Fam Proc 24:213-224, 1985

Alternative Strategies for Creating "Relational" Family DataLAWRENCE FISHER, PH.D.a

RONALD F. KOKES, PH.D.b

DONALD C. RANSOM, PH.D.c

SUSAN L. PHILLIPS, M.A.d

PAMELA RUDD, PH.D.e

aChief, Psychology Service, Veterans Administration Medical Center, Fresno, California, and Professor Of Psychiatry, Family, andCommunity Medicine, University of California, San Francisco, Fresno-Central San Joaquin Valley Medical Education Program, 2615East Clinton Avenue, Fresno, California 93703.bDirector, Behavioral Medicine Unit, Veterans Administration Medical Center, Fresno, California, and Associate Professor ofPsychiatry, Family and Community Medicine, University of California, San Francisco, Fresno-Central San Joaquin Valley MedicalEducation Program.cAssociate Professor of Family and Community Medicine, University of California, San Francisco.dResearch Associate, University of California, San Francisco.eAssistant Clinical Professor of Psychiatry, University of California San Francisco.

A major problem facing family clinicians and researchers is creating data that will reflect the family as a unit. Toaddress this problem, we present a framework for family assessment based on three measurement strategies: individualfamily member assessment, relational family assessment, and transactional family assessment. Within this context, wepresent several categories of methods for combining individual family member data into "relational" scores that reflectthe couple or family as a unit. The problems and benefits of each method are presented, and it is suggested that thechoice of method is dependent upon the content of the assessment, the theory underlying the content, and the statisticalproperties of the individual family member scores.

A number of commentators have warned that most "family" research is based on data produced by individual familymembers who provide information about their families, rather than on data obtained by studying families directly orincluding several family members in the same data collection process. For example, in a review of major family journalsfrom 1969, to 1976, Hodgson and Lewis (14) reported that the marital dyad was the object of study in only 7 per cent of theso-called family studies published. Further, they reported that in only 10 per cent of the cases was the family the unit ofanalysis.

There is little question that the methods used to collect family data define and limit the statements that can be generatedfrom those data (12). Given this state of affairs, it would appear that most "family" data are really data generated fromindividuals who also happen to be family members. Below, we shall present several alternative strategies for creatingfamily-based datadata reflecting the family as a unitfrom scores generated by individual family members. It is ourcontention that a creative and thoughtful use of such strategies has been neglected and that the field could benefit fromadditional imaginative applications of this kind. Consequently, the purpose of this paper is to increase awareness of theseapproaches and strategies.

To set the stage for this discussion, we shall begin by describing a framework for categorizing family data so as tofacilitate the correspondence between the category of family data employed and the statements that can be generated fromthe data. Three logical levels will be described: individual, relational, and transactional.

A Classificaton of Family DataWhen data from a single family member are utilized with no reference to the views, perceptions, or actions of other

family members, such measurement occurs at the "individual" level of assessment.1 Individual data may include responsesto self-report scales or interview questions or may refer to individual, noncontingent behaviors occurring in interactiontasks. Strictly speaking, statements derived from such data reflect only the particular respondent's view or behavior and donot apply to a quality or characteristic of the family unit or system.

For example, in a detailed study of marital structure, tension, and balance based on field theory, Herbst (13) asked 96children from 10 to 12 years of age to complete a questionnaire regarding the division of responsibilities and activities oftheir parents. The children's self-report of their family's hierarchy of responsibilities became the primary data for the study.

_____________________________________________________________________________________________________________

1

These self-reports were useful statements based on the views of each child, but one cannot extrapolate from theseindividual views to statements about the family itself.

As suggested by this example, individual level scores may have considerable value in their own right and may be highlypredictive of a range of outcomes and related variables. But such individually based scores reflect the actions of a singlemember of a family system, without reference to that system. Consequently, they reflect only that element's behavior orperspective as an isolated object and in this sense are not to be viewed logically as "family" data in the strictest sense of theterm (24).

We have termed the second logical level "relational." In this case, individual-level data collected from two or morefamily members are "related" to each other in some way by the investigator. From these two or more individual data sets, asingle, new score representing a previously defined characteristic of the contributing individuals' combined scores iscomputed (e.g., mean, difference, etc). In this manner, the responses of individual family members are combined in someway to derive a score that reflects some attribute or characteristic of the family unit. It should be kept in mind that thisderived score is no longer a reflection of a single family member, as in individual level data, but instead is descriptive of thecombined products of individual family members. Statements made from these data refer to characteristics or views of thecontributing members. They are statements about the family in so far as they refer to a quality of family members'perceptions of some event or to aspects of their history, attitudes, or attributes. For example, Van der Veen (33) suggestsaveraging husband and wife scores to create a family adjustment score on his Family Concept Q-Sort, and Moos and Moos(23) recommend using both spouse mean and discrepancy scores on their Family Environment Scale.

Transactional data, the third logical level, refer to a measure of the family. They reflect some product of the system orbehavioral interchange among system members that indicates the transactional unification of the system's elements into awhole that is significantly different from the sum of its parts. It is derived from the functioning of the entire unit and is not areflection of the separate contributions of family members, individually or in combination. Data obtained from naturalisticobservation and contingent, structured interaction tend to reflect transactional level assessment, as here defined. Examplesinclude Strauss and Tallman's (32) SIMFAM procedure as well as the various techniques developed by Mishler andWaxler (21), Wynne and Singer (35), and Reiss (29).

In sum, "individual level" data refer to statements by a single family member about the family out of the context of thefamily. Although useful in some respects, by our definition such data are not to be viewed as "family" data or, moreappropriately, as data generated by a family. "Relational" data are derived from the contributions of family memberscombined or contrasted in some way to indicate a characteristic of the unit. Such data yield descriptive statements about thefamily. "Transactional" family data are obtained directly from the actual contingent behavior of the dyadic or family unit andare not a reflection of the "separate elements" of the system as distinct individuals. Such data reflect a measure of the family.Each of the three levels of assessment yields data from which different statements about the family can be made. Within thiscontext, we shall review several alternative approaches for creating relational data.

Approaches to Creating Relational DataIn relational assessment, the researcher decides upon a formula for combining the various family members' responses

into a score from which statements about the family can be made. It is our contention that insufficient attention has beendevoted to (a) the selection of an "appropriate" method of combining data produced by individual family members into arelational score, and (b) the statistical properties of the selected method as they reflect the eventual distribution of scores.We shall review the pros and cons of various methods of combining the individual scores of two or more family membersto create relational data. Although we present several of the strengths of each method, our focus is on the potentialproblems and dilemmas that must be addressed in the selection of each of these approaches.

For simplicity, this discussion is based on the following assumptions: 1. Individual level data have been collected (e.g., paper and pencil, structured interview). 2. Each scale or dimension within a scale comprises more than one item or unit of responses. 3. The resulting data reflect a continuous scale approaching ordinal level measurement. 4. Scores are available for more than one family member in each family.

Arithmetic MeanThe arithmetic mean of two or more individual scores has been used widely in both current and past research (e.g., 23,

33). It is easy to compute, and it reflects the arithmetic center of the distribution of family member scores.There are three primary problems in the use of the arithmetic mean as a relational score. First, the mean, although a

legitimate measure of central tendency of a distribution of scores, may not yield a score that has sufficient conceptualmeaningfulness to be useful. For example, mean scores do not reflect differences among family members based on age ordevelopmental stage (19); parents' and children's scores are weighted equally (34); and they do not reflect differences of

_____________________________________________________________________________________________________________

2

power, influence or investment of family members in the task. Although weighting family members' scores may make sensein some cases, such a process requires careful conceptual justification.

The second major drawback to the use of mean scores concerns the problems of regression to the mean and reduction ofscore variance. Arithmetically, the distribution of the mean scores of several families will be less variable than thedistributions of the original individual scores. Furthermore, by reducing the variance of such a distribution, one reduces theinfluence of deviant families in the sample as well as deviant members within families. The larger the number of familymembers, the less impact each deviant member will have on the arithmetic mean. Consequently, there is a directrelationship between family size and the influence of deviant family members on the final "family" score.

Third, the arithmetic mean does not take into account the differences between or among the contributing scores. Forexample, on a fictional self-report couple scale with a range of 0 to 50, let us suppose that two spouses scored 48 and 2. Letus further suppose that a second couple scored 26 and 24, respectively. Using the arithmetic mean as the sole relationalscore, both couples would end up with the same "couple" score of 25.

Fourth, the arithmetic mean does not take into account the order of the scores; that is, whether the wife or the husbandhas the highest or lowest score. For example, Draper (9) mentioned that husband's occupation is most frequently used as ameasure of the dyad's social status. It is likely, however, that a couple with a wife's occupational score far above herspouse's will be quite different from a couple with a wife's occupational score far below her husband's. In this case, theordering of the two spouse scores can be of considerable importance above and beyond the arithmetic mean of the two.Clearly, no single aggregate score can reflect level, discrepancy, and order at the same time and still be practical. Meanscores, however, severely reduce the information contained in the original scores and potentially distort the data if takenalone.

The arithmetic mean may be used appropriately, however, when the discrepancy between spouses or among familymembers' scores is small and there is a similarly restricted range of intrafamilial scores within other families in the sample.It also may be useful when the purpose is to classify respondent couples on level of response alone. The use of mean scores,however, without accounting for the differences among the scores or other aspects of the separate distributions of familymembers' scores, reduces the power of any subsequent statistical analysis. Olson et al. (27) reported essentially zero-ordercorrelations between couple mean and difference scores across a number of scales, indicating the relative independence ofinformation contained in each. Mean scores should be used only with caution and with careful consideration for both theirstatistical and conceptual strengths and weaknesses.

Sum of ScoresThe advantages and disadvantages of summation scores are similar to score averages. Score sums, however, may

increase the range and variance of couple scores. This can be particularly useful in scales where the range of the scale islow. Several problems emerge, however. The sum of two spouse scores may be above the real upper limit of the scale. Onemight ask what a couple score of 65 means theoretically if the actual scale is limited by a score of 50 (10 questions, eachrated on a 5-point scale). That is, can scores derived from two individuals be viewed theoretically as additive?

Depending upon the data and the purpose for combining family members' scores, a variant of simple summation may behelpful in some cases. For example, we have used the "nonredundant sum" of family members' scores in the recording ofnegative family life events. Olson et al., (27) refer to this method as a "maximized couple score." On a list of negative lifeevents given to all family members over age 12, a potentially stressful event is noted as occurring in the family if one ormore members checked the item. The rationale is that if one member reported the occurrence of the event, that member wasforced to cope with a situation that, in turn, was also likely to affect the family. There are other areas where nonredundantfamily sums could be used as well: number of disagreements; complaints; and certain events such as doctor's visits, workchanges, or contacts with social or legal agencies. For example, Kosa (18) has reported that the general health of familymembers is made up of so many separate events that their recall by any single family member is impossible. Consequently,interviewing several family members and combining the data are suggested. The assumption in the use of a summary scoreof this type is that an event or situation occurred if indicated by one or more family members, regardless of the response ofother family members. By and large, these scores are most useful in recording discrete, situational events or circumstancesrather than family members' views, opinions, or behaviors.

Extreme ScoresDeviant scores from family members can point to pathological processes within the family group. Therefore, there are

times when it makes sense to choose the more deviant or extreme family member score as that score which best reflects thevariable of interest. Although extreme scores reflect level only and not the differences among family members' scores, givencertain measurement needs, an extreme score format can be quite helpful. For example, in a study in progress we askedcouples to rate the severity of impact on the family of certain life events, events they had previously categorized as negative.

_____________________________________________________________________________________________________________

3

Spouse agreement on these 0- to 5-point ratings varied considerably. For each of the 58 items, however, we selected themost extreme spouse rating, which, when summed across all items for the couple, yielded a score reflecting "the maximumlevel of stress affecting the couple over the last three months." As a second example, Klein and Hill (17) entered into theiranalysis the score of the family member who was least satisfied with the solution in a family problem-solving task. Theyargued that "the effectiveness of a problem solution depends on its consequences for family functioning, and that onesufficiently disenchanted family member is enough to upset the effective implementation of a solution" (p. 520).

The use of the extreme couple or family score may be helpful if (a) extreme reactions are to be noted; (b) variationsamong family scores are high and a mean score is undesirable; or (c) it fits with the conceptual model being employed.

Difference ScoresDiscrepancy or congruence scores reflect the difference among family members on a measured dimension and have been

used to predict a variety of outcome variables. For example, greater discrepancies have been associated with lesssatisfaction (1, 4), more conflict (3), and more stress (22). Relational discrepancy scores can be calculated as the simplearithmetic difference among the individual scores or the "relative balance" in the scores as computed by a ratio (spousescore/husband + wife score (27).

Discrepancy scores can be conceptually meaningful, but they raise several interesting methodological problems. First,difference scores do not reflect score level. For example, a discrepancy of 10 points could occur between spouses anywherealong a scale with a 50-point range. There may be profound differences, however, between a couple with a 10-pointdiscrepancy score between spouse scores of 38 and 48 and another couple with a similar 10-point discrepancy score butwhose spouse scores were 4 and 14. A simple discrepancy score format reflects the fact that differences between spousesare similar all along the entire range of the scale but may be conceptually misleading.

Second, because discrepancy or delta scores tend to be less reliable than any contributing score, their distributionscontain less variance, and some degree of attenuation of the distribution occurs. Because such distribution characteristicsoften reduce the possibilities for reaching statistical levels of significance, many statisticians recommend scoretransformations or weights when using a discrepancy score format. Cohen and Cohen (5) offer a more complete discussionof the use of raw difference scores.

Last, in and of themselves discrepancy scores do not contain additional information above and beyond what is alreadypresent in the correlations between separate family member scores and a dependent variable. For example, suppose wegathered husbands' and wives' scores on some measure of family life as well as on a dependent measure. We could thencompute three separate correlations: husband's score with dependent measure, wife's score with dependent measure, andthe sum of husband's and wife's score with the dependent measure. If we were then to compute a correlation between thespouse's discrepancy score and the dependent measure, that correlation would not add additional information to thevariance accounted for by the three correlations already computed. Mathematically, the correlation utilizing the discrepancyscore would tell us nothing in addition to what we already know when using the individual scores. This is of particularconcern in multiple regression or other multivariate analyses where redundant predictors add nothing to the R yet reducethe degrees of freedom.

To circumvent some of these difficulties, several authors use the correlation between spouses' scores instead of rawdifference scores (e.g., 31, 33) in analyzing their data. The rationale is that high correlations imply low discrepancy oragreement and low correlations imply high discrepancy or disagreement. This rationale, however, may be more apparentthan real unless a careful review of the scatterplot is undertaken prior to the computation of the correlation. For example,suppose a correlation is computed between spouse scores on a measure of family life with a range of zero to 100. The naïveinvestigator obtains a high correlation and assumes general agreement between spouses. A subsequent review of thescatterplot, however, reveals that wives' scores varied from 60 to 80 and husbands' scores from 10 to 30, indicatingsignificant disagreement between spouses. The obtained high correlation reflected only the degree of covariation within therespective range of husbands' and wives' scores and, in this case, did not reflect the degree of couple agreement. Such amisleading outcome can occur particularly when a moderating variable, such as sex of spouse or employment of spouse,interacts with the dependent variable under consideration.

An exemplary use of difference or congruence scores is found in Bell and Bell (2). These authors looked at the effects ofscapegoating and cross-generational coalitions on female adolescent maturity. They operationalized the first two variablesinto relational scores by calculating the differences between mother, father, and daughter responses on summed individualitems from the Family Environment Scale. They then obtained ratios of difference or congruence by comparing the threepossible pairs of difference scores between parents and daughter. Thus, a "scapegoating score" was calculated bycomparing the difference scores of the adolescent from both parents relative to the difference between the parents: (HA +WA)/HW. This score was used to represent the adolescent's isolation from both parents. The more distant she was fromboth parents relative to the interparent distance, the higher the scapegoating score. Similarly, a "coalition score" wascalculated by finding the imbalance in the child's difference from each parent relative to the spousal difference (HA -

_____________________________________________________________________________________________________________

4

WA)/HW). The closer the adolescent was to one parent relative to her distance from the other parent, the higher hercoalition score (p. 525).

A classification scheme of families and adolescents was then constructed by recomputing the three family memberdifference scores (HW, HA, WA) as a percentage of the total differences among scores (HW + HA + WA). Families inwhich the smallest distance was within 10 percentage points of the largest distance were classified as "balanced." Theremaining families were classified as "scapegoating" (the adolescent was farther from both parents than parents were fromeach other), "coalition" (the adolescent was closer to one parent than the parents were to each other), or "adolescent close"(the adolescent was closer to both parents than the parents were to each other). This classification scheme proved fruitful inaccounting for differences in the level of adolescent maturity based on other measures in this thoughtful study.

Combined ScalingIt is often desirable to combine the magnitude of each family member score with the differences between scores into a

single or combined score without the use of multivariate procedures. Possibilities begin with weighting the mean couplescore by the discrepancy between spouse scores or, alternatively, weighting the discrepancy score by the mean couplescore. Second, simple arithmetic addition or subtraction of these two relational scores, with or without weights, can beused. Third, families can be statistically equated on discrepancy when mean scores are being used or equated on meanscore level when discrepancy scores are being used through covariance analysis, partial correlation, or other similartechnique. Any such choice depends, of course, on the specific question being investigated and the nature of the measuresand the contributing scores.

When the investigator has need to account for differences between family members at various levels of the scale, inaddition to level and disagreement, a simple contingency table approach may be helpful. For example, let us assume that asample of spouse scores on a marital satisfaction inventory is collected. Both the magnitude of each spouse's score (level ofsatisfaction) and the difference between spouse scores ("satisfaction gap") are of interest. The investigators also feel thatdiscrepancies of equal size at different points along the satisfaction scale express different meanings, and they wish toaccount for them in their analysis. Can these three factors be accounted for in a single score?

One solution is to construct a contingency table accommodating both level and discrepancy. Assume a large sample isinvolved. We divide the couples into three categories on each dimension and classify them into a subsequent 9-fold tablefor comparison with a dependent variable (see Table I).

Table IA Contingency Table Format

DISCREPANCY SCORE

High Medium Low

LEVEL ON SCALE Both High a b c

Mixed d e f

Both Low g h i

This tactic is useful in partitioning couples into convenience types for further analysis. The rationale is the same thatbehind efforts to construct more fundamental family types based on a combination of variables or styles, such as Olson'scontingency patterning of "cohesion" and "adaptability" in the Circumplex Model (e.g., 25). Although comparisons andsubsequent analyses are no longer based on continuous-score data (from the original quasi-interval scale) and one runs therisk of obtaining empty cells, combining information from two or three dimensions into a single category highlightssubgroups of couples that might be of special interest. For instance, couples in cell "c" might be proposed to be at oppositeextremes from couples in cell "g" in terms of risk for divorce.

There are two related drawbacks to this approach. First, in the technical sense, important information can be lost whencontinuous data are reduced to categories. Second, this "lost information," because of a restriction of range, reduces powerand the chances of obtaining statistical significance utilizing samples of equal size. There may be meaningful differencesamong couples that go undetected because the original scores were "condensed" into narrowly defined categories.

In sum, one needs to be particularly aware of the conceptual and statistical problems involved in a combined scalingapproach. Experimentation with various possibilities before settling upon a specific choice is encouraged. The nominationof conjoined interval data has the virtue of simplicity in execution and in communication of results, which in somecircumstances may be reason enough to favor such an approach over the multivariate methods to be discussed next (see 5,p. 525).

_____________________________________________________________________________________________________________

5

Multivariate ApproachesMultivariate statistical procedures provide a range of options for combining family scores beyond the approaches

described above. These can be broken down into two general types: those that permit accounting for score level, difference,and order, using a multiple regression format, and those that permit grouping families on the basis of patterns of members'scores.

Multiple regression procedures permit the entrance of level, discrepancy, and order scores into an equation to predict adependent or criterion variable. Stepping all three scores into the equation permits the independent contribution of each tobe assessed and submitted separately so that any shared variance is not duplicated. Consequently, level and discrepancyscores contribute to the resulting correlation to the degree that each is relevant and important to a particular dependentvariable. A problem with using this method is that, although it accounts for the quantitative properties of the family'sscores, it is atheoretical. Based on an empirical model alone, the data themselves determine the relative contribution of eachcomponent. Also, the relative proportion of the contributing variance of these scores may vary both from sample to sampleand as a function of the particular dependent variable used. In some sense, this may be considered a strength of the methodin that it reflects the actual empirical relationships within the data. On the other hand, there may not be a theoreticalexplanation for why a score was a significant predictor in one situation and not in another. This becomes a strategic issue,however, reflecting one's approach to a methodological problem. With all multivariate methods, sample size and theavailability of replication samples are of utmost importance.

Modifications of basic multiple regression formats have led to the creation of other innovative strategies for creatingrelational data. For example, Draper (9) described a technique called scatterplot regression. Rather than use a least-squarefit to a given line as in traditional regression analysis, he suggested utilizing a "least-square fit between the appropriatepoints of two standardized bivariate normal distributions which share a common point of origin." Briefly, Draperrecommended standardizing husbands' and wives' scores separately, plotting the wives' scores along the y axis and thehusbands' along the x axis, and utilizing the score along the line drawn between the two spouse scores that is closest to theorigin of the two axes. This score, he suggested, "preserves the unique information about each couple" and enables theresearcher to extend the model by overlaying the grid onto several couple scores at the same time.

Multiple regression methods are easiest to use and most straightforward in correlational studies. In studies in whichdiscrete groups are directly compared (e.g., treatment vs. nontreatment), however, covariance procedures, multiple analysisof variance, and other techniques are more appropriate, and some modification of strategy is required. In sum, multipleregression permits the independent inclusion of data on level, discrepancy, and order scores and should be seriouslyconsidered for use in a variety of settings.

Cluster, or factor analytic, techniques provide a second multivariate strategy by classifying couples as units based onindividual family members' scores. Although somewhat dependent upon the type of technique and the algorithm selected,similar couples in the sample can be grouped on the basis of the pattern or relationship among their level, discrepancy, ororder scores. Such groupings create so-called "true types" or categories of couples or families that tend to be useful inhypothesis-generation research as opposed to hypothesis-testing research. For example, although not employingmultivariate statistical procedures per se, Peachey (28) classified families on the basis of the frequency and pattern offamily members' reports of illness, as culled from medical records. Surveying 21 families, she defined four family types andsuggested relating these types to a variety of structural and social aspects of family life.

Cluster, or factor analytic, techniques are atheoretical and based on an empirical model with the data alone dictating theoutcome. One runs the risk that the differences found among classified family groups may not make conceptual sense, sincethey are determined exclusively by the statistical properties of the data. In addition, relatively large samples are required forreplication, and the classification schemes may vary across different samples. The procedures are statistically complex andrequire extensive experience with multivariate methods.

New statistical procedures recently have enabled researchers to determine if their data fit an a priori model or theory.One could conceivably circumvent the atheoretical nature of multivariate methods, as discussed above, through the use ofconfirmatory factor analysis and structural modeling. This is a complex procedure that is beyond the scope of the presentpaper. The reader is referred to Kenny (16) for a complete discussion of this method.

The use of multivariate techniques provides at least a partial solution to the dilemma presented by many of the othermethods described above. By working with configurations of scores or score patterns rather than the scores themselves andby stepping to a higher level of abstraction, one creates a potentially powerful alternative research strategy.

Multiple Empirical MethodsOne last strategy for combining family member data is promising, although it should be employed with caution: using

several of the approaches listed above on the same data and seeing how each turns out. For example, we might find thatusing a discrepancy score with a given scale yields higher correlations with a particular health outcome dependent variable

_____________________________________________________________________________________________________________

6

than does a couple mean or a combined scaling format. Consequently, we might choose to use such a discrepancy score as amore "reliable" correlate in future studies. Such an approach is logical, and it provides a degree of flexibility in dealing withthe data. Also, in some studies several approaches may yield similar results, indicating that the type of technique used forcombining spouse data is unimportant. The need for a labored decision as to which method to employ can thereby beaverted.

On the other hand, such an empirical, "shotgun" approach may yield outcomes idiosyncratic to the specific data set onhand. That is, the outcome may change when another sample of data is collected and analyzed. Another drawback is that itnot a theoretically grounded approach to the relationship between independent and dependent variables. Moreover, it maynot be clear why statistical significance is reached when using one technique and not reached when another is used, perhapsadding confusion and uncertainty to the analysis. With this in mind, multiple empirical methods might best be used forpurposes of exploration after initial analyses have been designed and executed.

ConclusionsA variety of approaches have been described for creating relational data that represent the couple or family as a unit. We

do not propose that the groups of methods described exhaust all possibilities, but they do provide an outline for thinkingabout this important issue.

It is apparent that the selection of a method for combining spouse or family data should not be an arbitrary process. Ourexperience indicates that no one method is applicable in every case and that selection depends upon the scales employed,the specific question being addressed, and the predictive utility of the procedure. There is no "correct" or single "best" wayto combine individual family member scores into a relational measure. We emphasize the need to think through the issuesat hand clearly before selecting any method, because results may vary considerably as a function of the formula employed.

As the family literature expands with more and more studies using objective, empirical measurement, theoretical issuesof level of assessment (e.g., individual vs. relational vs. transactional) and technical issues of type of assessment (e.g.,self-report, family interaction, etc.) will assume increased emphasis. Because family research is based on the measurementof an elusive, multi-individual, transactive unit, issues of measurement theory and technique will require continueddiscussion and rethinking. We see the problems and issues in creating relational data as just one facet of this continuing andchallenging process.

REFERENCES

1. Arias, I. and O'Leary, K. D., "Prediction Accuracy, Perceived Similarity and Actual Similarity in MaritalSatisfaction," Paper presented at the 15th Annual Convention of the Association for Advancement of BehaviorTherapy, Toronto, Canada, November, 1981.

2. Bell, L. G. and Bell, D. C., "Family Climate and the Role of the Female Adolescent: Determinants of AdolescentFunctioning," Fam. Relations, 31, 519-527, 1982.

3. Billings, A., "Conflict Resolution in Distressed and Non-distressed Married Couples," J. Clin. Consult. Psychol.,47, 368-343, 1979.

4. Birchler, G. R. and Webb, L. J., "Discriminating Interaction Behaviors in Happy and Unhappy Marriages," J.Consult. Clin. Psychol, 45, 341-343, 1977.

5. Cohen, J. and Cohen, P., Applied Multiple Regression/Correlation Analysis for the Behavioral Sciences,Hillsdale, N.J., Lawrence Erlbaum, 1983.

6. Cronbach, L. J., Essentials of Psychological Testing, New York, Harper, 1960. 7. Dewey, J. and Bentley, A., Knowing And The Known, Boston, Beacon Press, 1949. 8. Doane, J. A., "Family Interaction and Communication Deviance in Disturbed and Normal Families: A Review of

Research," Fam. Proc., 17, 357-373, 1978. 9. Draper, T. W., "On the Analysis of Dyadic Data: Proposal of an Innovation," Paper presented at American

Psychological Association Meetings, Washington, D.C., 1982. 10. Fisher, L., "Transactional Theories but Individual Assessment: A Frequent Discrepancy in Family Research,"

Fam. Proc., 21, 313-320, 1982. 11. Guilford, J. P., Fundamental Statistics In Psychology and Education, New York, McGraw Hill, 1956. 12. Haley, J., "Family Experiments: A New Type of Experimentation," Fam. Proc., 1, 265-293, 1982. 13. Herbst, P. G., "The Measurement of Family Relations," Hum. Relations, 5, 3-36, 1952. 14. Hodgson, J. W. and Lewis, R. A., "Pilgrim's Progress III: A Trend Analysis of Family Theory and Methodology,"

Fam. Proc., 18, 163-174, 1979. 15. Jacob, T., "Family Interaction in Disturbed and Normal Families: A Methodological and Substantive Review,"

_____________________________________________________________________________________________________________

7

Psychol. Bull., 82, 33-65, 1975. 16. Kenny, D. A., Correlation and Causality, New York, John Wiley, 1979. 17. Klein, D. M. and Hill, R., Determinants of Family Problem-Solving Effectiveness, in W. R. Burr, R. Hill, F. I. Nye

and I. L. Reiss (eds.), Contemporary Theories About the Family, vol. I., New York, Free Press, 1979. 18. Kosa, J., Alpert, J. J. and Haggerty, R. J., "On the Reliability of Family Health Information: A Comprehensive

Study of Mothers' Reports on Illness and Related Behavior, Soc. Sci. Med., 165, 48-69, 1967. 19. Litman, T. J., "Health Care and the Family: A Three-Generational Analysis," Med. Care, 9, 67-81, 1971. 20. Mead, G. H., The Nature of the Past, in J. Gross (ed.), Essays in Honor of John Dewey, New York, Holt, 1929. 21. Mishler, E. G. and Waxler, N. E., Interaction In Families, New York, John Wiley, 1968. 22. McCubbin, H. I., Joy, C. B., Cauble, A. E., Compeau, J. K., Patterson, J. M. and Needle, R. H., "Family Stress and

Coping: A Decade Review," J. Marr. Fam., 42, 855-871, 1980. 23. Moos, R. H. and Moos, B. S., "A Typology of Family Social Environments," Fam. Proc., 15, 357-371, 1976. 24. Olson, D. H. and Cromwell, R. E., "Methodological Issues in Family Power,"in R.E. Cromwell and D.H. Olson

(eds.), Power in Families, Beverly Hills, Calif., Sage, 1975. 25. Olson, D. H., Sprenkle, D. H. and Russell, C., "Circumplex Model of Marital and Family Systems: I. Cohesion

and Adaptability Dimensions, Family Types, and Clinical Applications," Fam. Proc., 18, 3-28, 1979. 26. Olson, D. H., Russell, C. S. and Sprenkle, D. H., "Circumplex Model of Marital and Family Systems: VI.

Theoretical update," Fam. Proc, 22, 69-83, 1983. 27. Olson, D. H. and McCubbin, H. I., Families: What Makes Them Work, Beverly Hills, Calif., Sage Publications,

1983. 28. Peachy, R., "Family Patterns of Illness," G.P., 27, 82-89, 1963. 29. Reiss, D., The Family's Construction of Reality, Cambridge, Mass., Harvard University Press, 1981. 30. Spiegel, J. and Papajohn, J., Transactions: The Interplay Between Individual, Family and Society, New York,

Science House, 1971. 31. Stephen, T. and Markman, H., "Assessing Development of Relationships: A New Measure," Fam. Proc., 22,

15-5, 1983. 32. Straus, M. A. and Tallman, I., SIMFAM: A Technique for Observational Measurement and Experimental Study of

Families," in J. Aldous (ed), Family Problem Solving, Hinesdale, Ill., Dryden Press, 1971. 33. Van der Veen, F., Huebner, B., Jorgens, B. and Neja, P., "Relationships Between the Parents'Concept of the

Family and Family Adjustment," Am. J. Orthopsychiatr, 34, 45-55, 1964. 34. Weimer, S. R., Hatcher, C. and Gould, E., "Family Characteristics in High and Low Health Care Utilization," Gen.

Hosp. Psychiat., 5, 55-61, 1983. 35. Wynne, L. C., Singer, M., Bartko, J. and Toohey, M., "Schizophrenics and Their Families: Recent Research on

Parental Communications," in J. Tanner (ed.), Developments in Psychiatric Research, Kent, England, Hodden andStroughton, 1977.

Manuscript received February 29, 1984; Revisions submitted August 30, 1984; Accepted November 21, 1984.

1In an earlier paper (10), we suggested that reports made by individual family members taken out of context of the family systemshould be seen as "relational" family assessment (p. 316). In light of our current thinking, however, we wish to modify this earlierview by suggesting that single family member data, as defined above, should be subsumed under the conceptual umbrella ofindividual level assessment.

_____________________________________________________________________________________________________________

8