Dunbar Some Aspects of Research Design and Their Implications in the Observational Study of Behaviour

7/28/2019 Dunbar Some Aspects of Research Design and Their Implications in the Observational Study of Behaviour

1/22

Some Aspects of Research Design and Their Implications in the Observational Study of

BehaviourAuthor(s): R. I. M. DunbarSource: Behaviour, Vol. 58, No. 1/2 (1976), pp. 78-98Published by: BRILLStable URL: http://www.jstor.org/stable/4533756 .

Accessed: 10/06/2013 12:26

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .http://www.jstor.org/page/info/about/policies/terms.jsp

.JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of

content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms

of scholarship. For more information about JSTOR, please contact [email protected].

.

BRILL is collaborating with JSTOR to digitize, preserve and extend access to Behaviour.

http://www.jstor.org

This content downloaded from 164.15.128.33 on Mon, 10 Jun 2013 12:26:30 PMAll use subject to JSTOR Terms and Conditions
http://www.jstor.org/action/showPublisher?publisherCode=baphttp://www.jstor.org/stable/4533756?origin=JSTOR-pdfhttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/stable/4533756?origin=JSTOR-pdfhttp://www.jstor.org/action/showPublisher?publisherCode=bap


2/22

Behaviour, LVI1I, 1-2-

SOME ASPECTS OF RESEARCH DESIGN AND THEIRIMPLICATIONS IN THE OBSERVATIONALSTUDY OF BEHAVIOURby

R. I. M. DUNBAR')(Departmentof Psychology, University of Bristol, England)

(Ace.5-VIII-I975)

INTRODUCTIONThe success of any research project depends to a large extent on the

amount of forethought devoted to decision-making before the observer evenbegins to look at his animals. In any given case, there are five basic deci-sions which the investigator has to make, whether he does so consciouslyor not (J. ALTMANN, 974). These may be summarised as follows:

i. The problem to be investigated has to be defined;2. The kinds of behavioural parameters most appropriate to answering

that particular question must be identified;3. A sampling strategy which will provide an unbiassed estimate of these

parameters must be chosen;4. A method which is both suitable for recording the-data and practicable

in the field must be selected;5. The most appropriate statistical test(s) for analysing the data in the

form obtained must be chosen.Each of these decisions poses its own problems and difficulties. Further-

more, the decision reached at any one of these points places constraintson the options which are open to the investigator at other points. It is wellknown, for instance, that an inappropriate choice of statistical test caninvalidate the conclusions drawn from a set of data. It goes without saying,therefore, that the data must be recorded in a form which is suitable for

i) I am grateful to the many people with whom I have discussedthe problemswhichhave been the subject of this paper. I am especially indebtedto Mrs Jeanne ALTMANN,Dr J. H. CROOK,Professor R. A. HINDE, Jenny INGRAMand Dr E. M. RUSSELL fortheir comments on an earlier draft of the manuscript; their criticisms greatly helpedto clarify many of the issues. The field work was made possible by grants from theScience Research Council, London,and the Wenner-Gren Foundation for Anthropologi-cal Researchwhile the author was in receipt of an S. R. C. Overseas Scholarship.

http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp


3/22

RESEARCH DESIGN IN OBSERVATIONAL TUDIES OF BEHAVIOUR 79analysis by an available est. It is perhaps, ess often appreciated, owever,that iinappropriateolutionsat any of the other stages may have equallyunfortunateresults (see, for instance, the discussionof samplingmethodsby J. ALTMANN, I974).The increasing emphasisplaced on the quantificationof data in recentyears has forcedthe observer o consider hese five points of researchdesignmore seriously than has formerly been the case. This is reflected in thenumber of papers recentlypublished which have been concerned purelywith research methodology (e.g. HUTT & HUTT, I970; HINDE, I973;J. ALTMANN, I974). Unlike these latter works, the present paper is notintended o be an observer'sguide; rather,I aim simply to draw attentionto some important spectsof researchdesign whichhave not been discussedin any detail elsewhere.Whereas most methodologicalpapers have beenconcernedwith points (3) and (4) above, I shall be concernedprimarilywith point (2), that is, the choice of behaviouralparameterswhich arerelevant o the researchproblem.This shouldnot, however,be takento implythat this aspect of researchdesign is more importanthan the others.The particular roblem hat prompted his paperwas how best to quantifysocial relationships monggroupsof monkeys (in this case gelada baboons,Theropithecus elada). However, althoughI shall be concernedmainlywiththe effects of using different parameters o quantify social relationships,manyof the conclusionsdrawnapply to other areasof behavioural esearch.

I. THERESEARCH ROBLEM: ATIONALENDAIMSMany field workersfind it usefulto describe he overallpatternof social

relationshipsn the groupof animalsunderstudy. By doing so, the observercan identify social subgroups, he compositionand organisationof whichcan be related o the structureand dynamicsof the groupas a whole.Thus,rather than treating dominancehierarchies n vacuo, we may be able todeterminehow dominance elationships elateto preferences n socialpart-ners. Do, for instance,animalsthat frequentlyengage in "friendly"socialinteraction upporteach otherwhen attackedby otherindividuals?In studyingsocial relationshipsas such, we commonlysupposethat weare investigatingsomethingwhich is relevant to the animals themselves.Since, on the whole, animalstend not to interact at random,we assumethat when they do interactwith a particularndividual, n some real sensethey choose to do so. In other words,we assumethat the interactioncon-cernedis not some coincidental esult of a randomseries of events. Someauthorshavetermedthese relationships"socialaffinities" or "socialbonds"



4/22

8o R. I. M. DUNBAR(see, for instance, KUMMER, I968; ALDRICH-BLAKE et al., I97I; SADE,I973; DUNBAR & DUNBAR, I975). In general, these terms are used inan operational sense: that is to say, a social bond is defined in terms of aparticularly high rate of interaction between two individuals. Naturally,we can say little about what the animals themselves (to put it anthropomor-phically) "think" since we are restricted to considering only the observableaspects of their behaviour. Nonetheless, many authors have found a descrip-tion of the "communication network" within the group (i.e. the channelsor routes of most frequent communication) to be heuristically useful insofaras it provides a quantitative, first level approach to identifying the under-lying structure of the system (e.g. S. ALTMANN, I968; KUMMER, I968;SADE, I973).

However, different field workers have used different quantitative methodsfor determining such relationships. Consider the following few examples:KUMMER (I968) in his, study of Papio hamadryas used the number of I minintervals in which individuals interacted; SADE (I965, 1973) used the fre-quencies with which individuals groomed each other in his study of Macacamulatta, while S. ALTMANN (I968), in studying the relationships betweenage-sex classes in this species, used the frequencies with which social signals(or acts) were exchanged; LESKES & ACHESON (1971) used the frequencieswith which individual Colobus guereza exchanged social acts; and in ourown study of Theropithecus gelada we used the number of 2 min instan-taneous, scan-samples in which individuals were interacting as well as thefrequencies with which age-sex classes interacted (DUNBAR & DUNBAR,I975).On the other hand, some authors have used spatial criteria, an approachoriginally suggested by CARPENTER (I964) who argued that the metricdistance between any two individuals could be used as an index of thestrength of the "bond" between them. I list a few examples of the bewilderingarray of s.patialmeasures which have been used. KUMMER (I968) used themetric distance between two individuals as well as relative spatial relation-ships (his "nearest-neighbour"method); DEAG (1974) also used a "nearest-neighbour" method in his study of Macaca sylvanus; CHALMERS (I968)recorded the frequencies with which Cercocebus albigena age-sex classeswere within 3 ft of each other, while BRAMBLETT (I972), in a captivitystudy of T. gelada, recorded the rnumberof time intervals in which in-dividuals were within io ft of each other; STRUHSAKER (m67) recO(l-cthe frequencies with which Cercopithecus aethiops individuals spent thenight in the same sleeping subgroups, and DUNBAR & NATHAN (1972)adopted a comparable strategy in the case of Papio papio; ALDRICH-BLAKE



5/22

RESEARCHDESIGN IN OBSERVATIONAL TUDIES OF BEHAVIOUR SI

et al. (I97I) analysed the frequencies with which Papio anubis individualsfollowed each other during linear progressions.

All these different procedures are ostensibly measuring the same thing,namely the strengths of the social relationships between different individualsor classes of individuals. All, explicitly or implicitly, assume that individualswho interact (or sit together) often have a stronger social relationship or"bond" than individuals who do so only rarely. While the assumption thatfrequent interaction reflects a strong "social bond" in the psychological senseremains open to question, we can at least assert that some animals interactor sit together more often than do others.

However, we may justifiably ask whether the different measures whichhave been used provide data which are comparablein form. Do animals whosit together often necessarily interact with each other or groom each otheroften? What difference would it have made to the resulting picture of socialrelationships if the observer in question had used a different measure?There is some indication in the literature that different measures do produceresults which are sometimes radically different. KUMMER (I968), forexample, found that patterns of interaction did not necessarily correspondto patterns of spatial proximity. Likewise, SIMPSON (I973) found thatdifferent measures of the distribution of grooming among adult male chim-panzees did not necessarily correlate. This raises the critica-lquestion as towhether we can legitimately compareidata for different species if they wereobtained in different ways.

Our purpose, then, will be to examine the effect of using different be-havioural parameters and sampling methods to estimate social relationshipsamong a set of animals. Naturally, time and space preclude an attempt toinvestigate all possible methods, and I have chosen those which are eithercommonly used or representative of a group of methods. My aim is not toshow that any one procedure is absolutely better than any other, but simplyto illustrate the effects of choosing one particular approach and to try toshow why this is so.

2. BEHAVIOURAL PARAMETERSAs J. ALTMANNI974) has pointed out, the way in which the research

question is phrased both reflects the kinds of problems in which the observeris interested and determines to a large extent the kinds of data which willbe relevant to that particular question. This in turn may restrict the ways inwhich the behaviour in question can be sampled. Our present concernrequires that we have as many different measures as possible in order to beable to compare them. I have selected three types of relationship to examine.

6



6/22

82 R. I. M. DUNBARThese are (a) the extent to which animals interact with each other, (b) theextent to which they groom each other and (c) the extent to which theyare in a particular spatial relationship. Any given behaviour can be quanti-fied in two quite different ways, namely, either by counting events (fre-quency or rate measures) or by scoring states as a proportion of the totalsample time (durational measures) and these are by no means equivalent(see J. ALTMANN,974). Each of the three behavioural parameters listedabove can be quantified in terms of both frequency and duration. Forconvenience, I shall postpone the discussion of the actual measures used tothe following section. Before doing so, however, the kinds of behaviourwhich are to be sampled in each case must be specified.In the case of (a) above, only non-agonistic "associative" behaviourswill be analysed. Interactions which are agonistic in character will be ex-cluded from consideration on the grounds that they reflect the existence of"negative" relationships, although some authors have not made this distinc-tion. In practice, agonistic interactions are rare compared with non-agonisticinteractions in most species and the amount of distortion due to their in-clusion may be minimal. However, when dealing with individuals rather thanage-sex classes, there is an increased risk of finding individuals who interactrelatively frequently but always in an agonistic manner. To include agonisticinteractions might lead Jto the conclusion that there was a strong socialrelationship between the individuals concerned. Clearly this would be erro-neous since the relationship between them is qualitatively quite different tothat between two individuals who only groom each other, even though theyinteract less frequently.

I do not distinguish between the different kinds of non-agonistic inter-action (e.g. grooming, play, sex, etc.) on the grounds that, at this stage,there is no a priori reason to suppose that social relationships are exclusivelyexpressed in terms of a particular behaviour. To assume that the relativestrengths of social relationships are reflected in the extent to which indi-viduals groom each other would bias against those individuals or classes ofindividuals, such as juveniles, who normally play rather than groom. This isnot to say that grooming relationships themselves might not be of interest,but the analysis of particular types of interaction should be considered asecond order approachonce the general outlines have been established. I have,however, included grooming interactions (b above) as a separate measure,partly because they are commonly used as an index of social relationshipsand partly because grooming is often said to be the most important form of"friendly" behaviour among primates.

With regard to the spatial measure (c above), we shall record the extent



7/22

RESEARCHDESIGN IN OBSERVATIONAL TUDIES OF BEHAVIOUR 83to whichanimalsare withinarm'sreach of eachother.The usual rationalefor this approach is that, since the animalsdo not lnormallymix at random,the fact that two individualsare spatially close reflects the existence ofsome "attraction" etween them.By implication, herefore, it is necessaryto ensure that data of this kind are not collectedat times when there isreason to believe that this assumptiondoes not hold. Such circumstances relikely to include occasionswhen the animalsare moving rapidly (especiallyif fleeing from potential danger) and when the group is highly disturbedby agonisticencounters f a particularlyevere kind.Under such conditions,the animalsmay be assortedmore or less at random ince their main concernmay be to get away from the source of the disturbance ather than whomthey are next to. Both KUMMER (I968) and DEAG (I974), for instance,recordeddata of this kind only from resting troops,and in the presentcasea similarconditionwas imposed.

3. SAMPLING TRATEGIESSamplingmethodshave been discussed n 'detailby J. ALTMANN1974)andI shall heresimplydescribe he variousstrategiesused.In all, seven different measuresor estimatorsof social relationships ave

been taken.These result from different combinations f behaviouralpara-meters and samplingmethodsand are as follows:(I) The numberof social contactsbetween individuals(i.e. the numberof timesthey interacted, rrespective f the numberof social acts exchangedin each interaction);(2) The numberof socialacts exchangedbetweenindividuals;(3) The numberof 30 sec time intervals n whichindividuals nteracted

(One-Zero sampling);(4) The number of 6o sec instantaneousor point samples in whichindividualswere interacting;(5) The numberof groomingboutsexchangedbetweenindividuals;(6) The numberof 6o sec instantaneous r point samples n which indi-vidualswere grooming;(7) The number of 30 sec time intervals in which individuals were

withinarm's reach of each other(One-Zero sampling).The samplingstrategiesand some of the terms used requireexplanationand definition.A "socialcontact"s defined as the exchangeof an uninter-rupted series of social acts between two individuals, being defined asterminatingwhen the individuals concernedcease to exchange such actsand do somethingelse (e.g. feed, move apart,interactwith anotheranimal,etc.). In strategy (i), we simply count the number of such contacts that



8/22

84 R. I. M. DUNBAR

occur and ignore the number of social acts exchanged in any given case. Bydefinition, only interactions involving the exchange of non-agonistic signalswill be considered.

A "social act" refers to any element of the species' signal repertoire whenused in a "face-to-face" interaction between a pair of animals. In strategy(2), we count the number of such acts exchanged between interacting dyads.Individual social acts include grooming, touching, presenting, mounting, thevarious play behaviours, etc. Agonistic acts (threats, aggressive contactbehaviours, submissive behaviours) and signals such as alarm and contactcalls which are not generally directed at a specific individual are discountedin the present context.

A "grooming bout" is defined as a continuous series of grooming actionsby one individual. As in the case of social contacts, a grooming bout isdefined as terminating when the groomer stops performing grooming actionsand does something else (e.g. sits looking around, presents to be groomed,moves away or is groomed by his partner). I include among the actions ofgrooming not only searching through the fur and removing particles, butalso eating things picked from the fur or skin and lipsmacking (see SPARKS,I967). A more detailed discussion of this is given elsewhere (DUNBAR &DUNBAR, I975). in strategy (.), we count the number of individualbouts of grooming exchanged between two individuals (irrespective ofwhich animal is the groomer).

An example may help to clarify the differences between these measures.Supposethe following sequencewas recordedfrom a single subjectdenotedby animal A:A approachesB and (i) presentshis rear to him. B (2) touchesA's rear, (3) peersclosely at it and then (4) briefly mounts him with (5) vigorous lipsmackirig.When

B dismounts,A turns and (6) begins to groom B. After grooming for some time,A stops and (7) presents his side to B; then B (8) grooms A. Later, B stopsgrooming and sits looking around.A moves away a few metres and begins to feed.After feeding for a while, A walks back over to B and (g) begins to groom him.When B stops grooming, A (Io) presents his back to B and B (iI) grooms Aagain. A few minutes later, B walks away and A begins to feed.

In this example, two separate social contacts occurred which were separatedby a clear break in the exchange of social signals between the animals, namelywhen A moved away to feed and B sat looking around between (8) and (9).During these two interactions, four separatebouts of grooming were recorded(at points 6, 8, 9 and i i) and a total of i i social acts were exchanged.

These three strategies (I, 2 and 5) are event frequency (or rate) measuresand the data are properly expressed, even if unstated, as the number ofevents per unit time. The remaining four strategies, on the other hand, are



9/22

RESEARCH DESIGN IN OBSERVATIONAL STUDIES OF BEHAVIOUR 85durational measures, the data being strictly expressed in terms of theproportion f time spentin a given state.

In strategies3 and 7, the samplingprocedure s based on One-Zeroorinterval sampling.The sampleperiod (in this case 30 min) is divided upinto a series of equal time intervals (here 30 sec). So long as the relevantrelationshipoccurs within any given 30 sec interval,a "plus" s scored; ifthe relationshipdoes not occur, a "minus" s scored. Only a single plus isscoredfor a particularpairof interacteesn any given interval, rrespectiveof how manytimes they actually nteracted or sat together)in thatinterval.This samplingstrategy has been referredto by a numberof terms, but Ishall follow J. ALTMANN (i974) in termingit "One-Zero" ampling. (ForpresentpurposesI have consideredOne-Zerosampling o be an estimateofthe proportionof time. Some authors have used a One-Zero samplingstrategy o estimateeventfrequencies,but, as ALTMANN & WAGNER (1970)and FIENBERG (I972) have pointed out, the strategy does not give a truefrequencyestimateand I have refrained from samplingevents in this way.For a fullerdiscussionof this, see, HINDE (I973) and J. ALTMANN (I974).)In strategies3 and 6, the data are also recorded n the form of +'s and-'s, but in this casesamples ake the formof instantaneousamplescarriedout at predeterminedntervals of time (here 6o sec). Every 6o sec wedeterminewhetheror not two individuals reinteractingn theway specified.This strategyhas likewise been referredto by a numberof terms,but forsimplicity shall use the term"point" ampling.

METHODSi. Recording methods

The data were obtained from a set of eleven 30 min samples of individual geladababoons Theropithecuselada)recorded or this purposeduringour I97I-2 studyofthis species in the Simien Mountains National Park, Ethiopia. The recording methodwas so designed that data were simultaneouslyobtainedon each of the seven differentmeasures. In each case, a single subject animal was selected at random from the herd,and his behaviourwas recordedon a checksheetgraduatedinto sixty 30 sec intervals.Everything the subject did (or had done to him) and the age-sex class of his interacteewas recorded n the interval in which it occurred.Persistent behaviourssuch as groomingwere recordedin each interval in which they occurred,the end of a bout being markedby a vertical line behind the last entered symbol. At 6o sec intervals, an instantaneoussample of what the subject was doing (and, where appropriate,with whom) was made.In addition,the exact times of onset and terminationof all groomingbouts were noted tothe nearest sec (with a possibleerror of 2-3 secs) and the age-sex classes of all animalswithin arm's reach of the subject during any part of each 30 sec inerval were recorded.Observationdistances were of the order of 3-IO m, visibility conditions in all casesbeing excellent. Inter-observer reliability when recording samples of this kind wasexceptionally high (95%: see DUNiTBAR& DUNBAR, 1975). Eleven age-sex classeswere distinguishedin this analysis: these are describedin full in DUNBAR & DUNBAR(I975).



10/22

86 R. I. M. DUNBARNo attempt was made to sample either from a single known individual or from indi-viduals of the same age-sex class. Although either of these might have been preferablein terms of realism, we aimed to sample a broad spectrum of all age-sex classes exceptinfants. This has the advantage of including in the sample a variety of individualswhose characteristic nteractionsare likely to differ, hence emphasising any differenceswhich might exist between the measures.

2. Analysis of data.The major analyses of the data will be in terms of the proportional distribution ofsocial interactions (however measured) among the various age-sex classes of interactee.For these purposes, the eleven samples bave been pooled and treated as though theywere obtained from the same individual or class of individuals.It is importantto note,therefore, that no tonclwions concerning the behaziour of gelcda baboons should bedrawn from the data presented below. In analysing the distributionof social interactions,

all triadic or multipartite nteractionshave been treated as a set of dyads.Spearman rank correlation tests have been used to determine whether the varioussXnpling strategies give the same ordinal ranking to the data. These analyses representthe kind of qualitative and semi-quantitativecomparisons that are often made, namelythose of the "A-interacts-more-with-B-than-with-C"ind. However, since a large num-ber of comparisons are made, significance levels cannot be attached directly to theresulting r, values (SIEGEL, 1956). The results of the tests are used simply to showmore clearly the extent to which the individual measures agree in their rankings ofthe data. The overall degree of correlation between the measures is tested for signifi-cance using Kendall's coefficient of concordance.Some direct tests of absolutenumerical comparabilityhave been carried out using theWilcoxon matched-pairs signed-ranks test. In general, however, most numerical com-parisons will be made without the use of statistical tests since the differences aregenerally sufficiently extreme not to require statistical verification. (There is, however,the additional problem that only goodness-of-fit tests are appropriate n some of thesecases, but, since independence f the data cannotbe assumed n six of the seven measures,such tests cannot be used. Only in the case of the frequencyof social contacts measurecan independenctbe demonstrated: or a discussionof this, see Appendix C of DUNBAR &

DUNBAR, 1-975).RESULTS AND DISCUSSION

I. ESTIMATING SOCIAL RELATIONSHIPSThe frequencies with which the sulbjects (all samples pooled) interacted

(or sat with) each of the various age-sex classes as determined by the sevensampling strategies, are given in Table i. In each case, the raw number ofunits of interaction or association with each class is given. Strictly speaking,estimators 3, 4, 6 and 7 should be interpreted in terms of the proportionof the total number of samples taken (330 in the case of strategies 4 and 6,66o in the case of strategies, 3 and 7). However, for simplicity the rawscores are entered into the appropriate columns rather than the proportionof samples.

As a first step, each pair of measures may be compared to determinewhether or not they rank the interactees in the same order. The results ofthe Spearman rank correlationtests are given in Table 2. It can be seen that,



11/22

RESEARCH DESIGN IN OBSERVATIONAL STUDIES OF BEHAVIOUR 87TABLE I

Frequencies wth which subjects "interacted" with the various age-sex classesas determined in different ways (all subjects combined)

Age-sex class of interacteeEstimator Ad X 6yr& syrd! 4yrd Ad y 3yrd 3yr? 2yrd 2yry Yearl Inf Total1.Number of contacts 3 3 I 2 8 II 3 3 2 I0 I 472. Number of acts I9 IO 3 7 47 34 8 I5 8 I7 I I693. One-Zero interacting 34 33 3 I7 99 49 II 39 9 I0 I 3054. Point interacting I5 I5 I 7 44 I9 3 I6 5 I - I265. Grooming bouts 16 9 I 5 38 I8 7 I5 2 3 - II46. Point grooming I5 I5 I 7 44 I7 3 I6 4 - - I227. Spatial measure 45 37 4 Ig III 5I I7 40 2I 67 3 4I5

TABLE 2Comparison of the results given by the various estimators in Table I bySpearman rank correlation test, the values of r8 being given in each case

One-Zero Point Grooming PointContacts Acts interacting interacting bouts groomingActs 0.900One-Zero interacting 0.724 0.8i6Point interacting o.862 0.778 o.986Grooming bouts 0.753 o.88o 0.973 0.932Point grooming 0.5ig 0.7I9 0.941 0.993 0.909Spatial measure o.888 o.96i o.673 o.685 o.8I8 o.607N II in each case.

on the whole, the correlations between the measures are quite high. Of thetwenty-one pairings, thirteen correlate at better than r = o.Soo, while in onlyone case is the value of r, below o.6oo.Such a generally high level of agreement between the seven measuresmight be taken to imply that, at least in the case of the species under study,which strategy is used to estimate overall social relationships is relativelyimmaterial so long as the data are ordinally scaled. Indeed, calculation ofKendall's coef ficient of concordance between the seven measures givesW = 0.840, which indicates a significant degree of agreement between themeasures (x2 = 58.8, p


12/22

88 R. I. M. DUNBARvarious measures, and it can be seen that this class accounts for much ofthe discrepancy in many cases.

These discrepancies imply that, although the measures co-vary, they donot do so monotonically. This can readily be demonstratedby converting theraw scores to percentages of the total number of samples in each case. Theseare given in Table 3 in the form of probabilities, each row summing tounity.

TABLE 3Relative strength of relationship zwitheach age-sex class as determined bythe different estimators given as a probability (based on data in Table i)

Age-sex class of interacteeEstimator Ad d 6yrS 5yrd 4yrd Ad V 3yr& 3yrY 2yrS 2yr? Yearl InfI. Number of contacts .o64 .o64 .02I .043 .170 .234 .o64 .o64 .043 .2I3 .02I2. Number of acts .II2 .o59 .oi8 .041 .278 .20I .047 .o89 .047 IO' .oo63. One-Zero interacting iii .io8 .OIO .o56 .325 .i6i .036 .128 .030 .033 .0034. Point interacting .II9 .IIg *oo8 .056 .349 I151 .024 .127 .040 .oo8 .ooo5. Grooming bouts .140 .079 .009 .044 .333 I58 .o6i .132 .oi8 .o26 .ooo6. Point grooming .123 .I23 .oo8 .057 .36I J.39 .025 .13I .033 .000 .0007. Spatial measure .I08 .o89 .010 .046 .267 .123 .041 .o96 .05 I .i6i .007

(In the case of the four durational measures, there is a major difficultyin this respect. In order to transform the raw data into a form comparableto that for the three frequency measures, the total number of samples takencannot be used as the denominator since the rows would not then sum tounity. On the other hand, since a subject could be interacting or sittingwith more than one animal during any given sample, the number of intervalsor point samples in which he was interacting cannot be used either, sincethe row totals would then be greater than unity. It has therefore beennecessary to use the sum of the appropriaterow in Table i as the denominatorin each case. This raises some difficulty over the precise interpretation ofthe probabilities, since they no longer represent either the proportion oftime or the proportion of interaction time that the subject spent interacting.Perhaps the best interpretation to place on these probabilities is one in termsof the proportional distribution of the subject's social contacts or spatialassociations.)

Consideration of these data clearly illustrates the different pictures ofgelada social relationships which would result from use of the differentmeasures. Frequencies of social contact, for instance, would suggest that6.4% of the subject's social activiitywas with adult males and 2I.3% withyearlings. On the other hand, the number of acts exchanged would give



13/22

RESEARCHDESIGN IN OBSERVATIONAL TUDIES OF BEHAVIOUR 89values of I1.2% and io.i% respectively, while the number of groomingbouts exchanged would give values of I4.0% and 2.6%. Even more extremewould be the number of point samples spent interacting: I I.9% as againsto.8%.

Hence, although we might be able to compare different animal's socialrelationships using ordinally scaled data with a reasonably high expectationof concordance between different measures, we cannot necessarily make amore direct comparison between the relative probabilities of interaction.

2. DIFFERENCES BETWEEN MEASURESThese differences draw attention to one important aspect of social rela-

tionships. In a complex social system, "social bonds" (in the psychologicalsense) may be expressed in a variety of different ways. Hence, whicheverindex we select to estimate the strengths of these relationships will providedata which reflect only certain facets of this system.

The mean number of acts exchanged per social contact, for instance, isnot the same for all classes of interactee. Interactions with some classesaverage five or six acts per contact, while those with others average onlyone or two (see Table I). Hence, by taking frequencies of social contact asan index of the strength of the relationships between individuals, the dataare biassed in favour of those classes who exchange fewest acts per socialcontact, while the reverse is the case if the number of acts exchanged isused as an index. Likewise, consideration only of grooming bouts will biasthe results against those individuals or classes among whom grooming is arare form of social expression. Conversely, it would tend to overemphasiserelationships between classes who do little else but groom when they dointeract, even though they may not be especially sociable.Again, if we record the proportion of time spent interacting, we mayfind that similar biasses emerge due to the fact that interactions with someclasses tend to be of shorter duration than those with others. Not only maythese measures differ from the distribution of social contacts, but theymay also differ from the distribution of acts if animals tend to exchangedifferent kinds of signals in their interactions with different individuals.

The strongest correlations obtained were between the three estimatorswhich measured the proportion of time individuals spent interacting andgrooming (strategies 3, 4 and 6). This can be attributed to the fact that,in the present case, a large proportion of the social acts exchanged weregrooming (see Table I). Since grooming tends to account for a great dealmore time than other behaviours, it is not too surprising to find that ameasure of the proportion of time spent interacting correlates well with the



14/22

90 R. I. M. DUNBAR

proportion of time spent grooming. This may not, however, necessarily beso in all cases, since the interactions of some classes, such as juveniles,tend to involve a great deal more play than grooming, particularly whenthey are interacting with peers. In this case, the assumption that groomingprovides an accurate estimate of the strengths of social relationships maybe seriously misleading.

Similarly, a measure of spatial association may lead to a rather differentpicture of the relationships among a group of animals than an interactionalmeasure would give. The reason for this is fairly obvious: an animalmay spend a great deal of time close to certain individuals, but interactwith them only rarely. This could be due to one of several reasons. Infants,for instance, might often be found near adult males because their motherscommonly associate with this age-sex class: nonetheless, infants mightinteract with adult males only rarely. The assumption that there is anykind of "social bond" between adult males and infants purely on the basisof frequent spatial proximity would be erroneous, since the apparent relation-ship is in reality a consequence of the strong relationship between malesand females. On the other hand, some animals may actively seek out theproximity of certain individuals, but be unable to interact with them as oftenas they might wish. KUMMER (I968) found such a relationship betweenjuvenile females and adult males in Papio hamcadryas,and similar resultswere obtained in the case of gelada baboons (DUNBAR & DUNBAR, I975).Thus, in this case, spatial proximity might reflect the juvenile's choice ofpreferred social partner rather more closely than frequencies of interaction.A third possible interpretation might be that spatial proximity as a wholeis irrelevant. Animals might mix more or less at random when not inter-acting, moving through the group to seek out their preferred social partnersas and when they wished to interact. There was evidence to this effect amongcertain age-sex classes of gelada (DUNBAR & DUNBAR, 1975). In thissituation, spatial proximity would be a useful index of social preferencesonly when the animals were actually interacting. It remains an untestedpossibility that in many species neighbourhood relationships among a groupof animals change according to the nature of the predominant activity. It isnot unreasonable to suppose, for instance, that an animal's neighbours mightapproach a random distribution more often when the group is feeding thanwhen it is socialising.

3. DIFFERENCESBETWEENFREQUENCYAND DURATIONThe differences between frequency and duration can be illustrated more

clearly by examining the relationship between the frequency of grooming



15/22

RESEARCHDESIGN IN OBSERVATIONAL TUDIES OF BEHAVIOUR 9I

bouts and the amount of time spent grooming. These data are given foreach subject in Table 4, together with the mean duration of grooming boutsin each case.

The number of grooming bouts correlates significantly with the totalamount of time spent grooming (r, = o.872, po.io, two-tailed, N = II), indica-ting that the most frequent groomers are not necessarily the most persistent.Not only this, but animals may also differ in the length of time for whichthey are prepared to groom different individuals. SIMPSON (I973), forinstance, found this to be the case among adult male chimpanzees.

Consequently, when choosing an appropriate measure for estimating theextent to which individuals interact, we should bear in mind the differencesbetween frequency and duration. In an absolute sense, neither can be con-sidered better than the other, since each examines a different aspect of thesituation, although in any given set of circumstances one may be moreappropriate in terms of the investigator's inlterestsand the requirements ofthe research question. It may often be valuable, however, to record bothkinds of data.



16/22

92 R. I. M. DUNBAR

4. DIFFERENCES BETWEEN ONE-ZEROAND POINT SAMPLINGAlthough I do not intend to discuss the merits and demer'itsof the various

sampling methods in any detail, it does sieem worthwhile examining someof the disadvantages of One-Zero sampling insofar as we have data availableto do this.

Throughout the preceeding analyses, I have used One-Zero sampling asa durational measure. Strictly speaking, One-Zero sampling is so designedthat it incorporates features of both frequency and duration. Ostensibly, thestrategy would sleem to have much to recommend it since it may permit theobserver to record event frequencies while at the same time obtaining dataon duration. The number of studies which have used this particular methodattests to this. Regrettably, however, the method leaves much to be desiredsince considerable error may be incorporated into one or both kinds of data.A more detailed discussion of One-Zero sampling is given by J. ALTAIANN(I974). In the presen't context, I sihall consider the differences betweenOne-Zero and point sampling in the extent to which they provide an accurateestimate of the proportion of time spent doing something. I shall again usethe proportion of time the subjects spent grooming. That I consider onlythis aspect of One-Zero sampling should not be taken to imply that I con-sider it a suitable strategy for recording event frequencies. I have alreadyremarked that this strategy does not give a true estimate of the frequencyof event occurrence (see also ALTMANN & WAGNER, 1970; HINDE, I973;J. ALTMANN, I974).

Table 5 gives the proportion of time each subject spent grooming basedon the actual number of seconds (from Table 4), together with estimatesof this obtained by One-Zero and point sampling with different time bases.In the final column is given ithe overall proportion of time spent groomingby all subjects combined.

Considerationof the data shows that point sampling provides a consistentlymore accurate estimate of the true function than does One-Zero sampling,irrespective of the length of the sample interval. Comparison of the truefunction and the estimated function in each case using the Wilcoxonmatched-pairs signed-ranks test shows that none of the point sample esti-mates differ significantly from the true function (T not below 2I.0,p>o.o5 in each case), whereas all the One-Zero samples are significantlydifferent (T = o, p


17/22

RESEARCHDESIGN IN OBSERVATIONAL TUDIES OF BEHAVIOUR 93TABLE 5

Estimates of the proportionof time individualsubjects spent grooming(basedon One-Zeroand pointsamplingwith different timebases) compared

with theactualproportionof time spent groomingSubjectEstimates I 2 3 4 5 6 7 8 9 I0 II Overall %Number of seconds 58.3 39.8 39.1 37.4 34.6 26.I i6.6 I3.4 6.i i.i o.8 24.8One-Zero sampling based on intervals of:

5 sec 63.1 41-4 40.8 38.9 356 28.1 i8.i 14.4 6.9 I.3 I.I 26.3I5 sec 69.2 45.0 43.3 42.5 37.5 33-3 21.7 I6.7 8.3 I.7 o.8 29.I30 sec 75.0 5I.7 46-6 45.0 40.0 38.3 26.7 20.0 I0,0 3.3 I.7 32.66o sec 8o.o 6o.0 56.7 50.0 50.0 43.3 30.0 23.3 I6.7 3.3 3.3 37.9

I20 sec 86.7 66.7 73.3 53.3 53.3 66.7 46.7 26.7 20.0 6.7 6.7 46.IPoint sampling based on intervals of:

15 sec 67.5 39.2 38.3 37.5 34.2 25.8 I7.5 I3.3 5.8 o.8 o.8 25.530 sec 6o.0 4I.7 35.0 36.7 33.3 25.0 I6.7 II.7 5.0 1.7 0.0 24.26o sec 53.3 43.3 36.7 36.7 36.7 20.0 I0.0 I3.3 6.7 0.0 0.0 23.9

I20 sec 46-7 46.7 46.7 33.3 40.0 26.7 I3.3 I3.3 6.7 0.0 0.0 24.8the extent of the over-estimationbecoming greater as the length of thesample intervalincreases.One-Zerosamplingyields a reasonablyaccurateestimate only when the sample interval is very small relative to the typicalduration of the behaviour in question.

The cause of this is not difficult to see. In One-Zero sampling, werecord simply whether or not the behaviour in question occurs, irrespectiveof how much of the interval it actually occupies. Since the moments of onsetand termination of any given bout are most likely to occur in the middle ofan interval, each bout will be credited with lasting up to two full time unitslonger than it did. Clearly, the longer the length of the sample interval,the greater will be the over-estimation. The extreme case will occur whenthe sample interval is the same as the sample period (in the present case30 min). An animal who spent only IO sec grooming during this periodwould in this case be credited with spending ioo%v of his time grooming,whereas in fact he spent less than o.6%. It can readily be seen that anaccurate estimate of the proportion of time spent grooming (or indeedanything else) can only be obtained when the sample interval approximatesthe actual unit of measurement (i.e. one second). In this event, it wouldobviously be a great deal easier to record transition times and calculate thedurations from these. As J. ALTMANN (I974) points out, One-Zero sam-pling can only provide an upper limit on the proportion of time accountedfor by a given state. As we have seen, this upper limit may often be absurdlyhigh: at best, it will simply be misleading.



18/22

94 R. I. M. DUNBAR

This may be particularly critical when comparing species or individualswho differ in the mean duration of their grooming bouts. We cannotnecessarily assume that the extent of the over-estimation resulting fromthe use of One-Zero sampling will be consistent in each case. This can beseen from Table 5. Although, for instance, the absolute difference betweenthe true function and the 30 sec One-Zero estimate correlates significantlywith the actual amountof time spent grooming(r, = 0.79I, p


19/22

RESEARCH DESIGN IN OBSERVATIONAL STUDIES OF BEHAVIOUR 95choice of samplingmethodsince data which incorporate amplingbiassesofunknownproportions annotbe rescuedby any amountof statisticalwizar(l-ry. In point of fact, the five decisionpoints are not independentof eachother, for the options which are open to us at point 5 may restrict the kinidsof decisions we can make at points 3 and 2, and vice versa. Suppose wewished to investigate the relationship between dominance and groomingrelationships. To do so we have to be able to compare two behavioural para-meters, and before we can even decide what we are going to record, letalone how to quantify it, we are confronted with the spectre of a stage 5decision. The availability of suitable statistical tests may limit the kinds ofsampling strategies we can use, and these in turn may limit the behaviourssampled. Normally, of course, it is ilogical to proceed in the reverse orderand make the necessary changes to previous decisions as we come up againstinsurmountable problems. However, the point serves to illustrate the factthat the constraints may come not only from above, but also from below.

Our main concern in this paper has been with stage 2 decisions, namelychoosing the appropriate behavioural parameters to quantify in order toanswer the research question posed. The critical problem at this point liesin deciding which particular behaviours are relevant to the question in hand.In many cases, the inherent difficulties are not at all obvious. Often it mayseem that the topic we wish to investigate, for instance grooming relation-ships, leaves little room for doubt concerning the belhaviours which oughtto be quantified. Grooming, we suppose, is simply grooming, but there areat least three quite distinct problems here.

Firstly, we have to decide precisely what should be included in the activityof grooming, i.e. we have to arrive at an operational definition of grooming.Preferably, our definition should be related to those used by other authors,since the use of different definitions for the same term (especially whenunstated) may lead to confusion in the literature.

The second problem concerns external validity. Are the behaviours wechoose to quantify relevant to the question posed? In cases where theresearch question is very precisely defined there may be little problem here.However, where the research question is of a rather vague nature, the in-vestigator runs the considerable risk of recording data which are inappro-priate. As we have seen, there are at least three types of behaviour whichmight be considered relevant to a study of "social relationships", and whichwe choose to quantify will affect the results we get. Strictly speaking, eachof these three parameters answers a different question. Analysis of groom-ing relationships answers the question "How do the animals distribute theirgrooming?", while an analysis of spatial relationships answers the question



20/22

96 R. I. M. DUNBAR"Do animals prefer to be spatially closer to some individuals than to others ?".Neither necessarily tells us anything about the way in which the animalsdistribute their social contacts.

Thirdly, and finally, there are two quite distinct ways of quantifyingany given behaviour, namely by frequency of occurrence (rate measures)and by duration (proportion of time measures). Which we choose may makea considerable difference both to the kinds of statistical analyses we canmake and to the kinds of statements we can make about the animals'behaviour. We cannot necessarily assume, for instance, that grooming rela-tionships are completely described by an analysis of the frequencies withwhich individuals groom each other. To the contrary, the lengths of timefor which inidividuals are prepared to groom may be at least as interestingand probably just as significant.

In conclusion, then, four general points can be made relating to thechoice of behavioural parameters to quantify. Firstly, the research questionneeds fine definition. If the problem is not clearly defined, we may findourselves collecting data, which although admirable in themselves, do notanswer the question we suppose ourselves to be asking. Secondly, the dictumthat ethological studies should begin with a substantial period of naturalisticobservation before any attempts are made to quantify aspects of a socio-biological system must be emphasised. Only in this way can the externalvalidity of the research project be guaranteed. This applies both to thechoice of a suitable problem to study (and, hence, an adequate definitionof the research question) and to the choice of the corre'ctbehavioural para-meters to quantify. Thirdly, the fact that there are often several behaviouralparameters relevant to a given research topic suggests that it may often beuseful to examine more than one facet of the problem. Finally, the inter-dependence of the five decisions makes it necessary to bear in mind, evenwhen defining the research problem, aspects which might seem at first sightto be irrelevant at such an early stage. Not only must the availability ofsuitable statistical procedures be borne in mind, but so must the limitationsof the various sampling strategies and the exigencies of the actual fieldsituation (including the species being studied). Certain sampling strategiesmay be too difficult or too time-consuming to employ in a particular fieldsituation, with the result that the kinds of data which can be quantified maybe restricted. This, in turn, may limit the kinds of questions we can askabout the animals' behaviour.

SUMMARYThis paper has been concerned with examining the effects of choosing differentbehaviouralparameters o investigate a particularresearchproblem.Social relationships



21/22

RESEARCHDESIGN IN OBSERVATIONAL TUDIES OF BEHAVIOUR 97among a set of animals were estimated in seven different ways. Three behaviouralparameters (social contacts, grooming and spatial associations) were selected and anumber of sampling strategies were used to quantify them. Although the measuresyielded results which correlatedsignificantly, there were marked numericaldifferencesbetween them due to the fact that they measureddifferent aspectsof a complexbiologicalsystem. The differences between frequency and durationalmeasures were investigatedand the accuracy of two common durational measures was determined. One-Zerosampling was found to provide a poor estimate of the proportionof time spent grooming,whereas instantaneousor point sampling always gave a reliable estimate. Finally, therelationshipsbetween the different stages of research design were discussed.Emphasiswas laid on four aspects of choosing parameters o quantify,namely, the precise delinea-tion of the research problem, the validity of the parameters n relation to the researchquestion, the differences between the various measuresand the requirementsand limita-tions of the other key features of research design.

REFERENCESALDRICH-BLAKE,. P. G., BUNN, T. K., DUNBAR,R. I. M. & HEADLEY,P. M. (1971).Observations on baboons, Papio anubis, in an arid region in Ethiopia. - Foliaprimatol. I5, P. I-35.ALTMANN, J. (1974). Observationalstudy of behaviour: samplingmethods.- Behaviour

48, p. I-4I.ALTMANN, S. A. (I968). Sociobiology of rhesus monkeys. III. The basic communicationnetwork. - Behaviour 32, P. I7-32.& WAGNER, S. S. (1970). Estimating rates of behaviourfrom Hansen frequencies.- Primates ii, p. I8I-I83.BRAMBLETT,. A. (1970). Coalitions among gelada baboons. - Primates II, P. 327-334.CARPENTER, . R. (I964). Social behaviour of non-human primates. - In: NaturalisticBehaviour of Non-Human Primates, C. R. CARPENTEREd), p. 365-385. Pennsyl-vania State University Press, Pennsylvania.CHALMERS,N. R. (I968). The social behaviour of free living mangabeysin Uganda.-Folia primatol. 8, P. 263-28I.DEAG, J. M. (I974). A study of the social behaviour and ecology of the wild BarbaryMacaque,Macaca sylvanus L. - Ph.D. Thesis, University of Bristol.DUNBAR, R. I. M. & DUNBAR E. P. (1975). Social dynamics of gelada baboons.Contrb.Primatol. Vol. 6. Basel: Karger.& NATHAN, M. F. (1972). Social organizationof the Guineababoon, Papio papio.

- Folia primatol. I7, P. 32I-334.FIENBERG, S. E. (1972). On the use of Hansen frequencies for estimating rates ofbehaviour. - Primates I3, P. 323-325.HINDE, R. A. (I973). On the design of checksheets. - Primates I4, P. 393-406.HUTT, S. J. & HUTT, C. (1970). Direct observation and measurement of behaviour. -Thomas, Springfield.KUMMER, H. (I968). Social organizationof Hamadryas Baboons. - Karger, Basel.LESKES, A. & ACHESON, N. H. (197I). Social organization of a free-ranging troop ofblack-and-whitecolobus monkeys (Colobus abyssinicus). - Proc. 3rd. int. Congr.Primatol. 3, P. 22-3I. Karger, Basel.SADE, D. S. (I965). Some aspects of parent-offspring and sibling relations in a groupof rhesus monkeys, with a discussion of grooming. - Amer. J. phys. Anthrop. 23,

P. I-I7.(I972). Sociometrics of Macaca mulatta. I. Linkages and cliques in groomingmatrices. - Folia primatol. i8, p. I96-223.SIEGEL, S. (1956). Nonparametricstatistics for the behaviouralsciences. - McGraw-Hill, New York.

7



22/22

98 R. I. M. DUNBARSIMPSON,M. J. A. (I973). The social grooming of male chimpanzees. In: ComparativeEcology and Behaviourof Primates, R. P. MICHAEL& J. H. CROOK Eds), p. 4II-

505. Academic Press, London.SPARKS, J. (i967). Allogrooming in primates: a review. - In: Primate Ethology,D. MORRISEd), p. I48-I75. Weidenfield & Nicholson, London.STRUHSAKER,T. T. (I967). Social structure among vervet monkeys (Cercopithecus

aethiops). - Behaviour 29, p. 83-I21.RP-SUMf

Dans cette note, on examine les effets du choix des comportementsdivers pourl'investigation quantitative d'un probleme socio-biologique particulier. Les rapportssociaux entre un groupe d'animaux ont ete determines par sept methodes differentes.Trois comportements interactions sociales, episodes du lechage et associations spatiales)ont ete choisis et quelquesmethodes ont ete employees afin de les quantifier. Quoiqueles methodesdonnent des resultatsqui s'accordent,on releve entre eux quelquesdifferen-ces numeriquesparce que chaque methode considere l'aspect different d'un systemebiologiquecomplique.On examine les differences entre les mesures de frequence et lesmesures de duree, et on determine a precision de deux mesures de la duree.La methode,,One Zero" donne une estimation imprecise de la proportion du temps que le sujetpasse a lecher, mais la methode instanitanee("point sampling") donne une estimationprecise dans tous les cas. Enfin, on discute des relations entre les aspects divers de lam'thode de recherches.On souligne quatreaspects du choix des parametres"aquantifier,a savoir la definition precise du probleme des recherches, la validite des parametresen relationavec la question posee, les differences entre les diverses mesures et les exigen-ces et les limitations des autres caracteristiquesdes recherchesentreprises.

Documents

Dunbar Some Aspects of Research Design and Their Implications in the Observational Study of Behaviour