One motive I like Bayesian software program (comparable to JAGS, and so on.) is for its capability to precise novel fashions that are not prepackaged in canned stats packages. In some latest analysis, I had the chance to create a novel development mannequin and estimate its parameters in JAGS.
We had knowledge as graphed in Fig. 1, and I had to consider a mannequin to explain the traits
| Fig.1. Information to be modeled. |
The variable on the vertical axis is a ranking of emotion (comparable to disappointment, happiness, and so on.) briefly tales. The horizontal axis is the retelling of the story, such that 0 is the unique story, 1 is a retelling of the unique, 2 is a retelling of the first retelling, and three is a retelling of the 2nd retelling. Retellings are inclined to lose plenty of info however however retain some data too. Do they preserve feelings? Every curve within the graph corresponds to a special unique story. I assumed the traits within the knowledge appeared just like the completely different unique tales have been converging towards (or diverging from) a typical backbone, as in Fig. 2:
 |
| Fig. 2. Mannequin predictions. |
So, I invented a easy development mannequin to precise that concept, and I programmed it in JAGS. As a result of the rankings have been on an ordinal scale, I used an ordered-probit response distribution on a latent scale that adopted an underlying linear backbone with exponential convergence, as steered in Fig. 3:
 |
| Fig. 3. The mannequin, with latent scale in left panel and ranking scale in proper panel. |
You’ll be able to learn all the small print of the mannequin within the HTML doc at https://osf.io/nbuxg/ (obtain the HTML doc after which view it in a browser). The revealed article describing the analysis is titled Serial copy of narratives preserves emotional value determinations by Fritz Breithaupt, Binyan Li, and John Okay. Kruschke. It may be obtained from https://doi.org/10.1080/02699931.2022.2031906 and the ultimate pre-publication manuscript is at https://osf.io/hwvza/.
