I thinks its worth mentioning that there are two levels of black box models too. ML can memorize the expected value at each set of variables (at 1 rmp crank wheel rotates at 2 rpm) or it can ‘generalize’ and, for this example, tell us that the wheel rotates at 2x speed of crank. To some extent ‘ML generalization’ provides good ‘out of distribution’ predictions.
rain8dome9
Karma: 12
There is no “Wikipedia for predictive models” that I know of. No big repository to easily share and find predictive scientific models other than the relevant domain’s scientific literature, which is not optimized for these tasks: it is not organized by the variables being predicted, it is not generally available as reusable and modular software components, it is usually not focused on predictive work, some of it is paywalled, etc.
Have you tried www.openml.org?
Prototypical example: imagine a scientific field in which the large majority of practitioners have a very poor understanding of statistics, p-hacking, etc. Then lots of work in that field will be highly memetic despite trash statistics, blatant p-hacking, etc. Sure, the most competent people in the field may recognize the problems, but the median researchers don’t, and in aggregate it’s mostly the median researchers who spread the memes.
Complicated analysis (like going far beyond p-values) is easy for anyone to see and it is evidence of effort. Complex analysis usually coocurs with thoroughness so fewer mistakes. Complicated analysis coocurs with many concurrent tests so less need to produce positive results so less p-hacking. Consequently, there is a fairly simple solution to researchers with mediocre statistical skills gaining too much trust: more plots! Anyway, I find correlation graphs and multiple comparison impressive. Also I am usually more skilled in data analysis than the subject of a paper so can more easily verify that.
Is this a paper? Has it been published anywhere?
Relevant quote from Dragonfired by J. Zachary Pike. “Brokers make money by knowing key information; they make fortunes by ensuring that other brokers remain unaware or unsure of the same information until after critical trades.”
In ggplot (R statistical language) the defaults include a subtle grid and no axes. They also put in the extra random space.
Here is some code in case someone else using R wants to try out things discussed here:
library(ggplot2)
qplot(wt, mpg, data = mtcars, colour = factor(cyl)) +
theme(axis.line.x = element_line(color=”black”, size = 0),
axis.line.y = element_line(color=”black”, size = 1)) +
scale_x_continuous(expand = c(0, 0), limits = c(0,8)) +
scale_y_continuous(expand = c(0, 0), limits = c(0,36))
Might be able to use Multi-Armed Bandit-like sampling for this, even? Hm…
Effects may take time and may require time to build up to detectable levels. This is why Winters increased the length of each intervention till they lasted some weeks. If the placebo causes a different self report rating then its a bad placebo and should be Blinded out but if it causes a psychological improvement then why not use it?
so non-X days will be more likely measured as being high in X-effect. But that’d mean that X days are more likely followed by non-X, which with random order is not the case.
Yes but it will still make the effect size much less.
Could you elaborate on this a bit
Lag and build up is mentioned above. Training effect is when you get better at something just by doing it, so later interventions look better. At the same time there may be drift of self report. In other words effect of slowly growing change on memory making user think there is no change. For all these reasons plot the time series with time on X results on Y and make each point the color of intervention or placebo. Do not connect the dots with lines but do make a smooth loess-like line. You will be able to see some of the issues if they occur. Some more on all the issues.
The more important an effect is usually the stronger it is so starting many of the experiments but for a short time might yield results much faster. May be possible to overlap the non blinded experiments and run many at the same time with varying periodicity so the same interventions do not always happen on top of each other.
Your statistical method is similar to two sample t test right? Well that does not account for several possible issues of time series and dependence between data points of one variable. Lag and training effects for example. So be sure to control all other possible independent variables and plot the data timeline and when you do do not connect data points with lines!
In all experiments, I will be using the statistical method detailed here, code for it here, unless someone points out that I’m doing my statistics wrong.
Links lead no nowhere?
Will you try running the two notebooks on your data? I am starving for feedback and attention.
Really thorough statistical analysis of Anki (flashcard app) data
rpubs.com/rain8/1100036 Its a work in progress with only two steps finished. Not exactly an addon because its in R not Py. So far the project does many little things like find bugs in user’s collection, describe the growth of their collection and text mining. Ultimate goal is to hopefully be able to use anki as continuous cognitive tester and allow users to learn about and optimize their memorization process. Instructions to run on your own data : github
I am not sure data in anki could really be used as a continuous cognitive health test. Probably requires removing lots of artifacts and other influences and then finding outside influence that definitely relates to cognition. Lit review.
I am willing to be a test subject. Evidence that I am serious is I have 119k reviews on Anki and am analyzing the data hoping it will be a psychometric test.
https://wiki.openhumans.org/wiki/Finding_relations_between_variables_in_time_series This is the link I meant to post.
Thank you that was enlightening.
An analytic framework that takes multiple comparisons etc. into account and lets you see if any correlations are statistically significant.
Blinding.
Two issues, one of which I did not think of, out of like 20.
EDIT: I suspect, including from my own experience, that many problems can be solved without resorting to advanced statistics. Often by using through experimental procedure instead. Like eliminating a food type for a month then not doing an intervention for a month. Repeat. Trying out medications sounds like it should be done safely. This safety can only be achieved by monitoring vital signs and analyzing them using advanced statistics.
Is there a way to help users collect and analyze the data without needing to be a statistics expert?
Collection is really just a matter of finding the right devices and taking the time to use them. Analysis outside of immediate obvious effect can become difficult. If the effect is subtle and drowned in other effects, or hard to measure. If the intervention is not something user can easily or wants to reproduce. If the effect take long time to build up, or is shifted in time from intervention. If the successful effect only happens under several conditions or several interventions together. If the spray and pray approach is dangerous. If the spray and pray approach only hits gold once in a while. Multiple comparison problem (see wikipedia). If user is bad at keeping records. There are probably more. There are many many apps that just do correlation and none that do anything more. Here is a list of both problems and apps.
In my case it turned out to be manufactured food and gluten. This post is very similar to Quantifed Self movement.
Also please remember that side effects and drug interactions are a thing. Anything with a real effect can hurt you. I gave a very caveated suggestion of BosPro to someone on Twitter and it caused something akin to niacin flush in them. This is the same brand that does nothing to me but makes me better at digestion and uninterested in sugar.
What if the problem or the negative consequence of some intervention is hard to detect? I know this is not a popular opinion but anyone trying spray and pray really really should track their basics like HRV, Mood and cognitive ability.
EDIT: The other way to spray and pray in one huge chunk is to move to a different country. A large number of variables change when you do. In my case moving to Mexico helped because there are many restaurants that make food from scratch and bakeries that seem to make food without gluten.
So this will be on sept 21 right?
Excuse me for the necro. I think saying all the synonyms is better than letter based constraining. If the word that fits the constraint is found later than most other synonyms, the act of checking for the constraint takes longer than just listing. According to the 20 rules of formatting knowledge by Wozniak, its better for the mind to follow a set path even if it is longer, and that is the act of making a list. It is probably good to have sets of synonyms memorized for writing. Adding a constraint makes the question longer, which is something Wozniak advises against.
Then wolfram alpha?