Toward a general theory of diary panels.

This article represents an integration of findings reported in seven articles on diary panels that the authors have published, based on the AT&T database. There are four major issues involved with diary panels, to wit: respondent noncooperation involving item nonresponse and attrition, bias leading to estimation errors, mathematical artifacts involving regression toward the mean, and the conditioning effect of being observed. An integrated conceptual framework for diaries is advanced, consisting of five independent variables (questionnaire design, length of participation, level of aggregation, duration of usage, and defining usage rate with length used), four mediating variables (degree of difficulty, participation fatigue, natural mortality, and level of involvement), and six dependent variables (item nonresponse, attrition, accuracy, regression toward the mean, the conditioning effect, and estimation regression). This attempt at a general theory of diary panels is admittedly incomplete, but is meant to serve as a useful conceptual framework for further research on longitudinal studies involving record keeping and reporting.

Averaging to minimize or eliminate regression toward the mean to measure pure experimental effects.

This paper explains how regression toward the mean can contaminate diary data, making it difficult to measure the pure effects of an experimental variable over time. Using a large scale real-life database collected by AT&T, a method of measuring this mathematical artifact is advanced. It is shown to manifest very quickly as a result of a spontaneous reaction toward happenstance, with the most extreme initial values gravitating most toward the mean. Then averaging over longer and longer periods of time to define use categories is shown to dilute happenstance increasingly, and therefore progressively minimizes or eliminates regression toward the mean. Finally, regression toward the mean is very pervasive and very persistent.

Social desirability bias in diary panels is evident in panelists' behavioral frequency.

The purpose was to analyze a large-scale database, a national sample of 1108 heads of households collected by AT&T, to show that behavioral frequencies of the activities of consumer diary panelists may regress toward the population mean during the diary-keeping period given social desirability bias produced by the conditioning effect of keeping diaries. This effect is distinguished from regression toward the mean, a statistical phenomenon reflecting happenstance of extreme initial values. Social desirability bias is demonstrated in two ways-by observing decreasing coefficients of variation over time and by detecting a greater proportion of panelists' behavioral frequencies moving toward the population mean than moving away from it. Both cannot be explained by regression toward the mean. Social desirability bias was manifest only during the early stages of the diary-keeping period and only for activities high in involvement. The presence of social desirability bias in diary panels implies that, when people are subjected to observation, diary observations may be contaminated, leading to the mistaken impression that the population is more homogeneous than it actually is. Thus it is important for researchers to monitor the diary panel carefully to detect social desirability bias when engaged in longitudinal studies.

Empirical tests on regressive estimations: implications on surveys and diaries.

Using a large real-life corporate database initially consisting of 3,990 heads of households stratified on the basis of various demographic and geographic variables, and whose communication activities (long distance telephone calls, letters, cards, and visits) were surveyed and monitored, this study investigated the direction as well as magnitude of estimation errors in survey responses and diary entries. Supporting the 1994 Fiedler and Armbruster psychometric formulation and conjecture, we show that estimation errors in reports of the frequency and duration of people's own communication activities exhibit a consistent tendency to regress toward the mean. This regressive estimation is greater for those who are further away from the mean in actual behavior and is proportional to the actual deviation from the mean. Furthermore, this regressive estimation is inversely related to the average frequency across behavioral categories. An important implication of our findings is that the distribution of estimated behavioral frequencies and durations appear more concentrated in surveys than they actually are in the general population, although the general shape of the distribution is unaffected.

Exploring the dimensionality of the Originality subscale of the Kirton Adaption-Innovation Inventory.

The Kirton Adaption-Innovation Inventory, which is a widely used measure of innovative (as opposed to adaptive) cognitive individual style, is believed to have three dimensions: Sufficiency of Originality, Efficiency, and Rule/Group Conformity. Several studies have raised concerns regarding its construct validity, specifically with respect to the Sufficiency of Originality subscale. Within this subscale, exploratory factor analysis identified two distinct subdimensions, Idea Generation and Preference for Change. In this study, we used a sample of 356 household participants (with an average age of 56 yr., average income of 39,700 dollars, and average of 15 yr. of education) from the Arkansas Household Research Panel. We then employed Bollen and Grandjean's approach based on confirmatory factor analysis to assess whether there are actually two distinct subdimensions instead of one. Our study shows that within the Sufficiency of Originality subscale, there are indeed two distinct subdimensions, Idea Generation and Preference for Change. Further analyses indicate that dropping double-loaded items identified through exploratory factor analysis significantly improves the fit statistics. Also, allowing correlated errors for the measurement items that belong to the same subdimension can also significantly improve the overall fit of the model based on chi-square statistics.