,
22 tweets,
33 min read
Read on Twitter
Thank you all for the invite and attendance! It was awesome! What an awesome group! While I didn't post my slides since I have others' images and screenshots in them, I'll do a quick tweetorial here, also as a way of recognizing resources/people who have influenced me in DS 1/n
My #1 rule, which was also the only rule to be numbered, is "Always back up your work!" Do this either with a continuous option like Dropbox or make sure you do a daily backup to Github, Google Drive etc 2/n
Next, I emphasized that data analysis is a many-stage process and gave the example of the "data pipeline" from @rdpeng's and @jtleek's great commentary nature.com/news/statistic…, noting that I do a lot of work on data cleaning, exploratory data analysis, statistical modeling 3/n
@rdpeng @jtleek Which bring me to... how do you organize your data or ask collaborators to organize it? Here I basically shared @kwbroman's amazing resource kbroman.org/dataorg/, now also available as an article tandfonline.com/doi/full/10.10… by Karl and @kara_woo 5/n
@rdpeng @jtleek @kwbroman @kara_woo Much of this is related to the concept of "tidy data" and "tidy data analysis," as introduced by @hadleywickham in jstatsoft.org/article/view/v…. There are many packages for tidy data analysis at tidyverse.org - Principles are valuable whether or not you use these 7/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham I then moved on to "naming things." After all, we'd much rather not be in this scenario right? phdcomics.com/comics/archive… I ❤️ @JennyBryan's "low-tech" wisdom at speakerdeck.com/jennybc/how-to…. As with many things, a little upfront planning now can save heartache/headache later. 8/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan Also, Jenny is so awesome that she even uses the ISO 8601 standard for items in her freezer!
_So_ useful! 9/n
_So_ useful! 9/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan For collaborative papers, it can help to use a system like Google docs or Overleaf (for LaTeX), though it depends on your collaboration/number of people. At least have a system that makes sense for you so you're not drowning in poorly named versions. 10/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan In terms of organizing the entire analysis, try to separate code, data, R objects etc. I still use this approach that my PhD adviser Giovanni Parmigiani introduced me to way back when. 11/n 
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan I also discussed a bit about reproducibility - there's obviously so much here, so I just referenced the original paper by @rdpeng science.sciencemag.org/content/334/60… and mentioned R markdown. I also love this motivation for reproducibility 12/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan I then spent a bunch of time on exploratory data analysis cause it's _so important_. My major overview is below. I would also recommend @rdpeng's book leanpub.com/exdata and maybe a look at my checklist at siminab.github.io/2018/09/05/omi… 13/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan Exploratory data analysis can help check for diff. kinds of confounding healthknowledge.org.uk/e-learning/epi… Ex: Socioeconomic factors can confound relationship between diet and disease risk. Remember: Correlation is not causation! (true even if you have lots of data, “big data” etc!) 14/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan Why plot? What do you consider when plotting? Here is my take. Also, @jtleek has some great examples in leanpub.com/datastyle 15/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan My overview on visualizing big data: Remember though, also we see plots in 2D, adding point color, type, and size can add layers that you can think of as "extra dimensions." 16/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan I personally find this one of the most beautiful examples of PCA plots ncbi.nlm.nih.gov/core/lw/2.0/ht… (genes mirroring geography in Europe) from ncbi.nlm.nih.gov/pmc/articles/P…. Maybe it's just that I'm a bit of a genetics/history nerd! 17/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan Another cool example of PCA and MDS: Gene expression data from post-mortem prefrontal cortex samples. Note the clear age dependence! ncbi.nlm.nih.gov/pmc/articles/P… (panel C) from ncbi.nlm.nih.gov/pmc/articles/P… 18/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan Also an example from my own paper 😂 journals.plos.org/plosone/articl… I usually do PCA plots to check for artifacts so was very happy to see that cases and controls _did not cluster separately_ (which could indicate batch effects) and also there was no clustering by study site 19/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan In response to a question about what processing should be done before PCA, I recommended @SherlockpHolmes's paper journals.plos.org/plosone/articl…. This paper + the convo we had on it convinced me to stop scaling omics data prior to PCA 20/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan @SherlockpHolmes Finally, my tips on modeling - just one slide since it was already running long/I was talking too fast and there are so many great resources on this (also you can get a PhD in stat/biostat if you really want to go in depth 😂)! 21/n
@rdpeng @jtleek @kwbroman @kara_woo @hadleywickham @JennyBryan @SherlockpHolmes And a plug for our new "Health Informatics and Data Science" 1-year MS program @ICBI_Georgetown healthinformatics.georgetown.edu as well as for our awesome annual informatics symposium, to be held on October 18th this year! icbi.georgetown.edu/symposium/ Thank you again!! 22/22
