I am a PhD graduate from NYU's Wilf Department of Politics and current Postdoctoral Fellow at Vanderbilt's Data Science Institute. I specialize in Methods and Political Psychology. My primary research interest is understanding the relationship between memory organization and political attitudes. My medium: natural language. Building on the premise that patterns in natural language are informative of the content and organization of memory, in my dissertation I develop methods to model memory representations of politically charged concepts (e.g. welfare, abortion, justice) and study their role in attitude judgments. My research draws in significant ways from multiple disciplines including cognitive science, computer science, linguistics and social psychology.
In joint work with David J. Halpern (NYU, Psychology PhD candidate) we propose a framework to evaluate models of memory retrieval estimated using semantic fluency data - a task wherein subjects are provided a cue and must list associations in a limited amount of time (see working paper). To facilitate data collection we designed an RShiny App that can be used in conjunction with online crowdsourcing platforms or on mobile survey devices (soon to be made publicly available at @prodriguezsosa). Building on this work we estimate memory representations of politically charged concepts (e.g. welfare) and evaluate whether partisan differences in representations are predictive of attitude judgments as long suggested by constructivist theories of attitudes. The output of this ongoing interdisciplinary collaboration –several working papers including a proceedings publication (CogSci 2018 Proceedings)- lays the theoretical and methodological groundwork to quantitatively explore the relationship between memory organization and attitudes with important implications for long-standing literatures in political science including priming, political attitudes and survey design.
On a related project, I employ deep learning methods to estimate word representations from large collections of text. In particular my efforts have focused on extending existing methods to identify group differences in representations (e.g. PolMeth 2018 Poster) and developing new metrics to validate computer-estimated representations. This includes joint work with Arthur Spirling (NYU) in which we propose a novel take on the classic Turing test to evaluate computer-estimated representations. Preliminary results suggest current state-of-the-art word embedding models produce representations that approximate human performance (working paper).
In addition to the above projects I also work on network methods. This includes joint work with Jennifer M. Larson (Vanderbilt) exploring the role of social networks in spreading information and motivating action. In a study conducted in rural Uganda we find social networks play a role in spreading both information and behavior but in different ways (paper currently under review). In other work, also with Jennifer M. Larson, I use simulated and real network data to evaluate how the common practice of using the union of different types of social ties to study network effects can negatively bias and ultimately mask effects that are in fact present (working paper).
According to memory-based models of attitude judgments when subjects are asked to evaluate a political object they sample a limited number of relevant considerations from memory, at least some of which are valenced, and then proceed to aggregate the valences through some function and respond with sign and/or magnitude of function output. Although memory retrieval and organization are central to these models, they have not been the direct object of study. In this project we develop a methodological framework to quantitatively study memory-based models of attitudes.
We consider the properties and performance of word embeddings techniques in the context of political science research. Beyond reporting extensive technical findings, we provide a novel crowd-sourced "Turing test"-style method for examining the relative performance of any two models that produce substantive, text-based outputs.
Network data often include more than one type of tie among actors -for instance, social network data may include ties defined by blood relation, friendship, trust, past economic transactions, etc. A typical approach is to aggregate these ties, using the union of all measured networks for analyses that consider the extent to which ties in the network affect some outcome. However, aggregating ties when diffusion is network-specific and overlap across types of ties is low will negatively bias and potentially mask network effects that are in fact present.