Evaluation of social network measurement instruments
Introduction
The quality of survey data from complete social networks can be affected by many characteristics of the measurement instrument. Some of the factors that can affect response behavior, e.g., are (a) the wording of questions, (b) the response scale, (c) the data collection method, and (d) the context of the questions. The overall quality of measurement in a given network thus depends upon several factors whose impact on data quality has seldom been thoroughly analyzed and controlled. This state of affairs results from a lack of systematic research into the issues of data quality in the field of social network analysis. Research work on measurement issues mainly focuses on the questions of measurement validity, reliability, accuracy and measurement error (Wasserman and Faust, 1994, p. 56).
There are, however some studies which have focused mainly on the question of network data quality (Hammer, 1984; Sudman, 1985, Sudman, 1988; Hlebec, 1993; Brewer and Webster, 1997). These papers are especially important for the experimental design used in this study, which focuses on the test–retest stability of complete networks. Each of these studies has compared two basic data collection techniques frequently used in the survey collection of network data: free recall and recognition. In all above mentioned studies, free recall and recognition yielded different measured egocentric networks. Previous research indicated little difference (when the networks were small) between the free recall and recognition techniques in the assessment of (a) important ties, (b) most recent contacts, and (c) most frequent contacts. The recognition technique gave much better results when networks were larger and ties were weaker. One aim of this study is to test whether the recognition data collection technique is more stable than that of free recall, since differences in reliability could significantly alter the interpretation of results.
Much work has been done on topics such as respondent accuracy (e.g., Killworth and Bernard, 1976; Bondonio, 1998; Casciaro, 1998), characteristics of the measured networks (e.g., Burt, 1984; Marsden, 1987; Wellman and Wortley, 1990), comparison of the measured networks using different network generators (e.g., Bernard et al., 1987, Bernard et al., 1990; Campbell and Lee, 1991), and characteristics of the measured ties (e.g., Marsden and Campbell, 1984; Burt, 1986). Owing to the complexity of the data structure, there is still need for an extensive and systematic evaluation of survey measurement instruments in terms of test–retest reliability of measurement in the field of social network analysis. Nevertheless, in evaluating survey measurement instruments when measuring variables, there are some approaches which are applicable to social network analysis. The first such evaluation of survey measurement instruments, by two stage meta-analysis, was by Andrews (1984), who analyzed the quality of American and Canadian surveys. Together with Willem Saris and several other European social scientists, they established an international group on methodology and comparative survey research (IRMCS). Their extensive and fruitful work (Saris and van Meurs, 1990, Ferligoj and Hlebec, 1998; Ferligoj et al., 1995; Saris and Münnich, 1995; Scherpenzeel, 1995) contributed substantially to knowledge about the quality of survey measurement instruments. Their results inspired the work of Ferligoj and Hlebec (1995), Ferligoj and Hlebec (1998), who first used the multitrait–multimethod (MTMM) approach to estimate the reliability of complete network measurements. They reported that the binary scale is the least reliable scale, at least when compared with an 11-point ordinal scale or a line drawing scale, regardless of the order of presentation. In the present study, two five-point ordinal scales have also been included.
Saris and Münnich (1995) and Ferligoj and Hlebec (1998) have also reported that factors, such as the order of repetition and the time between two successive presentations of the measurement instrument, can also have substantial impact on the quality estimates. Therefore, the order of presentation and the time between presentations have been included in the experimental design of this study. Finally, the content of the network name generators — social support — was selected on the basis of the characteristics of the experimental groups: eight classes of third year high school students.
This paper presents and discusses the results from eight experiments that were designed to analyze systematically the impact of different measurement characteristics on the reliability and validity of complete network data. In the first phase of this study, estimates of test–retest reliability, validity and method effects are obtained for each set of relationships in each of eight classes, using the MTMM approach. In the second phase, the effects of the characteristics of the measurement instruments used in different classes are analyzed to explain the variability of the estimates for the reliability, validity and method effects. A secondary analysis of MTMM results is done by multiple classification analysis (MCA). The two-stage procedure described is similar to that used by Saris and Münnich (1995).
Section snippets
Experimental design
In this study, data were collected regarding social support relationships among third year students in a high school in Gimnazija Bežigrad in Ljubljana, the capital city of Slovenia. On average, there were 31 students, aged 17, in each of eight classes. Four name generators2 (traits) were used — exchange of study materials, exchange of information in the case of long-term illness, invitation to a birthday party, and discussion of important personal matters.
Mean levels of data quality
In Table 3, summary statistics for the validity and the reliability coefficients over eight classes are presented. Within each class, 12 reliability and 12 validity estimates were obtained for both the original and the reversed measures. The overall mean reliability of 12 reliability coefficients for the original questions is 0.881, and 0.882 for the reversed. The mean overall validity coefficient is 0.981 for original questions, and 0.974 for the reversed.
As is evident in Table 3, the validity
Conclusions
The results from four meta-analyses show that the domain of social support, as measured by the binary scale, is the least reliable when the recognition data collection technique is used. In contrast, when free recall is used, a smaller number of dyads is reported and all scales are equally reliable. It seems that when a full list of membership is available, it should be used in any measurement procedure to simplify the reporting task for respondents, and to increase the number of reported ties.
Acknowledgements
The authors are grateful to the members of IRMCS for their critical comments and especially to Willem Saris and Brendan Bunting for their useful suggestions on the text.
References (28)
- et al.
Comparing four different methods for measuring personal social networks
Social Networks
(1990) Predictors of accuracy in perceiving informal social network
Social Networks
(1998)- et al.
Name generators in surveys of personal networks
Social Networks
(1991) Seeing things clearly: social structure, personality, and accuracy in social network perception
Social Networks
(1998)Explorations into the meaning of social network interview data
Social Networks
(1984)Experiments in the measurement of the size of social networks
Social Networks
(1985)Experiments in measuring neighbour and relative social networks
Social Networks
(1988)Construct validity and error components of survey measures: a structural modeling approach
Public Opinion Quarterly
(1984)- et al.
How much of a network does the GSS and RSW dredge up?
Social Networks
(1987) - Brewer, D.D., Webster, C.M., 1997. Patterns in the recall and recognition of friends by residents of a university...
Network items and the general social survey
Social Networks
A note on sociometric order in the general social survey network data
Social Networks
Cited by (67)
Social networks, secondary syringe exchange, and opioid agonist therapy retention among people who inject drugs in Hartford, CT
2024, International Journal of Drug PolicyFriends, family, and family friends: Predicting friendships of Dutch women
2022, Social NetworksCollecting survey-based social network information in work organizations
2022, Social NetworksCollecting large personal networks in a representative sample of Dutch women
2021, Social NetworksModeling policy networks: The case of Natura 2000 in Croatian forestry
2019, Forest Policy and Economics
- 1
E-mail: [email protected].