Correlation, Causation, Motivation, and Second LanguageAcquisition

R. C. GARDNER

Department of PsychologyUniversity of Western Ontario

CPA Award for Distinguished Contribution to Psychol-ogy in Education and Training (1999) —Prix de la SCP pour contribution remarquable a1'education et la formation en psychologic (1999)

AbstractIt is well known to anyone who has taken even an introduc-tory course in statistics that "correlation does not meancausation." This is a truism! The present article examinesfour truisms about correlation, demonstrating that they arenot always true. For example, under certain conditions,correlation can imply causation, though these conditionsare seldom satisfied in most applications. Nonetheless,there are many of us who are interested in investigatingindividual differences, and in making inferences of thetype that this individual difference variable is related to,mediates, moderates, or even causes or influences thatindividual difference variable. Generally speaking, theanalytic procedures we use involve the correlation coeffi-cient in one form or another. I propose four steps thatresearchers can follow to accumulate evidence that in-creases one's confidence in the validity of a particularcausal model. These steps are illustrated by reviewingresearch on individual differences in second languageacquisition. This approach is not conclusive, of course, butit does force one to examine the implications of the model,thus leading to further insights and research. Although thefocus here is on second language acquisition, the general-izations apply to other areas of research that are concernedwith individual differences.

When I was told that I was to receive the CPA Educationand Training Award and was to give a talk in conjunctionwith it, I thought long and hard about a possible topic.My teaching interests involve statistics and data analysis,and my research interests are concerned primarily withthe role of attitudes and motivation in second languageacquisition. I decided, therefore, to focus attention on a

Canadian Psychology/Psychologic canadienne, 41:1

statistical and conceptual issue that has troubled me overthe years, and to discuss how I have resolved the issue inmy own mind in the context of my research interests.The issue is that of the correlation coefficient andcausation. To some, this is a nonissue. You cannot infercausation from correlation. Case closed!

To those interested in individual differences, however,such a fatalistic conclusion is tantamount to concludingthat there is no possible way of ever drawing a causalinference based on individual differences. One approachis to accept the canon that correlation does not implycausation, then go on to talk about prediction as op-posed to causation (though as we shall see this stillimplies causation), and rely on causal (i.e., structuralequation) modelling, and the like. The point is thatindividual difference research involves covariation, andregardless of which analytical procedure one adopts (i.e.,multiple regression, factor analysis, discriminant func-tion analysis, or even structural equation modelling,etc.), the basic statistic involves co-relation in one formor another. In the end, many of us believe that we haveidentified causal associations, even though we willconcede that other interpretations are possible. That is,we believe that personality causes, or accounts for somebehaviours, intelligence plays a role in academic achieve-ment, anxiety disrupts performance, etc.

The solution I propose is to direct attention not somuch to the relationship but to the underlying process,accepting the causal interpretation that seems mostappropriate to the relationship and then expanding theimplications, continually refining and evaluating them inthe context of a research program. This is similar to thenotion of construct validity, but more inclusive since thefocus is not so much on the validity of a test or measure,but rather the elaboration of a conceptual model that isbased on research sometimes using different instrumentsin different contexts. The focus in this instance is on thevalidity of the causal hypothesis explaining the relation-ships among a series of variables. I will attempt to illus-trate this by considering our research program onattitudes and motivation in second language acquisition.Any one study is correlational, but when a number ofstudies, viewing the area from a number of perspectives,produce compatible results, this increases the probabilitythat the presumed causal sequence is valid.

The purpose of this presentation, therefore, is todiscuss the issue of inferring causation from correlation.I approach this duly with fear and trepidation because

Correlation, Causation, and Motivation 11

we all know that you cannot infer causation from correla-tion. This is a truism. Consider, however, the followingfour truisms about correlation. Each is true, to be sure,but the point is that they are not always true, at least inexecution.

THE FOUR TRUISMS

1. The Pearson product moment correlation coefficient variesfrom -1 to +1. We all know this. We also know that oneformula that defines the correlation coefficient is:

and that the variance of a standard score is defined as:

= 1.0

Thus, die only way r^ can be equal to 1.0 is if Zy = zx,and the only way this can happen is if the two distribu-tions are identical in standard score form. That is, theymust have identical shapes, even diough diey need nothave identical means or standard deviations in their rawscore form. Similarly, the only way two variables can havea correlation of - 1 is when the two variables have mirror-image distributions such diat each Zy =- zx. Thus, diegeneralization about die limits of r being ± 1.0 is alimited one. It is true only when die two distributions aresymmetrical (not necessarily Normal) and identical inshape. In preparation for this presentation, I generatedtwo random exponential distributions, each widi 2000observations. One was positively skewed widi a highdegree of kurtosis (2.986 and 15.113, respectively), andthe odier was negatively skewed widi a slighdy smallerlevel of kurtosis (-2.624 and 10.117, respectively). Ireordered die values so dial diere was a maximumpositive correlation between die two variables. This valuewas only .44. Thus, depending on die nature of dieunderlying distributions, diere can be quite an influenceon die maximum value that can be obtained. We shouldalways, therefore, consider die distributions underlyingdie variables we are investigating.

2. When the null hypothesis of independence between twovariables, X and Y, is true, the value of the correlation in thepopulation is 0. This is true, but it sure doesn't seem to be.Clearly when you have two variables diat are independ-ent in die population, die population correlation is 0,but diere would be a sampling distribution of correlationcoefficients for a given sample size diat would tend to benormally distributed around 0. And as a consequence, if

1 These definitions are based on considering the standard devia-tion as a biased estimate. If an unbiased estimate is used, n wouldbe replaced with (n-1), but the generalization would still apply.

die null hypothesis is true, 5% of die sample correlationsare larger in magnitude dian would be expected on diebasis of chance if a two-tailed Type I error rate of .05 hasbeen adopted for die test statistic. This has even beenshown to be true regardless of die shapes of die distribu-tions of die variables being correlated (see Havlicek &Peterson, 1977). There really is nodiing to argue widihere. The problem is widi die application of the statistic.

When we are interested in determining die correla-tion between two variables, X and Y, die first tiling wehave to do is to measure die two variables. This is not aminor point, and my contention is that we must alwaysbe aware of die distinction between die "variables" ondie one hand and die "measurement of die variables" ondie other. In traditional test dieory, die measurement ofa variable can be considered to be composed of die truescore plus error, where die error is defined as anydeviation of die measurement from die true value (cf.Nunnally, 1978, p. 201). In traditional test dieory, thiserror is considered to be random. When considering twovariables (i.e., X and Y), however, it is reasonable toconsider die measurement of both variables to becomposed of die respective true scores, plus randomerror plus error common to die two measures. That isdie two measures X' and Y' can be seen to be as follows:

X' - T + « + <> Y1 - T + f + f~ Y X X(Y} ~ Y Y Y( X)

where ex(Y) and ey(X) are conceptualized as error commonto die two assessments.

Given this conceptualization, it is quite possible diattwo variables, X and Y, might be independent, but diattheir measures X' and Y' may not be independent. Therecould be a number of reasons for this, but for the sake ofsimplicity, we can lump all of diese under die termcommon measurement error as distinct from randommeasurement error. That is, to die extent diat twomeasures have common sources of measurement error,die measures will be correlated even when die variablesthemselves are not.

The problem, dien, is not diat when die null hypothe-sis is true die value of die correlation in die populationis not 0, but radier diat often because of die operationswe perform in our investigation, die null hypodiesis widirespect to our measures is not true. Most often when wecompute a correlation coefficient between two variables,we take observations on both variables. The more similar

2 The notion of more than one component of error in any mea-surement operation has been postulated before, as for example inLee Cronbach's Generalizability Theory (though not for the samepurpose or in the same form as described here) (see Shavelson &Webb, 1991 for an introduction to Generalizability Theory). Thesimple form proposed here is meant only to highlight the pointthat the measurements of two independent variables can nonethe-less be correlated.

12 Gardner

the nature of the measurement operations for these twovariables, the greater the possibility of correlated mea-surement error, and thus the greater the opportunity ofrejecting the null hypothesis when the null hypothesiswith respect to the variables, as opposed to the measuredform of these variables, is true. We should always, there-fore, consider the possibility of correlated measurementerror.

3. Given a large enough sample size the correlation will besignificant. This is obviously an extension of item number2. The point is, that if the null hypothesis is true that thepopulation correlation is 0, dien the probability ofrejecting the null hypothesis in any study of this relation-ship is .05 (at the .05 level), regardless of sample size.This isn't simply a mathematical axiom, a logical argu-ment, or something that can occur only in Monte Carlosimulations, as has been claimed in the literature (Har-ris, 1997; Hunter, 1997; Schmidt, 1996). It is true bydefinition.

Of course, if the null hypothesis is not true, then thistruism is true. And this is an instance of power. If thepopulation correlation is not 0, then as sample sizeincreases, it is more likely that we will reject the nullhypothesis. Thus, considering the point made above, ifthere are common measurement errors resulting in acorrelation of .10 in the population of measures (whenthe correlation for the underlying variables is 0), thenwith a sample size of 1000, the probability of rejectingthe null hypothesis is .89 with a two-tailed Type I errorrate set at .05 (Cohen, 1988, p. 93). For a sample size of100, the probability of rejecting the null hypothesis isonly .17. That is, with large sample sizes, we might wellobtain significant correlations because of commonmeasurement error even when the two variables underinvestigation are independent. Thus, using a largesample size does not necessarily uncover meaningfulassociations.

4. Correlation doesn't mean causation. It is taught in allintroductory statistics classes that correlation doesn'timply causation, and I certainly don't mean to startsuggesting otherwise. On the other hand, causation doesresult in covariation. This is a basic principle underlyingthe experimental method. In research, causation meansdiat variation in one variable, the dependent variable,can be attributed to variation in another variable, theindependent variable. Thus, if we were to conduct astudy in which we randomly assigned individuals to two(or more) groups, administered different treatments tothe groups, and made assessments of the participants, wewould attribute any differences we obtained on thedependent variable to the treatments administered. Wewould state that the differences in the treatments caused

the differences in the dependent variables. That is,causation means simply that variation in one variable isresponsible for variation in another.

It isn't so much that correlation does not implycausation as it is that correlation is seldom used in thecontext of random assignment to conditions. Correlationcan imply causation under the right conditions. Ifparticipants were randomly assigned to a continuousindependent variable (X), and observations were madeon another variable (Y), we could calculate the correla-tion between X and Y. Moreover, we could describe thenature of the causal relationship in terms of a regressionequation. If the relationship were simple we could makeuse of a linear regression equation, but if the relation-ship were more complex, we could make use of polyno-mial regression, or some other means of describing thecausal relationship between the two variables.

CORRELATION AND THE INFERENCE OF CAUSATIONTo elaborate on the issues involved in inferring causationfrom correlation, consider the following scenario.Assume that we were interested in asking the question,"Does amount of alcohol consumed in the afternooninfluence how much one sleeps that night?" We couldconduct such a study as follows.

A random sample of individuals could be invited tothe laboratory, and asked to participate in a study. Aftertheir informed consent was obtained, the experimentercould randomly assign them to drink a given number ofounces of alcohol. Participants could then be kept in thelaboratory for the rest of the day and that night, andcareful records could be kept on how long they slept. (Ifthe researcher wanted to eliminate the effects of anumber of variables associated with social interaction,participants could be tested individually, or isolated, etc.On the other hand, if the researcher was interested inallowing social factors to be part of the variation in thestudy, the participants could be allowed to socialize. Thatis, the researcher could control or not a number of othervariables, depending on the purpose of the study.) Theimportant point is that if the researcher obtained asignificant correlation between these two variables, itwould be perfectly reasonable to conclude that thealcohol level caused the amount of sleep. This is becausethe level of alcohol each participant ingested was ran-domly determined.

Any variation on preingestion variables would berandomly associated with level of alcohol consumed andthus could not be used to explain the relationshipobtained. Any postingestion variables would either berandom or dependent on the level of alcohol consumedby that individual and thus attributable to the level ofalcohol consumed by that individual. A postingestionvariable might be identified as a mediator variable (cf.

Correlation, Causation, and Motivation 13

16

14

12

10

W•5 8e3 6

8 10 12

Ounces of Alcohol

Figure 1. Regression of Number of Hours of Sleep on Amount ofAlcohol Consumed

Baron & Kenny, 1986), but this does not discount thecausal influence of the level of alcohol. Moreover, a testof mediation is based on co-relation, thus all of the issuesdiscussed above apply equally to mediation.

Note that having concluded that level of alcoholconsumption is responsible for the amount of sleep doesnot require that the correlation be overly high. We havealready seen that the shape of the distributions of thevariables can influence the values of the correlationobtained. Moreover, diere are many factors that influ-ence amount of sleep, and there are large individualdifferences in reactions to alcohol, tolerance to alcohol,etc.

Figure 1 presents a scatter plot of 100 pairs of observa-tions that might be obtained in such a study. It shows therelation between two variables, the amount of alcoholadministered, and the amount of sleep experienced.Both variables are relatively normally distributed, withmeans and standard deviations of 3.89 and 2.14 forAmount of Alcohol Consumed, and 8.52 and 2.94 forHours of Sleep. These data are fictitious and are asample of observations taken from an infinite populationfor which the correlation was .30. As can be seen, thereis a positive relation between the two variables. As theamount of alcohol ingested increases, the length of sleepincreases. In fact, the correlation between these twovariables was calculated to be .36. With 98 degrees offreedom, this correlation is significant at the .0002 level.It might also be noted that, assuming that the true valueof the correlation in the population is .30, the power(i.e., die probability of rejecting the null hypothesis atthe .05 level, two-tailed) is .86 (Cohen, p. 93). That is, ifthe assumption that the correlation in the population is.30 is correct, 86% of die time a correlation based on asample of 100 observations would be significant at the.05 level.

The question can be asked, therefore, whether suchresults would indicate that Amount of Alcohol consumedcaused the Amount of Sleep if this were a true experi-ment. My contention (allowing for the possibility of aType I error, of course) is yes! And, I could be quiteconfident that other researchers would similarly obtainsignificant results if they were to replicate the study.Since the number of ounces of alcohol each participantconsumed was randomly determined and administeredby the researcher, it is logical to conclude that theamount of sleep is dependent on the number of ouncesof alcohol consumed. Note too, that in this study, thereis no way that you could argue that hours of sleep wereresponsible for the amount of alcohol consumed. Thecausal effect is unidirectional.

In the present instance, the model is a simple linearone, and we could describe the nature of the causalrelation between the two variables by calculating theregression of Amount of Sleep on Amount of AlcoholConsumed. If we wished to focus only on a linear rela-tionship, the equation would be of the form:

Y1 = a + bX

For die present data, die equation is:

Y1 = 6.58 + .498X

That is, this regression equation is a causal descriptionof die nature of die influence of X (Amount of AlcoholConsumed) on Y (Amount of Sleep Obtained). Whendie Amount of Alcohol Consumed is 0, die prediction isdiat an individual will sleep 6.58 hours. For each unitincrease in die Amount of Alcohol Consumed, diere is acorresponding increase in Amount of Sleep of .498hours. If an individual consumed 3.89 ounces of alcohol,die prediction is dial he/she would sleep for 8.52 hours.Thus, some very clear causal statements can be made.

We could also of course compute die regression ofAmount of Alcohol Consumed on Hours of SleepObtained, but most individuals would consider this notto be meaningful because dial causal direction does notmake sense. As a minor point we might note diat if wehad a record of how much diey had slept die nightbefore die experiment, it still wouldn't make any senseto regress die amount of alcohol administered on dienumber of hours slept die night before, even though dietime sequence is right, because die amount of alcoholadministered to each participant was randomly deter-mined. That is, time is a useful aid in determining diedirection of causation, but it isn't die only one.

Of course, die study probably wouldn't be conductedin diis fashion. A researcher would probably form two ormore groups, and administer different amounts ofalcohol to die groups and then perform a t-test or ananalysis of variance on die data. But die important thing

14 Gardner

to point out is that the generalizations made in bothcases about the causal effects of alcohol on sleep aremeaningful. Correlation does imply causation if theindependent variable is randomly assigned to the partici-pants.

Now consider die situation if diis study were con-ducted as described above, except that die experimenterpermitted die participants to pour dieir own drinks (i.e.,decide on how much diey would drink). We might verywell obtain results similar to diose presented in Figure 1(and, of course, we also might not!), but by adopting diisresearch strategy, we compromise our causal interpreta-tion of die data. Why? Well, because diere are a numberof diings dial might be responsible for the number ofounces of alcohol die individual pours, and also anumber of diings associated with diis act diat mightinfluence one's sleep diat night. All of diese, however,can be captured by die simple observation diat dieindependent variable was not randomly assigned. As aconsequence, die correlation between amount of alcoholconsumed and die amount of sleep may be higher orlower because of such extraneous variables, or diecorrelation may be essentially die same, and die datamight well also look like diose illustrated in Figure 1.Thus, it is possible diat even in diis situation, die numberof ounces of alcohol caused die amount of sleep. We justcan't say so unequivocally. There may be many odierpossible explanations of die relationship, but also theremay not be any odier explanations.

We could also conduct diis study in yet anotherdifferent way. We could ask participants, before they wentto bed, how much diey had to drink, and the next morningwe could ask diem how much diey slept. Following areexamples of die type of questions we might ask:

How many ounces of alcohol have you had to drink today?0 1 2 3 4 5 6 7 8 9 1 0 ounces

How many hours of sleep did you have last night?0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 hours

Since diese two items have very similar scale characteris-tics, it is possible diat diey may share common measure-ment error variance (i.e., die tendency to respond to dieright on bodi scales, die tendency to exaggerate, dietendency to understate, etc.) which could increase (ordecrease) dieir correlation as we discussed earlier.However, we might very well obtain results comparableto diose displayed in Figure 1. Researchers could argue,justifiably, diat such results cannot be used to concludediat amount of alcohol consumed influenced dieamount of sleep one had because of a myriad of con-founding factors from measurement error to odierextraneous substantive variables. But if die results diat

were obtained were comparable to diose shown in Figure1, isn't it possible diat die results do indicate diat diereis reason to conclude diat amount of alcohol consumeddid influence die amount of sleep?

Please, don't get me wrong. I'm not recommendingdiis as a viable approach to diis research question, butdie diing is diat when one is interested in die effect ofindividual difference variables (diat cannot be assignedto participants) on some odier individual differencevariable, diis is exacdy die type of situation in which onefinds oneself. You can attempt to control for all suchodier variables, solicit cooperation, and die like, but indie end you cannot draw an unequivocal cause-effectconclusion, simply because you did not make use ofrandom assignment of participants to die independentvariable. And diere is nothing you can do to circumventthis problem. Nonetheless, die point remains diat diecause-effect conclusion you may wish to draw may well bedie correct one. It doesn't help to simply agree diat diisis only a correlational study. You must establish proce-dures for checking die validity of diis conclusion. Anddiis means more dian simply replicating die study.

This dien is die problem facing many researchersinterested in individual differences who believe diat dieyare identifying processes in which some individualdifference variables are responsible for odiers. Take myown research, for example. For die past 40 years or so, Ihave been interested in die role of individual differencesin motivation on die learning of a second language.Because of our research as well as diat by many odierresearchers around die world, I believe diat it is valid toconclude diat differences in motivation, and oftenspecifically hi integrative motivation, are responsible fordifferences in how well individuals learn another lan-guage.

MOTIVATION IN SECOND LANGUAGE ACQUISITIONMy interest in motivation in second language acquisitionbegan at McGill University in 1956, in Wally Lambert'soffice. We were discussing what I should do for myMaster's diesis, and I was bemoaning die fact diat it wasdifficult for a monolingual like myself to do research inbilingualism, which was Wally's main interest at die time.In die course of our free-wheeling discussion, Wally wastelling me about John Carroll's research on languageaptitude (see, for example, Carroll, 1958, for an earlyinvestigation, and Carroll, 1990, for a more recentsummary of diis research), and I commented somethingto die effect diat aptitude may account for some successin learning a second language, but diat I couldn't seehow you could truly learn anodier language if you didn'tlike die people who spoke die language and wanted tocommunicate widi diem. Wally said, "Hey man, I thinkyou've got a diesis!" And so it began.

Correlation, Causation, and Motivation 15

Our initial studies focused attention on attitudestoward French Canadians, ethnocentrism, authoritarian-ism, reasons for learning a second language, motiva-tional intensity, and the like. The third study of my M.Sc.thesis involved a factor analysis of a set of variablesincluding measures of language aptitude, verbal ability,attitudes and motivation, and teacher ratings of Frenchachievement among high-school Anglophone students.We obtained four factors, two of which included appre-ciable loadings from the French Teacher's ratings ofFrench achievement. One we identified as a linguisticaptitude factor, and the other as a motivation factor.Since the teacher's ratings of French achievementloaded about equally on both factors, we concluded thatFrench achievement was related to two different compo-nents, linguistic aptitude, and motivation. We character-ized this motivation as reflecting a "willingness to be likevalued members of the language community" (Gardner& Lambert, 1959, p. 271).

For my Ph.D. dissertation, I conducted a larger-scalestudy with improved measures, and obtained a muchmore complex set of results, but basically found thatachievement in French, again among Anglophone high-school students, was associated with language aptitudeon the one hand, and attitudes and motivation on theother. Following that, Wally and I conducted very similarstudies in Maine, Louisiana, and Connecticut withEnglish-speaking samples in the three regions, andFranco-Americans learning French in Maine and Louisi-ana (Gardner & Lambert, 1972)). I would be fooling youif I said we obtained the same results in all of thesesamples. The results were much more complex andinformative than that, but there was evidence thatachievement in French was associated with both aptitudeand attitudinal/motivational variables in all samples.

Much of my research on this topic has been con-cerned with exploring the implications of these initialfindings, and elaborating on the motivation to learnanother language. In 1970, Pat Smythe and I initiated alarge-scale project designed to measure the majorattitudinal and motivational variables, and indices ofanxiety associated with learning a second language. Thisproject involved a major testing program in London,Ontario, and then in seven locations across Canada. Theprimary purpose of this project was to develop scales withhigh levels of internal consistency and test-retest reliabil-ity, and to assess the relationship of these variables toeach other and to indices of French achievement insamples of students in Grades 7 to 11. We named thecollection of tests the Attitude/Motivation Test Battery(AMTB) (e.g., Gardner & Smythe, 1981).

The results with this battery of tests were clear, andthough we noted variations that could be attributable toage, level of French instruction and/or region, it was

obvious that the basic associations between attitudes andmotivation on the one hand, and achievement on theother were quite stable. It was around then that weproposed our first formal causal model linking thesevariables. There have been revisions since then to besure, but the overall intent has been to identify thefunctional relation between attitudes, motivation, andachievement in the second language. It was also duringthis period that we formulated the basic components ofthe model, asserting that there were basically four classesof variables, as follows:

Integrativmess. This was viewed as an open interest in theother language group, and/or outgroups in general. Theintent of this concept was to capture the notion thatlearning another language involved taking on character-istics of the other language group and that, therefore, awillingness to identify with that group was necessary.Variables that were thought to influence this were suchattributes as Xenophilia (assessed in the AMTB by Interestin Foreign Languages), favourable attitudes toward theother language community and individual members ofdiat community, an interest in becoming closer to thegroup for the purpose of communication and interac-tion, (referred to as an Integrative Orientation) etc., aspositive instances, or Xenophobia, ethnocentrism,authoritarianism, and the like as negative instances. Fourmeasures — Attitudes toward French Canadians, Atti-tudes toward the European French, Integrative Orienta-tion, and Interest in Foreign Languages — were in-cluded in the AMTB to tap this component, and all butone (Attitudes toward the European French) havesurvived. Sometimes, however, other measures have beenused to tap this dimension (see, for example, Gardner &Maclntyre, 1993).

Attitudes toward the Learning Situation. This was viewed asan evaluative reaction to the language learning context.Broadly conceived, it could involve attitudes toward thegeneral school environment, reactions to the textmaterials, etc., but operationally it was defined primarilyin terms of evaluation of the language instructor and thelanguage course. The intent of this concept was tocapture variation attributable to the situation in whichlearning the language took place since it was realizedthat such emotional reactions could influence how wellan individual would acquire the language. In the AMTB,the primary measures of this concept were Evaluation ofthe French Teacher and Evaluation of the FrenchCourse.

Motivation. We theorized diat Integrativeness andAttitudes toward the Learning Situation would notdirectly promote second language acquisition, but would

16 Gardner

provide the basis for the individual's motivation to learnthe language. Our original formulation of motivationfocused on motivational intensity, or how hard a studentworked to learn the language. We quickly realized in ourinitial studies, that motivation involved more than this,and postulated that an additional component to motiva-tion was the desire to learn the language. Similarly, werealized that an individual who was truly motivated tolearn a second language would find the act of learningthe language rewarding, so we added a measure ofAttitudes toward Learning French (cf. Randhawa &Korpan, 1973). Thus, the AMTB assessed MotivationalIntensity, Desire to Learn French, and Attitudes towardLearning French. Basically, we reasoned that eitherattribute on its own (effort, desire or attitudes) did notcharacterize the motivated individual, but taken as aunit, all three would distinguish the motivated from theunmotivated individual.

Other Variables. We investigated a number of othervariables. One which has survived in our research wasInstrumental Orientation. This refers to studying thesecond language for die practical advantages of doing so.For example, one might study a second language inorder to get a good job. We tried on a number of occa-sions to expand this concept, but we were unsuccessful.More recently, however, the notion of InstrumentalMotivation has been expanded and shown to relate toachievement in a second language (Dornyei, 1990).Another variable that Pat Smythe and I introduced wasFrench Classroom Anxiety. In 1973, Richard Clementjoined our group, and initiated studies of Francophoneslearning English. Unlike most of our Anglophonesamples who lived in largely English-speaking environ-ments, his research participants often lived in environ-ments where both English and French were spoken.Thus, he introduced the variable, English Use Anxiety inhis research (see, for example, Clement, Gardner, &Smydie, 1977a), and we incorporated a measure ofFrench Use Anxiety into the AMTB to round out theanxiety concept. I will agree, however, that the conceptof anxiety does not play a major role in our model. Thethree additional scales of the AMTB, Instrumental Orien-tation, French Class Anxiety, and French Use Anxiety,brought the number of scales in the battery to 11.

The results of these various studies offered muchsupport for die notion that the various attitude, motiva-tion, anxiety, and language achievement measures werecorrelated in most instances. It has been noted (see, forexample, Au, 1988) that sometimes correlations are notsignificant but, as indicated above, one might expect thisfrom sampling fluctuations. The weight of the evidencefavours associations. In general, it would be reasonableto expect dial die relation between motivation and

achievement in a second language would be what Cohen(1988) refers to as a medium relationship (i.e., p = .30).If one identified significance using a two-tailed Type Ierror rate of .05, die power would be .57 for a sample sizeof 50 participants, and .86 for a sample size of 100. It isnot unreasonable, therefore, to find some studies dial donot obtain significant results, but the majority would.

In our research, we were careful to construct diescales of die AMTB to be as internally consistent aspossible, and to measure the attributes of interest. In dieearly development of die scales, we realized diat many ofour variables would be interrelated, and dius made useof different item formats (i.e., Likert, multiple choice,and Semantic Differential) to reduce common measure-ment error (see, for example, Gardner & Smythe, 1981).Over time, we have tended to make more use of dieLikert procedure, and have discarded die multiplechoice and Semantic Differential formats for diesubscales in order to facilitate test administration (see,for example, Gardner, Tremblay, & Masgoret, 1997), butwe have continued to maximize die internal consistencyof die scales.

A MODEL OF SECOND LANGUAGE ACQUISITIONFigure 2 presents a schematic representation of diesocio-educational model of second language acquisitionas currendy envisioned in our research. It is a slightvariation of die model we presented many years ago (see,for example, Gardner, 1985). In tiiis model, die twovariables, Integrativeness and Attitudes toward dieLearning Situation, are shown as correlated causes (orsupports) of Motivation, while Motivation and LanguageAptitude are seen as direct causes of Achievement in diesecond language. In die model, it is recognized diatodier factors might well support Motivation, and diat yetodier factors may well have a direct effect on Achieve-ment in die second language independent of Motivationor Language Aptitude. Odier researchers are investigat-ing these odier types of variables. For example, RichardClement has investigated die role of self-confidence indie language in influencing language achievement (see,for example, Clement, 1980; Clement, Dornyei, & Noels,1994). Peter Maclntyre and Elaine Horwitz, workingindependendy, have articulated die characteristics andconsequences of language anxiety (see Horwitz & Young,1991; Maclntyre, Noels, & Clement, 1997). ZoltanDornyei has investigated die role of an instrumentalmotivation in language learning (see, for example,Dornyei, 1990). And, Rebecca Oxford has identified anumber of learning strategies diat facilitate languageacquisition (Oxford, 1990).

In die model, die configuration of Integrativeness,Attitudes toward die Learning Situation, and Motivationare shown to form a complex variable diat we have

Correlation. Causation, and Motivation 17

Figure 2. Basic Model of the Role of Aptitude and Motivation in Second Language Acquisition

identified as Integrative Motivation. That is, we haveargued that individuals who exhibit high levels ofintegrative attitudes, a favourable evaluation of thelanguage learning situation, and heightened levels ofmotivation to learn the second language can be charac-terized as being integratively motivated in their languagestudy. We have not argued that this is the only type ofmotivation to learn the language, nor that it is necessar-ily the most effective form of motivation. It does seem,however, that for an individual to truly learn anotherlanguage, he/she must identify to some extent with theother language community, must find the learningsituation rewarding, and must be motivated to learn thelanguage. It is not necessary to have all three characteris-tics, of course, but if one does exhibit all three, then itseems meaningful to characterize that individual asbeing integratively motivated.

In our research, we have sometimes directed attentionto the individual scales (e.g., Attitudes toward LearningFrench, Motivational Intensity, etc.). Sometimes, how-ever, we have focused on the components and havemade use of aggregates of the relevant scales (e.g., wehave computed scores on Integrativeness, Attitudestoward the Learning Situation, and/or Motivation). Atother times, we have aggregated the three aggregates toobtain a total score on Integrative Motivation. Thenature of the variable used in an investigation depended

on the purpose of the investigation. These different usesare reflected in some of the discussion to follow.

STEPS TO INCREASE THE VIABILITY OF A CAUSAL MODELTo infer causation with individual difference datarequires a multifaceted approach. Four steps are pro-posed. Obviously, these four steps cannot guarantee thatthe causal relationships proposed are necessarily valid.Following them, however, will increase the probabilitythat a particular causal model is valid. They are pre-sented below, followed with examples of how they havebeen used in conjunction with our research.

1. Construct measures of the variables of interest that have goodmeasurement properties. The first requirement for establish-ing causal relationships with individual difference data isto establish good psychometric properties of the mea-sures in your research. This includes constructing testswith high levels of internal consistency reliability and,where applicable, high levels of test/retest reliability.Generally, the internal consistency reliabilities of ourtests are in the .80s and .90s, while the test-retestreliabilities are in the .60s to .70s, except for evaluationsof the teacher and course which tend to vary largely, Isuspect, because the teacher and the course changesfrom year to year. Also, since much of one's validationdata often involves correlations between tests developed

18 Gardner

for the research, it is advisable to use different measure-ment strategies, if at all possible. In our earlier research,we made use of Likert, multiple choice, semantic differ-ential, and Guilford (1954) forms of items in our mea-sures.

2. Assess the relationships of the variables with the majorcriteria using a variety of analytic procedures stick as bivariatecorrelation, factor analysis, and structural equation modelling,etc. The second step in increasing die probability diat agiven causal relationship is meaningful is to assess dierelationships of die variables widi die major criteria. Asan inidal step, diis involves an examination of diebivariate correlations of die variables of interest with diemajor criteria. In our research, we are concerned withthe relation of motivational variables with indices ofachievement in a second language. For example,Lalonde and Gardner (1985) summarized data for 39samples varying in size from 38 to 226 students in Grades7 to 11 in various regions across Canada. The resultsdemonstrated that die composite of Motivation tendedto correlate more highly with Grades in French (medianr= .39), and scores on objective tests of French achieve-ment (median r= .30) than did eidier of die compositemeasures of Integrativeness (median /s = .28, and .24respectively) or Attitudes toward die Learning Situation(median f s = .29, and .17).

Relationships among die collection of variables alsocan be investigated using Factor Analysis. In many of ourstudies, for example, die various indices ofintegrativeness, attitudes toward die learning situation,and motivation often defined one factor identified as anintegrative motive (see, for example, Gardner & Smydie,1981). These studies also showed relationships betweenthese attributes and indices of achievement. In studies,using multitrait multimediod approaches, it was cleardial integrativeness, attitudes toward die learningsituation, and motivation formed three ordiogonalfactors (Gardner & Maclntyre, 1993). Even so, scores ondie diree components, Integrativeness, Attitudes towarddie Learning Situation, and Motivation tend to correlatesubstantially with one another.

Finally, structural equation modelling studies can beused to provide direct tests of a causal model. Over dieyears in our research, diese tests have become morecomplex, varying from one considering die elements ofdie integrative motive, language aptitude, and Frenchachievement (Gardner, 1985) to one that investigatedodier variables as well such as self-confidence widi dielanguage, language learning strategies, and field inde-pendence (Gardner, Tremblay, & Masgoret, 1997). Wehave even investigated how different conceptualizationsof motivation such as Goal Salience and Attributions fitinto diis basic model (Tremblay & Gardner, 1995).

These tests have all provided support for die validity ofdie socio-educational model of second language acquisi-tion, but as indicated earlier, such support does not"prove" die model.

3. Assess the relationships of the variables with other variablesthat could be considered secondary criteria in the overall causalmodel. The diird step to strengthen die viability of acausal interpretation is to study die relationships of dievariables of interest widi other criteria. Thus, if integra-tive motivation is important because it causes (i.e.,facilitates) second language achievement, it seemsreasonable to expect diat integrative motivation will berelated to odier characteristics diat might be expected toaccount for differences in achievement. Examples ofsuch odier variables include behaviour in die languageclassroom, perseverance in language study, and participa-tion in bicultural excursion programs to die odierlinguistic community.

It is reasonable to expect diat individuals differing indieir level of integrative motivation would act differendyin die language classroom. In our research, we investi-gated diis in classes of students in Grades 9, 10, and 11studying French as a second language (Gliksman,Gardner, & Smydie, 1982). Students were assessed on dieAMTB during die first week of classes, and then on sixoccasions during die next four months they were ob-served in their classrooms. Students were classified asIntegratively or not Integratively motivated based on amedian split on an aggregate of Integrativeness andMotivation. It was found diat students classified asIntegratively motivated volunteered more, gave morecorrect answers, and were rated as being more interestedin class than students not integratively motivated. Thesedifferences were consistent over die three grade levels,and across die four month period.

We also reasoned diat, if integrative motivationreflected differences in motivation, diis would result ina relation between characteristics of motivation andperseverance in language study. As a consequence, weconducted research in which we tested individuals wididie AMTB in one year (i.e., Grade 9,10, or 11), and dienconducted a follow-up investigation to determinewhether or not die students enrolled in French diefollowing year (see, for example, Gardner & Smydie,1974). We found diat diose who continued widi theirstudy of French, as opposed to those who dropped out,scored significandy higher on die measures ofIntegrativenesss, Attitudes toward die Learning Situation,Motivation, Language Aptitude, and French Achieve-ment taken in die first year of die study. That is, therewere a number of factors diat influenced whether or nota student would continue French study the next year.Further analysis demonstrated diat, in general, the

Correlation, Causation, and Motivation 19

differences were greater for measures of Motivation andIntegrativeness than for the other indices. Differences inFrench achievement became more important in distin-guishing between Stay-ins and Drop-outs as grade levelincreased, but never matched the discriminatory powerof Motivation. Very comparable results were obtainedwith a sample of Grade 12 Anglophone students whenthose who continued their study of French in Grade 13were compared with those who dropped French(Gardner, Lalonde, Moorcroft, & Evers, 1987).

We have also conducted research on the role ofintegrative motivation on participation in biculturalexcursion programs. For example, Clement, Gardnerand Smythe (1977b) investigated attitudinal and motiva-tional attributes of three groupings of Grade 8 students,those who did not take part in an optional three-dayexcursion to Quebec City, those who did but reported ontheir return relatively little contact with French speakersin Quebec, and those who did and reported muchcontact. The results demonstrated an orderly increase ofscores on measures of attitudes and motivation from thenonparticipation to the low-contact to the high-contactgroups. Significant between group variability was ob-tained on the measures of Integrativeness, Motivation,and Evaluation of the French Course (but not theteacher).

In addition to these criteria, it is reasonable to expectthat differences in motivation would be related toretention of language proficiency after formal instruc-tion in the language ends, but this expectation was notsupported directly in a study by Gardner, Lalonde,Moorcroft, and Evers (1987). That is, they found nocorrelation between any of the attitude and motivationalmeasures and change in measures of achievement (Time1 minus Time 2 scores). Using causal modelling, how-ever, they did find that Language Attitudes influencedmotivation which in turn influenced achievement at theend of the course, thus supporting the basic socio-educational model. In addition, however, the modeldemonstrated that this motivation also influenced thereported use of French over the summer period, and theprior achievement and the reported use influencedachievement after the summer period. In short, theintegrative motive did influence retention, largelythrough its influence on the tendency to use the lan-guage after instruction ended. A subsequent study(Gardner, Moorcroft, & Metford, 1989) investigatedstudents in an intensive French-language summertraining program. One aspect of this study involved theadministration of the AMTB and the short form of theModern Language Aptitude Test (MLAT; Carroll &Sapon, 1959) at the beginning of the program and aseries of measures of French achievement at the end ofthe program. Two of these French achievement mea-

sures were measures of proficiency in speaking French.It was possible to test a sample of these students fivemonths later on the same oral production tests, and tocorrelate scores on the AMTB and MLAT measures withchanges in oral skill over the five-month period. Theresults demonstrated that Integrativeness, Motivationand Reported Use of French over the five-month periodcorrelated significantly with change in French oralproduction, while Language Aptitude, Attitudes towardthe Learning Situation, and Instrumental Orientationdid not. As before, there was a significant relationbetween Reported Use and Integrativeness, Attitudestoward the Learning Situation, and Motivation, suggest-ing that the association could reflect the role ofIntegrativeness and Motivation on Use.

All of these studies involving secondary criteriasupport the hypothesis diat differences in integrativemotivation are related to many variables that are impli-cated in learning a second language. All of the data arecorrelational, but it is clear that one interpretation thathelps to account for the relations is that integrativemotivation facilitates second language acquisition.

4. Make use of other procedures such as laboratory research toinvestigate aspects of the process believed to underlie the basiccausal model. The fourth step diat helps to strengthenone's confidence in a particular causal interpretation isto make use of other procedures to focus on the processthat is believed to underlie the causal links. Sometimesit is not possible to study processes in the classroom, andwe have found it necessary to conduct laboratory-basedstudies where we have more control over odier factors.

One of our hypotheses is that integrative motivationis important in that individuals who are integrativelymotivated will learn the material more quickly thanindividuals who are not so motivated. We have investi-gated this a number of times by having students learnlists of rare French-English pairs of words using a pairedassociates learning paradigm. In various studies, studentshave been administered a version of the AMTB to distin-guish between those who tend to score high on relevantcharacteristics of integrative motivation and those whoscore low. We then give them six trials to learn the pairsof words, noting the number correct on each trial.Generally, we find significant ordinal interactionsbetween motivation and trials in which the rate oflearning is steeper for the motivated students than forthose who are not so motivated. Two published studieshave classified the research participants as IntegrativelyMotivated or not (Gardner, Day, and Maclntyre, 1992;Gardner & Maclntyre, 1991). One study (Gardner,Lalonde, & Moorcroft, 1985) made the classification onthe basis of the Attitude Motivation Index (IntegrativeMotivation minus French Class Anxiety in diis study).

20 Gardner

60-iNumber Correct Number Correct

60-i

1 2 3 4

Trials—•—Lowlnteg —°— High Integ

A. Number correct by Integrative Motivation and Trials.

Figure 3. Number Correct as a Function of Motivation

Another study classified the participants based on anaggregate of state motivation measures taken on eachtrial (Tremblay, Goldberg, & Gardner, 1995). The resultswere comparable in all four studies, showing steeperrates of learning for the highly motivated participantsthan for those less motivated. In all studies, however, themeasure of motivation was based on a classification ofresearch participants based on an individual differencemeasure; thus, even here the causal interpretation of thedata is not unequivocal.

There is, however, indirect evidence supporting thecausal hypothesis that integrative motivation influencesthe rate of learning second language vocabulary. In oneof the studies (Gardner & Maclntyre, 1991), we con-trasted integrative motivation against what might betermed an instrumental motivation. Instrumentalmotivation was determined in this investigation byforming two groups, an incentive group who wereoffered $10 if they achieved a high number correct bythe last trial (i.e., at least 23 of the 26 pairs), and a no-incentive group. In this particular study, significanteffects were obtained for both the interaction of Integra-tive Motivation and Trials and Instrumental Motivationand Trials for the dependent variable, number correct.Figure 3 shows the two sets of results. Two points standout in these two figures. First, it is clear that the effects ofInstrumental Motivation are greater than those forIntegrative Motivation; this is reflected in the F-ratios andthe proportion of variance accounted for by the twointeractions (^5,440) = 8.35; Partial if = .087 for Instru-mental Motivation, and f(5,440) = 3.35; Partial n2 = .037for Integrative Motivation). Second, and more important

5 6

Incentive

B. Number correct by Incentive Condition and Trials.

1 2 3 .

Trials—•—No incentive

for the present discussion, the nature of the effects forthe two interactions are extremely similar. In both cases,the rate of learning is steeper for the participants classi-fied as high motivation in comparison with those classified as low motivation. In fact, for both interactions, posthoc tests of the means indicated that the differencesbetween high- and low-motivation conditions weresignificant at Trials 3, 4, 5, and 6. In the case of Instru-mental Motivation, we can identify the two differentslopes as being caused by the Incentive (i.e., the Instru-mental Motivation), because the incentive was randomlyassigned to the research participants. Note, however, thatalthough the results were very similar for IntegrativeMotivation, we can't, strictly speaking, make a similargeneralization about the causal effects of IntegrativeMotivation on learning because the levels were notdetermined by random assignment. Nonetheless, thesimilarity between the patterns is noteworthy. It is evenmeaningful to find that the effects were greater forInstrumental Motivation than for Integrative Motivation.The incentive condition was very specific to the task oflearning the word-pairs, whereas the index of integrativemotivation was much more general.

ConclusionsIn this presentation, I have stated, clearly I hope, thatcorrelation does not mean causation in situations wherewe are investigating the covariation between two individ-ual difference variables, where neither one has beenrandomly assigned to the participants. And this is most ofthe time. This is a truism that cannot be refuted. What Ihave said, however, is that although one can never

Correlation, Causation, and Motivation 21

demonstrate it unequivocally, it is certainly possible thatone individual difference variable could cause, or beresponsible for, another individual difference variable.

Most of our studies are based primarily oncorrelational methods making use of individual differ-ence data. In each instance, however, diere is clearlycovariation between integrative motivational characteris-tics and the criteria under investigation. And it seemslogical to conclude that the differences in integrativemotivation are responsible for die variation observed,even diough correlation does not mean causation. Thispoint is shown most clearly, I believe, in the nature of thegeneralizations that can be made from die resultspresented in Figure 3. My contention, then, is dialalthough individual difference data can never be used toidentify causal associations unequivocally, a multifacetedapproach where die process is investigated from anumber of perspectives does strengthen confidence incausal interpretations. Thus, we should not write offresearch "because it is only correlational." Instead, weshould continually test the limits of a causal interpreta-tion of individual differences by using a multifacetedapproach. For my part, I believe dial integrative motiva-tion does promote second language acquisition, and dialsuch motivation is relatively stable, being supported by ahost of social (background) variables. I might be wrong.But as I see it, if it looks like a duck, walks like a duck,and quacks like a duck, dien, for all practical purposes itis a duck. Even diough you can't ever demonstrate itunequivocally.

Preparation of this manuscript was facilitated by a grant(410-99-0147) from the Social Sciences and HumanitiesResearch Council of Canada. I would like to express myappreciation to Ljiljana Mihic and Paul Tremblay for theirassistance with its preparation. I am also indebted to Rich-ard N. Lalonde for his thoughtfulness in writing the letterto CPA that initiated this award, and am grateful to all myfriends and colleagues who supported his nomination, andto the executive of CPA who accepted it.

SommaireAyant appris que j'allais recevoir le Prix de 1'educationet de la formation de la SCP et que je devrais pronon-cer une allocution a cette occasion, j'ai longtempsreflechi au choix d'un theme. Etant donne que j'en-seigne la statistique et 1'analyse de donnees, et que mesrecherches portent surtout sur le role des attitudes etde la motivation dans 1'acquisition d'une langue se-conde, j'ai decide d'attirer 1'attention sur un problemestatistique et conceptuel qui me preoccupe depuis desannees, et de montrer comment je 1'ai resolu person-nellement, dans le cadre de mes interets en recherche.

Pour certains, ce probleme n'existe tout simplementpas: a leurs yeux, c'est un simple truisme que Ton nepuisse inferer la causalite a partir de la correlation et iln'y a rien d'autre a ajouter.

Toutefois, pour ceux qui s'interessent aux differen-ces individuelles, une conclusion aussi fataliste revienta conclure qu'il n'y aurait aucun moyen possible detirer un jour une inference causale a partir des diffe-rences individuelles. Une approche consiste a accepterle principe fondamental selon lequel la correlationn'entraine pas la causalite, puis de passer a la predic-tion en 1'opposant a la causalite, et enfin de se fondersur la modelisation causale (c.-a-d., de 1'equationstructurale) et sur des elements similaires. En fait, larecherche sur les differences individuelles englobe lacovariation et, quelle que soit la metiiode analytiqueadoptee (regression multiple, analyse factorielle,analyse discriminante, yoire modelisation d'equationstructurale, etc.), la statistique de base louche uneforme quelconque de relation mutuelle. En fin decompte, bon nombre d'entre nous croyons avoir deter-mine des associations causales, tout en admettant qued'autres interpretations soient possibles. Autrementdit, nous croyons que la personnalite cause ou en-traine certains comportements, que 1'intelligence joueun role dans le rendement scolaire, que 1'anxietederange le rendement, etc.

Dans cet article, je montre comment quatre truis-mes sur la correlation ne sont pas toujours exacts, aumoins en pratique, etj'indique comment j'ai resolupersonnellement le probleme causal, dans le cadre demes interets en matiere de recherche. Les quatretruismes sont les suivants : (a) le coefficient de correla-tion de Pearson varie de -1 a +1; (b) lorsque 1'hypo-diese nulle d'independance entre deux variables estverifiee, la valeur de la correlation dans la populationest de 0; (c) si 1'echantillon est de taille suffisante, lacorrelation sera significative; et (d) correlation n'estpas synonym e de causalite.

Je discute de chacun de ces truismes, soulignantquand ils sont vrais et quand ils sont faux. Parexemple, en ce qui concerne le probleme de causalite,on constate que, dans des conditions approprieesd'attribution aleatoire de valeurs d'une variable (paropposition a un echantillonnage aleatoire de celle-ci),les inferences causales sont appropriees. Je proposeensuite une solution partielle en quatre etapes auprobleme de causalite, et cette solution exige que Tonporte attention, non pas tant a la nature de la relation,qu'au processus sousjacent, en acceptant 1'interpreta-tion causale qui semble la plus appropriee a la rela-tion, puis en etendant les implications tout en lesperfectionnant et en les evaluant continuellementdans le contexte d'un programme de recherche. Ce

22 Gardner

processus ressemble a la notion de validite concep-tuelle, mais est plus inclusive, car on insiste moins surla validite d'un test ou d'une mesure que sur 1'elabora-tion d'un modele conceptuel base sur la recherche etutilisant parfois differents instruments dans diverscontextes. Dans ce cas, on insiste sur la validite deI'hypothese causale expliquant les relations entre uneserie de variables. Je tente d'illustrer ce fait en exami-nant notre programme de recherche sur les attitudeset la motivation dans 1'acquisition d'une langue se-conde. Toute etude individuelle est correlationnelle,mais lorsque plusieurs etudes, examinant le domainesous des angles divers, produisent des resultats compa-tibles, cela augmente la viabilite de la sequence causalepresumee.

Les quatre etapes que je propose peuvent aider acorroborer une interpretation causale particulierederivee des donnees sur les differences entre indivi-dus: (a) etablir des mesures pour les variables interes-santes ayant de bonnes proprietes de mesure; (b)evaluer les relations entre ces variables et les princi-paux criteres en utilisant diverses methodes analyti-ques comme la correlation entre deux variables, 1'ana-lyse factorielle, la modelisation d'equation structurelle,etc.; (c) evaluer les relations entre les variables etd'autres variables pouvant etre considerees commecriteres secondaires dans le modele causal global; et(d) utiliser d'autres methodes comme la recherche enlaboratoire pour etudier les aspects du processus appa-remment sousjacents au modele causal de base.

Je decris 1'application de chacune de ces etapes a larecherche que nous avons menee au cours des 40dernieres annees, sur le probleme que pose le role dela motivation dans 1'acquisition d'une langue seconde.Le modele socio-educatif de cette acquisition (Gard-ner, 1985) pose comme hypothese que deux conceptsfondamentaux, 1'integration et les attitudes envers1'apprentissage, soutiennent la motivation a apprendreune langue seconde, mais que la motivation et 1'apti-tude linguistique sont deux principaux facteurs quidetermineront le degre de succes obtenu dans 1'ap-prentissage de la langue. Tel que detaille dans lepresent manuscrit, d'autres facteurs entrent egalementen jeu, mais j'attire 1'attention sur le construit motiva-tionnel considere comme une motivation integrative.

L'application des quatre etapes demontre qu'enexaminant les proprietes de mesure des grandes varia-bles (c.-a-d., les variables caracterisant une motivationintegrative), il est possible d'elaborer des mesuresfiables et valides. En exposant la premiere etape et nosconclusions a son egard, je tiens aussi compte de lavaleur que presente 1'usage d'autres techniques demesure. La deuxieme etape expose certaines de nosconclusions sur la relation entre les trois compOsants

de la motivation integrative et les indices de progresdans la langue seconde, ainsi qu'une vue d'ensembledes resultats obtenus relativement aux facteurs analyti-ques, et un resume des applications de la modelisationd'equation structurelle que nous avons employeespour examiner le modele sous des angles divers.

Dans 1'expose sur la troisieme etape, je resumenotre recherche sur la relation entre les elements de lamotivation integrative et d'autres aspects de 1'appren-tissage d'une langue seconde, a savoir le comporte-ment en classe, la perseverance dans 1'etude de lalangue, la participation a des excursions biculturellesdans 1'autre collectivite linguistique, et la conservationde la langue seconde apres 1'etude linguistique. Laquatrieme etape insiste sur 1'usage d'autres methodesdestinees a examiner le processus considere commesous-jacent au modele causal de base. Un element duprocessus qui est important pour le modele socio-educatif de 1'acquisition d'une langue seconde est quela motivation integrative facilite 1'acquisition du Ian-gage parce que les etudiants ainsi motives apprendrontplus rapidement que les autres. Des etudes en labora-toire corroborent cette generalisation en demontrantque le taux d'apprentissage d'une langue seconde,mesure en une serie d'epreuves dans un paradigmed'apprentissage par paires associees, est plus eleve chezles participants possedant une motivation integrative.

Une etude (Gardner et Maclntyre, 1991) men-tionnee dans cette section est particulierement ap-propriee au probleme de causalite. Elle met en causedes participants a la recherche classes respectivementcomme possedant ou non une motivation integrative,selon les notes obtenues pour les indicateurs de diffe-rences individuelles de cette motivation, et commepossedant ou non une motivation instrumentale, selonqu'on leur avail ou non promis une recompense finan-ciere pour acquerir une competence elevee dans cettetache. Cette etude a demontre que les deux types demotivation ont des liens avec les epreuves, et que, dansles deux cas, les participants a la recherche tres moti-ves apprenaient plus rapidement que les autres. Etantdonne que le degre de motivation instrumentale etaitdefini en fonction de stimulants financiers determinesau hasard, on peut conclure que les differents niveauxde motivation instrumentale determinaient les diverstaux d'apprentissage. Toutefois, puisque le degre demotivation integrative etait defini selon une variabledesignant les differences individuelles, une tendancetres semblable peut etre consideree comme« uniquement correlationnelle ». II n'en semble pasmoins tres opportun de se demander si oui ou non lesresultats obtenus avec la motivation integrative peu-vent en fait etre attribuables aux differents niveaux demotivation, meme si Ton ne peut raffirmer sans equi-

Correlation, Causation, and Motivation 23

voque. Voila le probleme auquel se heurte le cher-cheur des differences individuelles qui, neanmoins,tente de comprendre comment les fluctuations d'uneou de plusieurs variables dans ce domaine influent surcelles d'une ou de plusieurs autres. La methode enquatre etapes proposee dans cet article ne peut suppri-mer entierement 1'ambiguite, mais contribue bel etbien a corroborer davantage un modele causal particu-lier.

ReferencesAu, S.Y. (1988). A critical appraisal of Gardner's social

psychological theory of second-language (L2) learning.Language Learning, 38, 75-100.

Baron, R.M., & Kenny, D.A. (1986). The moderator-media-tor variable distinction in social psychological research:Conceptual, strategic, and statistical considerations.

Journal of Personality and Social Psychology, 51, 1173-1182.Carroll, J.B. (1958). A factor analysis of two foreign lan-

guage aptitude batteries. Journal of General Psychology, 59,3-19.

Carroll, J.B. (1990). Cognitive abilities in foreign languageaptitude: Then and now. In T.S. Parry & C.W. Stansfield(Eds.), Language Aptitude Reconsidered (pp. 11-29). Engle-wood Cliffs, NJ: Prentice Hall Regents.

Carroll, J.B., & Sapon, S.M. (1959). Modern LanguageAptitude Test: MLAT. New York: Psychological Corpora-tion.

Clement, R. (1980). Ethnicity, contact, and communicative

competence in a second language. In H. Giles, W. P.Robinson, & P. M. Smith (Eds.), Language: Social Psycho-

logical Perspectives. Oxford: Pergamon.Clement, R., Dornyei, Z., & Noels, K.A. (1994). Motivation,

self-confidence, and group adhesion in the foreignlanguage classroom. Language Learning, 44, 417-448.

Clement, R., Gardner, R.C., & Smythe, P.C. (1977a). Moti-vational variables in second language acquisition: Astudy of Francophones learning English. Canadian

Journal of Behavioural Science, 9, 123-133.Clement, R., Gardner, R.C., & Smythe, P.C. (1977b). In-

ter-ethnic contact: Attitudinal consequences. Canadian

Journal of Behavioural Science, 9, 205-215.Cohen, J. (1988). Statistical Power Analysis for the Behavioral

Sciences. Hillsdale, NJ: Erlbaum.Dornyei, Z. (1990). Conceptualizing motivation in foreign-

language learning. Language Learning, 40, 45-78.Gardner, R.C. (1985). Social Psychology and Second Language

Learning: The Role of Attitudes and Motivation. London:Edward Arnold Publishers.

Gardner, R.C., Day, J.B., & Maclntyre, P.O. (1992). Integra-tive motivation, induced anxiety, and language learningin a controlled environment. Studies in Second Language

Learning, 14, 197-214.

Gardner, R.C., Lalonde, R.N., & Moorcroft, R. (1985). Therole of attitudes and motivation in second languagelearning: Correlational and experimental considera-tions. Language Learning, 35, 207-227.

Gardner, R.C., Lalonde, R.N., Moorcroft, R., & Evers, F.T.(1987). Second language attrition: The role of motiva-tion and use. Journal of Language and Social Psychology, 6,29-47.

Gardner, R.C., & Lambert, W.E. (1959). Motivational

variables in second language acquisition. CanadianJournal of Psychology, 13, 266-272.

Gardner, R.C., & Lambert, W.E. (1972). Attitudes and

Motivation in Second Language Learning. Rowley, MA;Newbury House Publishers.

Gardner, R.C., & Maclntyre, P.D. (1991). An instrumentalmotivation in language study: Who says it isn't effective?Studies in Second Language Acquisition, 13, 57-72.

Gardner, R.C., & Maclntyre, P.D. (1993). On the measure-ment of affective variables in second language learning.Language Learning, 43, 157-194.

Gardner, R.C., Moorcroft, R, & MetfordJ. (1989). Secondlanguage learning in an immersion program: Factorsinfluencing acquisition and retention. Journal of Lan-

guage and Social Psychology, 8, 287-305.Gardner, R.C., & Smythe, P.C. (1974). The integrative

motive in second-language acquisition. In S.T. Carey(Ed.), Bilingualism, BiculturaUsm & Education. Edmon-ton, AB: University of Alberta Press.

Gardner, R.C., & Smythe, P.C. (1981). On the develop-ment of the Attitude/Motivation Test Battery. The Cana-

dian Modern Language Review, 37, 510-525.Gardner, R.C., Tremblay, P.F., & Masgoret, A.-M. (1997).

Towards a full model of second language learning: Anempirical investigation. Modern Language Journal, 81,

344-362.Gliksman, L., Gardner, R.C., & Smythe, P.C. (1982). The

role of the integrative motive on students' participationin the French classroom. Canadian Modern Language

Review, 38, 625-647.GuilfordJ.P. (1954). Psychometric Methods. New York:

McGraw Hill.Harris, R.J. (1997). Significance tests have their place.

Psychological Science, 8, 8-11.Havlicek, L.L., & Peterson, N.L. (1977). Effect of die viola-

tion of assumptions upon significance levels of thePearson r. Psychological Bulletin, 84, 373-377.

Horwitz, E.K., & Young, D.J. (1991). Language Anxiety: FromTheory and Research to Classroom Implications. EnglewoodCliffs, NJ: Prentice Hall Regents.

Hunter, J.E. (1997). Needed: A ban on the significancetest. Psychological Science, 8, 3-7.

Lalonde, R.N., & Gardner, R.C. (1985). On the predictivevalidity of the Attitude/Motivation Test Battery. Journalof Multilingual and Multicultural Development, 6, 403-412.

24 Gardner

Maclntyre, P.D., Noels, K.A., & Clement, R. (1997). Biasesin self-ratings of second language proficiency: The roleof language anxiety. Language Learning, 47, 265-287.

Nunnally, J. (1978). Psychometric Theory, New York: McGrawHill.

Oxford, R. (1990). Language Learning Strategies: What EveryTeacher Should Know. New York: Newbury House.

Randhawa, B.S., & Korpan, S.M. (1973). Assessment ofsome significant affective variables and the prediction ofachievement in French. Canadian Journal of Behaviou-ral Science, 5, 24-33.

Schmidt, F.L. (1996). Statistical significance testing and

cumulative knowledge in Psychology: Implications fortraining of researchers. Psychological Methods, 1, 115-129.

Shavelson, R.J., & Webb, N.M. (1991). GeneralizabilityTheory: A Primer. Newbury Park, CA: Sage Publications.

Tremblay, P.F., & Gardner, R.C. (1995). Expanding themotivation construct in language learning. The ModernLanguage Journal, 79, 505-520.

Tremblay, P.F., Goldberg, M.P., & Gardner, R.C. (1995).Trait and state motivation and the acquisition of He-brew vocabulary. Canadian Journal of Behavioural Science,27, 356-370.