Sensitivity analysis on a model the social consequences of testimonial norms

This repo contains a python implementation of Kevin J. S. Zollman's "Science Cliques" agent based model. The original model was published in Zollman, Kevin J. S. “Modeling the Social Consequences of Testimonial Norms.” (Zollman, Kevin J. S. “Modeling the Social Consequences of Testimonial Norms.” Philosophical Studies 172, no. 9 (2015): 2371–83. doi:10.1007/s11098-014-0416-7.) Original source code in netlogo can be found here with original data output files here and here (original data files in QtiPlotformat)

Final write up for this project can be read in Science_Cliques_writeup.pdf.

Note that the displayed code coverage is inflated. This was a course project that I used as an opportunity to experiment with Codecov. Most tests I wrote do not actually test anything, thus the effective coverage is lower than the coverage displayed.

Motivation

Philosophers of science use Agent Based Models (ABMs) to study how collaboration and communication network structures emerge from decisions and actions made by individual scientists. While models in this field have allowed researchers to generate new hypotheses relating individuals' behaviors to emergent network structures, researchers rarely discuss the robustness of their results to changes in (parametric and non parametric) model elements, nor do researchers frequently characterize which elements have the greatest impact on emergent behavior.

Here I present an example of how a simple sensitivity analysis can be conducted on a philosophy ABM and what additional information can be gleaned from sensitivity analysis.

Model overview

number_of_individuals individuals exist in the model. Each individual has a reliability(which is pulled from a beta distrubution and averages .6 over the entire population). num_facts number of facts exist in the model. Individuals have a correct belief, an incorrect belief, or no belief about each fact. Individuals begin with a starting_knowledge number of beliefs. The accuracy of these initial beliefs is proportional to the individual's reliability (an individual with a reliability of .6 will have roughly 60% true beliefs and 40% false beliefs).

The model is advanced 500 time steps per run. At each time step, each individual progresses through three phases.

Investigate: With probability investigation_probability, independently generate a new belief about a fact the individual previously had no opinion about. The probability of this new belief being true is the individual's reliability.
Select Teachers: Select a set of number_of_neighbors unique individuals from whom to solicit testimony. Each teacher presents a random fact and opinion about that fact. Teachers only present facts about which they have an opinion.
Learn: Adopt the opinions of teachers if they offer an opinion about a fact this individual does not currently have an opinion about, otherwise ignore new teacher and keep previous opinion.

All individuals in a model follow one of four philosophies; skeptical, reid, direct, and indirect. The individuals' philosophy determines each individual's behavior in each phase. Skeptical individuals do not select teachers nor learn from teachers. They only learn new facts through investigation. Reid individuals select teachers randomly. Direct individuals select the individuals who have the highest percentage of true beliefs as their teachers, calculated using the equation below.

$\Small x=\frac{\mathrm{truebeliefs}}{(\textrm{truebeliefs}&space;+ &space;\textrm{falsebeliefs})}$

Indirect individuals select individuals who have the most beliefs in common with themselves as their teachers. All ties when direct or indirect individuals are selecting teachers are broken randomly.