For the effect of siblings on the exposure to adult speech, on the exact same data, the two algorithms produce inconsistent results without calibration. After calibration, their estimates are consistent. In addition, classification errors lead to 20-80% underestimation of the actual effect size!

To resolve the issue, we propose a Bayesian calibration approach relying on graphical models adequately capturing the behavior of the speech processing algorithm. The model is trained on validation data (human annotations serving as ground truth)

However, these algorithms make errors that introduce biases in downstream analyses by opening up biasing paths. For instance, the effect of siblings on the qty of speech heard by infants from adult is confounded by speech from siblings incorrectly classified as adult speech🫣!

In child-development research 👶, scientists have been using two algorithms (LENA and VTC) to detect and classify speech in child-centered longform recordings. These classify vocalizations according to four categories: the child wearing the device (CHI), other children (OCH), male/female adults.

Finally, I infer how scientists resolve conflicting preferences about which convention to use in co-authored publications. I find evidence that most often, the last-author's preference prevails: physicists appeal (implicitly) to leadership/seniority to solve this coordination problem

Interestingly, the Ising model (although unrealistic) can provide information about the actual process through which individuals form preferences: the plausibility Agent-based models of preference formation can be compared using the magnitude of local & global coordination as summary statistics!

I find that scientists' preferences are more aligned to those of their co-authors than expected by chance. Using the Ising model, I show that such alignment simultaneously arises through both local processes (endogenous to the social network) and global factors transcending the network structure.

Then, I investigate the social coordination of individual preferences using the Ising model.

I show that the Ising model can retrieve information about the underlying coordination game and the network infrastructure involved in the alignment of individual preferences.

Using a statistical physics approach, I establish that scientists have preferences, i.e. they tend to stick to their favorite choice across their publications. However, I show that they although they occasionally adapt to the target research area (the context) by deviating from their prefernce.

To illustrate these trade-offs, I study a sign convention from high-energy physics ("the metric signature"). This involves a choice between two equivalent options (the mostly plus and mostly minus metric signatures). I use textual/authorship/citation data from Inspire HEP and arXiv.

I identify three universal trade-offs affecting:
(I) individuals' decision-making in relation to competing conventions,
(II) the propagation of norms, and
(III) the resolution of conflicts.

I recently gave a lecture on inverse problems for computational philosophers and agent-based modellers interested in drawing connections with empirical data. The slides can be found here: lucasgautheron.github.io/assets/pdf/2...

Balancing Specialization and Adaptation in a Transforming Scientific Landscape
epjdatascience.springeropen.com/articles/10....

The latest inference suggests an unusually precise distance measurement for this event. Is this uncertainty estimate reliable?

Sauf que précisément, la définition de ce qui constitue l'électorat de la gauche est en train de changer.