Methodological research on changing values in society

Modelling longitudinal ranked data – social values and postmaterialism [Francis, Firth, Penn collaborating with Dittrich, Hatzinger, Katzenbeisser at WU-Vienna] (quantitative sociology, statistical modelling)

Ranked and partially ranked data are typically produced through social surveys, with respondents asked to rank a set of items into an order or partial order. However, longitudinal ranked data models have been relatively neglected. Here the correlation between responses over time needs to be taken into account, and other correlation structures between responses within a time point can also be considered. We aim to examine measures of world values and longitudinal changes. We will specifically focus on the Inglehart scale of post-materialism which is collected as a partially ranked dataset and is present as a series of questions in most social surveys, including the Euro-barometer and the BHPS.

Our focus is on the Bradley Terry model for paired comparisons.  A paired comparison experiment is where each of a set of individuals is asked to judge a number of pairs of objects, choosing which object is preferred.   We have shown that we can convert ranked data (where respondents are asked to rank a set of items) and partially ranked data (where participants are asked to rank the top two choices) into a set of paired comparison responses for each individual. While inconsistencies are allowed in a paired comparsion experiment ( A preferred to B, B to C but C to A) these are not allowed for ranked data by design, and the number of posible choices  for ranked is therefore smaller than for a paired comparison experiment.

It is also possible to consider likert data. If there are a set of likert questions , all measured on the same scale, with similar wording, then we can look at differences in likert items and analyse these differences as paired comparisons.  We have shown that the number of possible responses for likert data is again smaller than a full paired comparison experiment.

Longitudinal data is being modelled in two ways. Our first method is through a random effects formulation to account for individual specific factors over time. We have chosen a  non-parametric mass-point approach for the specification of the distribution of the random effect, which becomes a mixture of bradley terry log-linear models.  Another aproach under development with our Vienna colleagues is the inclusion of specific Markov-dependence terms for paired responses across time.