Analyses:

• For each condition, run PCA for each subject and reduce to N components (going to start with 10)
• With the time points x N components, divide into 2 groups and run corrmean_combine
• For each component, (1, 1 and 2, 1-3, etc) how well can we decode

• Redo the optimized weighted level analysis with pca as the reductions technique
• To do this, keep track of the groups when running ISFC but combine groups for reducing
• • set rfun to None and make a wrapper function that takes the infold and outfold correlations, combine, and reduce using PCA, the return the reduced smooth infold and outfield data

Sanity check 1) Check using eigenvector centrality - should return the same values as before
Sanity check 2) Scramble the time points for each person - should get chance decoding

Once timecorr is released, add timecorr to dockerfile.

• need list of potential reviewers
• which journal?

• set matheron as a dependency
• write a matheron wrapper (tc.predict or similar)

• @jeremymanning review prior to release

@lucywowen can you write a readme file for this repo? ideally you could model it after this one

• ramping correlations
• blocky ("event like") correlations

please make the kernel figure (Fig 2) fonts bigger to match the other figures

three panels showing weights over time (with times shown for t = 0...100 and the weights function evaluated at t = 50):

• delta function
• gaussian weights (with \sigma = 10, 25, 50, 100, 1000 each in a different color, plus a legend)
• laplace weights (with \lambda = 10, 25, 50, 100, 1000, each in a different color, plus a legend)

• include code for replicating all analyses
• include call to download pieman + synthetic data

note: where should we store the synthetic data?

This figure looks strange: https://github.com/ContextLab/timecorr-paper/blob/master/paper/figs/higher_order_sims.pdf

It looks like maybe there are two artboards...?

@lucywowen could you fix it and email the editor with the corrected figure?

• use HTFA, PCA, ICA, k-means clustering (or similar) to reduce dimensionality (hopefully HTFA, re-using fits from Manning et al. 2018)
• need per-subject T by K matrices (I think these are already stored on discovery)

some ideas:

• for each level, characterize proportion of connections (for intact condition) that are left-to-right, dorsal-to-ventral, and/or anterior-to-posterior. Do this for positive vs. negative connections.
• See which lobes, brain areas, and/or "networks" (defined by Yao et al. or another paper) tend to exhibit strong connections
• threshold connections (at some strength) and see how long connections persist (expressed as a proportion of the story)
• color different nodes according to how many strong connections involve that node (or sum up the duration proportions of all connections involving that node). then use this code to generate a .nii plot and display it in some nice-looking way (surface plot using pycortex?)

• @jeremymanning review prior to release

I'm using this notebook to debug timecorr and show how it can recover parameters from synthetic data.

However, in some cases it's hard to tell if the parameters are being recovered correctly, or under what circumstances it should be possible to recover them. Need to explore further...

• Include latex format with final submission
• Upload updated editorial checklist (signed)
• Upload revised reporting summary (signed)
• Remove bold from display items (so I'm assuming those are captions too)
• Make sure all math conform to guidelines
• Link repo to zenodo
• Check that all figures are referred in order
• Supplementary references at the end of supplementary file (J help please!)
• Figure out the source data

• methods figure (cartoon)
• kernels figure showing different weights functions (DONE)
• synthetic data recovery (correlations)
• timepoint decoding on synthetic and real data by level (real data: break down by condition)
  • add hypertools trajectory plots for each pieman condition/level?
• pieman: future prediction by level/condition

Figures:
-S1: ttests matrices (heatmap)
-S2: intact detailed (15 order) brain plots and terms
-S3: paragraph
-S4: word
-S5: rest

• some notes outlined in main.tex

• https://link.springer.com/epdf/10.3758/s13428-019-01217-1?author_access_token=W4ZtX4H9xAmG9sRsS-dch5AH0g46feNdnc402Wrhzyqi5vgx3EpFpIpZHqViL_QwpkIOXhFbisRxd5j1cyJYu2C75NkMPny0yGS_1Q307WfUGA-AJmBPCGhbP1_6ZYPHmgJJ_Y-EfWCOvMbwX6Q0Vw%3D%3D
• http://advances.sciencemag.org/content/5/2/eaat7603
• https://www.biorxiv.org/content/early/2017/08/23/178145
• https://psyarxiv.com/xtzre
• https://www.sciencedirect.com/science/article/pii/S1053811918321384
• https://www.sciencedirect.com/science/article/pii/S1053811917309606
• https://www.sciencedirect.com/science/article/pii/S1053811916307881
  And from that paper:

• Allen et al., 2014
• Barttfeld et al., 2015
• Chen et al., 2016a
• Handwerker et al., 2012
• Yang et al., 2014
• Marusak et al., 2016
• Miller et al., 2014
• Damaraju et al., 2014
• Zalesky et al., 2014
• Rashid et al., 2014
• Shakil et al., 2016
• Sourty et al., 2016a
• Sourty et al., 2016b
• Yaesoubi et al., 2015b
• Betzel et al., 2016
• Leonardi and Van De Ville, 2015
• Hutchison et al., 2013a
• Zalesky and Breakspear, 2015
• Chang and Glover, 2010
• Yaesoubi et al., 2015a
• Cribben et al., 2012
• Cribben et al., 2013
• Xu and Lindquist, 2015

another way to plot the optimization results:
1.) after optimizing the weights, compute the decoding accuracy (or error, or rank)
2.) for each level in turn, set the weights to 0 (leaving all others fixed at their optimized values). then re-normalize the weights to sum to 1. then compute the new decoding performance (it should be lower than the decoding performance without that level set to 0, if the optimization code worked.)
3.) then for each level we can plot the improvement in performance when we include that level vs. not.