Patterns of Attachment: A Psychological Study of the Strange Situation - Mary D. Salter Ainsworth 2015
An Examination of the Classificatory System
Measures and Methods of Assessment
A Multiple Discriminant Function Analysis
Introduction
Whereas the findings so far discussed have focused on normative patterns of behavior as they are linked to the situational properties of the sequence of episodes, the findings we now report focus on patterns of individual differences in behavior—patterns that occur commonly enough to be recognizable when they recur. The normative findings presented in the previous two sections depict certain features of the species-characteristic organization of attachment behavior in the human 1-year-old and its interplay with other behavioral systems. We consider that the normative findings substantially support Bowlby’s (1969, 1973) descriptions of the organization and function of infant attachment behavior.
As we have previously pointed out, however (Ainsworth, 1967, 1972, 1973), the infant—mother attachment relationship must be distinguished from the species-characteristic attachment behavior from which it develops and that continues to mediate it. To be sure, it is a characteristic of the human species to become attached to a mother figure. This attachment or bond has the same biological function as the attachment behavior from which it stems. Although the infant is predisposed to become attached, the attachment relationship develops only gradually and is influenced in its development by the specific patterns of interaction the infant has experienced with caregiving figures. It is a hypothesis impicit in ethological attachment theory that differences in early social experience will lead to differences in the development and organization of attachment behavior and hence in the nature of attachment relationships themselves.
In the classificatory system introduced in chapter 3, we presented certain salient features of strange-situation behavior in terms of which we identified three major patterns. These are hypothesized to represent qualitatively different attachment relationships on the basis of their correlates in behavior outside the strange-situation—correlates that are presented in later sections. We also presented data indicating satisfactory levels of interjudge agreement in the identification of these patterns. In this chapter we tackle the issue of reliability in more depth, using a multiple discriminant function analysis as our vehicle. The analysis was undertaken to test the significance of multivariate differences among the three groups and to test the hypothesis that the behaviors highlighted in the instructions for classification are indeed the major behaviors in terms of which three main groups differ. In later chapters we test the hypothesis that individual differences in the attachment relationship are stable across situations, in contrast to specific attachment behaviors, which our normative findings show to be sensitive to changes in context. This is accomplished by examining the relationship between patterns of behavior in the strange situation and patterns of behavior at home. We also test the hypothesis that the organization of the infant’s attachment to his mother is influenced in its development by his mother’s behavior in interaction with him. This is accomplished by examining the relationship between patterns of the infant’s behavior in the strange situation and his experience of his mother’s characteristic behavior at home.
Multiple Discriminant Function Analysis
Because we cannot assume that the reader is familiar with multiple discriminant function analysis (MDFA), we discuss certain of its features before reporting the way in which this procedure was used with our data. This procedure is useful when two or more groups are compared in terms of many variables, and when it is of interest not only to see whether the groups differ significantly from one another, but also to understand the nature of their differences. The MDFA (Tatsuoka, 1970, 1971; Tasuoka & Tiedman, 1954; Cooley & Lohnes, 1971) is a multivariate technique closely related to canonical correlation and to multivariate analysis of variance. For the purposes of evaluating our classificatory system, the MDFA has three features that offer distinct advantages over these other multivariate techniques, as well as over univariate techniques.
First, the MDFA allows us to test the significance of differences among groups of subjects with respect to multiple variables without the problems associated with repeated univariate tests (Tatsuoka & Tiedeman, 1954). Furthermore, it is not necessary to have equal numbers of subjects in each group.
Second, in the process of deriving and testing successive uncorrelated composites of the “predictor”1 variables, the MDFA generates weighted vectors or discriminant functions (DFs), which, like the factors yielded by a factor analysis, may be interpreted as dimensions underlying group differences. The number of possible functions, however, is defined as the number of groups minus 1, which in our analysis is 2.
A third advantage of MDFA is related to the classification of individual subjects. The analysis allows us to evaluate differences between groups not only in statistical terms but also in practical terms. This is accomplished by means of classification functions based on distributions of DF scores. The DF scores of individuals in each criterion group are compared to see whether these scores can reproduce the classifications based on a larger set of variables. Finally, and more important, individuals from an independent sample can be classified to see whether the discriminant functions derived from one sample are applicable to another sample. This test with an independent sample is labeled “cross-validation.” A high degree of classification and cross-validation success points to the generality of the descriptive aspects of the analysis on a subject-by-subject basis.
The MDFA procedure is a maximization technique—that is, it derives composite variables by maximizing the average degree of separation between groups relative to variance within groups. This being so, it has considerable potential for capitalizing upon variance that happens to be specific to a given sample. Just as in linear multiple regression (which is also a maximization procedure), the results of MDFA often “shrink” when functions derived from one sample are applied to another. In addition the descriptions of dimensions of group difference often vary from sample to sample. Two precautions can be taken to minimize the influence of sample-specific variance in the interpretation of the MDFA. The first precaution is to use large samples and to keep the number of predictor variables relatively small in relation to sample size. Tatsuoka (1970) suggested that total sample size should be at least two and preferably three times the number of variables used. He also suggested that the size of the smallest criterion group be no less than the number of groups used. These criteria prevent us from using MDFA to investigate the eight subgroups of our sample. Therefore we have concentrated on analysis of the three main groups. A second precaution that can help to reduce the influence of sample-specific variance is to employ cross-validation techniques to estimate the generalizability of results to an independent sample. The method of cross-validation that we employ is described presently.
Procedure
As a preliminary to the application of MDFA, we first checked to see whether the four component samples differed in the proportions of infants classified in the three main groups, A, B, and C. (See Table 7.) They were found not to differ significantly (Chi square = 3.337; df = p > .70). We also checked the distribution of sexes among the three groups. (See Table 8.) There were no significant sex differences (Chi square = .99; df = 2; p < .50). It was necessary to drop one subject2 from the sample for the MDFA, reducing the total sample to 105, consisting of 23 in Group A, 60 in Group B, and 13 in Group C.
TABLE 7 Distribution of Infants in the Four Samples Among the Three Strange-Situation Groups
TABLE 8 Distribution of Infants by Sex Among the Three Strange-Situation Groups
Sex of Infant |
|||
Groups |
Male |
Female |
Totals |
|
|||
A |
11 |
12 |
23 |
B |
41 |
29 |
70 |
C |
7 |
6 |
13 |
Totals |
59 |
47 |
106 |
The next preliminary step was the reduction of the number of variables from the original 73 that had been scored,3 to a smaller set consistent with Tatsuoka’s suggestions. Groups means were computed for each variable, along with one-way analyses of variance. Because only those variables in regard to which two or more groups differ can contribute to the discrimination among groups, we eliminated all variables for which the F-ratio yielded by the analysis of variance fell short of significance at the 1% level. This reduced the number of variables from 73 to the 30 shown, together with their group means and F-ratios, in Table 9.(See Appendix VI, Table 32, for the other 43 variables.)
A further reduction of the number of variables was desirable because the number of infants classified in Group C was so small. We did not wish to risk eliminating potentially important variables by specifying a more restrictive F-ratio as the criterion. Instead, all 30 variables were entered into 2-group discriminant function analyses—Group A vs. Group B, and Group B vs. Group C. (The rationale for omitting the A vs. C distinction was that the most interpretable differences among groups are those that contrast one of the smaller groups with the normative B group.) Any variable that did not contribute significantly to either one or other of these discriminations was eliminated from the analysis on the grounds that it contained little information about group membership that was not contained in the other variables, despite its significant F-ratio. It may be noted in Table 9 that each of the variables eliminated in this second step is represented in the analysis by its counterpart, as scored in other episodes, and by other variables scored in the same episode as the eliminated variable. This procedure reduced the set of independent variables to 22. No further reduction was attempted, for this could have defeated the descriptive goals of the analysis.
TABLE 9 Means and One-Way ANOVAs for Strange-Situation Variables That Distinguish Among Groups A, B, and C
Using these 22 variables, the multiple discriminant function analysis was run on the Statistical Package for the Social Sciences, SPSS 6.0 (Nie, Hull, Jenkins, Steinbrenner, & Bent, 1975) at the University of Minnesota.
Results
Table 10 gives the statistics relevant to the degree of separation among the groups, the significance of each discriminant function, its relative contribution to the among-groups variance, and the proportion of variance in the total set of 22 variables attributable to group differences.
The two discriminant functions were significant in distinguishing the three groups. The conventional approach to testing the significance of the functions relies on the fact that the first DF derived will be the one that yields the greatest average difference among the groups. Subsequent DFs are successively smaller. When the maximum number of DFs has been extracted and the total discrimination afforded by them is significant, then at least the first DF must be significant. The discrimination due to the first DF is then set aside, and the same inference is made with respect to the second, and so on until the residual discrimination is not significant, or until the maximum number of functions possible has been derived—in our case, two. It may be seen that each of the two DFs yielded by the analysis is significant at a very high level of confidence.
TABLE 10 Multiple Discriminant Function Analysis
Function I |
Function II |
|
|
||
Eigenvalue |
2.359 |
1.790 |
Canonical Correlation |
.838 |
.801 |
Percentage of Trace |
56.9 |
43.1 |
Wilks’s Lambda |
.107 |
.358 |
Group Centroids |
||
Group A |
−2.816 |
.426 |
Group B |
.917 |
.519 |
Group C |
.115 |
−3.507 |
Notes: N = 105 (23, 69, 13)
Number of groups = 3
Number of independent variables = 22
Maximum number of functions = (n groups − 1) = 2
Significances:
Functions 1 & 2; Chi square = 207.01; df = 44; p = 1.39 × 10−21
Function 2 alone; Chi square = 94.92; df = 21; p = 1.54 × 10−10
The canonical correlations (Rc) presented in Table 10 are correlations between the entire set of predictor variables and the criteria of group classifications. As may be seen, they are very high. The eigenvalues (ei) are related to the canonical correlations by the formula ei = R2ci. The squared canonical correlation is the proportion of the variance of the group-membership variables accounted for by the set of predictor variables. Thus the eigenvalues are indices of the amount of among-groups variance.
The percentage of trace shown in Table 10 is the percentage of the total among-groups variance correlated with the respective discriminant functions. This figure for the first DF is 56.9%, somewhat larger than that for the second DF.
Wilks’s Lambda is an inverse measure of the relative separation of the groups by the discriminant functions. It is distributed as an approximate chi square, with the degrees of freedom indicated. There was significant discrimination power in both functions.
Thus we may conclude that Groups A, B, and C differ significantly with respect to the set of predictor variables in the analysis, even when redundancies are removed. Evidence that the criterion groups are quite distinct is highly desirable in evaluating any classificatory scheme, and this one may be seen to meet the various tests satisfactorily.
Classification by Discriminant Scores and Cross-Validation
The economy of describing Groups A, B, and C in terms of two uncorrected functions instead of 22 correlated variables is considerable. Its practical advantage, however, depends on evidence that the two-factor description can reproduce the A-B-C classifications based on the criteria of our classificatory system. This was assessed by a centour analysis of the discriminant scores of each subject in each group, as plotted in Figure 10.
FIGURE 10 Centour plot of discriminant scores.
Note: Subjects whose scores were identical are represented by one symbol in this plot.
Development
Each subject was assigned three scores reflecting his proximity to the centroid of each group.4 Proximity was defined relative to the members of a particular group, to take into account different dispersions of subjects around the three group centroids. Infants were assigned to the group to which they were closest.
The results of this classification procedure are compared with our original classifications of each infant in Table 11. The extent to which the discriminant functions allow us to reproduce A-B-C classifications is presented in terms of “hit rates” and percentage agreement. Cohen’s (1960) index of nominal-scale agreement (Kappa) was computed by correcting the observed rate of agreement for the rate of agreement expected by chance. It was tested as described by Fliess, Cohen, and Everitt (1969). Obviously there is a very high degree of agreement between the original classifications and those derived from the discriminant functions.
It is clear that description in terms of two discriminant scores conveys as much information about a subject’s A-B-C classification as does description in terms of the 22 MDFA variables. The findings also suggest that the significant differences among Groups A, B, and C, as reported above, are not merely group trends. Indeed, most members of each group differ from most members of each of the other groups in the direction indicated by the discriminant functions.
Cross-Validation
Correct classification of the individuals of an independent sample would provide the best evidence that the discriminant functions developed on our sample have more general applicability. Because Groups A and C are small, however, there were not enough subjects in our total sample of 105 to provide both a “development” sample and an independent cross-validation sample for a 22-variable analysis.
TABLE 11 Discriminant Score Centour Analysis
Predicted Classification from Discriminant Scores |
||||
Actual Classification |
A |
B |
C |
% Correct Classification |
|
||||
A |
22 |
1 |
0 |
96% |
B |
5 |
63 |
1 |
91% |
C |
0 |
1 |
12 |
92% |
Total |
27 |
65 |
13 |
92% |
Note: Kappa = .854 (z = 11.8, p <.001).>
<.001).>
TABLE 12 Serial Cross-Validation of the Multiple Discriminant Function Analysis
Predicted Classification of Unknown Subject |
||||
Actual Classification |
A |
B |
C |
% Correct Classification |
|
||||
A |
12 |
1 |
0 |
92% |
B |
1 |
11 |
1 |
85% |
C |
0 |
4 |
9 |
69% |
Total |
13 |
16 |
10 |
82% |
Note: Kappa = .730 (z = 6.51, p <.001).>
<.001).>
As an alternative, we performed a “serial” cross-validation by repeatedly deriving classification equations for 104 subjects and applying them to a single “unknown” subject. The technique differs from independent cross-validation in that it provides an estimate of whether an independent cross-validation would reach statistical significance, rather than an estimate of the exact degree of success that would actually be achieved with an independent cross-validation sample.
Thirty-nine “unknown” subjects, 13 from each group, were used as cross-validation subjects. These included all of Group C and a random selection from each of Groups A and B. They were classified as previously described, using discriminant functions developed without reference to the scores obtained by the cross-validation subject on each of the 22 variables.
It may be seen from Table 12 that an almost entirely accurate match of actual classification and that predicted from DF scores was obtained for Groups A and B. The misclassification of 4 of 13 Group-C subjects suggests that the group is too small to yield the highly generalizable results obtained for the other two groups.
It may be seen in Chapters 9 and 11 that three investigations, including one by one of us (EW), used our sample of 105 infants as a development sample and then cross-validated the classifications on independent samples of their own. Although in each case the variables used for their MDFAs differed somewhat from our final list of 22, their findings nevertheless suggest that the level of generalizability indicated by our cross-validation procedure is indeed representative of the relevance of the classificatory system to samples of middle-class 1-year-olds.
The Contributions of Each of 22 Variables to Discrimination Among Groups
As implied earlier, the multiple discriminant function analysis provides an alternative to the vagaries of multiple univariate testing of group differences, by taking into account the intercorrelations among our 22 independent variables. The analysis also leads us to think in terms of a small number of factors underlying the wide range of individual differences in strange-situation behavior. The scatterplot of the 105 subjects’ scores on the two uncorrelated discriminant functions (Figure 10) points toward a clear relationship between these independent linear combinations of the behavioral variables and the A-B-C classificatory system. This relationship is evident from a correlation of −.918 between the first discriminant function (DF 1) and the dichotomy A vs. non-A, and from a correlation of −.852 between the second discriminant function (DF II) and the dichotomy C vs. non-C.
To take advantage of this relationship in the analysis of individual variables, and to use the discriminant functions for an economical description of the three groups, we must determine the relative contribution of each variable to each discriminant function. This is essentially the same problem that arises in the interpretation of factor loadings and, more exactly, in the interpretation of multiple-regression “weights”.5
Darlington (1968) has elaborated the difficulties involved in interpreting regression coefficients and has emphasized that there is no simple or single answer to the question: “What is an important variable?” In the present analysis, however, it would seem that an “important” variable would have some combination of the following characteristics:
1. The variable provides univariate discrimination between at least two of the classificatory groups—that is, it could be used to predict group membership.
2. The variable does not make a “trivial” contribution to group differences because of its correlations with or dependency on another variable. Thus, for example, smiling and exploratory variables reflect group differences, but these seem to be a product of the high negative correlations between these variables and crying, a variable to which they seem secondary—that is, crying babies do not smile or engage in exploratory play.
3. The variable is not largely redundant with information about group membership that is available from other variables. In the absence of such incremental validity, however, a variable that passes the first test (of providing discrimination) and the second test (of not being secondary or trivial) may be very important in summarizing the behavior of members of a given group, even though there is substantial redundancy with the information provided by other variables.
4. A variable that passes the aforementioned test may be especially interesting if a substantial proportion of its total variance is correlated with one or another or both of the discriminant functions (that is, if it is heavily saturated with the dimension in question).
The first characteristic of an “important” variable (univariate discrimination among groups) is reflected in the group means and F-ratios. Those for which the one-way ANOVA is significant at the 1% level or better were reported in Table 9; the nonsignificant data are reported in Appendix VI, Table 32.
The second characteristic (group differences not due to trivial dependencies) can be assessed by referring to the individual group means (in Table 9) and to the descriptions of behavior in each of the episodes of the strange situation reported in Chapter 4. We must draw, however, on what we know about infant behavior to deal with such obvious dependencies as the one just cited—the negative relationship between crying, on the one hand, and smiling and exploratory play, on the other.
In regard to our third test, when group differences on several correlated variables are not “trivial,” the interpretation of the partial discriminant coefficients presented in Table 13 can help to uncover redundancies in the information provided by the variables. For example, interactive behavior with the mother in different episodes can not be considered trivially dependent, even though the behaviors may be highly correlated. Under these conditions, any discrimination among groups or predictive relationship of the variable to groups’ membership will be “echoed” by the behavior as it appears in other episodes. The partial discriminant coefficients shown in Table 13 attempt to remove this type of redundancy by highlighting certain variables at the expense of their correlates. Thus a low value of the partial discriminant coefficient does not necessarily imply that the groups are indistinguishable in terms of this variable. Indeed the method of selecting variables for inclusion in the present analysis ensured that this was not the case. A low value of the partial discriminant coefficient may suggest that a subject’s score on the discriminant function cannot be predicted from the variable in question and/or that the variable adds little to the predictive power of the other variables with which it is correlated. In general, the variable highlighted by the partial discriminant coefficients is the one that makes the greatest contribution to discrimination among the groups, when the contributions of correlated variables are controlled for.6
The semipartial (or part) correlations shown in Table 13 reflect the correlation between the discriminant functions and that part of the variance of the variable that is uncorrelated with the other variables in the analysis.
The Pearson correlations of the 22 variables used in the MDFA with the two discriminant functions are shown in Table 14. Because the discriminant functions correlate so strongly with the A vs. non-A and the C vs. non-C dichotomies, it is clear that the correlation of a variable with either discriminant function reflects a mean difference between groups on that variable. The vector of correlations of variables with a discriminant function (not the vector of partial discriminant coefficients) provides the most descriptive summary of the behavioral correlates of the discriminant function.
TABLE 13 Multiple Discriminant Function Analysis of Strange-Situation Variables
Finally let us consider the communality statistic in Table 13. Because the discriminant functions are uncorrelated, the sum of the squares of the correlations of a variable with each function indicates the proportion of variance associated with the two functions.
Thus in the description that follows, the relevant statistics are: the group means and F-ratios in Table 9, the standardized partial discriminant function coefficients, and the communalities reported in Table 13; and the Pearson correlations of interactive behavior with the discriminant functions reported in Table 14.
TABLE 14 Correlations of Strange-Situation Variables with Two Uncorrelated Discriminant Functions
Correlations With |
||||
Variable |
Episode |
Persons Present |
DF I |
DF II |
|
||||
Interactive Behaviors with M |
||||
Proximity Seeking |
5 |
M, B |
.429a |
−.203 |
8 |
M, B |
.625a |
.177 |
|
Contact Maintaining |
5 |
M, B |
−.358a |
−.352a |
8 |
M, B |
.693a |
−.070 |
|
Avoidance |
5 |
M, B |
−.719a |
.148 |
8 |
M, B |
−.874a |
−.040 |
|
Resistance |
5 |
M, B |
.168 |
−.582a |
8 |
M, B |
−.313a |
−.542a |
|
Interactive Behaviors with S |
||||
Resistance |
3 |
S, M, B |
.029 |
−.529a |
4 |
S, B |
.183 |
−.699a |
|
7 |
S, B |
.281b |
−.416a |
|
Distance Interaction |
4 |
S, B |
−.323a |
.373a |
7 |
S, B |
−.525a |
.293b |
|
Exploratory Behavior |
||||
Exploratory Locomotion |
7 |
S, B |
−.525a |
.223 |
Exploratory Manipulation |
4 |
S, B |
−.358a |
.373a |
7 |
S, B |
−.468a |
.421a |
|
8 |
M, B |
−.386a |
.285b |
|
Crying |
2 |
M, B |
.022 |
−.560a |
3 |
S, M, B |
.041 |
−.413a |
|
5 |
M, B |
.297b |
−.457a |
|
6 |
B |
.335a |
−.619a |
|
8 |
M, B |
.314a |
−.556a |
ap < .001. b p < .01.
Characterization of the Discriminant Functions
The discriminant analysis highlights the importance of interactive behaviors in discriminating among the three classificatory groups. This is especially true of interactive behavior with the mother. The analysis also highlights the importance of the reunion episodes, numbers 5 and 8. These are also the behaviors and episodes that are most heavily emphasized in the criteria for classification (Chapter 3). Thus the multiple discriminant function analysis suggests that the instructions for classification do in fact dwell on the variables that convey the most discriminating information about an infant’s classification. As indicated below, the analysis also casts some light on the role of strange-situation crying as an indicator of the nature of the infant—mother attachment relationship.
Discriminant Function I (A versus non-A)
As indicated by the −.918 correlation between DF I and the dichotomy A vs. non-A, the first discriminant function serves to distinguish Group-A infants from infants in Groups B and C. Variables that correlate negatively with DF I typify Group-A subjects; variables that correlate positively with DF I typify non-A subjects, especially those in the large B group.
Avoidance of the Mother. The variables most highly correlated with DF I are avoidance of the mother in Episodes 5 and 8 (r = .719 and −.874, respectively). This matches the criteria for classification of Group A, which give the first emphasis to “conspicuous avoidance of proximity to or interaction with the mother in the reunion episodes.” Although avoidance in Episodes 5 and 8 are significantly correlated (r = .581, p < .001), their substantial partial discriminant coefficients (−.521 and −.922, respectively) indicate that they are by no means entirely redundant. Even though avoidance in Episode 8 is especially noteworthy, avoidance in Episode 5 still ranks as a highly important variable for the discrimination of Group-A from non-A babies.
Seeking To Gain and Maintain Proximity To and Contact With the Mother. Both proximity- and contact-seeking and contact-maintaining behaviors in the reunion episodes are positively correlated with DF I, and thus are shown to be more typical of non-A than of A babies. Their relationship to classification is especially clear in the second reunion episode (r = .625 and .693), in contrast with the first (r =.429 and .358). These behaviors are also featured in the criteria for classification. Presence of such behaviors in the reunion episodes is given first place in the instructions for identifying Group-B infants and is second only to resistance in the instructions for identifying Group-C infants. Relative absence of such behaviors is second only to avoidance in the instructions for identifying Group-A babies.
Proximity- and contact-seeking and contact-maintaining behaviors reflect activity directed toward a common goal, and hence are positively correlated in both reunion episodes (r = .593 and .538). Proximity seeking is significantly but not strongly correlated from Episode 5 to Episode 8 (r = .244, p < .01), reflecting the fact that infants who show weak proximity seeking in the first reunion episode may seek it strongly in the second, whereas some who seek it strongly in the first reunion may be too distressed to do more than signal for contact in the second—specifications detailed in the instructions for the classification of Group-B infants into subgroups. Contact-maintaining behavior, however, is somewhat more consistent across reunion episodes (r = .452), although very much more likely to occur strongly in Episode 8.
The low discriminant coefficients for proximity seeking in Episode 5 (.118) and Episode 8 (.058) and for contact maintaining in Episode 5 (−.284), as well as their small semipartial correlations, reflect the intercorrelations among these variables. They also reflect the fact that proximity seeking and contact maintaining in the reunion episodes are strongly and inversely correlated with avoidance (r = −.485 and −.444 respectively in Episode 5; −.615 and −.574 respectively in Episode 8). Despite these intercorrelations, contact maintaining in Episode 8 has a substantial discriminant coefficient (.577) and thus makes a relatively large contribution to the discrimination between A and non-A infants.
Resistance to the Mother. Resistance to physical contact or interaction with the mother is significantly but not strongly correlated with DF I, and then only in Episode 8 (r = −.313). The group means in Table 9 indicate that the correlation is primarily due to the absence of resistant behavior in the large B group (as specified in the instructions for classification) and to the presence of some resistance in Group-A babies. Although it was specified in the instructions for classification that A babies tend to lack resistant behavior, it was also specified that A2 babies, if picked up by the mother in Episode 8, tended to squirm to get down, a behavior scored as resistant.
Interactive Behavior With the Stranger. Relatively little weight was given to behavior with the stranger in the instructions for classification. It was mentioned, however, that it was characteristic of Group-B babies to be more interested in contact and interaction with the mother than with the stranger, and of C1 babies to be resistant to the stranger, whereas for Group A a tendency was noted to treat the stranger much as the mother is treated, although perhaps with less avoidance. Furthermore, it was specified that the A baby was not distressed when left with the stranger although he might be distressed when left alone.
The findings in Table 14 show that distance interaction with the stranger in the mother’s absence (Episodes 4 and 7) is characteristic of Group-A in distinction to non-A babies; it is significantly correlated with DF I (r = −.323 and −.525 for Episodes 4 and 7, respectively). It seems likely that this finding is secondary to the fact that Group-A infants are not distressed in Episodes 4 and 7; distance interaction with the stranger is not so much characteristic of A babies as it is of babies who are not distressed.
Exploratory Behavior and Crying. The instructions for classification placed relatively little emphasis on crying. As noted earlier, it was specified, however, that Group-A babies showed little or no separation distress, except possibly when left alone in Episode 6. For non-A babies the specifications in regard to crying differed from one subgroup to another. The instructions made no mention of exploratory behavior except by implication in the case of two subgroups—B4 and C2. Nevertheless the group means in Table 9 indicate that throughout the strange situation, Group-A infants explore more actively than either Group-B or Group-C infants and that they cry less in all episodes than the non-A infants, particularly less than Group-C infants. It could well be argued that lack of distress, whether due to the unfamiliarity of the physical environment and of the stranger or to separation, is the explanation of the ability of the A baby to sustain exploration throughout; crying babies do not explore. The dynamics of the Group-A pattern of behavior, however, are more complex than this, and are discussed after we have presented the findings of the behavior at home of infants and of their mothers in the three strange-situation groups.
In any event, even though the exploratory and crying variables have significant correlations with DF I, the standardized partial DF coefficients assigned to these behaviors are small, indicating that they offer little discriminative information not contained in the interactive variables. Furthermore, the signs of the coefficients for crying in Episodes 5, 6, and 8 contradict the direction of the group means and thus are best considered artifactual.
In summary, DF I is strongly correlated with the A vs. non-A dichotomy. Its strongest correlates are interactive behaviors displayed toward the mother in the reunion episodes, especially in Episode 8. It summarizes the dominant factor underlying group differences in terms of active avoidance of the mother in the reunion episodes (characteristic of Group A), and contrasts with proximity and contact seeking and contact maintaining in these same episodes (characteristic of non-A). In this respect the analysis confirms the match between the criteria for classification for Group A and the actual behavior of infants so classified.
Discriminant Function II (C versus non-C)
As indicated by the −.852 correlation between DF II and the dichotomy C vs. non-C, the second discriminant function serves to distinguish Group-C infants from infants in Groups A and B. Variables that correlate negatively with DF II typify Group-C infants; variables that correlate positively with DF II typify non-C infants, especially those in the large B group.
Resistance to the Mother. The criteria for classification of Group C give primary emphasis to “conspicuous contact- and interaction-resisting behavior.” These criteria are reflected in the negative correlations of resistance to the mother in Episodes 5 and 8 with DF II (r = −.582 and −.542, respectively). Resistance in Episode 5 is significantly but not strongly correlated with resistance in Episode 8 (r = .341, p <.001). neither="" variable="" is="" highlighted="" at="" the="" expense="" of="" other="" in="" standard="" partial="" df="" coefficients="">Table 13 (both are −.388). Thus resistance to the mother in each reunion episode adds to the information about group membership that is provided by the other. This belies the suggestion in the instructions for classification into Group that resistance might be especially telling in Episode 8.
<.001). neither="" variable="" is="" highlighted="" at="" the="" expense="" of="" other="" in="" standard="" partial="" df="" coefficients="">
Seeking to Gain and Maintain Proximity to and Contact With the Mother. The criteria for classification specify that Group-C infants show, in addition to resistance to the mother, “moderate to strong seeking of proximity and contact and seeking to maintain contact once gained.” The specifications for the two subgroups reflect the heterogeneity of Group C, however, stating that these behaviors are strong (active) in C1 babies, whereas the C2 babies, notable for passitivity, tend more to signal their desire for contact than to seek it actively. On the other hand, the criteria for classification specify that proximity seeking and contact maintaining is characteristic of Group-B infants. Therefore it is not surprising to find that these variables do not clearly distinguish C from B infants, as the group means in Table 9 attest. There is one exception, however—one that was not included in the instructions for classification. Contact maintaining in the preseparation Episode 3 is negatively correlated with DF II (r = −.328, p < .001). The discriminant coefficients in Table 13 indicate, however, that it is largely redundant with other variables—probably with crying in Episodes 2 and 3.
Resistance to the Stranger. The criteria for classification in specifying resistant behavior as characteristic of Group C did not limit the specification to the reunion episodes; and indeed in the case of subgroup C1 they were explicit in specifying that resistance was likely to be shown to the stranger as well as to the mother. The findings of the discriminant analysis are congruent with the notion of generalized resistant behavior as characteristic of Group C. Resistance to the stranger in both preseparation and separation episodes (Episodes 3, 4, and 7) is strongly correlated with DF II (r = −.529, −.699, and −.416, respectively). Resistance to the stranger in the first separation episode is correlated .415 (p <.001) with="" resistance="" to="" the="" stranger="" before="" separation.="" discriminant="" coefficients="" in="">Table 13 highlight the behavior during the first separation (−.538) rather than during the second. Because the means for these two episodes are not significantly different (4.23 vs. 4.12), this is not easily explained in terms of changes in the stranger’s behavior, but probably can be accounted for in terms of redundancy of information contributed by Episode-7 resistance. The only significant behavioral correlates of resistance to the stranger are crying and its correlates.
<.001) with="" resistance="" to="" the="" stranger="" before="" separation.="" discriminant="" coefficients="" in="">
Crying. The criteria for classification included crying or its absence, but in a way that split the sample into two rather than three groups. Infants who did not cry in the separation episodes were generally to be classified either in Group A or in Subgroups B1 or B2 (and perhaps occasionally in B3). Infants who cried in the separation episodes were to be classified in either Group C or in Subgroups B3 or B4 (or possibly in Group A if they cried only when left alone and otherwise fit into Group A). Furthermore, it was specified that B4 infants might be distressed even in the preseparation episodes.
Nevertheless, five measures of crying were significantly discriminating among groups on the basis of the one-way analysis of variance to be entered into the MDFA; and in the case of all of them, the greatest frequency of crying occurred in Group C (Table 9). Crying in Episodes 2, 3, 5, 6, and 8 correlated significantly with DF II (r = −.559, −.413, −.457, −.619, and −.368, respectively) as shown in Table 14. In addition, crying scores tended to be correlated across episodes.
Two crying variables were assigned substantial discriminant coefficients (Table 13). Crying when first introduced into the strange situation (Episode 2), although very rare, was assigned the highest discriminant coefficient of any variable of DF II (−.750). This implies that if an infant cries in Episode 2, he is likely to be best classified in Group C, even though not all Group-C babies do so. The correlation between crying in Episode 2 and in the subsequent preseparation Episode 3 was .311 (p< .001); but the latter was not assigned a significant discriminant coefficient (Table 13), probably because of redundancy with other variables. Crying in Episode 2 reflects inability to use the mother as a secure base from which to explore, and as such may be considered one of the ways in which Group-C infants may show the “maladaptive” behavior specified by the instructions for classification. Thus Group-C infants are clearly to be discriminated from non-C infants by the presence of even minimal unprovoked crying in the preseparation episodes. The other crying variable that received a substantial discriminant coefficient was crying in the first reunion episode 5; this is discussed later.
Crying during the separation episodes clearly distinguishes the C group from the others. Crying in Episodes 4 and 7 were, however, eliminated from the MDFA as a result of redundancies with the other crying variables. Crying in Episode 6, when the baby was alone, was retained, and as mentioned earlier was significantly correlated with DF II. It did not receive a large discriminant coefficient, however (.255). It was significantly correlated with crying in the reunion Episodes, 5 and 8 (r =.458 and .371).
Crying in the reunion episodes, while not explicitly specified as characteristic of Group C, is not inconsistent with the instructions for classification. This variable implies difficulty in being comforted when the mother returns—a difficulty that in part reflects extreme distress in the separation episodes, after which it takes a while to settle down, and in part reflects the ambivalence toward the mother that was specified for Group-C babies in terms of the simultaneous occurrence of both resistant and proximity- and contact-seeking behavior. In any event, crying in Episodes 5 and 8 is strongly associated with DFII (r = −.457 and −.556, respectively), and these two crying variables are significantly correlated with each other. It is clear that Group-C infants prolong the distress occasioned by separation into the reunion episodes and cannot be soothed easily by the mother’s presence. The discriminant coefficients are −.440 and .020 respectively, which suggests that crying in Episode 8 is largely redundant with crying in Episode 5, although together or separately they are “important” descriptive variables. Furthermore, it seems plausible to class prolonged reunion-episode crying as one of the “maladaptive” behaviors characteristic of Group C.
Exploratory Behavior. As noted in the discussion of DF I, the criteria for classification did not specify group differences in exploratory behavior, and indeed implied especial infrequency of such behavior with reference to two subgroups, B4 and C2. Nevertheless each of the exploratory behaviors included in the MDFA has a modest positive correlation with DF II, which indicates that Group-C infants explore less actively than do non-C infants. These correlations can be explained in part in terms of crying; crying babies do not explore. The correlations between crying and exploratory behavior are consistently negative and highly significant. Because of this redundancy, none of the exploratory variables in the analysis contributed in an important way to the discrimination between C and non-C babies.
In summary, DF II is strongly correlated with the C versus non-C dichotomy. Its strongest correlates are crying variables and resistance to both mother and stranger. It is difficult to summarize this second dimension underlying group differences in strange-situation behavior, except to repeat the rather imprecise and hence unsatisfying term “maladaptive,” which was used in the criteria for classification. The findings to be reported later in regard to group differences in the behavior of infants and their mothers at home suggest that the attachment relationship of Group-C (and also Group-A) infants with their mothers is anxious. The small number of C infants—both in Sample 1, for which home data are available, and for the total sample involved in the MDFA—together with the nature of the MDFA cross-validation results, suggests that classification criteria for Group C should be left open to refinement in the light of new subjects who may be studied in the strange situation and, ideally, also at home.
Conclusion
The multiple discriminant function analysis does not, of course, prove that our present classificatory system is the best of all possible descriptions of individual differences in strange-situation behavior. It does, however, demonstrate that the ways in which the infants of our sample were classified is consistent with the stated specifications.
Furthermore, the MDFA helps us to assess the extent to which the range of individual differences was captured by classificatory system. It is clear that the groups A, B, and C differ markedly in terms of the dimensions described by the discriminant functions. It is also clear that the results of this analysis can be expected to generalize to new samples. Furthermore, the weighting assigned to each variable in the discriminant functions is for the most part highly congruent with the significance attached to each in the classificatory system. Only in regard to a few points relevant to the discrimination of C from non-C infants did the MDFA draw attention to the possible refinements that might be made in the classificatory system—preferably after obtaining a larger sample of infants potentially classifiable in Group C.
The MDFA has highlighted several interesting observations: the importance of interactive behavior, the importance of behavior in the reunion episodes (and indeed the importance of having a second separation so that there can be a second reunion episode), and the importance of assessing negative facets of interactive behavior, as well as the positive facets that may be classed as attachment behavior. These observations are not products of this analysis; they have been focal points of our strange-situation research for over a decade. The MDFA, however, removes from these observations the possible charge that they reflect our own bias in interpretation rather than the observable facts themselves. Ultimately, however, their importance and indeed the usefulness of the strange situation itself for the study of individual differences in infant-mother attachment and attachment behavior lie in the relationship between these observations and behavior outside the strange situation.
Notes
1 The “predictor” variables are used to predict the criterion variables. In our case the criterion is the classification into Groups A, B, and C.
2 One male subject had to be omitted because the recording equipment broke down in Episode 6. Although it was possible to classify this baby (in Subgroup B4) on the basis of a written record for the rest of the situation, insufficient detail could be included to permit us to score the measures for the later episodes.
3 The original 73 variables included separate scores for each episode and, of course, separate scores for behavior directed toward the mother and the stranger.
4 In classifying subjects from their scores on the discriminant functions—both for the development sample and in serial cross-validation—no a priori probability of membership in any group—A, B, or C—was specified. Thus, we did not take advantage of the fact that membership in Group B is most probable. As a consequence subjects were automatically classified in the group whose centroid their scores most closely approximated.
5 The standardized partial discriminant coefficients in Table 13 are in fact proportional to the regression coefficients relating the individual variables to the discriminant scores.
6 The discriminant coefficients are dependent upon the variables included in any given analysis. If new variables were added, these coefficients might change substantially, even though the simple correlations of the individual variables with the discriminant functions might not change significantly.