Assessing Mood in Daily LifeStructural Validity, Sensitivity to Change,
and Reliability of a Short-Scale toMeasure Three Basic Dimensions of Mood
Peter Wilhelm1 and Dominik Schoebi1,2
1University of Fribourg, Switzerland, 2University of California, Los Angeles, USA
Abstract. The repeated measurement of moods in everyday life, as is common in ambulatory monitoring, requires parsimonious scales,which may challenge the reliability of the measures. The current paper evaluates the factor structure, the reliability, and the sensitivityto change of a six-item mood scale designed for momentary assessment in daily life. We analyzed data from 187 participants who reportedtheir current mood four times per day during seven consecutive days using a multilevel approach. The results suggest that the proposedthree factors Calmness, Valence, and Energetic arousal are appropriate to assess fluctuations within persons over time. However, calmnessand valence are not distinguishable at the between-person level. Furthermore, the analyses showed that two-item scales provide measuresthat are reliable at the different levels and highly sensitive to change.
The repeated measurement of moods and emotions withhigh frequency is common in ambulatory psychologicaland psychophysiological assessment. Measurement sched-ules range from one assessment per day taken for severalweeks (e.g., Cranford, Shrout, Iida, Rafaeli, Yip, & Bolger,2006) to high-frequency assessment within a 24 h period(e.g., Ebner-Priemer & Sawitzki, 2007; Myrtek, 2004). Be-cause of the high repetition rate in such studies, the durationof a single assessment should be kept short to minimize theburden on participants. The higher the participants’ burdencaused by the frequency and duration of single assess-ments, the more likely their compliance and motivation togive valid responses will decline. Moreover, when partici-pants need to rate redundant items, additional effects likethe exaggeration of subtle differences between items mayoccur, compromising the psychometric properties of a scale(Bolger, Davis, & Rafaeli,2003; Fahrenberg, Leonhart, &Foerster, 2002; Lucas & Baird, 2006).
Consequently, some researchers have used single itemsto assess different facets of mood (e.g., Fahrenberg, Hütt-ner, & Leonhart, 2001; Myrtek, 2004). The use of singleitems, however, raises the problem that the reliability of thestate specific component of the measure cannot be deter-mined and separated from measurement error. Therefore, avariety of multi-item mood scales have been used, rangingfrom long item lists (e.g., Buse & Pawlik, 1996; Kubiak &Jonas, 2007) to specifically designed or adapted short
scales (e.g., Cranford et al., 2006). For these short scales,reliability coefficients and sometimes factor structureshave been reported, which are usually based on the analy-ses of the between-person variance (e.g., individuals’ av-erages over time). Yet, the within-person variance has oftenbeen ignored (for exceptions see e.g., Buse & Pawlik, 1996,2001; Cranford et al., 2006; Schimmack, 2003; Zelinski &Larsen, 2000; Zevon & Tellegen, 1982).
The goal of this article is to evaluate the psychometricproperties of a parsimonious six-item mood measure thatwas developed to assess three basic dimensions of mood inpeoples’ daily lives. We do so using a multilevel modelingapproach to investigate the variance and covariance be-tween items at the between-person and the within-personlevel simultaneously.
What Are Moods?
Moods are rather diffuse affective states that subtly affectour experience, cognitions, and behavior. They operatecontinuously and “provide the affective background, theemotional color to all that we do” (Davidson, 1994, p. 52).Moods can be consciously experienced as soon as they gainthe focus of our attention, and are then characterized by thepredominance of certain subjective feelings.
Moods should be distinguished from emotions. Al-though the definition of emotions depends heavily on the-oretical frameworks (e.g., Ekman & Davidson, 1994; Lew-
is & Haviland-Jones, 2000), most researchers would agreethat emotions are short-term reactions to events or stimulithat manifest themselves in different subsystems of the or-ganism (expression and behavior, physiology, subjectiveexperience, and cognitions). In contrast to emotions, moodsare not necessarily linked to an obvious cause that can berelated to an event and its specific appraisal. They showlittle synchronization of the different subsystems, do notinterrupt ongoing behavior, and do not prepare immediateactions (Scherer, 2005). Usually the intensity of moods islow to medium and they may last over hours and days.
How Can Moods Be Conceptualized andMeasured?
During the last two decades competing two-dimensional ap-proaches have dominated the discussion about the structureof mood and affect. One model, proposed by Russell, as-sumes that the core affect of a feeling is a “single integralblend” of the independent dimensions valence and arousal(Russell, 2003, p. 148). Russell, Weiss, and Mendelsohn(1989) introduced an “affect grid” to assess valence andarousal simultaneously via two items. Its brevity makes theaffect grid very attractive for ambulatory assessment research(Reicherts, Salamin, Maggiori, & Pauls, 2007). However, be-cause each dimension is assessed with one item only, mea-surement error cannot be determined for a single occasion. Inaddition, Schimmack and Grob (2000) criticized that the la-bels of the activation dimension are not close to commonlanguage and experience. In contrast to Russell, Thayer(1989) argued that two basic arousal dimensions need to bedistinguished to describe a mood state, namely tense arousal(relaxation–tension), and energetic arousal (tiredness–wake-fulness). In Thayer’s view, valence is not a separate dimen-sion, but a mix of his basic arousal dimensions.
Watson and Tellegen (1985) proposed that affects can bedescribed by two uncorrelated basic dimensions, which arecalled positive affect (PA) and negative affect (NA). Theydeveloped the Positive and Negative Affect Schedule(PANAS) to measure each dimension with 10 unipolar items.According to Watson and Vaidya (2003, p. 356) the PANAShas gained much popularity because of “the rich body ofpsychometric data that have established the reliability andvalidity of the scales.” However, the validity of the theoreticalconception of the PA and NA dimensions, the factorial solu-tion on which they are based, and the difficulty in interpretingthe scores and relating them to commonly experienced feel-ings were criticized (e.g., Fahrenberg, 2006; Russell & Car-roll, 1999a; Schimmack, 1999). Moreover, some critics re-jected the basic assumption that affect can be sufficientlydescribed by two orthogonal dimensions (Matthews, Jones,& Chamberlain, 1990; Schimmack & Grob, 2000). They ad-vocated a model in which valence (V; ranging from unpleas-ant to pleasant), calmness (C; ranging from restless/undertension to calm/relaxed), and energetic arousal (E; rangingfrom tired/without energy to awake/full of energy) form thebasic dimensions. Although these dimensions are substantial-ly correlated (cf. Table 1), they cannot be reduced to a two-dimensional model. In addition, different experimental ma-nipulations, such as taking sedative drugs or sleep depriva-tion, caused different patterns of changes in the three mooddimensions, which would not have been captured by the two-dimensional approaches discussed above (Matthews et al.,1990). Different instruments exist to measure the three mooddimensions: The UWIST Mood Adjective Checklist (Mat-thews et al., 1990) assesses each dimension with eight unipo-lar items; the German-language Multidimensional MoodQuestionnaire (MDMQ) provides short-scales consisting offour unipolar items (Steyer, Schwenkmezger, Notz, & Eid,1997). Schimmack and Grob (2000) used six unipolar items,which they combined into three bipolar items per dimension.
Table 1. Correlations between the three basic dimensions Valence (V), Energetic arousal (E), and Calmness (C) in differentstudies
Study 2: 135 American students, two times .47 .57# .20#
Schimmack & Reisenzein (2002, p. 415)c,a§
710 American and Canadian students.46 .65# .28#
Steyer et al. (1997, p. 14)b
503 German participants; 47% students, four times.50 to .62 .66 to .72 .43 to .53
Matthews et al. (1990, p. 25)c
388 British participants, mostly students.43 .37# .04#
Notes:aThe original dimensions were labeled as follows: “pleasure – displeasure” ≈ V, “awake –tiredness” ≈ E, “tension – relaxation” ≈ –C:bThe original dimensions were labeled as follows: “good – bad mood” ≈ V, “wakefulness –tiredness” ≈ E, “calmness–uncalmness” ≈ C:cThe original dimensions were labeled as follows: “valence” / “hedonic tone” ≈ V, “energetic arousal” ≈ E, “tense arousal” ≈ –C§Correlations were between latent factors and, therefore, adjusted for measurement error#Calmness was coded the other way around, such that high values indicated high tension. To ensure comparability with our coding system, thesigns of the original correlations were reversed.
P. Wilhelm & D. Schoebi: Assessing Mood in Daily Life 259
In addition to measures based on dimensional models ofmood and affect, various instruments have been developedto assess qualitatively distinctive mood states (e.g., the re-vised Multiple Affect Adjective Check List (MAACL-R),the PANAS-X or the Profile of Mood States (POMS); thelatter assesses, e.g., tension-anxiety, depression-dejection,anger-hostility, and others). The general problem withthese approaches is that neither the nature nor the numberof distinguishable mood states is clear. Moreover, the pro-posed specific mood-states are usually highly correlated(Schimmack, 1999; Watson & Vaidya, 2003).
Methods to Evaluate the PsychometricProperties of Scales Used in AmbulatoryAssessment Studies
Earlier approaches to demonstrate the factor structure ofrepeated measurement data followed Cattell’s suggestionto factorize the between-person correlations, which are re-peated time by time (R-technique) separately from thewithin-person correlations, which are repeated person byperson (P-technique; e.g., Zevon & Tellegen, 1982). Con-temporary approaches use structural equation models(SEM) or multilevel models (MLM) to estimate the facto-rial structure between and within persons simultaneously(see data analysis).
Specific reliability coefficients for ambulatory assess-ment measures have been calculated in various ways. Al-though the computational details differ, all of these meth-ods decompose the total variance into trait, state, and errorcomponents. To obtain indicators for the within-person re-liability, Buse and Pawlik (2001) correlated test halvesacross occasions for each participant (Cattell’s P-matrix)and averaged those coefficients across participants. To ob-tain indicators of the aggregate reliability, which is basedon the between-person variance, the odd-even method wasapplied (e.g., Buse & Pawlik, 1996; Perrez, Schoebi, &Wilhelm, 2000).
Cranford et al. (2006) decomposed the variance of theirmeasures into variance between persons, days, items of thesame scale, the two way interactions, and residuals. Usinggeneralizability theory they combined the variance compo-nents to demonstrate high aggregate reliability and satis-factory within-person reliability for their three-item moodscales. A similar but less formalized approach was pro-posed by Fahrenberg et al. (2002).
Other approaches to obtain specific reliability estimatesare based on structural equation modelling (SEM). One im-portant class of models in this framework are latent-statelatent-trait (LST) models. In LST theory (e.g., Steyer,Schmitt, & Eid, 1999) the total variance of a variable at agiven occasion is partitioned into three components: a la-tent-trait component, which does not change over occa-sions and indicates true consistency, a latent-state residual,which captures the occasion-specific deviation from the
trait and indicates true variability between occasions, andmeasurement error. In LST theory reliability is defined asthe ratio of true variance (latent-trait variance + latent-stateresidual variance) to total variance at a given occasion. An-other variance decomposition, which takes the serial de-pendency of repeated measures into account, was proposedby Kenny and Zautra (2001). In their model, the total vari-ance is divided into a stable trait, an autoregressive trait,and a state component, which contains situational influenc-es and error, and is supposed to vary randomly over time.
The Current Study
The purpose of this study was to evaluate the psychometricproperties of a short mood measure designed to assess threebasic mood dimensions in peoples’ daily lives. Data werecollected from a sample of 187 participants who reportedtheir mood state four times a day over the course of a weekby means of the current mood measure. Using a multilevelapproach, the three-factorial structure that was proposed byMatthews et al. (1990), Schimmack and Reisenzein (2002),and Steyer et al. (1997) was simultaneously tested betweenpersons and within persons. We further showed how errorvariance can be separated from latent variance at the dif-ferent levels to obtain level-specific reliability coefficientsand evaluate each scale’s sensitivity for measuring truechange over time.
Method
Participants
Ninety-eight Swiss couples were recruited to participate ina 1 week diary study either in undergraduate psychologyclasses or through private acquaintances of graduate stu-dents. Because of technical failures of the handheld com-puters, data of nine persons were lost. Thus, data of 93women and 94 men from 97 heterosexual couples could beanalyzed. Age of participants ranged between 19 and 36years (M = 25.6, SD = 3.2); half of them were students.
Electronic Diary Procedure
Four times a day over the course of a week, participants wereasked to rate their current mood and a series of other questionsnot relevant to this paper. The diary questions were imple-mented on Palm Tungsten T and T5, programmed with a pilotversion of IzyBuilder (http://www.izybuilder.com). Thequestions could be answered by using a stylus on a touch-screen. Around 11 a.m., 2:30 p.m., 6 p.m., and 9:30 p.m. thecomputer gave an acoustic signal. Signal time points wererandomized in a time window of ± 20 min around the intend-
260 P. Wilhelm & D. Schoebi: Assessing Mood in Daily Life
ed times to prevent participants from anticipating the exactbeginnings of the report.
Measures
Mood
To measure the basic mood-dimensions V, C, and E in peo-ple’s daily life, we developed a six-item short scale thatrelied on the Multidimensional Mood Questionnaire(MDMQ), a German-language mood scale (Steyer et al.,1997). The MDMQ provides consistent four-item scales tomeasure each dimension (Cronbach’s αs of the three scalesranged from .73 to .89 over four repeated measures). Dur-ing each observation participants responded to the state-ment “At this moment I feel:” by means of six bipolaritems, which were presented in the following order on onedisplay: tired–awake [müde–wach] (E+), content–discon-tent [zufrieden–unzufrieden] (V–), agitated–calm [unru-hig–ruhig] (C+), full of energy–without energy [energiege-laden–energielos] (E–)1, unwell–well [unwohl–wohl](V+), relaxed–tense [entspannt–angespannt] (C–). Thescales had seven steps. Their endpoints 0 and 6 were asso-ciated with the label “very.” Answers were given by mov-ing a slider from the start position 0, at the left end of ascale, to the position which corresponded best to the currentstate. To make sure that participants responded by movingthe slider rather than browsing through the allocation, atleast one of the two items belonging to a dimension had tobe moved to proceed to the next question. Prior to the anal-yses, data from three items were reverse coded, to ensurethat higher scores indicate higher positive V, higher E, orhigher C.
Data Analysis
We used multilevel analyses (e.g., Raudenbush & Bryk,2002; Goldstein, 2003) to investigate the variance and co-variance of the mood items. With MLMs, confirmatory fac-tor analyses (CFA) and regression models can be computedsimultaneously for the within- and the between-person partof the data. Compared with SEMs, they are better suited toanalyze hierarchically structured, unbalanced data setswith missing observations, such as are typically obtainedin ambulatory assessment. A shortcoming of MLMs is thatunlike SEM, they do not provide established fit indices.Recently Bauer (2003) and Curran (2003) have demon-strated that nested structures of unbalanced data can alsobe modeled with SEMs. However, the treatment of such
data is computationally easier with MLMs. We, therefore,used an MLM approach and the program MLwiN 2.02(Rasbash, Steele, Browne, & Prosser, 2005) to analyze thedata. MLwiN provides an iterative generalized least squarealgorithm to obtain parameter-estimates. At convergence,these estimates are maximum likelihood. The procedureyields a deviance-statistic (–2 log likelihood) that indicateshow well the specified model fits the data. If two modelsare nested, the difference of their deviances has a χ² distri-bution, with degrees of freedom equal to the difference inthe number of parameters estimated in the models. Thisstatistic can be used to test whether two models significant-ly differ in their fit.
Because of the large number of cases in our data set thepower was high to reject a more constrained model al-though its fit was not substantially worse. Therefore, the αlevel to evaluate the fit-difference of two models was setto p = .001.
Results
The raw data consisted of 4,577 observations provided by187 persons. Because of technical problems, the percentageof missing observations during the first 7 consecutive dayswas high (on average 20.4%, SD = 31.7). However, manyparticipants compensated for these technical failures by ex-tending the observation period, resulting in a satisfying av-erage number of 24.5 observations per participant (SD =5.9; range 6 to 44). Ten observations were excluded be-cause they contained contradictory extreme responses, andtherefore, a total of 4,567 observations were analyzed.
Item Variances and Covariances Betweenand Within Persons
In a first step, the item covariation at the within- and thebetween-person level was explored. A model with threelevels was set up, in which mood-items (Level 1) were nest-ed within observations (Level 2), which were nested withinpersons (Level 3).2 In the basic model, each of the six mooditems was identified by a dummy-coded indicator variablefor which a fixed effect and random effects at Level 2 andLevel 3 were estimated, according to the following equa-tion:
� This item is not part of the MDMQ. It was included because of positive characteristics in previous diary studies of our research group (Perrezet al., 2000, Wilhelm, 2004).
� To keep the models as simple as possible, we do not take into account that feeling states reported by romantic partners are positivelycorrelated. The consequence of not modeling the similarity between partners is that significance tests are too liberal at the between-personlevel. However, this bias is marginal when the number of couples is rather large – as in our study (see Kenny, Kashy, & Cook, 2006) and,therefore, does not compromise our conclusions.
Thus, the response yijk given on a particular item (subscripti) at a particular time (subscript j) by a particular person(subscript k) was modeled as a function of each item’s over-all mean βi, from which deviation was allowed. The esti-mate for ϖik captures the extent to which a person’s averageresponse k on item i deviates from the overall mean of thisitem (variation between persons). The estimate for the ran-dom effect at Level 2, uijk, captures the extent to whichresponses given at different times j deviate from each per-son’s average response k on a particular item i. Thus, thisestimate captures variation within persons, reflecting dif-ferences between observations over time. The random co-efficients of the six items were allowed to covary at eachlevel.
The fixed coefficients of Model 1 were 4.56 for content,4.53 for well, 4.41 for calm, 4.30 for relaxed, 3.42 forawake, and 3.51 for full of energy, indicating that on aver-age, participants were in a positive and relaxed state andabove a medium energy level. Results of the random partof Model 1 are shown in Table 2. As can be seen from thediagonals, the variances between observations (Level 2)are approximately 3 to 4 times larger than the variancesbetween persons (Level 3). This indicates that the biggestpart of the total variation in each item is the result of dif-ferences between observations and error. Below the diago-nals, the correlation coefficients between items are shown.At Level 2, the pattern of correlations indicates that theitems that belong to a common factor show the highest as-
sociations. However, the contrast between items that be-long to the same factor and items that belong to differentfactors was substantial only for the items full of energy andawake (which form the factor E). For the other items thisdifference was small. At Level 3, correlations were higherthan they were at Level 2, but the pattern was quite similar.
Factor Structure Between and WithinPersons
In the next step, a model was specified in which the vari-ances and covariances of the three postulated factors wereestimated at Level 2 and Level 3 (Model 2). Each factorwas identified by a dummy variable.3 At Level 2 and Level3, the variances and covariances of these factor variableswere estimated. In addition, each single item was allowedto vary, but the covariances between items and the covari-ances between factors and items were constrained to bezero. As before, fixed effects were estimated for each item.4
Compared with the saturated Model 1, Model 2 fit thedata significantly worse, χ²(18) = 179.4, p < .001. We,therefore, tested a modified model in which item residualswere allowed to be correlated. In order to keep this modelsimple, correlations between residual item variances wereonly allowed when their corresponding Wald-test was p <.01. The resulting Model 2r was not significantly differentfrom the saturated Model 1, χ²(9) = 14.0, p = .122.
Table 2. Random part of Model 1: Variances, covariances, and correlations of the mood items at the between and within-person level
� For example, the dummy variable of the factor C was coded 1 if a response corresponded to the items calm or relaxed and 0 if it correspondedto other items.
� Conceptually Model 2 is equal to the estimation of two CFAs in a SEM framework: one for the between-person data (the 187 individuals’means computed over time for each item) and another for the within-person data (4,567 cases in which the variables were centered aroundeach individual’s mean). In each CFA the variances and covariances of the three latent factors were estimated, as well as the residual varianceof each item. The loadings of the factors on the items were either constrained to 1 or 0. Each of the two CFA would then have 9 df.
The estimates of the random part of Model 2r are dis-played in Table 3. Between persons (Level 3) V and C werealmost perfectly correlated with each other and were bothhighly correlated with E. This indicates that persons whoreported a high average level of pleasure during the obser-vation week also reported a high average level of C and E.High correlations existed also between the three residualitem variances, probably because of stable response pat-terns in the use of items. Within persons (Level 2), V washighly correlated with C and moderately with E, but thecorrelation between C and E was close to zero. This patternof correlations indicates that changes over time were highlysynchronized between V and C, and slightly between V andE. Correlations between residual item variances were allpositive and of small to moderate size.5
In the next steps, we tested whether the three-factormodel above could be simplified to a more parsimonioustwo-factor model (at each level). We first forced the factorsof V and C at Level 2 to form a common factor. This model(Model 3a) fit the data significantly worse than Model 2r,χ²(3) = 197.2, p < .001, and was rejected. The same pro-cedure was then applied to the random part of Level 3. The
fit of this model (Model 3b) was not much worse than thefit of Model 2r, χ²(3) = 12.0, p = .007. The variance of thecommon V-C factor was 0.398 and its correlation with Ewas r = .65. We also tested whether a one-factor solutionwould be appropriate at Level 3, but the fit of this one-fac-tor model was clearly worse, χ²(2) = 104.3, p < .001.
In summary, the results show that the theoretically pos-tulated three correlated factors are necessary to describe thewithin-person variations of mood over time (Level 2).However, to describe rather stable differences in the week-ly averages of mood between persons (Level 3), two cor-related factors appear to be sufficient.
Sensitivity to Change
To evaluate sensitivity to change, one can directly comparethe relative size of the within-person variances of the moodfactors with the between-person variances (Table 3). In Table4, these variances are shown again in the first column. How-ever, we further decomposed the within-person variance be-cause temporal patterns within days differ from temporal pat-
Table 3. Random part of Model 2r: Estimated variances (diagonals) and correlations (below the diagonals) of the threemood factors and the residuals at the between-person Level 3 and the within-person Level 2
� If the residual item variances were not allowed to covary (Model 2) the Level 2 correlations were slightly higher (r V-C = .88, r V-E = .36, rC-E = .10) than in Model 2r.
terns between days. This was done by computing a four-levelmodel: Level 1 was constituted by the mood items, answeredat each observation (Level 2), which were repeated at differ-ent days of the week (Level 3) by different persons (Level 4).In analogy to Model 1, the full covariance matrix was esti-mated at Level 2, 3, and 4. The covariances between the itemsbelonging to the same factor were taken as estimates of thelatent factor variances. (Alternatively the factor Model 2rcould have been extended to four levels.)
As can be seen from Table 4, between two thirds andthree quarters of the total latent variance was the result ofvariation within persons. This clearly demonstrates thatthe measures of the three mood dimensions are highly sen-sitivity to capture change over time. A closer examinationof the within-person variance reveals that for V and C,half of the true variation was within days and about 15%was the result of changes between days. The proportionswere different for E, indicating that this dimensionchanged predominantly within days, whereas changes be-tween days were small. The distributions of the error vari-ances at the different levels6 show that measurement errorwas concentrated within days. As a consequence, sensi-tivity to change would be overestimated when measure-ment error was not controlled (see Table 4).
Level-Specific Reliability Estimates
To obtain level-specific reliability coefficients, the proportionof latent variation to total variation was computed (last col-umn of Table 4).7 At the between-person level, reliability was.92 for V, and .90 for C and E. Thus, the internal consistencyof the average (across observations) of each of the three mooddimensions was very satisfactory. As was suggested by Mod-el 3b, V and C can be merged to a common dimension atLevel 3. The reliability coefficient of this common score was.95. The estimated within-person reliability was .70 for V andC and .77 for E. This suggests that the internal consistency isalso sufficient when the score of a mood dimension is com-puted for a single observation from which the stable between-person variation (each person’s average over the week) hasalready been removed. The reliability for measuring the av-erage mood at a given day (based on four observations, fromwhich the person’s average has been removed) was very sat-isfactory for V and C (= .88) and good for E (.81). Even ascore obtained at a single observation, from which the day’sand the person’s average have been removed, still has anacceptable to satisfactory reliability (V and C = .66, E = .77).For a score based on a single observation, which has not beendecomposed, reliability was .76 to .80.
Table 4. Decomposition of the total variance into latent variance and error variance between persons and within persons
Latent variance Error variance Total variance Latent/total var.
– Between observations, within daysb 1.247 72 0.376 86 1.623 75 0.77
Total 1.726 100 0.438 100 2.164 100 0.80
Notes: aVariance components were obtained from the Model 2r; bA four-level model was estimated to obtain the variance components betweendays, within-persons and between observations, within days. In this model the between-persons variance was estimated to be slightly smallerthan in Model 2r.
264 P. Wilhelm & D. Schoebi: Assessing Mood in Daily Life
� The error variance is the mean of the residual item variances divided by the number of items. To obtain, for example, the within-personerror of V the mean of the residual variation of the items content and well from Table 3 (Level 2) is computed (0.725 + 0.512)/2 = 0.618and divided by 2 = 0.309.
� The proportion of latent to total variance leads to results equivalent to the computation of Cronbach’s α for each level.
Discussion
In this article, we evaluated the psychometric properties of ashort mood scale to assess fluctuation of mood states in dailylife. We did this by investigating the within- and between-per-son variance and covariance of six bipolar items that werechosen to measure three basic dimensions of mood, namely,valence (V), calmness (C), and energetic arousal (E).
Examination of the factor structure revealed that there wasevidence for the three-dimensional model proposed by Mat-thews et al. (1990), Steyer et al. (1997), and Schimmack andcolleagues (Schimmack & Grob, 2000; Schimmack & Rei-senzein, 2002) at the within-person level. The correlationsbetween the dimensions indicated that fluctuations of V andC over the course of a week were highly synchronized,whereas E was moderately associated with V, but not remark-ably with C. The comparison of our results with the correla-tions reported in other studies (see Table 1 and Table 3) sug-gests differences in the size of single coefficients rather thana different pattern of associations. The current study showsthat the three-dimensional model holds when correlationswere computed within persons across time. At the between-person level, a different pattern of correlations was found. Cand V converged into a common well-being factor, whichwas highly correlated with E. Thus, the three-factor modelcould not be confirmed for persons’ average scores.
High correlations between affect measures, which wereaggregated over many observations, have repeatedly beenfound in diary studies (e.g., Schimmack, 2003; Wilhelm,2004; Zelenski & Larsen, 2000). Watson and Vaidya (2003)attributed such correlations mainly to systematic responsebiases, in particular to the tendency to respond similarly todifferent items. They concluded that “general ratings ulti-mately appear to have superior construct validity, andtherefore should continue to be viewed as the gold standardin trait affect assessment” (p. 371). Although we agree thathigh correlations in the aggregated affect measures are, inpart, the result of stable response styles, we do not see ev-idence for Watson and Vaidya’s conclusion. First, responsestyles operate in conventional trait-affect scales, too. Sec-ond, and more important, there is striking evidence thatreports of feelings experienced in general or during a longertime period (e.g., last year), which are assessed in trait-af-fect questionnaires, are prone to many sources of distortion,like retrospective recall biases, and mood congruent- andautobiographic memory effects (e.g., Gorin & Stone, 2001;Fahrenberg, Myrtek, Pawlik, & Perrez, 2007; Robinson &Clore, 2002). Hence, these results question the validity oftrait-affect questionnaires to measure the actually experi-enced general state during a certain time period. Trait-af-fect scales may validly assess the participants’ current con-cept about their general affective state, yet this is a differenttheoretical concept (Perrez, 2006).
Mood averages are stable over time (Buse & Pawlik,1996) and are substantially correlated with affect-relatedtraits, like e.g., neuroticism (Fahrenberg et al., 2001;
Klumb, 2004; Schimmack, 2003). We, therefore, assumethat besides response styles, mood averages contain a largeportion of valid information about the affective dispositionof a person. The high correlations that we obtained betweenthe three mood dimensions can, thus, be meaningfully in-terpreted. They indicate that people who are in a bad moodoverall generally also feel more tension and less energy.Such a pattern is quite plausible and would be typical forlonger lasting dysphoric or depressed episodes.
In sum, the results of the structural analysis show that dif-ferent constructs are measured at the between- and within-person level: On the between-person level, V and C are equalindicators of well-being, which is highly associated with thesubjectively experienced general level of energy. Within per-sons, the scores of V, C, and E capture three differently syn-chronized mood states. The state quality of each dimensionwas clearly demonstrated, given that at least about two thirdsof the total latent variation was the result of fluctuations overtime. Thus, the three measures are highly sensitive to capturechanges in mood states, which is an essential requirement forthe assessment of mood.
Our approach allowed us to decompose the variance at eachlevel of the data into latent variance and error variance. Wecould, therefore, compute level-specific reliability coeffi-cients. They show that weekly aggregates provide highly reli-able measures of well-being and general E. Also, daily aver-ages were reliable. Moreover, even the obtained reliability ofscores computed at single observations was still in an accept-able range. These coefficients are especially satisfying whenwe take into account that they are based on two items only.
We used bipolar items to assess mood states. This opera-tionalization corresponds to our theoretical conceptualizationof mood as an ongoing affective coloring of the current ex-perience, which can be described on basic bipolar dimen-sions. In doing so, we are in accordance with Russell andCarroll (1999a, 1999b), who concluded that a bipolar modelof the valence dimension of affect fits well to the data report-ed in the literature, when response formats of the items andmeasurement error are adequately taken into account. Corre-spondingly, Steyer and Riedl (2004) confirmed the bipolarityassumption for the dimensions of the three-factorial moodmodel. Using an LST approach for ordinal data, they showedthat the latent-state residuals of the corresponding unipolaritems of the MDMQ – which we combined to obtain ourbipolar measures – were almost perfectly negatively correlat-ed. However, whether affect dimensions are uni- or bipolarand whether feelings of pleasure and displeasure are mutuallyexclusive or can both be experienced at the same time hasbeen intensively debated and is not yet resolved (e.g., Russell& Carroll, 1999a, 1999b; Watson & Tellegen, 1999; seeSchimmack, 2005, for arguments in favor of a two-dimen-sional conceptualization of pleasure and displeasure that af-fords unipolar response formats).
Some limitations of this study deserve attention. First, al-though we found that the three-dimensional model fit thewithin-person data much better than a more parsimonioustwo-dimensional model, the latent factors of V and C were
P. Wilhelm & D. Schoebi: Assessing Mood in Daily Life 265
highly correlated. It would, therefore, be worthwhile to inves-tigate whether there are other items that are appropriate indi-cators of V and C, respectively, but discriminate these twodimensions better than the ones used here. Second, althoughwe could demonstrate high sensitivity to change, which is abasic requirement for a mood scale, it is necessary to considerthat sensitivity to change only indicates that there is changeover time, but it does not indicate that such change is in ac-cordance with theoretical assumptions. Therefore, differentfacets of criterion validity need to be demonstrated. For ex-ample, E should increase until midday and decline in theevening, tense arousal should be higher during a conflict sit-uation, and V should be more positive when leisure activitiesare performed together with friends. Finally, the results arebased on a 1 week assessment of a nonrepresentative sampleof young Swiss couples. Because partners’ moods are usuallycorrelated (Cranford et al., 2006; Wilhelm, 2004, Wilhelm &Perrez, 2004), one might argue that the between-person vari-ance is probably underestimated compared to a sample ofsingle individuals. Analyses computed separately for womenand men revealed that compared to the whole sample (seeTable 4) the latent between-person variance was slightly larg-er for women but smaller for men. This suggests that sensi-tivity to change might only be slightly overestimated – if atall. We, therefore, believe that the results can be generalizedto young German-speaking adults. However, it remains to beinvestigated whether results would differ in other populations(adolescents, older people, patients), when other time sched-ules are used, or when participants experience demanding orstressful circumstances (e.g., see Cranford et al., 2006).
Nevertheless, the current article illustrates that a sensitiveand reliable measurement of the basic dimensions of dailymood is possible with a short set of only six bipolar items.
Acknowledgments
The preparation of parts of this article was supported by afellowship grant of the Alfried Krupp Wissenschaftskolleg,Greifswald, Germany, to Peter Wilhelm. Dominik Schoe-bi’s work on this article was supported by fellowshipPA001-108998 from the Swiss National Science Founda-tion. We are grateful to Andrea Conrad, Lukas Erpen, Car-men Faustinelli, Miriam Künzli, Annette Meier, JacquelineNagel, Adelaide Notter, and Melanie Sarbach for their en-gagement in recruiting, instructing, and coaching the par-ticipants of our study. We thank Ulrich Ebner-Priemer andThomas Kubiak for their suggestions on this article, andSiegfried Macho for his comments on the analytical strat-egy and the conceptual interpretation of our results.
References
Bauer, D.J. (2003). Estimating multilevel linear models as struc-tural equation models. Journal of Education and BehavioralStatistics, 28, 135–167.
Bolger, N., Davis, A., & Rafaeli, E. (2003). Diary methods: Cap-turing life as it is lived. Annual Review of Psychology, 54,579–616.
Buse, L., & Pawlik, K. (1996). Ambulatory behavioral assessmentand in-field performance testing. In J. Fahrenberg & M. Myr-tek (Eds.), Ambulatory assessment (pp. 29–50). Kirkland, WA:Hogrefe & Huber.
Buse, L., & Pawlik, K. (2001). Computer-assisted ambulatoryperformance-tests in everyday situations: Construction, eval-uation, and psychometric properties of a test battery measuringmental activation. In J. Fahrenberg & M. Myrtek (Eds.), Pro-gress in ambulatory assessment (pp. 3–23). Kirkland, WA: Ho-grefe & Huber.
Cranford, J.A., Shrout, P.E., Iida, M., Rafaeli, E., Yip, T., & Bol-ger, N. (2006). A procedure for evaluating sensitivity to with-in-person change: Can mood measures in diary studies detectchange reliably? Personality and Social Psychology Bulletin,32, 917–929.
Curran, P.J. (2003). Have multilevel models been structural equa-tion models all along? Multivariate Behavioral Research, 38,529–569.
Davidson, R.J. (1994). On emotion, mood, and related affectiveconstructs. In P. Ekman & R.J. Davidson (Eds.), The nature ofemotion. Fundamental questions (pp. 51–55). New York: Ox-ford University Press.
Ebner-Priemer, U.W., & Sawitzki, G. (2007). Ambulatory assess-ment of affective instability in borderline personality disorder:The effect of the sampling frequency. European Journal ofPsychological Assessment, 23, 238–247.
Ekman, P., & Davidson, R.J. (Eds.). (1994). The nature of emo-tion: Fundamental questions. New York: Oxford UniversityPress.
Fahrenberg, J. (2006). Self-reported subjective state – Singleitems or scales like AD-ACL and PANAS? European Networkof Ambulatory Assessment. Statements and Open Commen-taries. Retrieved June 21, 2007, from http://www.ambulatory-assessment.org.
Fahrenberg, J., Hüttner, P., & Leonhart, R. (2001). MONITOR:Acquisition of psychological data by a hand-held PC. In J. Fah-renberg & M. Myrtek (Eds.), Progress in ambulatory assess-ment (pp. 93–112). Kirkland, WA: Hogrefe & Huber.
Fahrenberg, J., Leonhart, R., & Foerster, F. (2002). AlltagsnahePsychologie: Datenerhebung im Feld mit hand-held PC undphysiologischem Mess-System [Psychological assessment ineveryday life: Data gathering in the field by handheld personalcomputers and physiological monitoring systems]. Bern:Huber.
Fahrenberg, J., Myrtek, M., Pawlik, K., & Perrez, M. (2007). Am-bulatory assessment – Monitoring behavior in daily life set-tings: A behavioral-scientific challenge for psychology. Euro-pean Journal of Psychological Assessment, 23, 206–213.
Goldstein, H. (2003). Multilevel statistical models (3rd ed.). Lon-don, UK: Edward Arnold.
Gorin, A.A., & Stone, A.A. (2001). Recall biases and cognitiveerrors in retrospective self-reports: A call for momentary as-sessments. In A. Baum, T.A., Revenson, & J.E. Singer (Eds.),Handbook of health psychology (pp. 405–413). Mahwah, NJ:Erlbaum.
dinal data. In L.M. Collins & A.G. Sayer (Eds.), New methodsfor the analysis of change (pp. 243–263). Washington, DC:American Psychological Association.
Klumb, P.L. (2004). Benefits form productive and consumptiveactivities: Results from the Berlin Aging Study. Social Indica-tors Research, 67, 107–127.
Kubiak, T., & Jonas, C. (2007). Applying circular statistics to theanalysis of monitoring data: Patterns of social interactions andmood. European Journal of Psychological Assessment, 23,227–237.
Lewis, M., & Haviland-Jones, J.M. (Eds.). (2000). Handbook ofemotions (2nd ed.). New York: Guilford.
Lucas, R.E., & Baird, B.M. (2006). Global self-assessment. In M.Eid & E. Diener (Eds.), Handbook of multimethod measure-ment in psychology (pp. 29–42). Washington, DC: AmericanPsychological Association.
Matthews, G., Jones, D, M., & Chamberlain, A.G. (1990). Refin-ing the measurement of mood: The UWIST Mood AdjectiveChecklist. British Journal of Psychology, 81, 17–42.
Myrtek, M. (2004). Heart and emotion. Ambulatory monitoringstudies in everyday life. Göttingen, Germany: Hogrefe & Huber.
Perrez, M. (2006). Plädoyer für theorieadäquate Methoden in ge-wissen Domänen der Psychologie [Plea for a theory-adequatemethodology in certain domains of psychology]. Verhaltens-therapie und Psychosoziale Praxis, 38, 319–330.
Perrez, M., Schoebi, D., & Wilhelm, P. (2000). How to assesssocial regulation of stress and emotions in daily family life? Acomputer-assisted family self-monitoring-system (FASEM-C). Clinical Psychology and Psychotherapy, 7, 326–339.
Rasbash, J., Steele, F., Browne, W., & Prosser, B. (2005). A user’sguide to MLwiN version 2.0. Bristol, UK: Centre for MultilevelModelling, University of Bristol.
Raudenbush, S.W., & Bryk, A.S. (2002). Hierarchical linearmodels: Applications and data analysis methods (2nd ed.).Thousand Oaks, CA: Sage.
Reicherts, M., Salamin, V., Maggiori C., & Pauls, K. (2007). Thelearning affect monitor (LAM): A computer-based system in-tegrating dimensional and discrete assessment of affectivestates in daily life. European Journal of Psychological Assess-ment, 23, 268–277.
Robinson, M.D., & Clore, G.L. (2002). Belief and feeling: Evi-dence for an accessibility model of emotional self-report. Psy-chological Bulletin, 128, 934–960.
Russell, J.A. (2003). Core affect and the psychological construc-tion of emotion. Psychological Review, 110, 145–172.
Russell, J.A., & Carroll, J.M. (1999a). On the bipolarity of posi-tive and negative affect. Psychological Bulletin, 125, 3–30.
Russell, J.A., & Carroll, J.M. (1999b). The phoenix of bipolarity:Reply to Watson and Tellegen (1999). Psychological Bulletin,125, 611–617.
Russell, J.A., Weiss, A., & Mendelsohn, G.A. (1989). Affect grid:A single-item scale of pleasure and arousal. Journal of Per-sonality and Social Psychology, 57, 493–502.
Scherer, K.R. (2005). What are emotions? And how can they bemeasured? Social Science Information, 44, 695–729.
Schimmack, U. (1999). Strukturmodelle der Stimmungen: Rück-schau, Rundschau und Ausschau [Structural models of mood:Review, overview, and outlook]. Psychologische Rundschau,50, 90–97.
Schimmack, U. (2003). Affect measurement in experience sam-pling research. Journal of Happiness Studies, 4, 79–106.
Schimmack, U. (2005). Response latencies of pleasure and dis-pleasure ratings: Further evidence for mixed feelings. Cogni-tion and Emotion, 19, 671–691.
Schimmack, U., & Grob, A. (2000). Dimensional models of coreaffect: A quantitative comparison by means of structural equa-tion modeling. European Journal of Personality, 14, 325–345.
Schimmack, U., & Reisenzein, R. (2002). Experiencing activa-tion: Energetic arousal and tense arousal are not mixtures ofvalence and activation. Emotion, 2, 412–417.
Steyer, R., & Riedl, K. (2004). Is it possible to feel good and badat the same time? New evidence on the bipolarity of mood-state dimensions. K. van Montfort, J. Oud, & A. Satorra (Eds.),Recent developments on structural equation models: Theoryand applications (pp. 197–220). Dordrecht, The Netherlands:Kluwer.
Steyer, R., Schmitt, M., & Eid, M. (1999). Latent state-trait theoryand research in personality and individual differences. Euro-pean Journal of Personality, 13, 389–408.
Steyer, R., Schwenkmezger, P., Notz, P., & Eid, M. (1997). DerMehrdimensionale Befindlichkeitsfragebogen. Handanwei-sung [The Multidimensional Mood Questionnaire (MDMQ)].Göttingen, Germany: Hogrefe.
Thayer, R.E. (1989). The biopsychology of mood and activation.New York: Oxford University Press.
Watson, D., & Tellegen, A. (1985). Toward a consensual structureof mood. Psychological Bulletin, 98, 219–235.
Watson, D., & Tellegen, A. (1999). Issues in the dimensionalstructure of affect – Effects of descriptors, measurement error,and response formats: Comment on Russell and Carroll(1999). Psychological Bulletin, 125, 601–610.
Watson, D., & Vaidya, J. (2003). Mood measurement: Currentstatus and future directions. In I.B. Weiner (Series Ed.) & J.A.Schinka & W.F. Velicer (Volume Eds.), Handbook of psychol-ogy, Vol. 2: Research methods in psychology (pp. 351–375).Hoboken, NJ: Wiley.
Wilhelm, P. (2004). Empathie im Alltag von Paaren. Akkuratheitund Projektion bei der Einschätzung des Befindens des Part-ners [Empathy in couples’ everyday lives: Accuracy and pro-jection in judging the partner’s feelings]. Bern: Huber.
Wilhelm, P., & Perrez, M. (2004). How is my partner feeling indifferent daily-life settings? Accuracy of spouses’ judgmentsabout their partner’s feelings at work and at home. Social In-dicators Research, 67, 183–246.
Zelenski, J.M., & Larsen, R.J. (2000). The distribution of basicemotions in everyday life: A state and trait perspective fromexperience sampling data. Journal of Research in Personality,34, 178–197.
Zevon, M.A., & Tellegen, A. (1982). The structure of moodchange: An ideographic/nomothetic analysis. Journal of Per-sonality and Social Psychology, 43, 111–122.
Peter Wilhelm
Department of PsychologyUniversity of FribourgRue de Faucigny 2CH-1700 FribourgSwitzerlandE-mail [email protected]
P. Wilhelm & D. Schoebi: Assessing Mood in Daily Life 267
