Tweetovi

Blokirali ste korisnika/cu @KyleCSN

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @KyleCSN

  1. proslijedio/la je Tweet
    5. velj

    I have 3995 followers. With five more I will break 4000 and my children will finally respect me. Could today be my day? Plz retweet.

    Poništi
  2. 4. velj

    I was not expecting to see causal inference specifically called out in this profile of Caltech alumn Prabhakar and her new innovation nonprofit Actuate. It's popping up more and more.

    Poništi
  3. proslijedio/la je Tweet
    21. sij

    I have post-doc openings in my lab. Do you want to use empirical methods, design experiments, and collaborate with both large and small tech firms for research and social impact? Have interest in ed tech, fin tech, charitable giving, and social sector effectiveness? Apply! 1/3

    Prikaži ovu nit
    Poništi
  4. 5. sij

    Oops. Pentagon officials should have made a principle-agent model with probability>0 of "crazy" type and solved for the optimal choice set.

    Poništi
  5. 8. pro 2019.

    The reason no intelligent alien life has been found in the universe--> In every industrialized civilization eventually so many plastic bags are stuffed into a single, other plastic bag a supermassive black hole forms, destroying the local star system.

    Poništi
  6. 4. pro 2019.

    Probability lessons that look great and have much attention to detail. I like the extended discussion on how to faithfully plot pdfs, pmfs, and cdfs.

    Poništi
  7. 1. pro 2019.

    Open Access copy of Sheldon Axler's new real analysis and measure theory textbook (author of the linear algebra classic). Probably a good supplement for 1st year PhD students.

    Poništi
  8. 1. pro 2019.

    Francois Morellet painted this using an innovative random number generator in 1960: His wife and children reading numbers from the phone book. It looks just like something from a computer vision class.

    Poništi
  9. 1. pro 2019.

    Notebook that produces the simulations and results. /end

    Prikaži ovu nit
    Poništi
  10. 1. pro 2019.

    Lesson: Think about your prior for the effect size. Don't just pick your significance level by convention. Debates about significance levels apply here as much as research.

    Prikaži ovu nit
    Poništi
  11. 1. pro 2019.

    Flip side: If treatment is on average slightly worse than control, then the environment is favorable to the hypo test rule. It clearly outperforms a coin flip and is close to the Bayesian rule.

    Prikaži ovu nit
    Poništi
  12. 1. pro 2019.

    The coin flip wins in this DGP where the treatment is on average slightly better than than control. The hypo test is too conservative in this environment. The coin flip has better performance even with a 50% type I error rate!

    Prikaži ovu nit
    Poništi
  13. 1. pro 2019.

    Hypo testing with the strictest alpha=0.01 under performs a coin flip for this DGP where the treatment group is on average slightly better. This environment is bad for the conservative significance test. Flipping a coin or just choosing treatment every time is therefore better.

    Prikaži ovu nit
    Poništi
  14. 1. pro 2019.

    The Bayesian rule also racks up small wins that the hypothesis testing rules miss due to their conservatism (lack of power). The tests with alpha=0.10 perform better (and alpha=0.50 will agree with the Bayesian rule in this data-generating process).

    Prikaži ovu nit
    Poništi
  15. 1. pro 2019.

    False positive rates ("Type I error"), the hypo tests have rates below their level, as expected. The Bayesian rule has a much higher error rate. When it makes errors it makes very small ones.

    Prikaži ovu nit
    Poništi
  16. 1. pro 2019.

    If the prior means of the two groups are equal (Beta(5,10) in this case), then hypothesis testing gives better results than a coin flip. The Bayesian rule is optimal.

    Prikaži ovu nit
    Poništi
  17. 1. pro 2019.

    I simulated 10k AB tests. The conversion probability in each group was sample from a Beta dist. Performance of a given rule = Average probability we end up with after running the 10k experiments and using rule X to pick one group.

    Prikaži ovu nit
    Poništi
  18. 1. pro 2019.

    Decision strategies in experiments: *Flipping a coin* can outperform hypothesis testing. Why? Conventional alpha levels can be inappropriately conservative. Note: I don't recommend to actually flip a coin (but it might make you happier). Explanation of simulation in thread. 1/N

    Prikaži ovu nit
    Poništi
  19. 23. stu 2019.

    TIL: A simple but clever practice that allows each engineer to veto the operation of a dangerous machine while they are working on it.

    Poništi
  20. 17. stu 2019.

    We can add precision with annotations on the plot, but the viz easily becomes busy/ugly. /end

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·