ correlations with BMI and explore whether BMI predominantly correlated with affective, behavioral, cognitive, or motivational item content. BMI correlated with many nuances, most prominently those reflecting immoderation, lack of orderliness, talkativeness, leadership tendencies, anger, traditionalism, and preference for routine. The highest nuance-level correlation was .21, compared to .11 for the Five-Factor Model domains. BMI correlated most strongly with nuances predominantly reflecting behaviors. Nuance-based approaches can thus reveal the strength, multitude, and content-nature of personality–outcome correlations that can potentially remain hidden in broader traits. If personality traits become relevant in the prevention or treatment of obesity, a focus on narrow behavioral traits may be especially warranted. Relevance Statement Exploring personality nuances’ associations with BMI in three large datasets, we found that BMI has stronger and more numerous correlations with personality traits than studies have shown. Personality may be more relevant in obesity than thought. Key Insights BMI correlated with many nuances across the personality space. Nuance-level correlations tended to be stronger than domain-level correlations. Analyses with domains cannot adequately represent trait–BMI associations. Some variability in correlations was observed across inventories. BMI was most strongly related to behavioral item content.

Personality traits are known to correlate with a variety of life outcomes. Among these are various health outcomes such as excess adiposity, a major risk factor for conditions like type 2 diabetes (Bjerregaard et al., 2018), cardiovascular disease (Van Gaal et al., 2006), and mortality (Lewis et al., 2009). If these correlations reflect causal influences between personality traits and overweight as is sometimes assumed (Kim, 2016;Sutin & Terracciano, 2017) and as some empirical findings also suggest (Arumäe et al., 2021), studying them could pave the way to an improved understanding of the development of overweight and obesity on the one hand and their contributors to personality differen ces on the other. Among other things, understanding these associations could inform us about whether and how weight-management interventions could benefit from the consideration of personality traits or, perhaps, influence them.
So far, it has been shown that broad personality trait domains like conscientiousness, as well as numerous narrower facets of these domains, correlate consistently with body mass index (BMI), the most commonly used proxy for adiposity (Jokela et al., 2013;Sutin et al., 2018;Vainik et al., 2019). Specifically, as facet-level analyses have repeatedly shown, relying on domains that aggregate various facets can mask the correlations of narrower traits with BMI, whereas facets can give a clearer picture of these associa tions (Sutin et al., 2018;Terracciano et al., 2009;Vainik et al., 2019). It therefore seems essential to use such narrower facets to describe personality traits' correlations with BMI. Although the idea that narrower traits provide incremental prediction of important outcomes over domains is not new (Dudley et al., 2006), only in recent years have researchers started to focus on even narrower traits than facets: nuances (McCrae, 2015;Mõttus et al., 2017;Revelle et al., 2021). Here, we aimed to provide a detailed picture of how personality traits intersect with body weight by using these nuances.

BMI and Personality Item Content: Affect, Cognition, Motivation, and Behavior
When the associations between personality traits and BMI have been described on an appropriately narrow level, a next step is to clarify which types of processes give rise to these associations. For instance, it is believed that the relations between traits and outcomes of interest are mediated by the thoughts, feelings, and behaviors that the traits are associated with (Roberts & Yoon, 2022). If there is a way to isolate these personality trait-related manifestations, it will be possible to test which, if any, of these expressions of personality traits is primarily associated with BMI. And this is indeed possible: al though personality inventories do not explicitly differentiate such components of traits, it is possible to distinguish the so-called ABCD components of personality traits-affect (A), behavior (B), cognition (C), and motivation (or desire, D)-in personality inventories by having knowledgeable raters assess to what extents items contain these components (Pytlik Zillig et al., 2002;Wilt & Revelle, 2015).
Nuances and the domains they make up differ in their ABCD content. The items included in the extraversion and conscientiousness domains are, on average, somewhat more behavioral, items of neuroticism are more affective, those entailed in openness tend to reflect cognition, and items in the agreeableness domain have the most balanced ABCD content (Pytlik Zillig et al., 2002;Wilt & Revelle, 2015); however, there remains a substantial variability within the domains in their items' ABCD content. Distinguishing the content of items and domains makes it possible to test whether items that correlate highly with BMI also have higher-than-average levels of affective, behavioral, cognitive, or motivational content-in other words, whether any of those components primarily account for the correlations BMI has with personality traits. This could potentially suggest how, or through which pathways, personality traits may relate to body weight.
It may be the dominant opinion among researchers that personality traits are related to BMI through behaviors, including those related to health: for instance, conscientious ness may go with higher frequency of health-protective behaviors and lower frequency of health-damaging or risky behaviors, therefore resulting in healthier body weight as well as other health outcomes (Keller & Siegrist, 2015;Lunn et al., 2014). However, BMI may also relate to the affective content of personality traits as both positive and negative emotions have been found to be related to increased food intake (Canetti et al., 2002;Evers et al., 2018;Sultson et al., 2017). If BMI is found to correlate with higher propor tions of behavioral content, for instance, this would be consistent with personality traits relating to BMI mainly through (health-relevant) behaviors; if it correlates most strongly with affective, cognitive, or desire content, this would instead suggest that processes related to emotions, thoughts, or motivations, respectively, are driving personality trait-BMI correlations. Therefore, the degrees to which items with high proportions of certain types of content correlate with BMI may shed light on the types of processes that link personality traits to body weight.

The Present Study
The current study aimed to provide a detailed description of BMI's cross-sectional associations with personality traits. First, we correlated over 700 personality items of three different inventories with BMI to clarify which of them associate with BMI most strongly. Second, we assessed the correlations between BMI and ABCD content to see whether body weight tends to relate to any certain type of personality item content. Most studies on BMI-personality trait associations have focused on one sample and one inventory (most commonly a version of the NEO Personality Inventory, i.e., NEO-PI-R or NEO-PI-3; McCrae et al., 2005). Here, we include three independent datasets, each of which includes personality traits measured using a different inventory: either NEO-PI-3, IPIP-NEO (Goldberg, 1999), or a recently developed item pool intended to measure a broad range of nuances (which we call here "A Hundred Nuances of Personality"; 100NP). Because a broad representation of personality nuances is necessary for the identification of the most relevant ones for an outcome, we focused on datasets where personality had been measured with a large number of items; the sample sizes were also sufficiently large to allow detecting non-trivial personality trait-BMI correlations while controlling for multiple testing. Including different inventories allowed us to investigate a broader set of personality nuances than any individual inventory would have, and therefore increase the possibility of finding the nuances most strongly related to BMI, as well as assess whether ABCD content calculated based on different sets of items relate to BMI similarly. Because the magnitude of correlations between BMI and nuances can best be interpreted in comparison to broader traits, we also report the correlations between BMI and the domains of the Five-Factor Model in the three inventories.

Materials
Datasets NEO-PI-3 Data -The NEO-PI-3 is a 240-item inventory that measures the five domains of the Five-Factor Model, as well as 30 facets, six of which belong within each domain. Answers to personality items in this inventory are provided on a 5-point Likert scale ranging from 0 (strongly disagree) to 4 (strongly agree). We used NEO-PI-3 personality data of a subsample of the Estonian Biobank of the Estonian Genome Center, University of Tartu (Leitsalu et al., 2015). This sample represents a cohort of Estonian adults from whom DNA samples as well as various other medical data have been collected, including height and weight based on which BMI could be calculated. BMI was calculated as weight in kilograms divided by height in meters squared. Within the sample, 3,562 individuals (2,147 female) had the necessary data available to calculate the BMI-personality item correlations while accounting for the covariates used throughout Arumäe, Vainik, & Mõttus this study (age, age 2 , and sex). Mean age of these individuals was 48.80 years (SD = 17.00, range 18-91); mean BMI was 25.98 kg/m 2 (SD = 4.87). Descriptive statistics and Cronba ch's alphas for the personality domains of all samples are provided in Table 1 and Table  S1, respectively. Note. For IPIP-NEO, descriptive statistics were found in subsamples of individuals who had responded to a minimum of 20 items within the domain (Ns ranged from 2,209 to 2,480 individuals). a In NEO-PI-3, answers are given on a scale from 0 to 4; in IPIP-NEO and 100NP, they are given on a scale from 1 to 6. For the sake of comparability between the inventories, NEO-PI-3 descriptives are provided not only as measured on its original scale, but also after converting the scores to a scale of 1 to 6 (in parentheses) with the formula (raw score / 4) × 5 + 1.

IPIP-NEO Data -
The IPIP-NEO is a 300-item inventory that assesses the same five domains as well as 30 facets conceptually similar to those of the NEO-PI-3. Ten items represent each facet; answers are given on a scale from 1 (very inaccurate) to 6 (very accurate). We used a publicly available dataset (Condon et al., 2019) where subsets of 696 personality questions of the International Personality Item Pool had been randomly administered to over 48,000 people online. From this data pool, we selected the 300 items included in the IPIP-NEO for which the ABCD ratings were available. After excluding respondents with missing data on BMI, age, sex, or country, a total of 37,971 participants from 191 countries remained (23,940 female) with a mean age of 26.34 years (SD = 11.41, range 14-89) and a mean BMI of 24.88 kg/m 2 (SD = 5.91). Because of the diversity of the respondents' geographic locations, we additionally accounted for the continent of the participants' countries in the IPIP-NEO data. Specifically, the countries were recoded as continents as follows: North America (n = 27,767), Europe (n = 5,286), Asia (n = 3,081), Oceania (n = 979), Africa (n = 438), and South America (n = 429). Due to the use of the Synthetic Aperture Personality Assessment (SAPA) method, the IPIP-NEO data had high (~ 90%) missingness; specifically, the participants had responded to 35.06 items on average (SD = 27.97, range 1-158 items) and each item had responses from an average of 4,437 participants for whom BMI was also available (SD = 1,081.97, range 3,787-9,751).
100NP Data -100NP is a 257-item item pool developed to measure a broad range of personality nuances. The items were largely drawn from the publicly available In ternational Personality Item Pool (Goldberg, 1999) and Synthetic Aperture Personality Assessment item pool (Condon, 2018). The items were responded to using a scale from 1 (very inaccurate) to 6 (very accurate); detailed information on the item pool is available at https://osf.io/tcfgz/. The 100NP differs from classical personality inventories in that it was not designed with the purpose of assessing any broad-level traits, but many of its items can nevertheless be mapped to the domains of the Five-Factor Model. For the purposes of the current study, the items were aggregated into these five domains using principal component analysis: items with loadings of .40 or greater on one compo nent were assigned to the domain that the component represented. Thirty-two items represented agreeableness, 25 conscientiousness, 33 extraversion, 40 neuroticism, and 17 openness, with the remaining 110 items falling outside of these domains. Participants for the 100NP dataset were recruited from Prolific Academic. After removing participants with missing data on BMI, age, sex, or country, the sample consisted of 1,101 persons (595 female) from 37 countries with a mean age of 27.83 years (SD = 10.88, range 12-77) and a mean BMI of 24.20 kg/m 2 (SD = 5.09). Because a majority of participants were from the United Kingdom, geographic location was coded as United Kingdom (n = 748) or other (n = 353) and used as a covariate in statistical analyses.

Item Content Ratings
Item content ratings for the three inventories were obtained from various sources, but the procedure was similar in all cases. Specifically, for each item of each inventory, the raters were instructed to assess to what extent, in their opinion, each item contained affective, behavioral, cognitive, and motivational content. These four ratings had to add up to 100%. A detailed description of the procedure is provided by Wilt & Revelle (2015).
Ratings for NEO-PI-3 Items -ABCD ratings for NEO-PI-3 items were collected from eight respondents: six female, two male; all respondents had university degrees, six of them in psychology. Ratings were collected for all items included in the NEO-PI-R and NEO-PI-3 (i.e., a total of 277 items), but only ratings for the latter inventory were used in the current study. The ratings are publicly available and can be accessed via the Supplementary Materials. Average-rater intraclass correlations (ICCs; interpretable similarly to Cronbach's alphas) for the ABCD components of NEO-PI-3 items were .90, .87, .83, and .67, respectively.

Ratings for IPIP-NEO Items -
Ratings for the IPIP-NEO items were obtained from two sources. Namely, six respondents' ratings were collected by Wilt & Revelle (2015); ratings from three additional persons were collected separately in the process of another ongoing study. Of the nine raters, six were female and three were male; most had university degrees and a background in psychology. The aggregated ICCs of the ABCD components across the nine raters were .88, .83, .88, and .78, respectively.
Ratings for 100NP Items -Ratings for the 100NP items were collected from ten persons using a similar procedure for the ongoing study mentioned above. All raters had a university degree and a background in psychology; the ABCD components' ICCs were .87, .88, .87, and .80, respectively.

Statistical Analyses
The 20 items most strongly correlated with BMI in each inventory were found by calculating Spearman's ρ between BMI and personality items after residualizing BMI and every personality item for age, age 2 , and sex, as well as geographic region in IPIP-NEO data and 100NP data. Geographic location was included as a covariate as a proxy for environmental or cultural differences which can moderate personality-BMI associations (Sutin & Terracciano, 2017). Geographic location was not corrected for in the NEO-PI-3 data because all respondents were Estonian residents. Because we were interested in the strongest correlated items, not necessarily the directions of the associations (which depend on the precise wording of the item), the 20 items most strongly correlated with BMI were chosen based on the absolute value of the correlation coefficient; correlations are reported in the direction implied by item content (i.e., they have not been reversed to match any broader traits). Analyses were repeated after additionally residualizing BMI and personality items for level of education. The covariates were included as potential influences on both personality traits and BMI which could inflate estimates of BMI-per sonality trait correlations; however, should personality traits also (or instead) influence level of education, controlling for education would result in statistical overcontrol and underestimation of effects (Kim, 2016). Because the influences between BMI, personality traits, and the covariates are insufficiently understood, we report results of analyses with education as an additional covariate in a supplementary document.
To facilitate comparison of the three datasets as well as the strength of BMI's asso ciations with domains and nuances, we also report correlations between BMI and the five domains of personality in each inventory. These correlations were calculated the same way as BMI-item correlations, but after aggregating the items into domains. Two items related to eating were excluded from the neuroticism domain in the NEO-PI-3 and IPIP-NEO datasets prior to calculating the domain's correlation with BMI to avoid inflation of the correlations.
Because the IPIP-NEO and 100NP items were both drawn from a common set of items (IPIP), the two questionnaires have 69 items in common. This overlap allowed us to additionally assess the similarity of the shared items' correlations with BMI in the two datasets, to test the replicability of the associations. To this end, the item-BMI correlations in the IPIP-NEO dataset were correlated with the item-BMI correlations in the 100NP dataset. Considering the overlapping items, our analyses included a total sample of 728 items.
To test whether BMI tended to correlate with any certain type of item content, we first correlated BMI with each item, adjusting for covariates as described above. We then correlated the absolute values of these correlations to the items' proportions of ABCD content. Absolute values were used to test whether the items' content was related to the strength of their correlations with BMI as the direction each specific item is phrased is arbitrary in this analysis. For each inventory, the sample size in these analyses was the number of items included in that inventory (i.e., 240-300); a fixed-effects meta-analysis where effect sizes were weighted by the inverse of standard errors was additionally con ducted for increased statistical power and to summarize which of the ABCD components associated with BMI across the three samples.
All analyses were run using R 3.6.1 using RStudio. All p-values were adjusted for false discovery rate with the number of comparisons being the number of associations tested in each inventory (i.e., five in case of domains and the number of items in the inventory in case of items).

Transparency, Openness, and Reproducibility
Analyses reported in the current study were not preregistered. The supplementary docu ment where analyses with education as an additional covariate are reported can be found in the Supplementary Materials. Table 2 shows the correlations between BMI and the five domains of personality in the three samples (correlations when additionally accounting for education can be found in Table S2). Each sample showed a significant negative correlation between BMI and conscientiousness. In some cases, however, correlations differed considerably between the three inventories, especially for neuroticism and openness, with estimates ranging from .02 to .11 for the former and from -.09 to .01 for the latter, suggesting that different samples and/or different inventories can lead to considerably different estimates of correlations.

Results
Correlations between BMI and the individual items by domain are shown in Figure  1. This figure illustrates the distribution of BMI's correlations, showing that the correla tions of BMI with items within each domain varied considerably in size and differed in direction. Note. Two eating-related items were removed from the NEO-PI-3 and IPIP-NEO before calculating the correla tions to avoid inflation of the estimates. For IPIP-NEO, the domains' correlations were found in subsamples of individuals who had responded to a minimum of 20 items within the domain (Ns ranged from 2,209 to 2,480 individuals). The domain-level correlations with NEO-PI-3 domains in overlapping samples have also been reported by Vainik et al. (2015) and Arumäe et al. (2022), and included in a meta-analysis by Vainik et al., 2019; as a difference, the correlation with neuroticism in the current study is calculated after excluding the two eating-related items from the scale. All p-values are adjusted for false discovery rate.

Individual Items' Correlations With BMI, by Domain
Note. Each colored dot represents an item within the specified domain. The items with reverse phrasing have been coded to match their broader domains. The eating-related items included in the NEO-PI-3 and IPIP-NEO are not shown. Dots with error bars reflect BMI's average correlations with the domain's items. Error bars denote standard deviations.

Items With the Strongest Correlations
Tables 3-5 show the items that most strongly correlated with BMI in NEO-PI-3, IPIP-NEO, and 100NP. Correlations after additionally adjusting for level of education are shown in Tables S3-S5. Because NEO-PI-3 is proprietary, we present the meaning of the items, but not the items themselves as they appear in the inventory, throughout the manuscript. Note. BMI and personality items were residualized for age, age 2 , sex, and geographic location. All ps are corrected for false discovery rate. The two eating-related items in NEO-PI-3 predictably occupied the first two positions in Table 3, but no other neuroticism items were found on the list, consistent with the absence of a significant domain-level association when excluding those two items. Although all domains were represented, it stands out that 10 of the 20 items belonged to the extraversion domain-items reflecting positive emotions (#5, #9, #16), talkativeness (#8, #18), assertiveness (#6, #12), activity (#15, #19), and excitement-seeking (#11), all related to BMI positively. These associations suggest that BMI is related to many aspects of extraversion as measured by the NEO-PI-3, but because the associations with other items vary in direction, the association tends to be "washed out" in typical domain-level analyses. The four conscientiousness items included in the list reflected orderliness (#4, #10, #17) and achievement striving (#3). The openness items reflecting a preference for variety related to BMI negatively (#7, #14) and the item regarding fantasy, positively (#20). Finally, an item related to trust (#13) correlated with BMI negatively. A breakdown of the number of represented items by domain and inventory is shown in Table 6 (see Table S6 for the breakdown after additionally adjusting for education). In IPIP-NEO, items of neuroticism were well-represented among the items most strongly correlated with BMI (12 items). Given the substantial correlation between BMI and the neuroticism domain in this inventory, this is not surprising, although it contrasts with the results found in NEO-PI-3 where impulsivity items other than eating-related did not appear among the items most highly correlated with BMI. Eight of the twelve items reflected immoderation in relation to food (#1, #4), money (#9, #14), and a general tendency to overindulge or an inability to resist temptations (#2, #3, #5, #7). A theme of self-consciousness also emerged (#17, #18). Items related to holding grudges or getting mad belonging to the agreeableness (#6, #12, #13) and neuroticism domains (#11, #19) also stood out. The openness items reflected conservative beliefs or values (#10, #15), including a lower preference for variety (#8) which was also seen in NEO-PI-3 data.
Although BMI was also correlated to conscientiousness, only one item of this domain was represented, reflecting orderliness (#20). Finally and unlike in the NEO-PI-3 data, only one extraversion item was included which reflected a preference for a leisurely level of activity (#16) and, also in contrast to NEO-PI-3 data, correlated with BMI positively. In the list for 100NP, as in that of IPIP-NEO, items related to immoderation or impulsivity were again present: inability to resist temptations (#1, #7), carelessness with money (#3, #9, #11), and impulsive behavior (#18, #19). Also similar to the results found with the IPIP-NEO data was the presence of items relating to self-consciousness (#15, #16), orderliness (#6), and making enemies (#17). Similarly with NEO-PI-3, on the other hand, items related to talkativeness (#5) and assertiveness (#10) were included. Items related to self-perceptions of appearance (#12, #14) and health (#2) also stood out, as well as following traditional masculine gender roles (#17), being forgetful (#8), knowing how to comfort others (#13), and a tendency for boredom (#20).
When additionally accounting for education, the items most strongly related to BMI largely remained the same as those reported in Tables 3-5. In NEO-PI-3 data, 18 of the 20 items overlapped regardless of whether education was accounted for. In IPIP-NEO data and 100NP data, 19 items overlapped. While there were slight changes in the magni tude of correlations and the order of the items, additionally controlling for education did not therefore make a major difference to the overall results.
Of the 69 items shared by IPIP-NEO and 100NP, three were present in both top lists ("Am able to control my cravings, " "Easily resist temptations, " and "Leave a mess in my room"). Two of the shared items were found in the list of 100NP ("Have a low opinion of myself, " "Act without thinking") and one was in the list of IPIP-NEO ("Yell at people"). Sixty-three were present in neither. To get a sense of the extent to which an item's correlations with BMI differ between the two samples, we estimated the correlation of the 69 overlapping items' associations with BMI in each sample. The vectors of the 69 items' correlations with BMI in the samples correlated at r = .66 (p < .001), suggesting that the correlations in different samples overlapped substantially.
Finally and as expected, in each inventory, various items correlated with BMI more strongly than domains did. In the NEO-PI-3 sample, the strongest domain-level correla tion was -.06; in IPIP-NEO, .11, and in the 100NP dataset, this was -.10. For single items, however, correlations were in many cases stronger (up to .13, -.19, and -.21, respectively, excluding eating-related items), suggesting that these narrow personality traits may be particularly relevant in body weight.

ABCD Components
First, we determined whether the proportions of ABCD content of the five domains were similar in the three inventories. We found that to be largely the case, although a few differences could be seen. For instance, among the NEO-PI-3 items, agreeableness had a higher proportion of cognitive content, whereas in the other inventories the domain was more balanced in its content; extraversion in the NEO-PI-3 was highest in affective content but behavioral in the other inventories. Across all items, the NEO-PI-3 had a somewhat higher cognitive content and slightly lower behavioral content when compared to the other inventories. However, the item content of the three inventories largely resembled the proportions reported elsewhere (Pytlik Zillig et al., 2002;Wilt & Revelle, 2015). Importantly for the current results, the overall similarity of the item content across inventories suggests that the three inventories were comparable in terms of ABCD content. Item content of the three inventories is illustrated in Figure S1. Although some items were relatively more "pure" in their content, not a single item was rated as reflecting just one component. Instead, items tended to be mixtures of all four components. The distribution of individual items' content by domain can be seen in Figure S2. Table 7 shows the correlations between each inventory's items' proportions of ABCD content and their correlations with BMI; analyses additionally accounting for education are reported in Table S7. The meta-analytic estimates showed that BMI was associated with items high in behavioral content regardless of accounting for education, and was additionally related to the proportions of cognitive and motivational content when ac counting for education. Personality traits' correlations with BMI may therefore mainly be mediated by the behaviors relevant to the traits, although cognitions and motivations may additionally be relevant in these associations. Note. Coefficients for the three inventories are Spearman's ρ with 95% confidence intervals. Two eating-related items have been removed from both NEO-PI-3 and IPIP-NEO to avoid inflation of the estimates. Correlations between BMI and personality items were adjusted for age, age 2 , sex, and, for IPIP-NEO and 100NP, geographic location. Meta-analytic p-values are adjusted for false discovery rate.

Discussion
This exploration of BMI's cross-sectional associations with personality nuances and their content adds to previous domain-and facet-level analyses in various ways. BMI correlated with many nuances across the personality space rather than being associated with only few narrow traits, and these nuance-level correlations were often stronger than domain-level correlations. Moreover, the correlations between BMI and the items in each domain varied considerably in their size and direction, suggesting that the domain level is often not optimal for describing personality traits' associations with BMI, at least when simplicity is not paramount such as in case of public engagement. It therefore seems that there is a lot in personality that is linked with BMI, but analyses with broader traits do not (and cannot) adequately represent the multitude and strength of nuance-level associations. BMI-personality trait associations appeared to be driven by behavioral item content, suggesting that traits may largely relate to adiposity through behaviors related to the traits.

BMI and Nuances of Personality
Facet-level analyses have previously shown that narrower traits correlate with outcomes like BMI irrespective of the broader traits that subsume them (Vainik et al., 2019). As we expected, the same held for nuances. Indeed, BMI had substantial correlations with items included in each Five-Factor Model domain, although it did not correlate with all domains. Nuance-level analyses identified that, in all three datasets, the items most strongly related to BMI included items related to orderliness or discipline (con scientiousness), immoderation and impulsivity (neuroticism/conscientiousness), getting angry and holding grudges (agreeableness/neuroticism), traditionalism and conservative beliefs (openness), self-consciousness (neuroticism), as well as leadership tendencies and talkativeness (extraversion). These items may or may not be included in the domains of particular inventories: if they are, the domains are more likely to be correlated with BMI as well. Some of the item-level correlations are well in line with previous domain-and facetlevel findings. For instance, the positive correlations with items reflecting leadership tendencies ("Act as a leader", "Emerges as a leader") and talkativeness ("Is the most talkative person in conversations", "Talk a lot") mirror the association between BMI and assertiveness (Sutin et al., 2011;Terracciano et al., 2009). The negative correlations between BMI and items like "Is meticulous" and "Leave a mess in my room" similarly resemble the repeatedly reported correlation with orderliness (Sawhney et al., 2020;Sutin et al., 2018). Further, items related to anger ("Seldom get mad", "Yell at people") and making enemies ("Hold a grudge", "Make enemies") seem to conceptually align with the positive association between BMI and hostility (Gerlach et al., 2015). Therefore, nuance-level analyses often capture similar associations as have been found in analyses with broader traits. This makes sense: after all, if a facet or domain is associated with an outcome (BMI), then this association must be driven by the items that make up these broader traits.
In some cases, however, the domain-and nuance-level correlations were less aligned with each other, as has been previously observed with some domain-and facet-level correlations (Vainik et al., 2019). For instance, despite the significant negative correlation between BMI and conscientiousness in the IPIP-NEO dataset, the 20 items with the strongest correlations only included a single item from this domain. This suggests that even if BMI most consistently associates with the domain conscientiousness, it has stronger correlations with items of other domains. As another example, despite there being no domain-level correlation with openness in the NEO-PI-3 dataset, items within this domain ("Vacations in different places", "Values sticking with principles rather than being flexible-minded", "Enjoys letting fantasies develop") did associate with BMI. In some cases, therefore, domains seem to be too broadly defined to detect associations, even if nuances entailed in them have substantial correlations with BMI. This supports the premise of the current study: in order to get a thorough understanding of what exactly in personality relates to BMI, narrow traits need to be considered.
Further, the results of the three inventories differed considerably in some cases. For instance, some domains were represented in very different numbers among the 20 stron gest-correlated items between the inventories. Most notably, extraversion was represen ted with ten items in the case of the NEO-PI-3 dataset and with only one in the case of the IPIP-NEO dataset. Although differences like this may be due to sample characteristics or the particular items in the inventories, it may also suggest that different trait config urations may correspond to similar levels of BMI-or, as has been stated previously: there is no single personality profile associated with obesity (Generali & De Panfilis, 2018). In the NEO-PI-3 dataset, higher BMI was associated with lower orderliness or discipline combined with higher scores on talkativeness, leadership tendencies, positive emotions, and traditionalism. In the IPIP-NEO dataset, in contrast, higher BMI appeared to associate with high immoderation combined with higher conservatism/traditionalism and a tendency towards feeling anger. The items most strongly related to BMI in the 100NP dataset varied the most in content, but this may not be surprising given that the inventory was designed to measure a broad range of nuances rather than a smaller number of internally consistent domains or facets. To conclude, however, it appears that people with high BMI are not alike: rather, overweight (as well as normal weight and underweight) are associated with different configurations of traits.

BMI and Personality Item Content
The meta-analytic results suggested that items most strongly related to BMI had, on average, higher behavioral content. This is consistent with personality influencing BMI through health-relevant behavior, a prominent hypothesis to explain their associations (Keller & Siegrist, 2015;Lunn et al., 2014). For instance, some traits have behavioral ex pressions that clearly relate to intake or expenditure of calories: traits related to activity levels entail keeping busy and moving vigorously (Costa & McCrae, 1992) which require energy; traits related to getting angry can lead to impulsive eating behavior (Canetti et al., 2002), increasing energy intake. If personality traits represent the frequency of trait-relevant expressions (Fleeson & Jayawickreme, 2015) such as certain kinds of behav iors, the current results are consistent with the view that personality traits can increase body weight through, for instance, altering the frequency of behaviors that increase or decrease intake or expenditure of calories.
While the finding that BMI is related to behavioral personality-related processes is consistent with previous literature, it is more surprising that no consistent association appeared between BMI and, for instance, affective item content. After all, studies have shown that positive and negative emotions alike can increase calorie intake (Canetti et al., 2002;Evers et al., 2018;Sultson et al., 2017). Because affective items represent all kinds of emotions regardless of valence, it seems likely that a propensity for affective experiences would lead to increases in weight. Of course, the current results do not rule out the influence of affective item content on BMI. Similarly, rather than suggesting that certain cognitions and motivations cannot be reliably linked to BMI, these results indicate that BMI does not seem to have any systemic associations with the strength, variety, or frequency of different cognitions or motivations. Indeed, the items reflecting cognition are very diverse in content, reflecting a variety of beliefs, opinions, cognitive abilities, and other peculiarities in information processing (see Wilt & Revelle, 2015, for an overview of items with high cognitive content). Therefore, many types of affective, cognitive, and motivational processes may still associate with BMI. Perhaps BMI relates to only specific types of affective, cognitive, and motivational processes while being unrelated to most, which is why we did not find such a consistent correlation as we did with behavior.
Still, when additionally accounting for education, cognitive and motivational content was also significantly related to BMI in addition to behavioral content. As a potential explanation, the variability of cognitions and motivations associated with different levels of education may suppress the association between BMI and personality item content, but certain types of thoughts and motivations may mediate the association between BMI and personality traits within groups of people with similar levels of education. However, because BMI's association with these types of item content was not consistent across analyses, these effects should be interpreted with caution.
It should be noted that we may have been limited in our ability to detect BMI's associations with item content due to the nature of the variables analyzed. Because the proportions of ABCD content in each item had to add up to 100%, a higher weight for one component meant lower weights for others and the proportion of ABCD content in each item indicated the relative importance of the components in regard to BMI. As also pointed out by Wilt & Revelle (2015), relying on percentage data can artificially introduce negative correlations between the ABCD domains. This dependency can ultimately man ifest in the ABCD components' correlations with BMI: for instance, if the component that correlates with BMI most strongly has a positive association with it, then BMI's correlations with other components may turn out negative. Therefore, the correlations should be interpreted as the contribution of a component to a trait's correlation with BMI in relation to the other components.

Implications
As discussed above (and as expected), BMI correlated with a broad range of narrow traits independently of and often more strongly than with their domains. Indeed, it also correlated with items in the 100NP dataset that did not belong in any domains. Thus, there seems to be a lot in personality that relates to BMI: results suggest that body weight correlates with items within each domain as well as items not covered by those domains, and its associations with various nuances appear stronger than its correlations with broader traits. Knowing which narrow traits are most strongly related to BMI can ultimately help pinpoint which specific trait-related expressions (e.g., behaviors) primarily account for their links with body weight. Indeed, it has been argued that all aspects of a broad trait are unlikely to be equally relevant in health outcomes, and when it comes to personality-focused health interventions, certain specific aspects of the traits may be the appropriate targets (Chapman, 2013;Murray & Booth, 2015). Our results are consistent with this view and suggest that if personality traits are ever to be considered in weight-management interventions, narrow traits and their behavioral manifestations may be the most appropriate targets. Nuance-level analyses may help pinpoint what have been called the "active ingredients" of personality traits (Chapman, 2013) that may most likely lead to improvements in health outcomes if the traits can be successfully altered.
But which items represent meaningfully unique nuances to qualify as the "active ingredients" of personality in relation to BMI? Neither the cross-sectional associations nor our intuition is the best guide in saying which items are meaningful or intervenable influences on BMI. This question can be addressed with data that allows for stronger causal inferences such as longitudinal or genetically informative data: nuances that ap pear to be causally linked with BMI are more likely to be the active ingredients whereas those that appear to simply (e.g., spuriously) correlate with or be influenced by BMI (Arumäe et al., 2021) are less likely to be the active ingredients.
However, the current results do suggest that the item-level associations should proba bly not be generalized to domains. If a domain correlates with an outcome, then all of its components (e.g., facets and nuances) should also correlate with the outcome for it to make sense that the outcome associates with the domain as a whole (Vainik et al., 2015). If, on the other hand, the components of a domain correlate with the outcome differently, then this would indicate that the components' correlations should not be generalized to the domain. In the case of BMI, it was fairly clear that the items in any of the five domains did not correlate with BMI uniformly regardless of the inventory, and therefore the associations seem to pertain to traits narrower than domains.
But beyond clarifying BMI's associations with personality traits, the results also suggest that different inventories and different samples can lead to different results. Although the strongest nuance-level associations highlighted certain narrow traits that appeared to correlate with BMI consistently, some associations appeared more dependent on the particular dataset. One noteworthy difference between the inventories is the absence of strong correlations with impulsivity-related traits in the NEO-PI-3 data (aside from the ones with the two eating-related items). Similar results have been ob served before and it has been concluded that the association between BMI and the trait impulsiveness is mostly driven by these eating-related items (Terracciano et al., 2009). In contrast, items related to impulsivity and immoderation were among the highest correlated with BMI in both IPIP-NEO and 100NP datasets, suggesting that people with high BMI are also impulsive in other ways beyond tending to overeat. Why exactly most of the items in the impulsiveness facet of NEO-PI-3 do not associate with BMI is unclear, but nevertheless, these results illustrate the importance of using different instruments, even to measure the ostensibly same constructs, to fully understand traits' associations with BMI and other outcomes.
Differences between inventories and samples were also evident in domain-level cor relations. Even though all three inventories measure the same domains, these domains' correlations with BMI were, at times, substantially different, ranging from .02 to .11 for neuroticism and from -.09 to .01 for openness. Whether these differences are due to the different items included in each inventory or due to sample characteristics, the correla tions suggest that relying on broader traits and a single inventory is likely insufficient to understand what in personality relates to BMI. If the differences in the domain-level correlations are due to the way different inventories measure the traits, it would be hard to say which of the estimates is the most "correct". However, in this situation, nuance-level analyses can bring some clarity. To conclude, these results illustrate the necessity of using inventories that cover a broad set of (narrow) personality traits.

Limitations
While using different samples and different inventories is a strength of this study, it is nevertheless hard to tell where domain-and nuance-level correlations differed between datasets due to the way particular inventories conceptualize or measure these traits and where they differed due to sample characteristics. Although the samples included people of diverse backgrounds, people of western countries were altogether overrepresented in the current study. The IPIP-NEO and 100NP samples were likely affected by self-selection bias; for instance, both samples had higher average levels of openness than the Estonian Biobank sample. Sample characteristics should therefore be considered when interpreting and comparing the results of the different datasets.
Further, it should also be considered that all items in each personality inventory may not be equally good representations of actual personality nuances. That is, items differ in the extent to which they capture unique variance in personality, with nuances that have higher cross-rater agreement also having, on average, higher rank-order stability and heritability . Even though the current study considered a larger num ber of personality traits than previous studies have, this is by no means a comprehensive set of personality items, and other nuances not included here may also have meaningful correlations with BMI.
Finally, although cross-sectional correlations can generate hypotheses, they are unin formative on the existence of causal relations: longitudinal or causal associations may not even match cross-sectional ones. Studies aiming to test if the nuance-BMI links are consistent with causal influences of nuances on BMI (or vice versa) could compare media tion models to test possible causal pathways and to rule out possible confounders (Deary et al., 2010), as well as make use of longitudinal and genetically informative designs. Although these possibilities are limited by the availability of nuanced personality data (e.g., large studies tend to use brief personality inventories), these kinds of data are more likely to be collected once the relevance of narrow personality traits in important life outcomes becomes more widely acknowledged.

Conclusions
To summarize, we found that BMI correlated with many narrow traits across the person ality space and that these correlations were often stronger than correlations with broad traits. Analyses with item content indicated that BMI may primarily relate to behavioral content of personality traits. Beyond describing how people with overweight differ from others, these results may have implications for weight-management interventions: if personality traits can be considered in these interventions, narrow traits and their behavioral manifestations may ultimately be the optimal level of personality-related phenomena to target. Finally, the results indicate that the clearest understanding of the personality profile(s) that accompany high BMI can be best understood by using inventories that cover a large number of traits.
Funding: Uku Vainik has been funded by Estonian Research Council's personal research funding start-up grants PSG656 and PSG759. The funder had no role in the study's design, execution, analysis, interpretation, or reporting.