In Sample 4, after the end of their conversation, participants completed the interaction evaluation questionnaire (see Appendix 7.4). The items of this post-interaction questionnaire were factor-analyzed to reduce the number of variables. Results of this analysis are presented in the following section.

To assess data quality, the level of agreement across data sources was calculated. First, it was tested whether persons who thought they were very understanding as interviewers actually had partners who felt understood during that interaction half. In addition, self- and partner-ratings were correlated with the aggregated MU observations by the student assistant judges. Table 10 shows the results of these analyses. As can be seen, all correlations were positive and at least marginally significant. Specifically, if interviewers reported being very understanding, their interaction partners felt better understood during the corresponding interaction half (rs = .31 and .26 for the first and second interaction halves, respectively). In addition, both partners’ impression of the degree of MU was corroborated by the observational ratings, with correlations ranging between .22 and .36 (all ps ≤ .10).

Note. IU = Self-rated understanding as interviewer, PU = felt understanding as partner, MU = mutual understanding rated by outside observers; index of interaction half in brackets
** p < .01, * p < .05, † p < .10

IU(2)

PU(2)

PU(1)

MU(1)

MU(2)

IU(1)

.12

.27*

.31**

.22†

.23†

IU(2)

.26*

.66**

.31**

.28*

PU(2)

.14

.22†

.36**

PU(1)

.34**

.30**

MU(1)

.74**

Table 10 Correlations Between Self-Reported and Observed Indices of MU

To obtain at a dyadic measure of MU, a factor analysis of both participants’ self-rated Interviewer and Interviewee Understanding, and the amount of observed MU for each interaction half (i.e., six variables) was conducted. Inspection of the scree plot suggested a one-factor solution, with the first factor explaining 45% of the variance, and factor loadings ranging between .50 and .76. Accordingly, the average of these six variables was taken as a composite index of MU, which had an acceptable reliability (α = .74), especially considering the heterogeneity of information sources. In addition, similar composite indices were created for each interaction half (because these consisted of only three items each, reliabilities were deflated to .54 and .57 for the first and second half, respectively).

An important issue in using dyadic data is that the individual members of the dyad can be interdependent. That is, for whatever reason, members may resemble each other in terms of certain characteristics. Failing to account for this interdependency may lead to biased conclusions (Kenny & la Voie, 1985). Because the dyads were formed unsystematically (with the exception of pretest intelligence) and did not have contact previous to the interaction, interdependence in Study 4 can only be due to the interaction itself (e.g., when both participants act in a relaxed way because of a nice conversation). Accordingly, no dyadic associations between partners on any of the pretest measures were expected.

In line with the expectation of non-interdependence, (intraclass) correlational analyses showed that correlations between participants’ crystallized and fluid intelligence scores ranged between -.11 and .03 and were not statistically significant, ps > .30. Similarly, openness to experience did not correlate across interaction partners, r = -.05, p = .69. Finally, both participants’ self-ratings and partner-ratings were uncorrelated between dyadic partners, r = -.05 and .02 (ps > .60), respectively. Accordingly, it was not necessary to account for dyadic interdependency regarding the current predictor variables.

Because the data were virtually non-interdependent, applying the formula by Burr and Nesselroade (1990) results in estimates of the reliability of the difference scores between two participants that approximate the reliability of the corresponding scales. The only exception is the difference between participants’ self-ratings of intelligence and their ratings of the intelligence of their interaction partners, which were significantly correlated (r = .40, p < .01). When the reliability of the single-item measure is estimated at .70, then the reliability of the (intra-individual) difference score does not exceed .50. Of course, the low reliability of the intra-individual difference score also affects the results of Studies 1-3.

To address the main effect hypotheses regarding fluid intelligence and openness to experience on the amount of MU in relationships (Main Effect Hypothesis 1 and 3, respectively), the ego-centered network data were analyzed using HLM. On the relationship-specific Level 1, the composite relationship quality index served as the dependent variable. The intercept (i.e., average level of relationship quality within a person’s social network) served as outcome variables in a participant-specific Level 2 regression. As predictors of the intercept

No covariates were included as predictors of the intelligence rating x relationship quality slope because the analysis focuses on main effects, not on moderator effects.

Table 11 shows the regression coefficients of the HLM analysis of main effects in Samples 1-2 and 3. As can be seen, the Table is divided into two horizontal halves covering the different samples. The first line of each half specifies the Level 2 intercept (the average relationship quality), followed by the slope of the association between rated intelligence and relationship quality. In lines 2-5, Level 2 moderators of the Level 1 intercept are displayed. For example, in Sample 1-2, an elevation of 1 SD in Openness increases the Level 1 intercept by .04 points.

As can be seen in Table 11, the parameter linking partner intelligence with relationship quality was significant and positive on Level 1, indicating that relationships with more intelligent network partners were rated as higher in quality. Thus, Main Effect Hypothesis 1 was confirmed regarding the relationship-specific Level 1. In contrast, self-rated intelligence on Level 2 had a significantly negative influence on the average level of relationship quality. That is, highly intelligent individuals reported less satisfying and understanding relationships, which is inconsistent with Main Effect Hypothesis 1. Finally, participants’ openness to experience was positively related to average relationship quality on Level 2, but this effect was not significant. Accordingly, Main Effect Hypothesis 3 was not supported.

Note. IQ = intelligence rating, corrected for age and gender. All continuous variables were standardized before entering in the analysis. In Samples 1-2 (Level 1, N = 7,863; Level 2, N = 410), random effects were estimated at .36, .20, and .86 for the Level 1 intercept, the IQ x relationship quality slope, and the residual variance, respectively (ps < .01). In Sample 3 (Level 1, N = 5,153; Level 2, N = 511), the corresponding estimates were .50, .22, and .79 (ps < .01).
** p < .01, * p < .05

Level 1 (relationship partner)

Relationship quality

intercept

Rated partner IQ

slope

Sample 1-2

Level 2

-Openness

-Self-rated IQ

-0.21**

-Network size

-0.07**

-Age

-0.01

-Female gender

0.03

Sample 3

0.41**

Level 2

-Openness

0.01

-Vocabulary

-Network size

-0.04*

-Age

-0.06*

-Female gender

0.28**

Table 11 HLM Regression Coefficients of the Effect of Rated Intelligence, Openness, and Vocabulary on Relationship Quality in Samples 1-2 and 3

In the current case, there were extreme differences between the Mensa (Sample 1.I) and alumni (Sample 2) in the mean level and range of the single item intelligence rating. Specifically, the average (single-item) self-rated intelligence was 19.7 for the Mensa members against 16.2 for the university alumni, a highly significant difference, F(1, 431) = 555.74, p = .01. In addition, Mensa members had a SD of .62 on the single item rating, whereas university alumni had a SD of 2.15. An F-test showed that this difference is significant, F(221, 219) = 3.49, p = .01. Because of these large differences in distribution, the single item intelligence rating is confounded with sample membership (i.e., a participant with a self-rating of 20 points almost certainly belongs to the Mensa sample).

Because of the potentially unmeasured selection bias accompanying Mensa membership (besides having a high intelligence), the single item intelligence rating scale is not ideal to test for the effects of intelligence on social relationships. Sample differences regarding the four-item intelligence self-concept scale were comparably smaller, though significant, F(1, 435) = 82.47, p < .01, whereas the SD of this scale did not differ between the Mensa members (SD = .68) and the university alumni (SD = .73), F(222, 223) = 1.09, p = .26. When the single-item intelligence ratings were replaced with the intelligence self-concept scale, no significant effect on the relationship quality intercept was found (p = .27), even though the trend was again negative (b = -0.03). Accordingly, the negative effect of intelligence self-ratings found for the single item measure was not replicated for the intelligence self-concept scale.

In sum, a negative/null association between participants’ self-rated intelligence and their social network’s average level of MU was found on Level 2. In contrast, a positive association between participants’ rating of their network partners’ intelligence and the quality of their relationship was found on Level 1. Thus, results differed between the relationship-specific (dyadic) Level 1 and the participant-specific Level 2. In interpreting this discrepancy, it should be noted that these two levels are mathematically independent of each other and represent different research questions. For example, whereas the Level 2 association between rated intelligence and MU is dependent of the relationship quality intercept, the association on Level 1 is independent of this parameter.

Like in Sample 1.I, 39 Mensa members filled out the NEO-FFI personality questionnaire, provided a list of their network partners and rated each partner’s intelligence and the quality of the corresponding relationship. In addition, a total of 172 network partners rated their own openness to experience and intelligence as well as the quality of the relationship with the Mensa member. Because the intelligence self-ratings of the Mensa members did not show any variation, they could not be used to assess main effects. In addition, it turned out that also the network partners’ ratings of the Mensa members’ intelligence were extremely right skewed (kurtosis -2.35, se .18, p < .01). As 42% of the network partners used the highest or second-highest intelligence rating (i.e., 19 or 20), many were apparently aware of the gifted status of the Mensa members, most likely because of their publicly known membership in an organization for the gifted. Because the intelligence ratings of the network partners were highly confounded with the availability of this information, they were not further used in the subsequent analyses.

Because two rating sources (Mensa members and network partners) for both dependent and independent variables were available, main effects were tested with four separate HLM analyses. The dependent variable in these regressions was the amount of relationship quality, predicted by ratings of partner intelligence while controlling for the total size of the social network and participants’ gender.

Different from the previous analyses, age was not included because the relatively high number of missing values for this variable would have reduced the already limited sample size.

Note. MM = Mensa member, NP = network partner, IQ = intelligence rating, corrected for age and gender. All continuous variables were standardized before entering in the analysis. Level 1, N = 172; Level 2, N = 39. Mean random effects (across all four analyses) were estimated at .26, .10, and .89 for the Level 1 intercept, the IQ x relationship quality slope, and the residual variance, respectively (ps ≤ .36). The table is divided into four horizontal parts, depending on the source (left of colon) and target , whereas the target (right of colon) of the intelligence ratings rating
** p < .01, * p < .05

Level 1 (relationship partner)

Relationship quality

intercept

Rated partner IQ

slope

MM: Quality/MM: IQ

0.24

0.24**

Level 2

-Network size

0.00

-Female gender

0.04

MM: Quality/NP: IQ

0.40**

-0.05

Level 2

-Network size

-0.03

-Female gender

-0.06

NP: Quality/MM: IQ

0.01

-0.04

Level 2

-Network size

-0.01

-Female gender

0.02

NP: Quality/NP: IQ

0.02

-0.07

Level 2

-Network size

0.00

-Female gender

-0.02

Table 12 HLM Regression Coefficients of the Effect of Ratings of Network Partner Intelligence on Relationship Quality in Sample 1.II

Like in Samples 1-2, a HLM analysis with vocabulary and openness as Level 2 main effect variables and rated network partner intelligence as Level 1 main effect variables was carried out. Instead of intelligence self-ratings (not assessed in this Sample), the vocabulary test score was included as a measure of crystallized intelligence. Like in the previous analyses, gender, age, and total network size were included as Level 2 covariates.

Table 11 (lower part) shows the results of the HLM analysis. As can be seen, neither the effect of vocabulary nor the effect of openness was significant on Level 2. As found in Samples 1-2, however, the main effect of rated partner intelligence on relationship quality was significant and positive. This is consistent with Main Effect Hypothesis 1, which predicts that intelligence and relationship quality are positively associated.

In sum, results for Sample 3 resulted in positive results for Main Effect Hypothesis 1, but no support for Main Effect Hypotheses 2 and 3. Consistent with Main Effect Hypothesis 1 and replicating results from Samples 1-2, participants’ ratings of their network partners’ intelligence (on Level 1) were positively associated with relationship quality. In contrast, vocabulary level as assessed with a psychometric test (on Level 2) was not associated with relationship quality, which disconfirms Main Effect Hypothesis 2. Finally, self-ratings of openness to experience (on Level 2) were not significantly associated with relationship quality, which disconfirms Main Effect Hypothesis 3.

According to the current Main Effect Hypotheses, MU should be positively related to intelligence, vocabulary, and openness to experience. In Sample 4, self-ratings, ratings by the interaction partner, and intelligence test results were available as information sources for fluid intelligence. In addition, vocabulary was measured with the help of a psychometric test (MWT), and openness with an established self-rating instrument (NEO-FFI). The level of observed and self-reported MU after the interaction served as dependent variable.

The intercorrelation matrix of the different predictor measures (see Table 13) was inspected to see whether it would be possible to create aggregated variables. As can be seen, self-ratings of intelligence were significantly correlated with psychometric numerical intelligence and openness to experience. Vocabulary was significantly correlated with openness to experience and marginally significantly with numerical intelligence. These correlations are consistent with previous research (e.g., Ashton et al., 2000; Paulhus et al., 1998), but they are not high enough to create composite measures.

Applying the Spearman-Brown formula shows that aggregating scales that are significantly correlated (e.g., self-ratings of intelligence and numerical intelligence) would result in inadequate reliability levels (i.e., alpha levels lower than .37).

Note. IQ SR = Self-rated IQ, IQ PR = Rating by interaction partner IQ, IQ NUM = numerical IQ, IQ FIG = figural IQ, IQ VOC = Vocabulary, O = Openness. Ns ranged from 139 (correlations with intelligence ratings) to 200 (correlations among pretest measures).
** p < .01, * p < .05, † p < .10

IQ SR

IQ PR

IQ NUM

IQ VOC

IQ FIG

O

IQ SR

1.00

.03

.23**

-.03

.04

.20*

IQ PR

1.00

-.01

-.03

-.11

IQ NUM

1.00

.13†

.10

.04

IQ VOC

1.00

.01

.17*

IQ FIG

1.00

-.05

O

1.00

Table 13 Correlations Among Predictor Variables in Sample 4 (Intelligence and Openness to Experience)

Main effects in Sample 4 were tested by a series of correlational analyses. As can be seen in Table 14, main effects were limited to partner-ratings of intelligence, and psychometrically assessed vocabulary. Interestingly, main effects differed between participants (i.e., first vs. second interviewer) but generalized across interaction halves (i.e., first vs. second half). That is, effects that were found for the first interviewer were not replicated for the second interviewer (and vice versa), but the effects (and lack of effects) of both persons’ personality were not constrained to one interaction half.

Across both interaction roles and interaction halves, two main effects of intelligence were significant. To begin with, the first interviewer’s level of vocabulary was significantly positively related to MU during the conversation. The positive effect of crystallized intelligence is consistent with Main Effect Hypothesis 2. However, no effect was found for the second interviewer’s vocabulary. Second, the rating of the first interviewer’s intelligence by his or her interaction partner was positively related to MU. However, no effect was found for the first interviewer’s ratings of the intelligence of the second interviewer. Besides the effects of the first interviewer’s vocabulary level and intelligence ratings by his/her partner, no other personality variables exerted a significant main effect on MU. That is, no significant main effects were found for self-ratings of intelligence, psychometric (figural, numerical) intelligence, or self-reported openness to experience.

Note. MU = Composite measure of self-reported and observed mutual understanding (interaction half in brackets)
** p < .01, * p < .05

Target of personality assessment

Interviewer(1)

Interviewer(2)

Source

MU

MU(1)

MU(2)

MU

MU(1)

MU(2)

-.12

-.19

-.01

.08

.15

-.01

.35**

.35**

.28**

-.07

-.09

-.04

Intelligence tests

-Numerical IQ

-.04

.00

-.08

.08

.07

.07

-Figural IQ

-.03

-.03

-.02

-.03

-.03

-.01

-Vocabulary

.30**

.27**

.27**

-.03

-.06

.01

Openness to experience

.13

.14

.09

.12

.15

.06

Table 14 Correlations Between Intelligence/Dispositional Valuations and MU in Sample 4

In sum, Main Effect Hypothesis 1 received weak support, with the ratings of the first interviewers’ intelligence by their partners being positively related to MU, but not the ratings of the second interviewers’ intelligence. Because the positive association between intelligence and MU was not replicated with other intelligence indices, this finding does not seem very robust, however. Second, Main Effect Hypothesis 2 received mixed support, with the level of vocabulary being positively associated with MU for the first but not for the second interviewer. Third, Main Effect Hypothesis 3 received no support since openness to experience was not related to the level of MU.

Because the extremely high intelligence level of the Mensa members makes it impossible to discriminate between main and dyadic effects in this sample (see Section 3.4.3), no tests of dyadic effects were conducted (for tests of main effects, see Section 4.2).

Because of the previously discussed problems with the self- and partner-ratings of intelligence of the Mensa members (see Section 3.4.3), these indices could not be used to test dyadic effects. In contrast, both dyadic partners completed the Openness to Experience scale of the NEO-FFI, so this information was available to calculate absolute difference scores that were used in three HLM analyses. First, an analysis was carried out to assess the effect of between-person differences in openness on a composite index of relationship quality (average of both dyadic partners’ ratings). However, because the degree of convergence across both raters was only modest (r = .24, p = .01), two separate HLM analyses were calculated to assess the impact of between-persons difference in openness on each participant’s idiosyncratic relationship quality ratings. As was done for the test of main effects in this sample (see Section 4.2.2), Mensa members’ gender and network size were included as control variables.

Inconsistent with Emergent Effect Hypothesis 4, results did not show any effects of between-person differences in openness on any of the three relationship quality indices. Specifically, the absolute difference score of the two dyadic partners’ openness values were neither associated with the average of both persons’ relationship quality ratings (b = -.07, p = .40), nor to partners’ idiosyncratic quality ratings (bs = -.02 and -.11 for Mensa members’ and relationship partners’ quality ratings, respectively, ps > .20).

In Sample 2, dyadic effects were studied in the following way. For every relationship, the difference between the university alumni’s self-reported intelligence and their ratings of their network partners’ intelligence was calculated. This difference score had an average of 2.7 (SD = 3.2), indicating that the alumni saw themselves as more intelligent than the people with whom they frequently interacted. To calculate the impact of between-person intelligence differences, the absolute value of this difference score was inserted in an HLM analysis as a Level 1 predictor of relationship quality.

Because the difference score varies across different relationships, it was deemed a Level 1 property. Note, however, that intelligence differences calculated in this way are not pure Level 1 measures because they are dependent on Level 2 information (i.e., the self-rated IQ of the university alumni).

Table 15 shows the results of the HLM analysis for the effects of absolute differences scores on the quality of social relationships. As can be seen, there was a significantly negative effect of absolute intelligence differences on relationship quality. Thus, when participants perceived larger between-person differences in intelligence, they also perceived the quality of the social relationship as lower, which is consistent with Dyadic Effect Hypothesis 1.

In only 4.3% of all relationships, participants rated their network partners as more than one scale point higher in intelligence than themselves (vs. 57.8% lower in intelligence). As a result, if there is a detrimental effect of intelligence differences when the network partner is rated as more intelligent, this is likely obscured by the sheer number of relationships where the opposite is the case. Because of the bias towards higher self-ratings, the difference score can also be seen as an index of the degree to which contact persons have a lower intelligence than the alumni.

To test the notion that communication is also hampered when a contact person has a higher intelligence level, two additional dummy variables were created. First, a dummy was created for all dyads in which participants rated a contact person as more than one scale point lower in intelligence than themselves (this concerned 5,006 dyads). The same was done for dyads in which participants rated a contact person as more than one scale point higher in intelligence (this concerned 376 dyads). Results showed that the dummy that specified relationships in which the network partner was rated as less intelligent was associated with lower levels of relationship quality (b = -.49, p = .00). In contrast, there was a non-significantly positive effect of the dummy that specified relationships in which the network partner was rated as moreintelligent (b = .15, p = .25).

Although the lack of significance for the dummy specifying relationships with more intelligent network partners may be the result of unequal cell sizes, the positive sign clearly disconfirms Dyadic Effect Hypothesis 1, which states that differences in intelligence should be related to lower levels of MU. Instead, the pattern of findings is more consistent with a positive main effect of participants’ ratings of their network partners’ intelligence.

In sum, the Dyadic Hypothesis that intelligence differences between persons are related to lower levels of MU received modest support. True, when participants perceived network partner as lower in intelligence than themselves, the relationship was rated as lower in quality. Yet the crucial prerequisite for a demonstration of a dyadic effect (lower quality ratings for relationships with network partners that are perceived as higher in intelligence) was not met in the current study. Accordingly, the significant „dyadic” effect of intelligent difference is more likely due to a main effect of rated intelligence on social relationship quality.

Note. IQ = intelligence rating, corrected for age and gender. All continuous variables were standardized before entering in the analysis. Level 1, N = 4,537; Level 2, N = 205. Mean random effects (across all four analyses) were estimated at .30, .19, and .90 for the Level 1 intercept, the IQ x relationship quality slope, and the residual variance, respectively (ps < .01).
** p < .01, * p < .05

Level 1 (relationship partner)

Relationship quality

intercept

Between-person IQ
difference slope

-0.29**

Level 2

-Network size

-0.08**

-Age

0.01

-Female gender

0.10*

Table 15 HLM Regression Coefficients of the Effect of Between-Person Differences in Rated Intelligence and Relationship Quality in Sample 2

In Sample 4, univariate between-person differences in intelligence and dispositional valuations were operationalized as the absolute difference between interaction partners’ scores. Second, the Euclidean distance between interaction partners’ personality profiles was calculated as an index of multivariate between-person differences.

Using the formula where d_ij is the Euclidean distance between persons i and j, m represents the number of personality dimensions, and x_ia and x_ja are the scores on personality variable a for person i and j, respectively.

To investigate the effect of perceived intelligence differences more fully, three different indices were created. First, an index of self-concept similarity was calculated by taking the absolute difference of both participants’ self-ratings of intelligence. Second, an index of target agreement, operationalized as the absolute difference between participants’ self-rating and the rating provided by their partners, was calculated. Third, an index of perceived similarity was calculated by taking the difference between individuals’ ratings of their own intelligence and their partners’ intelligence.

Table 16 contains correlations between MU (composite of participants’ ratings and behavioral observations) and between-person differences in intelligence/dispositional valuations. As can be seen, only one dyadic effect was marginally significant: The level of disagreement regarding the first interviewer was negatively correlated with the amount of MU. Thus, conversations between participants with very divergent opinions about the intelligence of the first interviewer were less understanding than conversations in which there was a large degree of agreement on this issue. The level of agreement regarding the second interviewer was not associated with MU.

Apart from the effect of intersubjective agreement regarding the first interviewer’s intelligence level, no other dyadic effects were significant. Between-person differences in psychometric intelligence tests and self-reported openness were unrelated to the level of MU. Also, profile similarity regarding the psychometric intelligence measures (i.e., numerical and figural intelligence, and vocabulary), the Big Five factors (excluding openness), the AIST interests scales, and the Rokeach value rankings was not significantly related to MU. Accordingly, the emergent effect hypotheses did not receive much support.

^a More positive values are indicative of less similarity
Note. MU = Composite measure of self-reported and observed mutual understanding (interaction half in brackets), FFM = Factors of the Five-Factor Model of personality description. N = 68-72.
** p = .01, * p = .05, † p = .10

Impact of absolute difference

MU

MU(1)

MU(2)

-Self-concept similarity ^a

-.06

-Target agreement (Interviewer 1) ^a

-.22†

-.17

-.23†

-Target agreement (Interviewer 2) ^a

-.06

.04

-.15

-Perceived similarity (Interviewer 1) ^a

-.05

-.05

-.04

-Perceived similarity (Interviewer 2) ^a

.07

.08

.04

Intelligence tests

-Numerical IQ

-.06

.05

-Figural IQ

-.09

-.01

-.16

-Vocabulary

-.01

-.03

.02

Openness

.04

-.08

.14

Profile similarity

- Intelligence + vocabulary

-.04

-.06

-.01

- FFM (excl. Openness)

.10

.09

.09

- Values

-.12

-.07

-.14

- Interests

-.10

-.11

-.07

Table 16 Correlations Between MU and Between-Person Differences in Intelligence and Dispositional Valuations in Sample 4

Multivariate analyses of variance (MANOVA) were carried out with sample membership (Mensa member, highly achieving alumni, and averagely achieving alumni) as between-subjects factor and the above described adjustment and MU indicators as dependent variables. Except for neuroticism, significant group differences were found for all variables (see Table 17 and Figure 6), multivariate F(6, 431) = 27.82, p < .01. Planned contrasts showed that this was due to differences between Mensa members and alumni (ps ≤ .01 for all variables except neuroticism). In contrast, differences between highly and averagely achieving university alumni were not significant except for self-esteem of opposite-sex relationships, with somewhat higher values for the high achievers (p = .02).

Note. Because of the lack of significant differences between high achievers and average achievers, these groups were pooled together in the comparisons with the Mensa members.
^a Based on the difference between Mensa members and university alumni
^b Calculated with Cohen’s formula d = (M ₁ - M ₂) / SD _p , where M ₁ is the mean for the Mensa group, M ₂ is the mean for the alumni, and SD _p is the pooled SD across the entire sample
** p < .01, * p < .05

Sample

Mean

SD

F ^a

d ^b

Neuroticism

Mensa members

2.65

0.84

2.01

0.15

High achievers

2.50

0.64

Average achievers

2.57

0.61

General self-esteem

Mensa members

3.72

1.03

5.88**

-0.30

High achievers

4.03

0.76

Average achievers

3.96

0.73

Opposite-sex self-esteem

Mensa members

3.50

0.95

-0.43

High achievers

4.00

0.74

Average achievers

3.74

0.83

Same-sex self-esteem

Mensa members

3.23

0.86

46.19**

-0.90

High achievers

3.95

0.70

Average achievers

3.93

0.72

Parents self-esteem

Mensa members

3.26

1.00

31.31**

-0.72

High achievers

4.00

0.85

Average achievers

3.88

0.92

Loneliness

Mensa members

2.30

0.70

49.37**

0.92

High achievers

1.73

0.44

Average achievers

1.79

0.42

Size of social network

Mensa members

16.41

9.06

29.86**

-0.71

High achievers

22.91

8.39

Average achievers

22.29

8.51

Table 17 Differences Between Mensa Members and University Alumni in Self-Reported Social Adjustment

An inspection of the effect sizes of the difference between the Mensa members and university alumni

High and average achievers were pooled together because they did not significantly differ from each other in their level of social adjustment.

Figure 6. Mean Level Differences in Self-Rated Social Adjustment Between Intellectually Gifted Mensa Members, Highly Achieving University Alumni, and Averagely Achieving University Alumni.

Group comparisons between Mensa members and university alumni are somewhat biased because these groups differed in a number of respects. For example, Mensa members were older and less educated, which may have affected the comparisons. To correct for this potential bias, GLM analyses with participant age as a covariate and gender, sample, questionnaire format, and their interaction terms as a between-subjects factors were carried out.

Findings showed that the multivariate effect of participant sample remained significant, F(6, 401) = 4.55, p < .01. Inspection of univariate follow-up tests showed that Mensa members no longer differed from the university alumni in terms of general self-esteem and total network size. However, the effects for the self-esteem of same-sex, opposite-sex and parental relationships, and loneliness remained significant (ps ≤ .03). Accordingly, the lower social self-esteem and higher loneliness of the Mensa members was not a by-product of differences in gender, age, education or questionnaire format. This result is consistent with the Extreme Group Hypothesis.

Finally, an especially stringent test was created to compare Mensa members and highly achieving university alumni in terms of their social adjustment. This test included the average absolute difference between participants’ and their network partners’ rated intelligence as a control variable (in addition to the covariates listed above). The inclusion of this control variable provides an indication whether the differences in adjustment between Mensa members and university graduates are due to larger between-person differences in intelligence (as the Simonton model would predict) or to other, unknown factors (e.g., selection bias). Results of this stringent analysis replicated the lower level of social adjustment for Mensa members, multivariate F(6, 318) = 3.25, p < .01. In contrast, the multivariate effect of intelligence differences became only marginally significant after controlling for Mensa membership, F(6, 318), = 1.90, p = .08. This clearly speaks against the notion that the lower social adjustment of the Mensa members is due to between-person differences in intelligence.

As described above, group comparisons between Mensa members and university alumni uncovered significant differences in terms of their social self-esteem and feelings of loneliness. These differences were found in spite of the absence of sample differences in more generalized emotional tendencies such as Neuroticism and general self-esteem. Because MU theoretically is an important determinant of the quality of social relationships, the lower relationship quality observed for gifted individuals indirectly supports the hypothesis that they experience a lower level of MU in their social relationships.

To investigate whether this hypothesis is also confirmed by the ego-centered network data, a HLM analysis was carried out. As covariates , network size

Because only 19% of Mensa members vs. 74% of alumni used the Internet version of the questionnaire, the cell sizes for this comparison were relatively unbalanced. Because of this, the inclusion of both Mensa membership and questionnaire format would lead to colinearity among predictors (r = .54, p = .01) and bias the linear HLM analysis. Note, however, that the most important difference between both questionnaire formats seems to be the smaller number of contact persons in the Internet condition. Because the total size of the ego-centered network was included as a covariate in the HLM analysis, concerns about this possible source of bias can be somewhat assuaged.

Results of the analysis are shown in Table 18. As can be seen, Mensa members did not differ from the university alumni in terms of the average relationship intercept. Thus, the level of MU in relationships with „typical”

In this case, „typical” refers to a relationship with a person who does not belong to any of the dummy-coded categories (i.e., who is not a family member, friend or partner).

Note. IQ = intelligence rating, corrected for age and gender. All continuous variables were standardized before entering in the analysis. Level 1, N = 7,808; Level 2, N = 405. Mean random effects (across all four analyses) were estimated at .35, .20, and .74 for the Level 1 intercept, the IQ x relationship quality slope, and the residual variance, respectively (ps < .01).
** p < .01, * p < .05

Level 1 (relationship-specific)

Intercept

IQ

Parent

Family

Partner

Friend

-0.60**

1.21**

0.81**

1.86**

-0.09

-0.01

-0.38**

-0.27*

-0.12

-0.02

0.00

-0.15*

0.06

-0.06

0.02

0.10**

-0.04

Table 18 HLM Regression Coefficients of the Effect of Ratings of Network Partner Intelligence and Mensa Membership on Relationship Quality Across Different Relational Categories

In light of the above-described social adjustment problems of Mensa members, it is interesting that (controlling for age and education) the current study did not find any differences between Mensa members and university alumni in terms of general self-esteem. This raises the possibility of a potential compensatory function of the Mensa membership. That is, it may be that gifted individuals who experience social adjustment problems somehow “compensate” this threat to their self-esteem by means of their higher self-concept of intelligence, of which the Mensa membership is the symbolic emblem.

Alternatively, it may be that the lower social self-concept of the Mensa members is a result of an intra-individual contrast effect resulting from their high self-concept of intelligence (Marsh, 1986). Note, however, that the highly achieving university alumni can also be expected to have a high self-concept of intelligence, yet their social self-concepts are significantly more positive when compared to the norm data reported by Schwanzer (2002; z-scores = .18 and .50 for the self-concept of peers of the same and opposite sex, respectively, ps < .01).

If the above interpretation is correct, then the Mensa members should have a lower social-concept, a higher intellectual self-concept, and an equally high general self-concept when compared to the university alumni. To test this notion, a GLM analysis with participant age and education as a covariate and Mensa-membership, gender, sample, questionnaire format, and their interaction terms as between-subjects factors were carried out (i.e., equivalent to the procedure described in Section 4.4.1). As dependent variables, the self-concept scale of intelligence and the mean of the self-concept with relationships with parents, same-sex, and opposite-sex peers were used. Results indicated that Mensa members (M = 3.33, SD = .66) had a significantly lower social self-concept than the university alumni (M = 3.95, SD = .53), F(1) = 19.33, p < .01. By comparison, Mensa members (M = 3.89, SD = .68) had a higher intellectual self-concept than the alumni (M = 3.28, SD = .74), F(1) = 35.10, p < .01.

Note that results did not change when the Mensa members were compared to the subgroup of highly achieving alumni (M = 3.34, SD = .78), F(1) = 23.04, p < .01

In sum, comparisons between Mensa members and university alumni resulted in a number of clearly delineated social differences. After controlling for age, gender, and education, Mensa members had lower self-esteem of relationships with peers of the same and opposite sex, a lower self-esteem of the relationship with parents, and a higher self-perceived loneliness. In terms of relationships with specific partners, these self-views were corroborated for relationships with family members and partners. Note that the differences in social adjustment were found despite a lack of effects for more generalized constructs such as Neuroticism or general self-esteem. In addition, Mensa members did not differ from university alumni in terms of their relationships with friends, colleagues, neighbors, and co-workers, so the differences found for family members and partners were not due to generalized response tendencies in completing the social network questionnaire. Because of these consistent social adjustment differences to the disadvantage of the Mensa members, the Extreme Group Hypothesis received strong support.