roberts et al. (2007). the power of personality

8/13/2019 Roberts Et Al. (2007). the Power of Personality

1/33

The Power of PersonalityThe Comparative Validity of Personality Traits,Socioeconomic Status, and Cognitive Ability for PredictingImportant Life OutcomesBrent W. Roberts, 1 Nathan R. Kuncel, 2 Rebecca Shiner, 3 Avshalom Caspi, 4,5 andLewis R. Goldberg 6

1University of Illinois,

2University of Minnesota,

3Colgate University,

4 Institute of Psychiatry at Kings College, London,

United Kingdom, 5 Duke University, and

6Oregon Research Institute

ABSTRACT The ability of personality traits to predict important life outcomes has traditionally been questioned because of the putative small effects of personality. In thisarticle, we compare the predictive validity of personalitytraits with that of socioeconomic status (SES) and cogni-tive ability to test the relative contribution of personalitytraits to predictions of three critical outcomes: mortality,divorce, and occupational attainment. Only evidence from prospective longitudinal studies was considered. In addi-tion, an attempt was made to limit the review to studiesthat controlled for important background factors. Resultsshowed that the magnitude of the effects of personality

traits on mortality, divorce, and occupational attainmentwas indistinguishable from the effects of SES and cognitiveability on these outcomes. These results demonstrate theinuence of personality traits on important life outcomes,highlight the need to more routinely incorporate measuresof personality into quality of life surveys, and encourage further research about the developmental origins of per-sonality traits and the processes by which these traits inuence diverse life outcomes.

Starting in the 1980s, personality psychology began a profoundrenaissance and has now become an extraordinarily diverse andintellectually stimulating eld (Pervin & John, 1999). However, just because a eld of inquiry is vibrant does not mean it ispractical or usefulone would need to show that personalitytraits predict important life outcomes, such as health and lon-gevity, marital success, and educational and occupational at-

tainment. In fact, two recent reviews have shown that differentpersonality traits are associated with outcomes in each of these domains (Caspi, Roberts, & Shiner, 2005; Ozer & Benet-Martinez, 2006). But simply showing that personality traitsare related to health, love, and attainment is not a stringent testof the utility of personality traits. These associations could bethe result of third variables, such as socioeconomic status(SES), that account for the patterns but have not been controlledfor in the studies reviewed. In addition, many of the studiesreviewed were cross-sectional and therefore lacked the meth-odological rigor to show the predictive validity of personalitytraits. A more stringent test of the importance of personality

traits can be found in prospective longitudinal studies that showthe incremental validity of personality traits over and aboveother factors.

The analyses reported in this article test whether personalitytraits are important, practical predictors of signicant lifeoutcomes. We focus on three domains: longevity/mortality,divorce, and occupational attainment in work. Within eachdomain, we evaluate empirical evidence using thegold standardof prospective longitudinal studiesthat is, those studies thatcan provide data about whether personality traits predictlife outcomes above and beyond well-known factors such as SESand cognitive abilities. To guide the interpretation drawn

from the results of these prospective longitudinal studies, weprovide benchmark relations of SES and cognitive ability withoutcomes from these three domains. The review proceeds inthree sections. First, we address some misperceptions aboutpersonality traits that are, in part, responsible for the idea thatpersonality does not predict important life outcomes. Second,we present a review of the evidence for the predictive validity of personality traits. Third, we conclude with a discussion of theimplications of our ndings and recommendations for futurework in this area.

Address correspondence to Brent W. Roberts, Department of Psy-chology, University of Illinois at Urbana-Champaign, 603 EastDaniel Street, Champaign, IL 61820; e-mail: [email protected].

PERSPECTIVES ON PSYCHOLOGICAL SCIENCE

Volume 2Number 4 313Copyrightr

2007 Association for Psychological Science


2/33

THE PERSONALITY COEFFICIENT: ANUNFORTUNATE LEGACY OF THE PERSONSITUATION

DEBATE

Before we embark on our review, it is necessary to lay to rest amyth perpetrated by the 1960s manifestation of the personsituation debate; this myth is often at the root of the perspectivethat personality traits do not predict outcomes well, if at all.Specically, in his highly inuential book, Walter Mischel(1968) argued that personality traits had limited utility in pre-dicting behavior because their correlational upper limit ap-peared to be about .30. Subsequently, this .30 value becamederided as the personality coefcient. Two conclusions wereinferred from this argument. First, personality traits have littlepredictive validity. Second, if personality traits do not predictmuch, then other factors, such as the situation, must be re-sponsible for the vast amounts of variance that are left unac-counted for. The idea that personality traits are the validityweaklings of the predictive panoply has been reiterated in un-

mitigated form to this day (e.g., Bandura, 1999; Lewis, 2001;Paul, 2004; Ross & Nisbett, 1991). In fact, this position is sowidely accepted that personality psychologists often apologizefor correlations in the range of .20 to .30 (e.g., Bornstein, 1999).

Should personality psychologists be apologetic for their modest validity coefcients? Apparently not, according toMeyer and his colleagues (Meyer et al., 2001), who did psy-chological science a service by tabling the effect sizes for a widevariety of psychological investigations and placing them side-by-side with comparable effect sizes from medicine and every-day life. These investigators made several important points.First, the modal effect size on a correlational scale for psy-

chology as a whole is between .10 and .40, including that seen inexperimental investigations (see also Hemphill, 2003). Itappears that the .30 barrier applies to most phenomena inpsychology and not just to those in the realm of personalitypsychology. Second, the very largest effects for any variables inpsychology are in the .50 to .60 range, and these are quite rare(e.g., the effect of increasing age on declining speed of infor-mation processing in adults). Third, effect sizes for assessmentmeasures and therapeutic interventions in psychology are sim-ilar to those found in medicine. It is sobering to see that theeffect sizes for many medical interventionslike consumingaspirin to treat heart disease or using chemotherapy to treat

breast cancertranslate into correlations of .02 or .03. Takentogether, thedata presentedby Meyer andcolleagues make clear that our standards for effect sizes need to be established in lightof what is typical for psychology and for other elds concernedwith human functioning.

In the decades since Mischels (1968) critique, researchershave also directly addressed the claim that situations have astronger inuence on behavior than they do on personality traits.Social psychological research on the effects of situations typi-cally involves experimental manipulation of the situation, and

the results are analyzed to establish whether the situationalmanipulation has yielded a statistically signicant difference inthe outcome. When the effects of situations are converted intothe same metric as that used in personality research (typicallythe correlation coefcient, which conveys both the direction andthesize of an effect), theeffects of personality traitsaregenerallyas strong as the effects of situations (Funder & Ozer, 1983;Sarason, Smith, & Diener, 1975). Overall, it is the moderateposition that is correct: Both the person and the situation arenecessary for explaining human behavior, given that both havecomparable relations with important outcomes.

As research on the relative magnitude of effects has docu-mented, personality psychologists should not apologize for correlations between .10 and .30, given that the effect sizesfound in personality psychology are no different than thosefound in other elds of inquiry. In addition, the importance of apredictor lies not only in the magnitude of its association withthe outcome, but also in the nature of the outcome being pre-dicted. A large association between two self-report measures of

extraversion and positive affect may be theoretically interestingbut may not offer much solace to the researcher searching for proof that extraversion is an important predictor for outcomesthat society values. In contrast, a modest correlation between apersonality trait and mortality or some other medical outcome,such as Alzheimers disease, would be quite important. More-over, when attempting to predict these critical life outcomes,even relatively small effects can be important because of their pragmatic effects and because of their cumulative effects acrossa persons life (Abelson, 1985; Funder, 2004; Rosenthal, 1990).In terms of practicality, the .03 association between takingaspirin and reducing heart attacks provides an excellent ex-

ample. In one study, this surprisingly small association resultedin 85 fewer heart attacks among the patients of 10,845 physi-cians (Rosenthal, 2000). Because of its practical signicance,this type of association should not be ignored because of thesmall effect size. In terms of cumulative effects, a seeminglysmall effect that moves a person away from pursuing his or her education early in life can have monumental consequences for that persons healthand well-being later in life (Hardarson et al.,2001). In other words, psychological processes with a statisti-cally small or moderate effect can have important effects onindividuals lives depending on the outcomes with which theyare associated and depending on whether those effects get cu-

mulated across a persons life.

PERSONALITY EFFECTS ON MORTALITY, DIVORCE,AND OCCUPATIONAL ATTAINMENT

Selection of Predictors, Outcomes, and Studiesfor This ReviewTo provide the most stringent test of the predictive validity of personality traits,we chose to focus on three objectiveoutcomes:mortality, divorce, and occupational attainment. Although we

314 Volume 2Number 4

The Comparative Predictive Validity of Personality Traits


3/33

could have chosen many different outcomes to examine, weselected these three because they are socially valued; they aremeasured in similar ways across studies; and they have beenassessed as outcomes in studies of SES, cognitive ability, andpersonality traits. Mortality needs little justication as an out-come, as most individuals value a long life. Divorce and maritalstability are important outcomes for several reasons. Divorce isa signicant source of depression and distress for many indi-viduals and can have negative consequences for children,whereas a happy marriage is one of the most important predic-tors of life satisfaction (Myers, 2000). Divorce is also linked todisproportionate drops in economic status, especially for women(Kuh & Maclean, 1990),and it canunderminemens health(e.g.,Lund, Holstein, & Osler, 2004). An intact marriage can alsopreserve cognitive function into old age for both men andwomen, particularly for those married to a high-ability spouse(Schaie, 1994).

Educational and occupational attainment are also highlyprized (Roisman, Masten, Coatsworth, & Tellegen, 2004). Re-

search on subjective well-being has shown that occupationalattainment and its important correlate, income, are not ascritical for happiness as many assume them to be (Myers, 2000).Nonetheless, educational and occupational attainment are as-sociated with greater access to many resources that can improvethequality of life (e.g., medical care, education) andwith greater social capital (i.e., greater access to various resources throughconnections with others; Bradley & Corwyn, 2002; Conger &Donnellan, 2007). The greater income resulting from higheducational and occupational attainment may also enableindividuals to maintain strong life satisfaction when faced withdifcult life circumstances (Johnson & Krueger, 2006).

To better interpret the signicance of the relations betweenpersonality traits and these outcomes, we have provided com-parative information concerning the effect of SES and cognitiveability on each of these outcomes. We chose to use SES as acomparison because it is widely accepted to be one of the mostimportantcontributors to a more successful life, including better health and higher occupational attainment (e.g., Adler et al.,1994; Gallo & Mathews, 2003; Galobardes, Lynch, & Smith,2004; Sapolsky, 2005). In addition, we chose cognitive ability asa comparison variable because, like SES, it is a widely acceptedpredictor of longevity and occupational success (Deary, Batty, &Gottfredson, 2005; Schmidt & Hunter, 1998). In this article, we

compare the effect sizes of personality traits with these twopredictors in order to understand the relative contribution of personality to a long, stable, and successful life. We also re-quired that the studies in this review make some attempt tocontrol for background variables. For example, in the case of mortality, we looked for prospective longitudinal studies thatcontrolled for previous medical conditions, gender, age, andother relevant variables.

We arenotassuming that personality traits aredirect causes of the outcomes under study. Rather, we were exclusively inter-

ested in whether personality traits predict mortality, divorce,and occupational attainment and in their modal effect sizes. If found to be robust, these patterns of statistical association theninvite thequestion of whyandhow personality traits might causethese outcomes, and we have provided several examples in eachsection of potentialmechanisms andcausal steps involved in theprocess.

The Measurement of Effect Sizes in ProspectiveLongitudinal StudiesBefore turning to the specic ndings for personality, SES, andcognitiveability, we must rst address the measurement of effectsizes in the studies reviewed here. Most of the studies that wereviewed used some form of regression analysis for either con-tinuous or categorical outcomes. In studies with continuousoutcomes, ndings were typically reported as standardized re-gression weights (beta coefcients). In studies of categoricaloutcomes, the most common effect size indicators are odds ra-

tios, relative risk ratios, or hazard ratios. Because many psy-chologists may be less familiar with these ratio statistics, a brief discussion of them is in order. In the context of individualdifferences, ratio statistics quantify the likelihood of an event(e.g., divorce, mortality) for a higher scoring group versus thelikelihood of the same event for a lower scoring group (e.g.,persons high in negative affect versus those low in negativeaffect). An odds ratio is the ratio of the odds of the event for onegroup over the odds of the same event for the second group. Therisk ratio compares the probabilities of the event occurring for the two groups. The hazard ratio assesses the probability of anevent occurring for a group over a specic window of time. For

these statistics, a value of 1.0 equals no difference in odds or probabilities. Values above 1.0 indicate increased likelihood(odds or probabilities) for the experimental (or numerator)group, with the reverse being true for values below 1.0 (down to alower limit of zero). Because of this asymmetry, the log of thesestatistics is often taken.

The primary advantage of ratio statistics in general, and therisk ratio in particular, is their ease of interpretation in appliedsettings. It is easier to understand that death is three times aslikely to occur for one group than for another than it is to makesense out of a point-biserial correlation. However, there are alsosome disadvantages that should be understood. First, ratio sta-

tistics can make effects that are actually very small in absolutemagnitude appear to be large when in fact they are very rareevents. Forexample, although it is technically correct that one isthree times as likely (risk ratio 5 3.0) to win the lottery whenbuying three tickets instead of one ticket, the improved chancesof winning are trivial in an absolute sense.

Second, there is no accepted practice for how to divide con-tinuous predictor variables when computing odds, risk, andhazard ratios. Some predictors are naturally dichotomous (e.g.,gender), but many are continuous (e.g., cognitive ability, SES).

Volume 2Number 4 315

Brent W. Roberts et al.


4/33

Researchers often divide continuous variables into some arbi-trary set of categories in order to use the odds, rate, or hazardmetrics. For example, instead of reporting an association be-tween SES and mortality using a point-biserial correlation, aresearcher may use proportional hazards models using somearbitrary categorization of SES, such as quartile estimates (e.g.,lowest versus highest quartiles). This permits the researcher todraw conclusions such as individuals from the highest categoryof SES are four times as likely to live longer than are groupslowest in SES. Although more intuitively appealing, the oddsstatements derived from categorizing continuous variablesmakes it difcult to deduce the true effect size of a relation,especially across studies. Researchers with very large samplesmay have the luxury of carving a continuous variable into veryne-grained categories (e.g., 10 categories of SES), which maylead to seemingly huge hazard ratios. In contrast, researcherswith smaller samples may only dichotomize or trichotomize thesame variables, thus resulting in smaller hazard ratios and whatappear to be smaller effects for identical predictors. Finally,

many researchers may not categorize their continuous variablesat all, which can result in hazard ratios very close to 1.0 that arenonetheless still statistically signicant. These procedures for analyzing odds, rate, and hazard ratios produce a haphazardarray of results from which it is almost impossible to discern ameaningful average effect size. 1

One of the primary tasks of this review is to transform theresults from different studies into a common metric so that a fair comparison could be made across the predictors and outcomes.For this purpose, we chose the Pearson product-moment cor-relation coefcient. We used a variety of techniques to arrive atan accurate estimate of the effect size from each study. When

transforming relative risk ratios into the correlation metric, weused several methods to arrive at the most appropriate estimateof the effect size. For example, the correlation coefcient can beestimated from reported signicance levels ( p values) and fromtest statistics such as the t test or chi-square, as well as fromother effect size indicators such as d scores (Rosenthal, 1991).Also, the correlation coefcient can be estimated directly fromrelative risk ratios and hazard ratios using the generic inversevariance approach (The Cochrane Collaboration, 2005). In thisprocedure, the relative risk ratio and condence intervals (CIs)are rst transformed into z scores, and the z scores are thentransformed into the correlation metric.

For most studies, the effect size correlation was estimatedfrom information on relative risk ratios and p values. For thelatter, we used the requivalent effect size indicator (Rosenthal &Rubin, 2003), which is computed from the sample size and p value associated with specic effects. All of these techniquestransform the effect size information to a common correlational

metric, making the results of the studies comparable acrossdifferent analytical methods. After compiling effect sizes, meta-analytic techniques were used to estimate population effectsizes in both the risk ratio and correlation metric (Hedges &Olkin, 1985). Specically, a random-effects model with nomoderators was used to estimate population effect sizes for boththe rate ratio and correlation metrics. 2 When appropriate, werst averaged multiple nonindependent effects from studies thatreported more than one relevant effect size.

The Predictive Validity of Personality Traits for MortalityBefore considering the role of personality traits in health andlongevity, we reviewed a selection of studies linking SES andcognitive ability to these same outcomes. This informationprovides a point of reference to understand the relative contri-bution of personality. Table 1 presents the ndings from 33studies examining the prospective relations of low SES and lowcognitive ability with mortality. 3 SES was measured using

measures or composites of typical SES variables including in-come, education, and occupational status. Total IQ scores werecommonly used in analyses of cognitive ability. Most studiesdemonstrated that being born into a low-SES household or achieving low SES in adulthood resulted in a higher risk of mortality (e.g., Deary & Der, 2005; Hart et al., 2003; Osler et al.,2002; Steenland, Henley, & Thun, 2002).Therelative risk ratiosand hazard ratios ranged from a low of 0.57 to a high of 1.30 andaveraged 1.24 (CIs 5 1.19 and 1.29). When translated into thecorrelation metric, the effect sizes for low SES ranged from .02to .08 and averaged .02 (CIs 5 .017 and .026).

Through theuse of the relative risk metric, we determined thatthe effect of low IQ on mortality was similar to that of SES,ranging from a modest 0.74 to 2.42 and averaging 1.19 (CIs 5

1.10 and 1.30). When translated into the correlation metric,however, the effect of low IQ on mortality was equivalent to acorrelation of .06 (CIs 5 .03 and .09), which was three timeslarger than the effect of SES on mortality. The discrepancybetween the relative risk and correlation metrics most likelyresulted because some studies reported the relative risks interms of continuous measures of IQ, which resulted in smaller

1This situation is in no way particular to epidemiological or medical studiesusing odds, rate, and hazard ratios as outcomes. The eld of psychology reportsresults in a Babylonian array of test statistics and effect sizes also.

2The population effects for the rate ratio and correlation metric were notbased on identical data because in some cases the authors did not report rateratio information or did not report enough information to compute a rate ratioand a CI.

3Most of the studies of SES and mortality were compiled from an exhaustivereview of the literature on the effect of childhood SES and mortality (Galobardeset al., 2004). We added several of the largest studies examining the effect of adult SES on mortality (e.g., Steenland et al., 2002), and to these we added theresults from the studies on cognitive ability and personality that reported SESeffects. We also did standard electronic literature searches using the terms socioeconomic status, cognitive ability, and all-cause mortality . We also exam-ined the reference sections from the list of studies and searched for papers thatcited these studies. Experts in the eld of epidemiology were also contacted andasked to identify missing studies. The resulting SES data base is representativeof the eld, and as the effects are based on over 3 million data points, the effectsizes and CIs are very stable. The studies of cognitive ability and mortalityrepresent all of the studies found that reported usable data.

316 Volume 2Number 4

The Comparative Predictive Validity of Personality Traits


5/33

T A B L E 1

S E S a n d I Q E f f e c t s o n M o r t a l i t y / L o n g e v i t y

S t u d y

N

O u t c o m e

Y e a r s

C o n t r o

l s

P r e d

i c t o r s

O u t c o m e

E s t . r

A b a s e t a l . ,

2 0 0 2

2 , 5 8 4 m e m

b e r s o f

t h e

M e d i c a l

R e s e a r c

h C o u n c i

l

E l d e r l y

H y p e r t e n s

i o n

T r i a l

A l l - c a u s e

m o r t a l i t y

1 1 y e a r s

L o w s c o r e s o n t h e

N e w

A d u l t R e a

d i n g

T e s t

( I Q )

H R

5

0 . 9 4 ( 0 . 8 6 , 1 . 0 2 )

p 5

. 1 6

r h r

5

. 0 3 a

r e 5

. 0 3 a

L o w s c o r e s o n

R a v e n s

P r o g r e s s

i v e

M a t r i c e s

( I Q )

H R

5

0 . 9 7 ( 0 . 8 8 , 1 . 0 6 )

p 5

. 5 3

r h r

5

. 0 1 a

r e 5

. 0 1 a

B a s s u k , B e r

k m a n , &

A m i c k , 2 0 0 2

9 , 0 2 5 m e n

f r o m

B o s t o n

A l l - c a u s e

m o r t a l i t y

9 y e a r s

A g e , s m o k

i n g ,

B M I , a l c o

h o l

c o n s u m p t

i o n , a c t i v i t y l e v e

l , s o c i a l

t i e s ,

h a v i n g a r e g u

l a r h e a

l t h c a r e

p r o v

i d e r , n u m

b e r o

f c h r o n

i c

c o n d

i t i o n s ,

d e p r e s s i v e s y m p t o m s ,

c o g n

i t i v e

f u n c t i o n , p h y s

i c a l

f u n c t i o n ,

h e a l t h s t a t u s

L o w a d u l t e d u c a t i o n

H R

5

1 . 3 2 ( 0 . 9 5 , 1 . 8 3 )

r h r

5

. 0 2

L o w a d u l t i n c o m e

H R

5

0 . 9 4 ( 0 . 6 5 , 1 . 3 4 )

r h r

5

. 0 0

L o w a d u l t o c c u p a t

i o n a

l

p r e s t i g e

H R

5

1 . 0 9 ( 0 . 8 6 , 1 . 3 9 )

r h r

5

. 0 1

6 , 5 1 8 w o m e n

f r o m

B o s t o n

A l l - c a u s e

m o r t a l i t y

9 y e a r s

A g e , s m o k

i n g ,

B M I , a l c o

h o l

c o n s u m p t

i o n , a c t i v i t y l e v e

l , s o c i a l

t i e s ,

h a v i n g a r e g u

l a r h e a

l t h c a r e

p r o v

i d e r , n u m

b e r o

f c h r o n

i c

c o n d

i t i o n s ,

d e p r e s s i v e s y m p t o m s ,

c o g n

i t i v e

f u n c t i o n , p h y s

i c a l

f u n c t i o n ,

h e a l t h s t a t u s

L o w a d u l t e d u c a t i o n

H R

5

0 . 7 4 ( 0 . 5 3 , 1 . 0 4 )

p

roberts et al. (2007). the power of personality

Documents