Content uploaded by Alastair J Wilson

Author content

All content in this area was uploaded by Alastair J Wilson on Apr 14, 2014

Content may be subject to copyright.

UNCORRECTEDPROOF

1

2

3ORIGINAL PAPER

4How stable are personalities? A multivariate view of behavioural

5variation over long and short timescales in the sheepshead

6swordtail, Xiphophorus birchmanni

8Kay Boulton &Andrew J. Grimmer &Gil G. Rosenthal &

9Craig A. Walling &Alastair J. Wilson

10

11

12 Received: 18 September 2013 /Revised: 4 February 2014 /Accepted: 4 February 2014

13 #Springer-Verlag Berlin Heidelberg 2014

14 Abstract Many studies have revealed repeatable (among-

15 individual) variance in behavioural traits consistent with var-

16 iation in animal personality; however, these studies are often

17 conducted using data collected over single sampling periods,

18 most commonly with short time intervals between observa-

19 tions. Consequently, it is not clear whether population-level

20 patterns of behavioural variation are stable across longer time-

21 scales and/or multiple sampling periods or whether individ-

22 uals maintain consistent ranking of behaviours (and/or per-

23 sonality) over their lifetimes. Here, we address these questions

24 in a captive-bred population of a tropical freshwater poeciliid

25 fish, Xiphophorus birchmanni. Using a multivariate approach,

26 we estimate the among-individual variance-covariance matrix

27 (I), for a set of behavioural traits repeatedly assayed in two

28 different experimental contexts (open-field trials, emergence

29and exploration trials) over long-term (56 days between ob-

30servations) and short-term (4-day observation interval) time

31periods. In both long- and short-term data sets, we find that

32traits are repeatable and the correlation structure of I is con-

33sistent with a latent axis of variation in boldness. While there

34are some qualitative differences in the way individual traits

35contribute to boldness and a tendency towards higher repeat-

36abilities in the short-term study, overall, we find that

37population-level patterns of among-individual behavioural

38(co)variance to be broadly similar over both time frames. At

39the individual level, we find evidence that short-term studies

40can be informative for an individual’s behavioural phenotype

41over longer (e.g. lifetime) periods. However, statistical sup-

42port is somewhat mixed and, at least for some observed

43behaviours, relative rankings of individual performance

44change significantly between data sets.

45Keywords Multivariate behaviour .Personality .

46Repeatability .Stability .Boldness .Imatrix

47Introduction

48It is now apparent that, within animal populations, individuals

49often exhibit differences in behaviour that are repeatable

50across time and context. This repeatable variation is taken as

51evidence for animal temperament (e.g. Boissy 1995; Réale

52et al. 2007), behavioural syndromes (Sih et al. 2004), coping

53styles (Koolhaas et al. 1999) or personality, the latter term

54reflecting parallels with research in human psychology

55(Budaev 1997b; Gosling 2001). A number of axes of

56among-individual behavioural variation condensed into “per-

57sonality traits”have been described, including boldness-

58shyness, exploration-avoidance and general activity (Réale

59et al. 2007). Understanding the evolution of personality has

60become a major field of study in behavioural ecology (Dall

Communicated by N. Dingemanse

Craig A Walling and Alastair J Wilson made equal contributions.

Electronic supplementary material The online version of this article

(doi:10.1007/s00265-014-1692-0) contains supplementary material,

which is available to authorized users.

K. Boulton :C. A. Walling

Institute of Evolutionary Biology, University of Edinburgh, West

Mains Rd, Edinburgh EH9 3JT, UK

A. J. Grimmer :A. J. Wilson (*)

Centre for Ecology and Conservation, Biosciences, College of Life

and Environmental Sciences, University of Exeter, Cornwall

Campus, Treliever Road, Penryn, Cornwall TR10 9EZ, UK

e-mail: A.Wilson@exeter.ac.uk

G. G. Rosenthal

Department of Biology, Texas A&M University, 3258 TAMU,

College Station, TX 77843, USA

G. G. Rosenthal

Centro de Investigaciones Científicas de las Huastecas “Aguazarca”,

Calnali, Hidalgo, Mexico

Behav Ecol Sociobiol

DOI 10.1007/s00265-014-1692-0

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

61 et al. 2004; Stamps and Groothuis 2010). There is now grow-

62 ing evidence that traits relating to personality contribute to

63 fitness variation and therefore may be both adaptive and

64 generally under selection (Q1 Smith and Blumstein 2008). How-

65 ever, if natural selection occurs through variation in lifetime

66 fitness, then an important question arises: just how stable are

67 personalities over individual lifetimes? Here, we address this

68 question in a captive population of fish. We do this using a

69 novel multivariate approach that characterises personality var-

70 iation as a latent character underpinning among-individual

71 (co)variation in a suite of observed behaviours.

72 While there remains considerable disagreement over how

73 best to define individual personality traits (Réale et al. 2007;

74 Toms et al. 2010;Carteretal.2013; see below), there is a

75 broad consensus that among-individual behavioural variance

76 is the statistical signature of animal personality. Typically, this

77 is quantified as the (among-individual) repeatability, defined

78 as the proportionof observed variance explained by individual

79 identity, of one or more observed behavioural traits. Thus,

80 partitioning of observed variance into among- and within-

81 individual components (the latter arising from individual plas-

82 ticity and/or measurement errors) from repeat observations on

83 individuals is crucial to empirical studies of personality

84 (Dingemanse et al. 2012b;Brommer2013;Araya-Ajoyand

85 Dingemanse 2014). In a meta-analysis, Bell et al. (2009)

86 concluded that on average, estimates of repeatability for ob-

87 served behavioural traits decreased as the interval between

88 sampling events increased. Consequently, it may be danger-

89 ous to assume that short-term studies reflect behavioural (and

90 by implication, personality) differences that are stable over the

91 lifetime of individuals. This is potentially important since

92 short-term repeatability estimates predominate in the litera-

93 ture, although the number of studies conducted over

94 timeframes that may be considered more representative of

95 natural life spans is growing (for more recent examples, see

96 Ronning et al. 2005; Bushuev et al. 2010; Chervet et al. 2011;

97 David et al. 2012; Kanda et al. 2012). However, few studies

98 have collected repeated observations over two distinct time

99 periods from the same individual (but see for e.g. Carere et al.

100 2005) that would allow the “repeatability of repeatability”to

101 be assessed. Here, we do this, but also extend our analysis to

102 the multivariate case to ask whether patterns of among-

103 individual behavioural (co)variation reflect an underlying per-

104 sonality trait that is stable across distinct long- and short-term

105 sampling periods.

106 In what follows, we investigate the temporal stability of

107 multiple behavioural traits in the freshwater poeciliid fish,

108 Xiphophorus birchmanni to answer two complementary ques-

109 tions. Firstly, at the level of the population, how stable are the

110 patterns of among-individual trait (co)variance generated by

111 underlying personality? Secondly, at the level of the individ-

112 ual, do short term studies reveal behavioural tendencies that

113 are stable across lifetimes? To answer these questions, we

114characterise behavioural variation along what we loosely con-

115sider to be an axis of shyness-boldness. Boldness is the most

116commonly studied axis of personality in fish (Toms et al.

1172010) and positively correlates with fitness-related traits in-

118cluding reproductive success, parental provisioning, growth,

119aggression, social dominance, dispersal and proactive re-

120sponses to stressors such as predation risk (Dingemanse

121et al. 2004; Brown et al. 2005; Bell and Sih 2007;Coteetal.

1222010; Rudin and Briffa 2011; Ariyomo and Watt 2012;

123Mutzel et al. 2013). There remains, however, a lack of con-

124sensus on how best to define boldness and how it should be

125assayed (Toms et al. 2010). This raises obvious potential for

126misclassification of personality traits (Carter et al. 2013)and/

127or disagreement over appropriate experimental design (Toms

128et al. 2010).

129The present goal is to investigate stability of a personality

130trait without adding further to existing debate over issues of

131definition. Consequently we do not attempt to define boldness

132or the best way to measure it a priori; rather, we follow the

133view of others that personality traits should be considered as

134latent variables that can best be uncovered by observing

135several measurable, correlated and potentially overlapping

136behaviours across contexts (Dochtermann and Jenkins 2007;

137Dingemanse et al. 2010; Dochtermann and Roff 2010). We

138therefore make a distinction throughout between “behavioural

139traits”that are observed directly, and “personality (traits)”,

140inferred from among-individual (co)variance in observed be-

141haviour(s). This exploratory approach, which follows

142Huntingford (1976) and others (Budaev 1997b;Moretz

1432003), is becoming more mainstream and allows the avoid-

144ance of difficulties that can arise if a single behaviour is

145chosen a priori to assay boldness. For example, a fish that

146swims a long distance in one behavioural trial may be classi-

147fied as willing to explore and therefore as “bold”;however,

148this behaviour could also plausibly be indicative of anxiety,

149with the animal’s“exploration”being driven by a search for

150refuge.

151Currently, the most common experimental paradigm used

152to measure boldness is that of the open-field trial (OFT), in

153which an animal is placed in an open arena and its behaviour is

154monitored for a predetermined observation period. Initially

155developed for rodent studies (Hall 1934; Walsh and Cummins

1561976), OFTs have long been applied to fish models (Warren

157and Callaghan 1975;Budaev1997b). Considered the most

158reliable way to assay boldness by some authors (Burns 2008),

159others have argued that OFTs risk conflating boldness with

160other axes of variation that are distinct (if sometimes correlat-

161ed) personality traits in their own right (e.g. exploration-

162avoidance, overall activity, Réale et al. 2007). If so, then

163simple modifications to OFTs such as providing a refuge from

164which an animal can choose to emerge and explore (emer-

165gence and exploration trial, EET) may be useful (Dingemanse

166et al. 2007).

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

167 In what follows, we use both types of behavioural trial

168 mentioned above (OFT and EET) to observe how fish behave

169 in these contexts and to characterise the repeatable component

170 of multivariate behaviour. We then assess the extent to which

171 one or more major axes of variance adequately depict ob-

172 served variation. In other words, we aim to describe the

173 behavioural trait variation first, and then consider the extent

174 to which its repeatable component fits within the paradigm of

175 a major axis of personality, i.e. the boldness-shyness axis

176 (Dingemanse et al. 2010; Dochtermann and Roff 2010). We

177 then go on to address three specific questions regarding the

178 temporal stability of personality. Firstly, we ask whether re-

179 peatabilities estimated from repeated measures of individual

180 behaviours over a short time period give a misleading view of

181 the importance of among-individual variance over longer time

182 periods. Secondly, by extending our analysis to the multivar-

183 iate case, we ask whether the structure of the between-trait

184 among-individual covariance matrix denoted I, following Wil-

185 son et al. (2013), is similar when estimated from short- and

186 long-term data; i.e. do repeated empirical analyses of a single

187 population actually reveal the same major axes of among-

188 individual variation? If so, then a final question concerns the

189 extent that individuals retain the same relative ranking for

190 repeatable behaviours, and hence personality, over their

191 lifetimes.

192 Methods

193 Study species and husbandry

194 One hundred wild adult X. birchmanni were caught in the

195 Arroyo Coacuilco near the town of Coacuilco, municipality

196 of San Felipe Orizatlán, Hidalgo, Mexico (elevation 314 m

197 lat/long 21.099-98.587), and imported to the UK in Febru-

198 ary 2010. Between August 2010 and May 2011, we collect-

199 ed an offspring generation (n=384) from 13 males and 27

200 females (mean (SE) brood size of 8.86 (0.541)). Gravid

201 females were isolated and, following birth, broods were

202 immediately netted and moved to one half of a partitioned

203 30-L tank; broods of more than six offspring were split with

204 each half of the family placed in different tanks. Fry were

205 fed twice daily on a mix comprising equal quantities of

206 crushed ZM spirulina and brine shrimp flake and

207 laboratory-prepared brine shrimp nauplii. At an average of

208 17 weeks (range 12 to 27), juveniles were tagged with a

209 single elastomer injection for individual identification pur-

210 poses and transferred to mixed family rearing groups of n=

211 8. Note it is not possible to determine sex at this age in this

212 species, and therefore the sex ratio was not controlled.

213 Eight rearing groups were then kept within each of six

214 sequentially set up stacks of tanks, each stack sharing a

215 common water supply and recirculating filtration system.

216As part of a parallel study of density effects on growth,

217rearing groups were initially housed under two different

218density regimes as follows. Within each stack, four groups

219were placed in 30-L (37×37 ×22 cm) glass tanks (low-

220density treatment) with the remaining four groups in 15-L

221half tanks (high-density treatment). Half tanks were created

222by placing a black-net-covered Perspex-framed partition

223down the centre of a full-size tank. Thus, establishing a

224stack required 64 fish (i.e. 8×8) to be available for tagging

225simultaneously and this accounts for the variation in tag-

226ging age within stacks. Fish were fed twice daily with a

227standardised ration of flake food as above (morning) and a

228mixofpreviouslyfrozenbloodwormanddaphnia

229(afternoon). On the days when behavioural data was to be

230collected, the morning feed was omitted in an attempt to

231encourage exploration tendencies. Temperature was main-

232tainedbetween22and24°Canda12:12-hlight-to-dark

233cycle imposed. After being housed in this manner for

23428 weeks, density was swapped for half of the tanks, thus

235creating four treatment effects with the total number of fish

236divided approximately equally between them as follows:

237low/low (n=93), low/high (n= 95), high/high (n=87) and

238high/low (n=93). Observations from individuals failing to

239reach sexual maturity by the end of the long-term study

240(50 weeks) were excluded from the analysis and the above

241breakdown (n=11).

242Behavioural data collection

243The trials were performed over two experimental study pe-

244riods, denoted long term (LT) and short term (ST). All avail-

245able fish contribute to the long-term data set (n=373) while a

246random subset of 32 fish from each of the four density treat-

247ments (low/low n=13, low/high n=4, high/high n=9,

248high/low n=6) was used for the short-term study (Table 1).

249Trials were of two types, open field (OFT) and emergence and

250exploration (EET) with multiple specific behavioural traits

251assayed in each trial type (Table 2). Overall, the long-term

252trials took 13 months to complete (May 2011 to May 2012),

253with data collected over an actual 30-week period for each

254fish. Each individual was subject to an OFT followed by an

255EET seven days later, a process that was repeated three times

256at 56-day intervals, thus yielding four OFTs and four EETs per

257fish. The short-term data set was collected in February 2013,

258with 32 individual fish subjected to alternating OFT and EET

259at 48-h intervals (i.e. 2 days between trials, 4 days between

260repeated trials of the same type) with each animal undergoing

261five trials of each type. For those 32 individuals used in both

262study periods, data was therefore collected over a timeframe

263with a mean (SE) of 531.4 (6.38) days. By comparison, the

264mean (SE) longevity of individuals with known birth and

265death dates under our laboratory conditions is 450.3 (8.10).

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

266 Experimental procedures

267 Open-field trial

268 An empty 45×25× 25-cm tank was filled to a depth of 8 cm

269 with room temperature water (22 °C). The tank was lit from

270 below and visually screened by a cardboard casing to prevent

271 external laboratory disturbance. Fish were caught individually

272 from their home tank with a dip net, quickly examined for

273 identification tags and immediately placed into the centre of

274 the OFT tank. Following a 30-s acclimation period, behaviour

275 was filmed for 300 s using a Sunkwang C160 video camera

276 fitted with a 5–50-mm manual focus lens suspended above the

277 tank. Data were then extracted from the video using the

278 tracking software Viewer II (http://www.biobserve.com/

279 products/viewer/index.html), which was set up to divide the

280 tank basal area into two approximately equal halves (middle

281 and perimeter zones) (Fig. 1a). Water was changed between

282 individual trials to prevent chemical cues affecting behaviour.

283 Emergence and exploration trial

284 A 45× 25×25-cm tank was physically divided into three sec-

285 tions with opaque Perspex, providing a right-hand, centre and

286 left-hand chamber. A small (5 cm) opening was cut in each

287 divider, starting 2 cm from the tank edge. The openings were

288 positioned at opposite sides of the tank. The chamber on the

289 right-hand side was designated as the refuge and equipped

290with a plastic plant and several small stones. A rising trapdoor

291was rigged to a pulley above the tank, positioned inside the

292refuge and covering the exit into zone 1 (Fig. 1b). Tanks were

293filled, emptied, lit and screened as above. Fish were individ-

294ually caught and examined as before and placed directly into

295the centre of the refuge where they were allowed 30 s to

296acclimate before the trapdoor was lifted. Filming then com-

297menced for 300 s (as above), but only behaviour outside the

298refuge (i.e. in zones 1 and 2) was tracked and extracted for

299analysis.

300Behavioural traits

301The behavioural traits recorded in this study were selected as

302those likely to reflect variation along a bold-shy type person-

303ality axis. For the OFT, we predicted that fish tending toward

304boldness would actively explore the novel environment of the

305OFT by leaving the tank sides and spending more time in the

306central zone than shy fish. OFT behaviour was therefore

307quantified by four traits: track length (TL), activity (Act), area

308covered (AC) and time in middle of the tank (TIM), which we

309predicted would be positively correlated with one another. In

310the EET, we expected bold fish to locate the doorway in the

311refuge and leave through it. We recorded two traits from the

312EET: whether or not the individual emerged from the relative

313safety of the refuge (emergence) and latency in seconds to do

314so. We predicted positive within-individual correlations be-

315tween emergence from the refuge and the OFT traits, with

316negative correlations between latency to emerge and all other

317traits. Note that the EET tank was set up with the area outside

318the refuge further divided into two zones (1 and 2; Fig 1b). In

319the EET, we had initially planned to use “latency to enter zone

3202”(distal to the refuge) as an additional trait in our analyses;

321however, in practice, this became a redundant trait due to a

322low frequency of fish entering this area.

323Statistical analyses

324All data were modelled using restricted maximum likelihood

325mixed effects models implemented in ASReml V3 (Gilmour

326et al. 2009). Prior to analysis, data for the OFT trait time in

t2:1Ta b l e 2 Behavioural traits recorded in OFT and EET

t2:2Trial type Measured trait Definition

t2:3OFT TL Distance swum (cm)

t2:4OFT Act Percentage of time moving

at a minimum 1.5 cm/s (%)

t2:5OFT AC Area of tank floor covered (%)

t2:6OFT TIM Time spent in zone 2 (s, see Fig. 1)

t2:7EET Em Whether or not the fish emerged

from the refuge (binary)

OFT open-field trials, EET emergence and exploration trials, TL track

length, Act activity, AC area covered, TIM time in middle, Em emergence

t1:1Ta b l e 1 Data

Q2 set for long-term and short-term studies. Number and sex

of individuals involved. Periods of data collection and intervals between

trial pairs. Number of trials conducted; N

LT

=2,448, N

ST

=320. Mean age

of fish in days at the start of each trial pair with standard error in

parentheses

t1:2Study Number Data collection period Number of trials Mean fish age (SE)

t1:3M F T Start End Days

between trials

OFTEET12345

t1:4LT 223 150 373 May 2011 May 2012 56 1,224 1,224 203 (26.35) 259 (26.44) 372 (27.15) 427 (27.13) NP

t1:5ST 16 16 32 Feb 2013 Feb 2013 4 160 160 715 (13.36) 719 (13.36) 723 (13.36) 727 (13.36) 732 (13.36)

LT long-term, ST short-term, Mmale, Ffemale, Ttotal, OFT open-field trial, EET emergence and exploration trial, NP trial not performed

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

327 middle were square-root-transformed to reduce positive skew.

328 Visual inspection of residuals suggested that the assumption

329 of residual normality was reasonable for the other traits re-

330 corded in OFT. All traits were rescaled to standard deviation

331 units prior to analysis to prevent trait scale effects from

332 influencing the structure of I (defined and estimated as de-

333 scribed below). Given that a large proportion of fish did not

334 emerge from the refuge (see results) the latency to emerge data

335 were heavily censored and we elected to use only the binary

336 variable of emergence in subsequent analyses. Emergence

337 was included in full multivariate models usingQ3 residual or

338 restricted maximum likelihood (REML) under an assumption

339 of (multivariate) residual normality. Statistical inferences on

340 this trait should therefore be treated with obvious caution.

341 While statistical approaches exist that allow non-Gaussian

342 trait distributions to be used (e.g. MCMC Bayesian ap-

343 proaches implemented in the R package MCMCglmm

344 (Hadfield, 2010)), they do not currently allow the error struc-

345 tures appropriate to our multivariate models (i.e. no definable

346 or estimable residual covariance between OFTand EET traits;

347 see below) and thus could not be used here. However, we

348 checked the validity of REML-based conclusions regarding

349 emergence by fitting additional univariate and bivariate

350 models using MCMCglmm. Specifically, we fitted a univari-

351 ate model of emergence to estimate the repeatability of this

352 trait and bivariate models of emergence with all other OFT

353 traits to estimate the covariance structure between these traits.

354 All models in MCMCglmm modelled emergence as a cate-

355 gorical trait with the residual variance fixed at 1 and all OFT

356 traits, as Gaussian. All MCMCglmm models were run for a

357 total of 1,050,000 iterations with a burn-in of 50,000 iterations

358 and a thinning interval of 1,000 iterations. The repeatability of

359 Emergence from MCMCglmm models was defined as the

360 intraclass correlation, calculated as V

I

/(V

I

+V

R

+π

2/3

), where

361 V

I

is the among-individual variance and V

R

is the residual

362 variance that in this case, is fixed to 1 (Hadfield 2010).

363 To test the hypothesis that among-individual variance for

364 behavioural traits is both present and repeatable in our fish

365 species, we first combined data from both collection periods

366and fitted a multivariate model of our observed behavioural

367traits. For each trait we included fixed effects of the mean, sex

368(a two-level factor determined from external morphology at

369maturation), home stack (a six-level factor accounting for

370differences between sets of fish sharing water supplies), trial

371number, density treatment and day order. Trial number is the

372cumulative number of trials experienced by an individual

373(fitted as a linear effect). Density treatment is a four-level

374factor describing density conditions experienced in the rearing

375stacks. Day order was modelled as a linear effect of the

376number of preceding trials conducted on any day and was

377used as a proxy for time of day. This was included to control

378for potential diurnal rhythms in fish behaviour. We also fitted

379an interaction term of trial number×density treatment, in case

380any systematic changes in observed trait means across trials

381(due to e.g., age effects, habituation etc.) are themselves

382treatment dependent. Wald Ftests were used to test the sig-

383nificance of fixed effects in the models.

384By including individual identity as a random effect, we

385then partitioned multivariate phenotypic (co)variance not ex-

386plained by the fixed effects into an among-individual and a

387within-individual (residual) component. The former is esti-

388mated as the variance-covariance matrix I, which contains

389estimates of the among-individual variance (V

I

)component

390for each trait on the diagonal and estimates of the correspond-

391ing covariance between trait pairs (COV

I

)offthediagonal.

392The within-individual component is similarly estimated as a

393residual variance-covariance matrix (R). We make the stan-

394dard assumptions that residual errors are normally distributed

395and uncorrelated across observations, and that (co)variance

396parameters in Iand Rare homogeneous across levels of the

397fixed effects (i.e. density treatments, trial number, stack etc.).

398Although the two experiment-specific sets of traits are not

399observed in the same trials, we grouped the data by trial period

400(e.g. OFT1 with EET1). Thus, we modelled a residual covari-

401ance term between OFT and EET traits observed within each

402trial period. Repeatability (R

I

) was then estimated for each trait

403as the among-individual variance (V

I

) divided by total pheno-

404typic variance (V

p

) (where V

P

is the phenotypic variance

Fig. 1 Setup of experimental tanks for aopen-field trials (OFTs) and b

emergence and exploration trials (EETs) as viewed from above. Both

tanks measured 45×25×25 cm and were filled to a depth of 8 cm. For

OFT, two zones of equal area were defined for analysis. For EET, the tank

was divided into three equal zones with fixed opaque material. The refuge

area contained a plastic plant and several small stones. A removable

doorway (hatched line) provided a means of access from the refuge to

the rest of the tank

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

405 conditional on the fixed effects i.e. V

P

=V

I

+V

R

). Between each

406 pair of traits (1, 2) the among-individual covariance (COV

I

)

407 was rescaled to give the corresponding correlation r

I

(where

408 r

I(1,2)

=COV

I(1,2)

/√(V

I1

×V

I2

)).

409 To test the statistical significance of among-individual be-

410 havioural variation, we compared the likelihood of our full

411 multivariate model to two further models. In the first of these,

412 we fitted Ias a diagonal matrix such that the model allows

413 among-individual variance V

I

for each trait, but assumes

414 COV

I

is zero between all trait pairs. In the second, a null

415 model, we removed the random effect of individual identity

416 completely. Comparison of the diagonal model with the null

417 model using likelihood ratio tests (LRTs) allows a global test

418 of the significance of among individual behavioural variance

419 (Wilson et al. 2010). Comparison of the full model with the

420 diagonal model, again by LRT, allows a statistical test of

421 whether Icontains significant between-trait covariance struc-

422 ture (Wilson et al. 2013). LRTs were performed by estimating

423 χ

2

nDF

as twice the difference in model log likelihoods, with

424 the number of degrees of freedom (n) equal to the number of

425 additional parameters to be estimated in the more complex

426 model.

427 The above analyses were then repeated using long- and

428 short-term data subsets to estimate the corresponding matrices

429 I

LT

and I

ST

and associated parameters. Note that, following the

430 conclusion of the LT, the density treatments were no longer

431 applied, and the 32 fish used in the STwere housed together in

432 the same stack. Therefore, the fixed effect stack was redundant

433 and omitted from the models for the short-term subset analy-

434 ses. To further investigate the structure of I,I

LT

and I

ST

,we

435 subjected each matrix to eigenvector (EV) decomposition.

436 This allowed us to examine the following: (a) how much

437 variance is captured by the first axis (eigenvector 1, EV1) of

438 multivariate behaviour in each case (b) whether factor load-

439 ings of individual traits onto EV1 are consistent with an

440 interpretation of “boldness-shyness”and (c) whether EV1 is

441 similar in I

LT

and I

ST

. To provide a quantitative measure of

442 how similar the multivariate behavioural axes emerging from

443 the long- and short-term data sets were, we calculated the

444 angle (θ) between the first eigenvectors of I

LT

and I

ST

.An

445 angle of θ=0° equates to the vectors being perfectly aligned,

446 meaning that EV1 i.e. the axes of multivariate behavioural

447 variation in I

LT

and I

ST

are identical. Conversely, an angle of

448 θ=90° would indicate the vectors are orthogonal (and thus

449 maximally differentiated) to each other across the two differ-

450 ent time periods (i.e. the major axis of behavioural variation

451 across the two studies are independent).

452 Uncertainty around the factor loadings for individual traits

453 on EV1 (for Imatrix) and around θwas estimated using a

454 parametric bootstrap approach (similar to that outlined in the

455 appendix of Morrissey et al. (2012)). We simulated 5,000

456 replicate draws of I,I

LT

and I

ST

from multivariate normal

457 distributions using the maximum likelihood estimates of these

458matrices as the means and the variance-covariance matrices of

459their elements to define the variances. In each case, the 5,000

460simulated matrices were subject to eigen decomposition. Un-

461certainty around the point estimates of trait-specific factor

462loadings was then described using the 95 % highest probabil-

463ity density interval the simulated values of these loadings (for

464I,I

LT

and I

ST

, respectively). Note that these intervals should be

465viewed as approximate as they are vulnerable to departures

466from multivariate normal assumptions. By comparing 5,000

467pairs of simulated LT and ST matrices, we similarly estimated

468the uncertainty around our point estimate of θ. Note however

469that since θcannot be less than zero, we also generated a null

470distribution for the estimator in the absence of any difference

471between (true) Imatrices. This was done by comparing the

472leading eigenvector of each of the 500 replicate draws of I

LT

473(simulated as described above), to the leading eigenvector of a

474second matrix, simulated with the same mean (i.e. the REML

475point estimate of I

LT

) but a variance equal to the estimated

476variance-covariance matrix from the short-term study. Thus

477the null distribution represents θestimates given that (i) the

478angle is zero since true Imatrices are identical (and equal to

479the REML estimate of I

LT

), but (ii) the second (short-term)

480matrix (and so its leading eigenvector) is estimated with

481greater uncertainty due to the lower sample size.

482Finally, we compared V

I

estimates in LT and ST data

483subsets, and tested the among-individual, across data subset

484correlations (r

I(LT,ST)

). For each behavioural trait (x), we used a

485likelihood ratio test to compare a bivariate model of x

LT

and

486x

ST

where V

I

is constrained to be equal, to a model where it is

487free to vary. This tests the hypothesis that among individual

488variance differs across data sets (note that since traits are

489analysed in observed standard deviation units, V

I

can also be

490interpreted as the repeatability estimate unconditional on fixed

491effects). We then expanded this model to estimate the among-

492individual, across data subset correlation (r

I(LT,ST)

) and tested

493this against null hypotheses of both r

I

=0 and r

I

=+1. Estima-

494tion of this correlation is possible since the 32 fish used in the

495short-term study were a subset of the long-term study. If r

I

=+

4961, then this indicates that the ranking of phenotypic merits (i.e.

497each individual’s repeatable component of the observed trait)

498is the same across data sets. However, if r

I

=0, then an indi-

499vidual phenotypic merit in the long-term study is uncorrelated

500with the repeatable component of that same behaviour ob-

501served over a short time period in later life.

502Results

503In total, 1,235 sets of behavioural observations were conduct-

504ed from a possible 1,492, the difference being due to mortality

505of some fish over the study period. Summary data for all

506behavioural traits are presented in the supplemental materials,

507Fig. S1. In EET, the number of fish emerging from the refuge

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

508 within the observation period was lower than anticipated

509 based on pilot data (LT=526/2,448, ST=100/318), resulting

510 in severe censoring of latency to emerge data. We therefore

511 elected to use only the binary emergence trait from this trial

512 type in our analyses.

513 Analysis of full data set

514 There was significant among-individual variance in multivar-

515 iate behaviour (diagonal model versus null model, χ

2

5

=125.6,

516 P<0.001), as well as among-individual covariance among

517 traits (diagonal model versus full model, χ

2

10

=101.8,

518 P<0.001). Estimates of individual repeatability (R

I

(±SE))

519 were low to moderate, ranging across traits from 0.055

520 (±0.024) for emergence (on the observed scale, estimated by

521 REML) to 0. 192 (±0.029) for time in middle (Table 3). Based

522 on univariate models, V

I

was statistically significant at

523 P<0.05 for all traits (Supplemental Table S2). The estimated

524 fixed effects are not directly relevant to present objectives;

525 however they are presented in full in the supplemental mate-

526 rials (Supplemental Table S3).

527 Between traits, the signs of all among-individual cor-

528 relations (r

I

) were positive, consistent with our a priori

529 expectations (Table 3). The OFT traits track length,

530 activity and area covered were all strongly correlated

531 (and nominally significant based on |r

I

|>two standard

532 errors); however, while time in middle was strongly

533correlated with area covered (r

I

=0.653± 0.075, Table 3),

534it was only weakly associated with the other OFT traits.

535The EET trait emergence was positively correlated with

536each OFT trait (r

I

estimates ranging from 0.304 with

537track length to 0.577 with activity, Table 3).

538Eigen analysis of I, estimated from the full data set revealed

539that the first two vectors explained 64 % (eigenvector 1, EV1)

540and 26 % (eigenvector 2, EV2) of the repeatable among-

541individual variation, respectively (Fig. 2). The trait loadings

542on the dominant vector EV1 are consistent with an interpre-

543tation of this axis of variation as boldness (or arguably explo-

544ration and/or general activity; see discussion). Thus, individ-

545uals that tended to emerge repeatedly in the EET swim longer

546distances, are more active, explore more area and spend more

547time in the middle of the OFT tank. By comparison, EV2 trait

548loadings show this axis to be dominated by the time spent in

549the middle of the tank. Track length and activity load on this

550vector to a lesser extent and with an opposing sign to time in

551middle, while the other traits show limited contributions to

552EV2 (Fig. 2b).

553As noted earlier, our REML analysis makes an assumption

554of (multivariate) residual normality that is violated by inclu-

555sion of the binary trait emergence. Univariate analysis of

556emergence using MCMCglmm, calculated following Equa-

557tion 15 of Nakagawa and Schielzeth (2010), yielded a slightly

558higher estimate of repeatability (on the liability scale) with a

559posterior mode of R=0.090, 95 % Q5highest probability density

t3:1Ta b l e 3 Among-individualQ4 variance/covariance matrix (I) from the multivariate analysis of (a) all data, (b) long-term study and (c) short-term study

t3:2TL Act AC TIM Em

t3:3(a) All Data

t3:4TL 0.130 (0.025)0.865 (0.033)0.750 (0.069)0.162 (0.117)0.304 (0.198)

t3:5Act 0.124 (0.024)0.159 (0.026)0.731 (0.065)0.241 (0.106)0.577 (0.182)

t3:6AC 0.097 (0.022)0.104 (0.022)0.128 (0.026)0.653 (0.075)0.414 (0.202)

t3:7TIM 0.026 (0.019)0.042 (0.020)0.102 (0.023)0.192 (0.029)0.540 (0.180)

t3:8Em 0.026 (0.017)0.054 (0.018)0.035 (0.018)0.056 (0.019)0.055 (0.024)

t3:9(b) Long term

t3:10 TL 0.143 (0.028)0.892 (0.030)0.777 (0.069)0.238 (0.118)0.272 (0.192)

t3:11 Act 0.137 (0.026)0.164 (0.028)0.708 (0.072)0.314 (0.106)0.539 (0.180)

t3:12 AC 0.108 (0.025)0.106 (0.025)0.136 (0.030)0.704 (0.075)0.458 (0.208)

t3:13 TIM 0.041 (0.022)0.058 (0.022)0.118 (0.026)0.207 (0.033)0.607 (0.181)

t3:14 Em 0.027 (0.020)0.058 (0.020)0.045 (0.021)0.073 (0.022)0.071 (0.028)

t3:15 (c) Short term

t3:16 TL 0.458 (0.155)0.926 (0.041)0.640 (0.182)−0.247 (0.256)1.070 (0.513)

t3:17 Act 0.381 (0.137)0.369 (0.134)0.812 (0.112)0.017 (0.274)1.001 (0.502)

t3:18 AC 0.188 (0.095)0.214 (0.097)0.188 (0.089)0.492 (0.222)0.545 (0.524)

t3:19 TIM −0.083 (0.089)0.005 (0.084)0.106 (0.079)0.248 (0.101)−0.667 (0.557)

t3:20 Em 0.165 (0.080)0.139 (0.073)0.054 (0.056)−0.076 (0.059)0.052 (0.066)

Estimates of variance (V

I

, ital) with among-individual between-trait covariances (COV

I

, bold) and among-individual between-trait correlations (r

I

,bold

ital). Standard errors are shown in parentheses for all parameter estimates

TL track length, Act activity, AC area covered, TIM time in middle, Em emergence

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

560 (HPD) interval 0.024–0.177, Table S1. While noting that

561 interval will never span zero since Ris constrained to lie in

562 positive parameter space, the posterior mode is nonetheless

563 “distinct”from zero (Supplemental material, Fig. S2). Bivar-

564 iate models (i.e. the use of one OFT trait plus emergence as the

565 phenotypic variates) also confirmed the presence of strong

566 positive among-individual correlations (r

I

) between emer-

567 gence and OFT traits. Thus, the MCMCglmm analyses cor-

568 roborate the results of the REML analysis for emergence

569 (Supplemental material, Table S1).

570 Comparison of long- and short-term results

571 In both the long- and short-term studies, the presence of

572 repeatable variance was statistically supported (comparisons

573 of null and diagonal model: LT χ

2

5

=77.0, P<0.001; ST χ

2

5

=

574 29.7, P<0.001) as was the presence of between-trait among-

575 individual covariance structure (comparisons of diagonal and

576 full multivariate model: LT χ

2

10

=95.0, P<0. 001; ST χ

2

10

=

577 54.9, P<0.001). Univariate models confirmed that V

I

was

578statistically significant for all OFT traits in both LT and ST,

579but not for emergence in ST (Supplementary Table S2).

580The estimate of I

LT

is very similar to that obtained using all

581data (as described above), not unexpected given that the long-

582term study contributes the bulk of the total data set. However,

583comparison of I

LT

and I

ST

(and derived parameters thereof)

584indicates some differences in the structure of among-

585individual behavioural variation as estimated from our long-

586and short-term studies (Table 3). Note that the smaller size of

587the short-term data set means that the estimates are less precise

588for this study; this is reflected in the larger standard errors

589associated with the parameters. Repeatability estimates (R)

590were higher in the short-term study across all traits. However

591the increased Rfrom ST was particularly striking for track

592length (Table 3,Fig.3). For this trait, along with activity and

593area covered the null hypothesis of equality of (V

I

)acrossdata

594sets could be rejected (comparison of bivariate models with

595homogeneous and heterogeneous V

I

,P<0.05, Fig. 3).

596The among-individual between-trait correlations (r

I

)reveal

597a broadly similar structure for the long- and short-term studies

598(Table 3). Thus, estimates for ST largely confirm our a priori

a)

b)

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

Eigenvector loading

ALL (64%)

LT (66%)

ST (73%)

-1.5

-1

-0.5

0

0.5

1

1.5

2

2.5

3

Track

Len

g

th

Activity Area

Covered

Time in

Middle

Emergence

Track

Length

Activity Area

Covered

Time in

Middle

Emergence

Eigenvector loading

ALL (26%)

LT (24%)

ST (26%)

Fig. 2 Eigenvector

decomposition of Ifor all data

combined (All), long-term (LT)

and short-term (ST) data sets,

with percentage of variance

explained in parentheses. Shown

are the trait loadings in standard

deviation units for the first (a)and

second (b) eigenvectors. Error

bars show 95 % HPD intervals

from the parametric bootstrap (see

text for details). Note that the

point estimates of EV1 loadings

on emergence in All and LT

datasets actually lie outside the

simulated intervals. This reflects

sensitivity of intervals estimates

to departures from multivariate

normality assumed in the

bootstrap

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

599 expectation of positive correlation structure between the OFT

600 traits and emergence. One qualitative exception to the expect-

601 ed pattern is provided by time in middle. In LT, this trait is

602 positively correlated with all other traits as expected; however,

603 in ST, the sign of r

I

is negative (but not significant) between

604 time in middle and track length and emergence (Table 3).

605 Eigen decomposition confirms the view that qualitative

606 differences between I

LT

and I

ST

are largely related to time in

607 middle. Thus, in both data sets, the first eigenvector again

608 dominates the variance in I(accounting for 66 and 73 % in

609 long- and short-term, respectively), consistent with an impor-

610 tant latent character underlying behavioural variation

611 (Fig. 2a). Time in middle has a strong positive loading on

612 EV1

LT

,consistentwithouraprioriexpectationthataboldfish

613 would spend more time in the middle of the open-field arena;

614 the corresponding loading coefficient is close to zero (in fact

615 slightly negative) on EV1

ST

. The angle (θ)betweenEV1

LT

616 and EV1

ST

is 34.63° (95 % HPD interval, 5.03–53.09°).

617 While the point estimate of 34.63° indicates at least some

618 divergence between the leading eigenvectors on a scale from

619 0 (no difference) to 90 (axes are orthogonal), it is not signif-

620 icantly greater than the angle expected by chance if the true

621 matrices are identical (95 % HPD of the null distribution for θ

622 generated by our parametric bootstrap is from 1.54 to 69.14°).

623 While we acknowledge that our null distribution indicates low

624 statistical power to reject the null hypothesis that θ=0 (see

625 Supplemental Fig. S3), our conclusion is however that EV1

LT

626 and EV1

ST

are broadly similar, with qualitative differences

627 largely attributable to the decreased loading of TIM on EV1

ST

.

628 This is further evidenced by a drop in θfrom 34.63° to just

629 11.15° for the corresponding comparison of Iestimates ex-

630 cluding time in middle. There are also some qualitative incon-

631 sistencies evident between EV2

LT

and EV2

ST

for the OFT

632 traits, due to greater loadings on track length (changes sign),

633 activity, area covered and time in middle, while the loading on

634 emergence is reduced (also changes sign) (Fig. 2b). The angle

635 (θ)betweenEV2

LT

and EV2

ST

=48.32° (95 % HPD interval

636 25.75–86.48°), which again is not significantly different from

637 null expectations.

638For those individuals tested in both long- and short-term

639studies, the among-individual correlations between LT and ST

640data sets were positive (although not always significant based

641on likelihood ratio tests) for OFT traits (Fig. 4), ranging from

6420.219 (±0.294) to 0.729 (±0.314). Estimates were significant-

643ly greater than zero for area covered and time in middle.

644However, we also found that the correlation was significantly

645less than 1 for the traits track length and activity. Thus, while

646phenotypic performance of an individual in one data set may

647be predictive of its behaviour in the other, there is also evi-

648dence that the ranking of individuals, at least for track length

649and activity, significantly differs between long- and short-term

650studies. For emergence, the corresponding among-individual

651correlation estimates between long- and short-term were ac-

652tually negative, though not significantly so. In fact, the esti-

653mate was characterised by so much uncertainty that despite

654being negative it was not possible to reject the null hypothesis

655of r=+1. We suggest this is a result of the low repeatable

656variation of emergence, and thus, little weight should be

657placed on this result.

658Discussion

659Data from our long-term (LT) and short-term (ST) studies

660provide evidence of among-individual variance in behaviour,

661both when considered separately and in combination. Of the

662five traits assayed in the two distinct types of behavioural

663trial—open field (OFT) and emergence and exploration

664(EET)—repeatabilities were statistically supported in all cases.

665In addition, our analyses support the presence of a significant

666among-individual correlation structure for behavioural traits in

667I. Correlation structure is found both within and acrosscontexts

668(i.e. trial types), indicating behavioural variation among fish

669that is consistent with accepted definitions of animal personal-

670ity. We found that repeatabilities of OFT traits were higher than

671the EET though not significantly so in all cases. Our results

672therefore support the assertion of Burns (2008)thattheOFTisa

673good and reliable test of boldness and exploratory behaviour in

-0.1

0.0

0.1

0.2

0.3

0.4

0.5

0.6

Track Len

g

th Activity Area Covered Time in Middle Emer

g

ence

Repeatability

LT

ST

**

*

*

Fig. 3 Estimated trait

repeatabilities from long-term

(LT) and short-term (ST) studies.

Error bars specify one standard

error. Pvalues (**P<0.01; *P

<0.05) indicate significant

differences between V

I

based on

likelihood ratio tests (see text for

detail)

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

674 small fish, although it is certainly possible that the EETcould be

675 better optimised to target the among-individual component. We

676 discuss the biological interpretation of (multivariate) variance

677 within these two trial types further below. However, here, we

678 note the pragmatic consideration that the binary distribution of

679 emergence data obtained from the EET is more difficult to

680 analyse and interpret while the censoring of latency to emerge

681 created a data distribution not readily modelled in any software.

682 Although such problems are likely surmountable by modifica-

683 tion of the behavioural assay (e.g. using an extended observa-

684 tion time to eliminate or at least reduce censoring), at least in

685 this case, it is not clear to us that the EET provides additional

686 biological insight.

687 Comparison of long- and short-term data sets suggested

688 that the patterns of individual (co)variance between traits

689 frequently used to define boldness are relatively stable.

690 Nevertheless, as predicted a priori we found a tendency

691 for the magnitude of R

I

to decrease with a higher interval

692 between observations, at least in OFT trials. For example,

693 repeatabilities for OFT traits ranged from 0.188 to 0.458

694 in the short-term data (with repeat observations at an

695 average interval of 4 days) but 0.136 to 0.207 in the

696 long-term data (average interval of 56 days). In a meta-

697 analysis of behavioural repeatability studies that included

698 either long-term (i.e. >1 year) or short-term (i.e. <1 year)

699 intervals between observations, the average (median)

700 across all estimates was 0.37 (Bell et al. 2009). Here,

701 our repeatability estimates pertain to correlated traits and

702 are therefore not independent. Nevertheless, apart from our

703 short-term study estimates for track length and activity, we

704 note that our estimates for all other traits were lower than

705 those of the meta-analysis average. Repeatability estimates

706 from short-term studies in the meta-analysis (Bell et al.

707 2009) outnumbered those from long-term studies by 11:1;

708 however, our study considers observations collected within

709 two distinctly separate periods across individual lifetimes.

710Arguably, the more important question to be asked of our

711long- and short-term data sets concerns the stability of corre-

712lation structure within the multivariate Imatrix and the inter-

713pretation of boldness from its eigenvector decomposition. As

714seen with the single-trait repeatabilities, the structure of I

LT

715mirrored that of Iestimated from all data combined. This is

716unsurprising given that the long-term data comprised a much

717greater number of individuals and will thus dictate patterns in

718the combined dataset. I

LT

is dominated by a single vector that

719is broadly consistent with our expectations of boldness. Sig-

720nificant within- and between-trial type correlations indicate

721that individuals emerging from the EET refuge are more likely

722to have high scores for all OFT traits, thus matching our

723expectation of bold behaviour.

724Though not statistically significant, qualitative differences

725between I

LT

and I

ST

were apparent. These differences were

726focussed around the sign and strength of correlations between

727time in middle and traits from both trial types, indicating that

728both bold and shy individuals from the short-term study spent

729a similar amount of time in the middle, whereas in the long-

730term study, shy fish had behaved in a more thigmotaxic

731manner. This pattern was reflected in comparisons of the

732major eigenvectors of long- and short-term data, where a

733moderate, albeit not statistically significant, angle (θ)between

734the first long- and short-term axes was estimated. Further-

735more, if time in middle is dropped from the calculation, the

736estimated angle is reduced by more than half. Thus, our

737interpretation is that both data sets reveal a major vector of

738among-individual (co)variance in observed behavioural traits.

739This vector is similar in the two data sets and can be

740interpreted as a latent personality trait—namely boldness. In

741both data sets, bolder individuals tend to swim longer dis-

742tances, be more active and explore more area (in the OFT) and

743are more likely to emerge from a refuge (in the EET). How-

744ever, tendency to spend more time in the middle of the OFT

745arena appears not be a reliable indicator of boldness as it was

-1.5

-1

-0.5

0

0.5

1

1.5

Track-length Activity Area Covered Time in Middle Emergence

Estimated among-individual correlations

* * † †

Fig. 4 Estimated among-

individual correlations (r

I

)

between LT and ST data sets for

each observed trait, with standard

error bars. Each correlation was

tested against two null hypotheses

of interest: (i) r

I

=1.0 (*P<0.05)

and (ii) r

I

=0.0 (

†

P<0.05), using

likelihood ratio tests to compare

unconstrained and constrained

models (see text for details)

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

746 only associated with this vector in the LT study. Indeed, this

747 trait was the major source of qualitative difference between

748 the two matrices.

749 In the current study, it is not possible to distinguish whether

750 higher repeatabilities and the changing structure of Iwith

751 regard to time in middle are a consequence of the sampling

752 period (long- versus short-term) or potentially reflect interest-

753 ing, possibly even species-specific, biological changes that

754 happen with age and/or trial experience. Note, however, that

755 our analyses control for any habituation effects on mean

756 behaviour, and that we found little statistical support for

757 individual-by-trial-number interactions (results not shown).

758 More generally, some authors have argued that individual

759 behaviour is likely to become more rigid and follow more

760 set patterns over time (Roberts and DelVecchio 2000). If so,

761 we would predict increasing repeatabilities with age (here,

762 confounded with time scale of data collection). Conversely,

763 others suggest that in the absence of any disturbance (e.g. in a

764 constant laboratory environment), expectations of changes to

765 individual patterns of behaviour formed in early life are ill

766 founded (Stamps and Groothuis 2010). While no overall

767 differences were found between juvenile and adult behaviour-

768 al repeatabilities in the Bell et al. (2009) meta-analysis, a

769 subset of data suggested juvenile behaviour to have higher

770 repeatability than that of adults. However, the meta-analysis

771 contained only three studies that included observations fol-

772 lowing individuals through from juvenile to adult status. Thus,

773 direct comparison of age classes is not straightforward. Clear-

774 ly, more empirical studies of how repeatability changes with

775 age would be valuable, as indeed would parallel studies ex-

776 ploring environmental dependence. Here, we assumed homo-

777 geneous variance structures across environments (density

778 treatments, stacks) and other fixed effects (sexes, day order)

779 for simplicity. These assumptions can be relaxed in the statis-

780 tical models to test for and quantify individual by environment

781 (I×E) as changes in the among-individual variance (or struc-

782 ture of Iin the multivariate case) (Dingemanse et al. 2010).

783 Here, post hoc analyses of the LT data set provides some

784 evidence of heterogeneous repeatabilities across density treat-

785 ment classes (see Supplemental Table S4). Though not ex-

786 pected to bias current conclusions (parameter estimates pre-

787 sented are effectively averaged across treatments), if robust,

788 this effect may certainly be biologically interesting.

789 The population level patterns of among-individual

790 (co)variances between traits were broadly similar between

791 LT

I

and ST

I

, albeit with some differences as described above.

792 However, by using the same individuals in both long- and

793 short-term studies, we were able to address the question of

794 whether the relative ranking of individuals with respect to

795 their behavioural tendencies was stable. The estimates of r

I

796 for each observed behavioural trait between the long- and

797 short-term datasets provide a mixed answer to this question.

798 Positive correlations for the OFT traits do show a degree of

799stability in (repeatable) behavioural tendencies across the data

800sets though statistical support was mixed and it appears indi-

801viduals were more likely to maintain a consistent ranking for

802some traits (e.g. area covered) than others (e.g. track length).

803We previously stated it is not our intention to be prescrip-

804tive about what boldness is or how it should be assayed.

805Nevertheless, a priori, we anticipated that in the OFT, bold

806fish would travel long distances and be willing to visit a large

807area of the tank including the central zone and that these traits

808would correlate significantly with whether individuals

809emerged in the EET. However, this depiction requires that

810the bold individual is also active and/or exploratory. Above,

811we have noted that the major axis of variation in Iis largely

812consistent with expectations of a bold-shy continuum as the

813terminology is used in the literature; however, the strength of

814among-individual correlations suggests that it could equally

815be called exploration or general activity in a novel environ-

816ment. Nevertheless, as qualitatively, almost all the variance

817loads onto this single axis of variance; we conclude that these

818continuums (personality axes) are, at least in our study spe-

819cies, either the same entity or so tightly correlated that

820attempting to distinguish between them may have little prac-

821tical value. Indeed, Burns (2008) concluded that emergence

822from a refuge was difficult to interpret strictly as either bold-

823ness or exploration, even though it has been described as

824boldness only by others (e.g. Budaev 1997a; Brown et al.

8252005). Exploring the functional significance of the conse-

826quences of this behavioural variance in wild populations is

827likely to yield more insight than further debate with regard to

828terminology (e.g. Dingemanse et al. 2012a; Kurvers et al.

8292012; Carvalho et al. 2013). Nonetheless, we have sufficient

830statistical support in our results to conclude that both trial

831types revealed behaviours characteristic of boldness, evident

832from the strong among-individual correlations between all the

833observed traits. This again leads us in the direction of Burns’

834(2008) view that in practice, the OFT offers the most useful

835test arena for this axis of personality. Here, we have obtained

836repeated measures of multiple behavioural traits during two

837test types and across two distinct sampling periods (long

838versus short term), something that has seldom been accom-

839plished in the literature. In practical terms, we conclude that

840the OFT is preferable to the EET as an experimental test for

841investigating boldness, and we show how eigen decomposi-

842tion of an Imatrix can usefully identify latent personality

843traits. This multivariate approach is broadly similar to that

844used in several other recent studies (Budaev 2010;Carteretal.

8452013; Araya-Ajoy and Dingemanse 2014). Our study also

846provides information about the stability of personality, both

847in terms of population level patterns and individual differ-

848ences. We find that observed behavioural traits are repeatable

849over long time periods as well as when observations are made

850over only a few weeks, although there is a tendency for short-

851term estimates to be higher. Taking a multivariate approach,

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

852 we show that Iis dominated by a single vector through

853 phenotypic space that is similar across the two study periods

854 and can be interpreted as boldness. We note, however, there

855 are at least some qualitative differences in the relationships of

856 observed behaviours to this vector. At the individual level, we

857 also find qualified support for the proposition that short-term

858 studies are informative for an individual’s behavioural pheno-

859 type over longer (e.g. lifetime) periods.

860 Acknowledgments The authors would like to thank Niels Dingemanse

861 and three anonymous reviewers for the helpful comments that have

862 helped to improve this manuscript. Additionally, we thank Jarrod Had-

863 field for advice with MCMCglmm software. This work was supported by

864 an EPSRC studentship to KB and a BBSRC fellowship to AJW. CAW

865 was supported by a NERC Junior Research Fellowship.

866

867 Ethical standards Ethical review committees at the Universities of

868 Edinburgh and Exeter approved all work in this study, which was carried

869 out under licence granted by the Home Office (UK) under the Animals

870 (Scientific Procedures) Act 1986.

871 References

872

873 Araya-Ajoy YG, Dingemanse NJ (2014) Characterizing behavioural

874 ‘characters’: an evolutionary framework. Proc R Soc B 281:

875 20132645

876 Ariyomo TO, Watt PJ (2012) The effect of variation in boldness and

877 aggressiveness on the reproductive success of zebrafish. Anim

878 Behav 83:41–46

879 Bell A, Sih A (2007) Exposure to predation generates personality in

880 threespined sticklebacks (Gasterosteus aculeatus). Ecol Lett 10:

881 828–834

882 Bell AM, Hankison SJ, Laskowski KL (2009) The repeatability of

883 behaviour: a meta-analysis. Anim Behav 77:771–783

884 Boissy A (1995) Fear and fearfulness in animals. Q Rev Biol 70:165–191

885 Brommer JE (2013) On between-individual and residual (co)variances in

886 the study of animal personality: are you willing to take the "individ-

887 ual gambit"? Behav Ecol Sociobiol 67:1027–1032

888 Brown C, Jones F, Braithwaite V (2005) In situ examination of boldness-

889 shyness traits in the tropical poeciliid, Brachyraphis episcopi. Anim

890 Behav 70:1003–1009

891 Budaev SV (1997a) Alternative styles in the European wrasse,

892 Symphodus ocellatus: boldness-related schooling tendency.

893 Environ Biol Fish 49:71–78

894 Budaev SV (1997b) "Personality" in the guppy (Poecilia reticulata): a

895 correlational study of exploratory behavior and social tendency. J

896 Comp Psychol 111:399–411

897 Budaev SV (2010) Using principal components and factor analysis in

898 animal behaviour research: caveats and guidelines. Ethology 116:

899 472–480

900 Burns JG (2008) The validity of three tests of temperament in guppies

901 (Poecilia reticulata). J Comp Psychol 122:344–356

902 Bushuev AV, Kerimov AB, Ivankina EV (2010) Estimation of heritability

903 and repeatability of resting metabolic rate in birds, with free-living

904 pied flycatchers Ficedula hypoleuca (Aves: Passeriformes) as an

905 example. Zh Obshch Biol 71:402–424

906 Carere C, Drent PJ, Privitera L, Koolhaas JM, Groothuis TGG (2005)

907 Personalities in great tits, Parus major: stability and consistency.

908 Anim Behav 70:795–805

909Carter AJ, Feeney WE, Marshall HH, Cowlishaw G, Heinsohn R (2013)

910Animal personality: what are behavioural ecologists measuring?

911Biol Rev 88:465–475

912Carvalho CF, Leitão AV, Funghi C, Batalha HR, Reis S, Mota PG, Lopes

913RJ, Cardoso GC (2013) Personality traits are related to ecology

914across a biological invasion. Behav Ecol 24:1081–1091

915Chervet N, Zöttl M, Schürch R, Taborsky M, Heg D (2011) Repeatability

916and heritability of behavioural types in a social cichlid. Int J Evol

917Biol 2011:321729

918Cote J, Clobert J, Brodin T, Fogarty S, Sih A (2010) Personality-

919dependent dispersal: characterization, ontogeny and consequences

920for spatially structured populations. Philos T R Soc B 365:4065–

9214076

922Dall SRX, Houston AI, McNamara JM (2004) The behavioural ecology

923of personality: consistent individual differences from an adaptive

924perspective. Ecol Lett 7:734–739

925David M, Auclair Y, Cezilly F (2012) Assessing short- and long-term

926repeatability and stability of personality in captive zebra finches

927using longitudinal data. Ethology 118:932–942

928Dingemanse NJ, Both C, Drent PJ, Tinbergen JM (2004) Fitness conse-

929quences of avian personalities in a fluctuating environment. Proc R

930Soc Lond B 271:847–852

931Dingemanse NJ, Wright J, Kazem AJN, Thomas DK, Hickling R,

932Dawnay N (2007) Behavioural syndromes differ predictably be-

933tween 12 populations of three-spined stickleback. J Anim Ecol 76:

9341128–1138

935Dingemanse NJ, Dochtermann NA, Wright J (2010) A method for

936exploring the structure of behavioural syndromes to allow

937formal comparison within and between data sets. Anim

938Behav 79:439–450

939Dingemanse NJ, Bouwman KM, van de Pol M, van Overveld T, Patrick

940SC, Matthysen E, Quinn JL (2012a) Variation in personality and

941behavioural plasticity across four populations of the great tit Parus

942major. J Anim Ecol 81:116–126

943Dingemanse NJ, Dochtermann NA, Nakagawa S (2012b) Defining

944behavioural syndromes and the role of ‘syndrome deviation’

945in understanding their evolution. Behav Ecol Sociobiol 66:

9461543–1548

947Dochtermann NA, Jenkins SH (2007) Behavioural syndromes in

948Merriam’s kangaroo rats (Dipodomys merriami): a test of competing

949hypotheses. Proc R Soc Lond B 274:2343–2349

950Dochtermann NA, Roff DA (2010) Applying a quantitative genetics

951framework to behavioural syndrome research. Proc R Soc Lond B

952365:4013–4020

953Gilmour AR, Gogel BJ, Cullis BR, Thompson R (2009) ASReml user

954guide release 3.0. VSNi, Hemel Hempstead, UK

955Gosling SD (2001) From mice to men: what can we learn about person-

956ality from animal research? Psychol Bull 127:45–86

957Q6Hadfield J (2010) MCMCglmm course notes. http://cran.r-project.org/

958web/packages/MCMCglmm/vignettes/CourseNotespdf

959Hall CS (1934) Emotional behavior in the rat. I. Defecation and urination

960as measures of individual differences in emotionality. J Comp

961Psychol 18:385–403

962Huntingford FA (1976) Relationship between anti-predator behavior and

963aggression among conspecifics in 3-spined stickleback,

964Gasterosteus aculeatus. Anim Behav 24:245–260

965Kanda LL, Louon L, Straley K (2012) Stability in activity and boldness

966across time and context in captive Siberian dwarf hamsters.

967Ethology 118:518–533

968Koolhaas JM, Korte SM, De Boer SF, Van Der Vegt BJ, Van Reenen CG,

969Hopster H, De Jong IC, Ruis MAW, Blokhuis HJ (1999) Coping

970styles in animals: current status in behavior and stress-physiology.

971Neurosci Biobehav R 23:925–935

972Kurvers RHJM, Nolet BA, Prins HHT, Ydenberg RC, van Oers K (2012)

973Boldness affects foraging decisions in barnacle geese: an experi-

974mental approach. Behav Ecol 23:1155–1161

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTEDPROOF

975 Moretz JA (2003) Aggression and RHP in the northern swordtail

976 fish, Xiphophorus cortezi: the relationship between size and

977 contest dynamics in male-male competition. Ethology 109:

978 995–1008

979 Morrissey MB, Parker DJ, Korsten P, Pemberton JM, Kruuk LEB,

980 Wilson AJ (2012) The prediction of adaptive evolution: empirical

981 application of the second theorem of selection and comparison to the

982 breeder’s aquation. Evolution 66:2399–2410

983 Mutzel A, Dingemanse NJ, Araya-Ajoy YG, Kempenaers B (2013)

984 Parental provisioning behaviour plays a key role in linking person-

985 alitywithreproductivesuccess.ProcRSocLondB280

986 Nakagawa S, Schielzeth H (2010) Repeatability for Gaussian and non-

987 Gaussian data: a practical guide for biologists. Biol Rev 85:935–956

988 Réale D, Reader SM, Sol D, McDougall PT, Dingemanse NJ (2007)

989 Integrating animal temperament within ecology and evolution. Biol

990 Rev 82:291–318

991 Roberts BW, DelVecchio WF (2000) The rank-order consistency of

992 personality traits from childhood to old age: a quantitative review

993 of longitudinal studies. Psychol Bull 126:3–25

994 Ronning B, Moe B, Bech C (2005) Long-term repeatability makes basal

995 metabolic rate a likely heritable trait in the zebra finch Ta en i o p y gi a

996 guttata. J Exp Biol 208:4663–4669

997Rudin FS, Briffa M (2011) Is boldness a resource-holding potential trait?

998Fighting prowess and changes in startle response in the sea anemo-

999ne, Actinia equina. Proc R Soc Lond B 279:1904–1910

1000Sih A, Bell A, Johnson JC (2004) Behavioral syndromes: an ecological

1001and evolutionary overview. Trends Ecol Evol 19:372–378

1002Stamps J, Groothuis TGG (2010) The development of animal personality:

1003relevance, concepts and perspectives. Biol Rev 85:301–325

1004Toms CN, Echevarria DJ, Jouandot DJ (2010) A methodological review

1005of personality-related studies in fish: focus on the shy-bold axis of

1006behavior. Int J Comp Psychol 23:1–25

1007Walsh RN, Cummins RA (1976) Open-field test—critical review.

1008Psychol Bull 83:482–504

1009Warren EW, Callaghan S (1975) Individual differences in response to and

1010open field test by guppy Poecilia reticulata (Peters). J Fish Biol 7:

1011105–113

1012Wilson AJ, Réale D, Clements MN, Morrissey MM, Postma E, Walling

1013CA, Kruuk LEB, Nussey DH (2010) An ecologist’s guide to the

1014animal model. J Anim Ecol 79:13–26

1015Wilson AJ, Grimmer A, Rosenthal GG (2013) Causes and consequences

1016of contest outcome: aggressiveness, dominance and growth in the

1017sheepshead swordtail, Xiphophorus birchmanni. Behav Ecol

1018Sociobiol

1019

Behav Ecol Sociobiol

JrnlID 265_ArtID 1692_Proof# 1 - 12/02/2014

UNCORRECTED PROOF

AUTHOR QUERIES

AUTHOR PLEASE ANSWER ALL QUERIES.

Q1. “Smith and Blumstein 2008”is cited in the body but its bibliographic information is missing.

Kindly provide its bibliographic information. Otherwise, please delete it from the text/body.

Q2. Please check if the modifications in Tables 1 and 2 are appropriate.

Q3. Please check if the captured expansion "residual or restricted maximum likelihood" for the

abbreviation "REML" is correct.

Q4. Kindly check if the modifications in Table 3 are correct.

Q5. Kindly check if the captured expansion "highest probability density" for the abbreviation "HPD" is

correct.

Q6. Please provide an access date for the reference "Hadfield J (2010)".