Structure of Deference Paper: Supplemental Material

Code

1. To run the deference simulator, first get the BayesACT code (v 0.5.1) from the BayesACT website or by clicking this zip file directly.
2. Then, download this deference.py Python program and put it in the same directory as the unzipped BayesACT source code
3. From that directory, you can run python deference.py -h which gives you this helpful message:

   BayesACT Deference score calculator (2 agents) usage:
    python deference.py
   	 -n number of samples (default 500)
   	 -f outputfilename prefix (output files will start with this prefix. Default: BayesACT)
   	 -a deference behaviour (default: defer_to)
   	 -b behaviours filename (default: fbehaviours.dat)
   	    behaviour files should be of type MEAN (see below)
   	 -i identities filename (default: fidentities.dat)
   	 -t type of identities (can be MEAN, COV or SD, default: MEAN)
   	    identity files are comma separated (CSV) and have the following format:
   		 MEAN: each line has identity followed by 3 numbers for E,P,A.  In this case the standard deviation used for identities is 0.1
   		 SD: each line has identity string followed by 3 values for mean EPA followed by 3 values for Standard Deviation EPA
   		 COV: each line has identity string followed by 3 values for mean EPA followed by 9 values for the covariance matrix row-wise as EPA
   
        

4. If you run it "as is" you will get the defaults which uses the default behaviour file (fbehaviours.dat, included with the BayesACT distribution, corresponds to the Indiana 2002-2004 dataset in Interact), and the default behaviour in that file of defer_to (EPA of E=-0.69, P=0.12, A=-0.9, which is not the same as in the paper (we used a different dataset where defer_to had an EPA rating of E=-.15,P=.45,A=-.44.
5. You also need an identities dataset. The default will be fidentities.dat (included with the BayesACT distribution, corresponds to the Indiana 2002-2004 dataset in Interact). You can edit this and create a smaller set, or use any other dataset you have.
6. Identities files have a mean EPA rating for each identity and then a specification of the covariance about that mean. This can be MEAN (so a diagonal covariance is used with a standard deviation in all three dimensions of 0.1), SD (a diagonal covariance with variances in the dimensions as specified), or COV (a full covariance matrix - all 9 components specified).
7. The paper used the University of Georgia 2015 dataset, which can be obtained from the ACT website.
8. A run, using "obey" instead of "defer to" and using an identity file called my_identities_file.csv with standard deviations in E,P, and A specified, would therefore look like:
   python deference.py -n 500 -f MyDirectory/MyFilePrefix -a obey -b my_behaviours_file.csv -i my_identities_file.csv -t SD
9. Output is three CSV files:
   • <FilePrefix>MeanDeflection.txt: gives the identities and their deference scores
   • <FilePrefix>SymmetricDeflection.txt: gives the identities and their symmetric deference scores (see this note)
   • <FilePrefix>DeflectionMatrix.txt: gives the full deference matrix
10. Need help? Email Jesse

Acknowledgements

See Also

Results

Variability Results

Deference Score Results

1. (≈ACT): a fixed diagonal covariance of 0.01 for all dimensions (this is approximately the same as deference scores based on ACT). Spearman correlation between this and the paper score is 0.92. R² of a linear fit between this and Harris poll scores is 0.625.
2. (Diag.): diagonal covariances from data (this is the same as in the paper). Spearman correlation between this and the paper score is 0.999. R² of a linear fit between this and Harris poll scores is 0.724.
3. (Full.): full covariances from data . Spearman correlation between this and the paper score is 0.99. R² of a linear fit between this and Harris poll scores is 0.726.