A new class of models for rating data - Marica Manisera, Paola Zuccolotto, September 4, 2013
0 likes342 views
The document discusses a new class of nonlinear cub models designed for rating data, focusing on the intricacies of modeling cognitive mechanisms such as feeling and uncertainty in responses. It presents the relationships between different parameters and illustrates the model's application with survey data on job satisfaction across European Union member states. The authors conclude with insights on improving the models and opportunities for further research.
A new class of models for rating data - Marica Manisera, Paola Zuccolotto, September 4, 2013
- 1. A new class of models for rating data SYstemic Risk TOmography: Signals, Measurements, Transmission Channels, and Policy Interventions Marica Manisera & Paola Zuccolotto University of Brescia ! New Castle UK - September 4, 2013
- 2. A new class of models for rating data Marica Manisera & Paola Zuccolotto University of Brescia, Italy Newcastle UK - September 4, 2013
- 3. Introduction Rating data are often encountered, e.g. when individuals’ perceptions are measured by their observed responses to survey questions with ordinal response scales In this contribution, rating data are modelled using the new class of the Nonlinear CUB models (Manisera & Zuccolotto, 2013) introduced to generalize the standard CUB (Piccolo, 2003; D’Elia & Piccolo, 2005)
- 4. CUB models The response of each subject is interpreted as the combination of • a feeling attitude towards the item being evaluated • an intrinsic uncertainty related to the circumstances surrounding the discrete choice The feeling and uncertainty components are considered in the CUB models by a mixture of a Discrete Uniform U and a Shifted Binomial V random variables
- 5. CUB models The observed rating r (r = 1,…,m) is the realization of a discrete random variable R with probability distribution given by with thus the parametric space is the (left open) unit square feeling uncertainty
- 6. CUB models Parameters are related to the latent components of the responses: 1 − ξ feeling with the item 1 − π uncertainty of the choice There is a one-to-one correspondence between a CUB random variable and θ=(π, ξ)' each CUB model can be represented as a point in the unit square (with coordinates 1−π, 1−ξ)
- 7. CUB models The CUB models have been extended in several directions (Iannario & Piccolo, 2011) Inferential issues (specification, estimation, validation) have been obtained also for the extended CUB models A program in R is freely available (Iannario & Piccolo, 2009)
- 8. Nonlinear CUB models The probability distribution of the discrete random variable R is given by It depends on a new parameter and l is a step function mapping into feeling uncertainty
- 9. Nonlinear CUB models This formulation is derived as a special case of a general framework for the decision process, which is assumed to be composed of two approaches: - the feeling approach, proceeding through T consecutive steps (feeling path): the last rating is obtained by summarizing T consecutive elementary judgments and transforming them into a Likert-scaled rating - the uncertainty approach, leading to formulate a completely random rating The final rating is derived by the feeling or the uncertainty approach, with given probabilities
- 10. Nonlinear CUB models - Example A person is asked to rate his/her job satisfaction on a Likert scale from 1 to m=5. His/her final decision could derive from an unconscious decision process Feeling approach (probability π ): he/she asks him/herself for T=m−1=4 times “Am I satisfied? Yes or No?”. Only a positive response allows to move from one rating to the next one. At each step, a provisional rating is given by 1 plus the no. of “Yes” up to that step. At the end, the last rating of the feeling path is given by 1 plus the Total no. of “Yes”
- 11. Nonlinear CUB models Uncertainty approach (probability 1 − π ): for many reasons, the respondent gives a completely random rating This example corresponds to the decision process modelled by the CUB models (T=m−1) In the Nonlinear CUB, a number of basic judgments T>m − 1 is required since the person could need to accumulate more than one “Yes” to move from one rating to the next one
- 12. Nonlinear CUB models The l function is essential in the decision process When T=m − 1 , CUB and Nonlinear CUB coincide The number gs of “Yes” needed to move from rating s to s+1 univocally determines l g1,…, gm can be considered parameters of the model
- 13. Transition probabilities The transition probability φt(s) is the probability of moving to rating s+1 at step t+1 of the feeling path, given that the rating at step t is s, with s=1,..,m They describe the state of mind of the respondents about the response scale. In general, we consider the average transition probability φ (s)(averaging over t) In Manisera and Zuccolotto (2013) , we defined the decision process linear or nonlinear according to whether φt(s) is constant or non-constant for different t and s
- 14. Nonlinear CUB models Also, we showed that, while CUB meets a sufficient condition for linearity, Nonlinear CUB do not. This is the reason for their name (we also gave a graphical explanation) When comparing different Nonlinear CUB, ξ cannot be fairly compared, since the l function can be different we introduce a new parameter µ= expected number of one-rating-point increments during the feeling path
- 15. Nonlinear CUB models: Estimation Following CUB, for Nonlinear CUB ML estimates can also be obtained We are still working on this issue, since it is a difficult task and presents identifiability problems Nevertheless, it is possible to obtain naïve estimates resorting to the algorithm used to estimate CUB: • fix maxs(gs) • maximize logLik with respect to ξ and π constrained to all the configurations g1, …, gm • choose the best model according to a proper goodness-of-fit criterion Results of simulations are encouraging
- 16. Case study Data come from Standard Eurobarometer 78, a sample survey covering the national population of citizens of the 27 European Union Member States (the average number of interviewees over the 27 Countries is 986) Focus on one question: How would you judge the current situation in your personal job situation? with possible responses: 1=“very bad” 2=“rather bad“ 3=“rather good“ 4=“very good”
- 18. Personal job situation (different l for each Country)
- 19. Conclusions and future research Results show how the Nonlinear CUB models allow us to model rating data resulting from cognitive mechanisms with non-constant transition probabilities The inferential issues of the Nonlinear CUB models are now our main challenge Also, the Nonlinear CUB models could be extended: • to include subjects’ covariates to relate feeling to respondents’ features • in a multivariate framework
- 20. Some references D'Elia A., Piccolo D. (2005) A mixture model for preference data analysis. Comput Stat Data An, 49, 917-934. Iannario M., Piccolo D. (2009) A program in R for CUB models inference, Version 2.0, available at http://www.dipstat.unina.it/CUBmodels1/ Iannario M., Piccolo D. (2011) CUB Models: Statistical Methods and Empirical Evidence, in: Kenett, R. S. and Salini, S. (eds.), Modern Analysis of Customer Surveys, Wiley, NY, 231–254. Manisera M., Zuccolotto P. (2013) Modelling rating data with Nonlinear CUB models, Submitted. Piccolo D. (2003) On the moments of a mixture of uniform and shifted binomial random variables, Quaderni di Statistica, 5, 85– 104.
- 21. ! ! ! This project is funded by the European Union under the 7th Framework Programme (FP7-SSH/2007-2013) Grant Agreement n° 320270 ! ! ! ! ! www.syrtoproject.eu ! ! This documents reflects only the author's view. The European Union is not liable for any use that may be made of the information contained therein