Introduction to the special issue

Design of the special issue: Researcher questions and expert answers

For this special issue, a call was posted to the listserv of the International Association for Relationship Research, asking researchers to submit specific data analytic questions/topics that they would like to see addressed in the special issue. From the questions submitted, we matched five to area experts and commissioned these experts to write an article on the topic. Authors were allowed to reword and expand the questions as appropriate for the scope of their papers. They were requested to review the literature on the question but also to give a strong perspective on how the question should be answered. Finally, authors were asked to write their articles at a level that would be accessible to the typical relationship researcher and to provide sample data and code so that readers could replicate the analyses described (this information is sometimes located in the paper itself and other times posted online as supplemental material). The manuscripts were then submitted for extensive peer reviews by other experts in each of the relevant areas.

In the first article, Lane and Hennes address how to conduct a power analysis for multilevel modeling of multiple time points and/or multiple individuals (e.g., dyads). They answer the questions: What information do you need to run a power analysis, how do you conduct the power analysis, what aspects of the study have the largest effect on power, and what do you do if you do not have all the needed information? The authors point out that because studies that use dyadic and/or longitudinal data are resource demanding, such studies are particularly vulnerable to being underpowered. They argue that a proper power analysis is essential for good decision-making about how to invest research resources. For example, a power analysis could clarify whether a dyadic or longitudinal study design is practical with the available resources (i.e., to avoid running an expensive study with low power to confirm a true hypothesis). A power analysis could also reveal ways in which aspects of the study could be altered to maximize power (e.g., recruiting more participants versus assessing more time points). In their paper, Lane and Hennes describe a Monte Carlo–based simulation approach to power analysis that handles a wide variety of configurations of multilevel models, and they provide syntax to conduct this analysis in a number of statistical packages.

Next, Castro-Schilo and Grimm discuss how to study change over time, commonly analyzed either as difference scores (Y _T2−Y _T1) or as residualized change (Y _T2 = b ₀ + b ₁ × Y_T1). They address the questions: Do difference score and residualized change approaches produce the same results? Which is more appropriate, under what circumstances? Why do difference scores have a bad reputation, and is it deserved? Are there additional analytic strategies that might be preferred over these two options? Castro-Schilo and Grimm argue that the two analytic approaches differ in their assumptions and consequently in the appropriate circumstances for their use. Specifically, they propose that the difference score approach is most appropriate when people who score differently on the variable used to predict change are from different populations (i.e., when there are group differences in the outcome variable at the first time point), whereas the residualized change approach is most appropriate when they come from the same population (i.e., when groups are the same on the outcome variable at the first time point). Using the example of cohabitation as a predictor of change in relationship satisfaction over two time points, Castro-Schilo and Grimm argue that the latter condition (favoring residualized change) holds for proper experiments (where groups are randomly assigned) but that for correlational research (where groups such as cohabitors versus non-cohabitors are preexisting), the former condition (favoring difference scores) is more likely to hold. Consequently, they argue that difference scores are preferable in correlational research should be more widely used than they currently are. To address concerns about the unreliability of difference scores, the authors favor a latent change difference score approach, which they describe and simulate with data.

Iida, Seidman, and Shrout discuss methods to analyze dyadic data, highlighting the utility of the dyadic score model (DSM), an alternative to the popular actor–partner independence model (APIM). They address the questions: What are alternative dyadic analyses to the APIM and how do you know when to use which? What happens if we want to examine relationships at the dyad level? The authors explain that although APIM uses dyadic data, it is fundamentally an individual-level approach (i.e., it models the unique effects of each individual on the self and the other in a dyad). They contrast the APIM approach with the DSM approach, which converts the scores of the individuals into two dyad-level scores: the sum of and the difference between the two partners on the same variable. Iida, Seidman, and Shrout emphasize that because APIM and DSM are alternative yet mathematically equivalent parameterizations, the two models will always fit the data equally well and therefore theoretical concerns should dictate which is used. They argue that researchers should use APIM when the variables and processes of interest occur at the individual level but DSM when they occur at the dyad level. The authors encourage researchers to think carefully about whether their theorizing suggests individual versus dyadic processes. Iida and colleagues provide data and syntax to conduct and compare both APIM and DSM models as well as other models that use dyadic data.

Stern and West describe methods to assess accuracy in dyads. They highlight the utility of the Truth & Bias (T&B) Model in the context of traditional discrepancy score and correlational accuracy measurement approaches. They address the questions: What types of accuracy can researchers assess in close relationships, and can these different types be examined in the same statistical model? How can researchers examine both accuracy and bias in perceptions of relationship partners? Stern and West overview two traditional approaches to measuring accuracy. The first—the mean level bias approach—uses discrepancy scores to index accuracy, and the second—the correlational approach—uses correlations/regressions to index accuracy. They highlight that the two approaches answer different research questions, and for each approach, they offer syntax for assessing accuracy at the sample level and individual level. They also discuss how multilevel modeling can be used to assess accuracy when there are multiple repeated measurements over time per dyad. Stern and West next overview the T&B model. They show how it encompasses both the mean-level bias and correlational approaches to measuring accuracy and describe how it additionally offers the opportunity to model predictors (i.e., moderators) of bias. The authors provide data and syntax for analyses using the mean level bias approach, the correlational approach, and the T&B model.

Rogers, Wood, and Furr discuss methods to measure similarity or agreement in dyads, with an emphasis on how to deal with the confounds associated with difference scores and correlations. They address the questions: How do the various ways of assessing similarity or agreement between members of a dyad differ? What are the most and least appropriate uses of each? The authors argue that difference scores and correlations, the most widely used indexes of correspondence in dyads, suffer from important confounds. On balance, Rogers and colleagues favor correlational over difference score approaches. When data are collected on one attribute of each dyad member, they recommend moderated regression analysis to study correspondence and its predictors. The authors additionally recommend response surface analysis: a newer and richer approach based on the same principles. For data collected on multiple attributes of each dyad member, the authors recommend profile correlations corrected for normative confounding (i.e., “distinctive” correspondence correlations). In the case where each person in the dyad serves as both perceiver and target, they recommend researchers to use the social accuracy model. They illustrate their arguments using data from a romantic couples data set and provide the data and syntax for all examples used in the paper.

There are many areas of overlap between the topics and viewpoints covered in this issue. Across papers, authors highlight the special challenges of collecting relationship data as well as the analytic challenge of collecting multiple observations, over time, across individuals, or both. The usefulness of MLM and SEM is frequently emphasized and a common theme is the importance of matching the analytic technique chosen to the theoretical question posed. There are also areas of non-overlap in the authors’ perspectives, and we think it is important to draw readers’ attention to one notable example: the authors’ view on the utility of difference scores. Castro-Schilo and Grimm as well as Iida and Shrout endorse strongly positive views of difference scores, whereas Rogers, Wood, and Furr endorse much more negative views. Stern and West fall in the middle, recognizing both the strengths and weaknesses of difference score approaches. While it has been recommended historically that the use of difference scores always be avoided (e.g., Cronbach & Furby, 1970), several methodologists have more recently demonstrated conditions under which difference scores are both valid and preferable (Gollwitzer, Christ, & Lemmer, 2014; Griffin, Murray, & Gonzalez, 1999; van Breukelen, 2013; Williams & Zimmerman, 1996), and we encourage interested readers to consult this work.

In closing, we thank all the authors and reviewers who played a part in this special issue. We believe that the pieces that follow will be accessible and useful resources to relationship researchers as they continue to work with data reflecting various types of interdependence and requiring sophisticated data analytic techniques.