|
On March 28 2014 06:02 killerdog wrote:The first, and (imo) relatively simple application of the data i would try, which reddit would probably like too (more then just raw statistics and p values) would be a "what should i pick" style program. Just choose a p value to use as a cutoff, then calculate all the counters/synergies of each individual champion. This'll give you two lists for each champion, 117 entries long (or however many champions there are -1.) delete all data with p values above your cutoff. Now each champion has an associated list showing what it's strongest with, and what it's best against. Now write a program where you can enter between 0 and five champions on the enemy team, and 0-4 champions on your team. The program will then calculate the probabilty of each available champion of winning in this situation, by adding the synnergy chances of your team and the "counter" chances of the other teams champions, (potentially weighting by some inverse function f(p)->x so lower p values are more important), and spit out 4-5 champions which have the highest chance of winning. Should be codeable in a day or two (plus however long creating data for each champion takes) and would be pretty interesting, plus reddit tends to love those kind of things, especially when the statistics behind it's creation are easily understandable. You can add specific roles if you want, or weight your lane opponent more heavily then the rest of their team or whatever. I'm not really a statistician, but find more "practical" applications like that much more interesting, even if they wont be 100% reliable. Show nested quote +On March 28 2014 05:50 krndandaman wrote:
I'm legitimately confused whether some people are parodying or actually serious.
plz stop i want my league discussion back Imoperator and roffles get temped, and GD immediately turns to discussing the merits of various statistical models for analyzing league meta-data.
This is the exact sort of thing that Yango, Kuomindong, and I have said that Sufficiency's data mining is not useful for, from various perspectives. Moreover, the fact that championselect.net is taken seriously by enough people to keep it running means this sort of modeling is fairly unnecessary as well if your gold is just to do something plebian
I think Sufficiency's project is very cool to see data, but you really shouldn't read too much into it because it suffers from the problems all historical data mining is suspect to.
|
On March 28 2014 06:48 xes wrote:Show nested quote +On March 28 2014 06:02 killerdog wrote:The first, and (imo) relatively simple application of the data i would try, which reddit would probably like too (more then just raw statistics and p values) would be a "what should i pick" style program. Just choose a p value to use as a cutoff, then calculate all the counters/synergies of each individual champion. This'll give you two lists for each champion, 117 entries long (or however many champions there are -1.) delete all data with p values above your cutoff. Now each champion has an associated list showing what it's strongest with, and what it's best against. Now write a program where you can enter between 0 and five champions on the enemy team, and 0-4 champions on your team. The program will then calculate the probabilty of each available champion of winning in this situation, by adding the synnergy chances of your team and the "counter" chances of the other teams champions, (potentially weighting by some inverse function f(p)->x so lower p values are more important), and spit out 4-5 champions which have the highest chance of winning. Should be codeable in a day or two (plus however long creating data for each champion takes) and would be pretty interesting, plus reddit tends to love those kind of things, especially when the statistics behind it's creation are easily understandable. You can add specific roles if you want, or weight your lane opponent more heavily then the rest of their team or whatever. I'm not really a statistician, but find more "practical" applications like that much more interesting, even if they wont be 100% reliable. On March 28 2014 05:50 krndandaman wrote:
I'm legitimately confused whether some people are parodying or actually serious.
plz stop i want my league discussion back Imoperator and roffles get temped, and GD immediately turns to discussing the merits of various statistical models for analyzing league meta-data. This is the exact sort of thing that Yango, Kuomindong, and I have said that Sufficiency's data mining is not useful for, from various perspectives. Moreover, the fact that championselect.net is taken seriously by enough people to keep it running means this sort of modeling is fairly unnecessary as well if your gold is just to do something plebian I think Sufficiency's project is very cool to see data, but you really shouldn't read too much into it because it suffers from the problems all historical data mining is suspect to.
I get that it won't be an actual viable tool for serious gameplay, it's not meant to be. It would just be a potentially very fun, and definitely interesting, use of the data. And it would be in a form everyone would appreciate, without the sometimes imtimidating aspect of p values and long numbers. I'd imagine a user friendly form like this would probably be a lot more popular on places like reddit, where he keeps posting it to.
As long as you explain your method, and explain any shortcomings in your method, then by definition there's nothing scientifically wrong with it at all.
|
On March 28 2014 06:29 Sufficiency wrote:Show nested quote +On March 28 2014 06:12 killerdog wrote:On March 28 2014 06:07 Sufficiency wrote:
Can someone tell me why the competitive scene always get Elder Lizard on Evelynn? It seems to me that the true damage DoT only applies when you deal physical damage, but her Q and R (her only AOE spells) deal magical damage? Well her q and e both have ad ratios equal or better then their ap ratios. You get 20 more ap from wraith then you get ad from lizard, so if you can proc the elder lizard with an auto or e once then you'd need like 5-6 more q's (without managing to auto or e once to reproc lizard) to get even the same damage with wraith as lizard would have given. Also While clearing you'll be autoing a bunch, using the extra ad + proccing the DoT. I see. It seems that her AD ratios are mostly the same as her AP ratios (same on E, AP ratio is only 0.05 smaller on Q). I guess Lizard gives her a quicker early game spike.... which makes sense. Show nested quote +On March 28 2014 06:02 killerdog wrote:The first, and (imo) relatively simple application of the data i would try, which reddit would probably like too (more then just raw statistics and p values) would be a "what should i pick" style program. Just choose a p value to use as a cutoff, then calculate all the counters/synergies of each individual champion. This'll give you two lists for each champion, 117 entries long (or however many champions there are -1.) delete all data with p values above your cutoff. Now each champion has an associated list showing what it's strongest with, and what it's best against. Now write a program where you can enter between 0 and five champions on the enemy team, and 0-4 champions on your team. The program will then calculate the probabilty of each available champion of winning in this situation, by adding the synnergy chances of your team and the "counter" chances of the other teams champions, (potentially weighting by some inverse function f(p)->x so lower p values are more important), and spit out 4-5 champions which have the highest chance of winning. Should be codeable in a day or two (plus however long creating data for each champion takes) and would be pretty interesting, plus reddit tends to love those kind of things, especially when the statistics behind it's creation are easily understandable. You can add specific roles if you want, or weight your lane opponent more heavily then the rest of their team or whatever. I'm not really a statistician, but find more "practical" applications like that much more interesting, even if they wont be 100% reliable. On March 28 2014 05:50 krndandaman wrote:
I'm legitimately confused whether some people are parodying or actually serious.
plz stop i want my league discussion back Imoperator and roffles get temped, and GD immediately turns to discussing the merits of various statistical models for analyzing league meta-data. Oh wtf lol http://www.teamliquid.net/forum/closed-threads/32696-automated-ban-list-latest-theanarchy?page=1699#33965What happened?
http://www.teamliquid.net/forum/viewpost.php?post_id=21070673 i think
|
On March 28 2014 07:21 killerdog wrote:Show nested quote +On March 28 2014 06:48 xes wrote:On March 28 2014 06:02 killerdog wrote:The first, and (imo) relatively simple application of the data i would try, which reddit would probably like too (more then just raw statistics and p values) would be a "what should i pick" style program. Just choose a p value to use as a cutoff, then calculate all the counters/synergies of each individual champion. This'll give you two lists for each champion, 117 entries long (or however many champions there are -1.) delete all data with p values above your cutoff. Now each champion has an associated list showing what it's strongest with, and what it's best against. Now write a program where you can enter between 0 and five champions on the enemy team, and 0-4 champions on your team. The program will then calculate the probabilty of each available champion of winning in this situation, by adding the synnergy chances of your team and the "counter" chances of the other teams champions, (potentially weighting by some inverse function f(p)->x so lower p values are more important), and spit out 4-5 champions which have the highest chance of winning. Should be codeable in a day or two (plus however long creating data for each champion takes) and would be pretty interesting, plus reddit tends to love those kind of things, especially when the statistics behind it's creation are easily understandable. You can add specific roles if you want, or weight your lane opponent more heavily then the rest of their team or whatever. I'm not really a statistician, but find more "practical" applications like that much more interesting, even if they wont be 100% reliable. On March 28 2014 05:50 krndandaman wrote:
I'm legitimately confused whether some people are parodying or actually serious.
plz stop i want my league discussion back Imoperator and roffles get temped, and GD immediately turns to discussing the merits of various statistical models for analyzing league meta-data. This is the exact sort of thing that Yango, Kuomindong, and I have said that Sufficiency's data mining is not useful for, from various perspectives. Moreover, the fact that championselect.net is taken seriously by enough people to keep it running means this sort of modeling is fairly unnecessary as well if your gold is just to do something plebian I think Sufficiency's project is very cool to see data, but you really shouldn't read too much into it because it suffers from the problems all historical data mining is suspect to. I get that it won't be an actual viable tool for serious gameplay, it's not meant to be. It would just be a potentially very fun, and definitely interesting, use of the data. And it would be in a form everyone would appreciate, without the sometimes imtimidating aspect of p values and long numbers. I'd imagine a user friendly form like this would probably be a lot more popular on places like reddit, where he keeps posting it to. As long as you explain your method, and explain any shortcomings in your method, then by definition there's nothing scientifically wrong with it at all.
Perhaps, but like Sufficiency already mentioned, the predictive power of these analytic metrics is unknown and questionable. In some of the matchup datasets, there is definitely "is this a thing or is it a statistical fluke?" where your usual notion of p-value isn't useful since those are meant for "finding needles in a haystack" while this kind of data dredging is "finding stuff that isn't hay in a haystack."
However, the metrics provided have been pretty interesting (like the pentakill one, and even this wintime differential [though i don't think that differential is entirely caused by lategame/earlygame per se]). Something I would like to see is "skillcap differential." LoLKing already has the data, but it isn't presented in a very good format. The premise would be an initial pass to see the winrate differential between Bronze V and Challenger of various champions, and out of the top5 see what the per league breakdown is.
For example, Mid lulu is extremely punishing when played well and fit into a team, while also being extremely punishing to play if you aren't effective with your spells. We would expect both individual skill, and relative team "sense-making" to go down as you go down in skill, and maybe that correlates with a drop in winrate.
|
On March 28 2014 06:48 xes wrote:Show nested quote +On March 28 2014 06:02 killerdog wrote:The first, and (imo) relatively simple application of the data i would try, which reddit would probably like too (more then just raw statistics and p values) would be a "what should i pick" style program. Just choose a p value to use as a cutoff, then calculate all the counters/synergies of each individual champion. This'll give you two lists for each champion, 117 entries long (or however many champions there are -1.) delete all data with p values above your cutoff. Now each champion has an associated list showing what it's strongest with, and what it's best against. Now write a program where you can enter between 0 and five champions on the enemy team, and 0-4 champions on your team. The program will then calculate the probabilty of each available champion of winning in this situation, by adding the synnergy chances of your team and the "counter" chances of the other teams champions, (potentially weighting by some inverse function f(p)->x so lower p values are more important), and spit out 4-5 champions which have the highest chance of winning. Should be codeable in a day or two (plus however long creating data for each champion takes) and would be pretty interesting, plus reddit tends to love those kind of things, especially when the statistics behind it's creation are easily understandable. You can add specific roles if you want, or weight your lane opponent more heavily then the rest of their team or whatever. I'm not really a statistician, but find more "practical" applications like that much more interesting, even if they wont be 100% reliable. On March 28 2014 05:50 krndandaman wrote:
I'm legitimately confused whether some people are parodying or actually serious.
plz stop i want my league discussion back Imoperator and roffles get temped, and GD immediately turns to discussing the merits of various statistical models for analyzing league meta-data. This is the exact sort of thing that Yango, Kuomindong, and I have said that Sufficiency's data mining is not useful for, from various perspectives. Moreover, the fact that championselect.net is taken seriously by enough people to keep it running means this sort of modeling is fairly unnecessary as well if your gold is just to do something plebian I think Sufficiency's project is very cool to see data, but you really shouldn't read too much into it because it suffers from the problems all historical data mining is suspect to.
I don't quite understand this. Are you saying the issue with this study is that this is an observational study?
|
On March 28 2014 05:27 ZERG_RUSSIAN wrote:Show nested quote +On March 28 2014 05:23 Sufficiency wrote:On March 28 2014 05:12 TheYango wrote:On March 28 2014 05:05 Sufficiency wrote:
If you wish, we can argue about the merits. No models are perfect, and there are always trade offs. A simple model for an exploratory analysis is very useful, in particular, before a more complex model is applied. The exploratory analysis is a lot less useful than you make out because in general the amount of rigor needed for something to be established within this sphere is far less than in formal scientific investigation. You need far less than a formal theory with significant data backing it up to convince a bunch of nerds on the internet playing video games of anything. Which in turn means the practical usefulness of such an approach is far less. You are saying my model is an oversimplification so much as to be useless. Unfortunately, doctors prescribe medicine based on statistical evidence that is even more oversimplified. Bankers invest money based on statistical models that barely make sense. Yet I am just here analyzing data from a video game. Actually tho doctors prescribe medication based on a good deal of understanding of pharmacology BACKED by statistical analysis idk about bankers
OK but do we believe that champion picks don't have an effect on the outcome of the game? The problem in doing this isn't on the theoretical side. We believe that champion picks matter, we know the system is binary with regards to outcomes and champions in the game, so we don't have to worry about complex functions. This pretty much locks down the potential theory to one which sufficiency should be using. So we have a theory, its as complex as it can and needs to be, and we want to measure how large the effect of champion counters are. No problem there.
The problem in going deeper than 1 or 2 interactions is that you just don't have enough data (and if you, probably not enough time to run the calculation). There are what, 117 champions and 5 slots for one side with 112 and 5 for the other? This is something like 22 quadrillion potential games which can be played. The number of dummy terms we would have would be far larger than that even because you want to know whether an effect is from the 2 champion or 3 champion or 4 champion interaction. Even if you had enough observations(which we don't), doing the math would break your computer.
My main problem with Sufficiency's work on champion counters is that I don't understand why the regression needs to be logistic (well, why it needs to be a regression at all, just look at win rates for one champion against another champion). I was under the impression that logistic regressions are valuable when you want to tie a binary dependent variable with a continuous independent variable and that it should be indistinguishable from a simple linear system or summary statistics when all of the predictors are binary(I.E. the situation that we find ourselves in). Additionally i wanted to know precisely what his model was because "its logistic" doesn't actually tell me much because it doesn't explain which terms are in it and in which manner... which is the important part of a model.
|
On March 28 2014 08:08 Goumindong wrote:Show nested quote +On March 28 2014 05:27 ZERG_RUSSIAN wrote:On March 28 2014 05:23 Sufficiency wrote:On March 28 2014 05:12 TheYango wrote:On March 28 2014 05:05 Sufficiency wrote:
If you wish, we can argue about the merits. No models are perfect, and there are always trade offs. A simple model for an exploratory analysis is very useful, in particular, before a more complex model is applied. The exploratory analysis is a lot less useful than you make out because in general the amount of rigor needed for something to be established within this sphere is far less than in formal scientific investigation. You need far less than a formal theory with significant data backing it up to convince a bunch of nerds on the internet playing video games of anything. Which in turn means the practical usefulness of such an approach is far less. You are saying my model is an oversimplification so much as to be useless. Unfortunately, doctors prescribe medicine based on statistical evidence that is even more oversimplified. Bankers invest money based on statistical models that barely make sense. Yet I am just here analyzing data from a video game. Actually tho doctors prescribe medication based on a good deal of understanding of pharmacology BACKED by statistical analysis idk about bankers OK but do we believe that champion picks don't have an effect on the outcome of the game? The problem in doing this isn't on the theoretical side. We believe that champion picks matter, we know the system is binary with regards to outcomes and champions in the game, so we don't have to worry about complex functions. This pretty much locks down the potential theory to one which sufficiency should be using. So we have a theory, its as complex as it can and needs to be, and we want to measure how large the effect of champion counters are. No problem there. The problem in going deeper than 1 or 2 interactions is that you just don't have enough data (and if you, probably not enough time to run the calculation). There are what, 117 champions and 5 slots for one side with 112 and 5 for the other? This is something like 22 quadrillion potential games which can be played. The number of dummy terms we would have would be far larger than that even because you want to know whether an effect is from the 2 champion or 3 champion or 4 champion interaction. Even if you had enough observations(which we don't), doing the math would break your computer. My main problem with Sufficiency's work on champion counters is that I don't understand why the regression needs to be logistic (well, why it needs to be a regression at all, just look at win rates for one champion against another champion). I was under the impression that logistic regressions are valuable when you want to tie a binary dependent variable with a continuous independent variable and that it should be indistinguishable from a simple linear system or summary statistics when all of the predictors are binary(I.E. the situation that we find ourselves in). Additionally i wanted to know precisely what his model was because "its logistic" doesn't actually tell me much because it doesn't explain which terms are in it and in which manner... which is the important part of a model.
The model is as follows:
ln E(win) = b0 + I(Champion A is on one side) * b1 + I(Champion B is on the other side) * b2 + I(Champion A is on one side and Champion B is on the other side) * b3
The estimates you see are for b3, since it is the interaction term. This eliminates the effect of the overall power of Champion A and B. Repeat this for any pairs of champions A and B.
I don't really want to explain to you why this has to be a logistic regression, or why logistic regression even makes sense. You can try reading a book on categorical data analysis to enlighten yourself.
|
Returned to IMop being banned. Best week off ever.
|
All those values are binary though so it's identical to a linear system.
|
I don't know which is a worse discussion, this or warwick support.
|
Sufficiency should open up his own thread on this so we can put our cancer in there.
On March 28 2014 09:01 Gahlo wrote:
I don't know which is a worse discussion, this or warwick support.
This is basically the equivalent of a Warwick support discussion the context of categorical data analysis.
|
On March 28 2014 08:48 Goumindong wrote:
All those values are binary though so it's identical to a linear system.
I hate to say this. but I think you should stop posting about this and read some books before you further embarrass yourself. Try this one, it's the classic:
http://www.amazon.com/Categorical-Data-Analysis-Alan-Agresti/dp/0470463635
Pay strong attention to 3-way contingency tables and inference on such tables.
Nothing against you in particular, but judging from your feedback so far, it seems that you have some degrees of training in basic applied statistics. But all you have been doing so far was throwing jargon at me, and it sounds to me that you have no idea what is actually happening beyond trying to impress the ordinary reader with a bunch of verbal diarrhea.
I chose to avoid jargon as much as possible so it's easy to grasp for any reader. I also chose to not disclose my educational background, training, experience, and publication records because I feel it's not useful and will only make me sound condescending.
|
On March 28 2014 09:01 xes wrote:Sufficiency should open up his own thread on this so we can put our cancer in there. Show nested quote +On March 28 2014 09:01 Gahlo wrote:
I don't know which is a worse discussion, this or warwick support. This is basically the equivalent of a Warwick support discussion the context of categorical data analysis.
Yes, please make this into a blog or something.
|
On March 28 2014 09:01 Gahlo wrote:
I don't know which is a worse discussion, this or warwick support.
I'm still wondering what the discussion is, so far it's just derp.
On March 28 2014 08:48 Goumindong wrote:
All those values are binary though so it's identical to a linear system.
Logistic regression is supposed to be identical to a linear system, what's your point?
|
On March 28 2014 09:13 GolemMadness wrote:Show nested quote +On March 28 2014 09:01 xes wrote:Sufficiency should open up his own thread on this so we can put our cancer in there. On March 28 2014 09:01 Gahlo wrote:
I don't know which is a worse discussion, this or warwick support. This is basically the equivalent of a Warwick support discussion the context of categorical data analysis. Yes, please make this into a blog or something.
I know, right? Can we please just get back to anecdotal evidence for backing up our theories? I mean, as evidenced by nearly every facet of modern day research, numbers are basically useless for analyzing anything.
|
On March 28 2014 09:41 petered wrote:Show nested quote +On March 28 2014 09:13 GolemMadness wrote:On March 28 2014 09:01 xes wrote:Sufficiency should open up his own thread on this so we can put our cancer in there. On March 28 2014 09:01 Gahlo wrote:
I don't know which is a worse discussion, this or warwick support. This is basically the equivalent of a Warwick support discussion the context of categorical data analysis. Yes, please make this into a blog or something. I know, right? Can we please just get back to anecdotal evidence for backing up our theories? I mean, as evidenced by nearly every facet of modern day research, numbers are basically useless for analyzing anything.
Numbers are fine, but when everyone starts shitstorming over. what seems to be scientific interest from Sufficiency (and the urge to share it), it doesn't prove anything apart from what is it (a freaking hobby thing that he wanted to share).
And the hilarious and yet sad part is, that Sufficiency seems to know what he's talking about, yet people keeps disbelieving, like what the fuck guys?
|
United States47024 Posts
On March 28 2014 09:41 petered wrote:Show nested quote +On March 28 2014 09:13 GolemMadness wrote:On March 28 2014 09:01 xes wrote:Sufficiency should open up his own thread on this so we can put our cancer in there. On March 28 2014 09:01 Gahlo wrote:
I don't know which is a worse discussion, this or warwick support. This is basically the equivalent of a Warwick support discussion the context of categorical data analysis. Yes, please make this into a blog or something. I know, right? Can we please just get back to anecdotal evidence for backing up our theories? I mean, as evidenced by nearly every facet of modern day research, numbers are basically useless for analyzing anything.
"Every facet of modern research" demands numbers because they are necessary for the precision and rigor required in their respective fields. Not to mention that they all went through hundreds of years of qualitative analysis and general theory before they reached the point where numbers could be practically applied.
We're talking about a game that has only existed for less than 10 years. Even when applied to a game like baseball there was at least some qualitative understanding of the statistics being mined (and like 100+ years of baseball theory) before something like Moneyball could happen.
Nobody has the qualitative understanding of the game necessary to draw proper conclusions from data like this and build meaningful models. The qualitative understanding of what a lot of numbers actually mean isn't there yet. The statistics we have are the analogues for stuff like batting averages that were proven to be useless. The complex aggregate statistics which are actually meaningful don't even exist yet--and many of those were developed through anecdotal impressions of their relevance before statistics showed them to be so.
|
On March 28 2014 09:02 Sufficiency wrote:Show nested quote +On March 28 2014 08:48 Goumindong wrote:
All those values are binary though so it's identical to a linear system. I hate to say this. but I think you should stop posting about this and read some books before you further embarrass yourself. Try this one, it's the classic: http://www.amazon.com/Categorical-Data-Analysis-Alan-Agresti/dp/0470463635Pay strong attention to 3-way contingency tables and inference on such tables. Nothing against you in particular, but judging from your feedback so far, it seems that you have some degrees of training in basic applied statistics. But all you have been doing so far was throwing jargon at me, and it sounds to me that you have no idea what is actually happening beyond trying to impress the ordinary reader with a bunch of verbal diarrhea. I chose to avoid jargon as much as possible so it's easy to grasp for any reader. I also chose to not disclose my educational background, training, experience, and publication records because I feel it's not useful and will only make me sound condescending.
The issue is that everything you're saying sounds the same to me.
If your independent variables take two possible values it doesn't matter if you take the log or not. The interpretation of your coefficients changes slightly but it's the same system. Because you have only two values a log transformation is indistinguishable from a linear transformation.
|
On March 28 2014 09:51 TheYango wrote:Show nested quote +On March 28 2014 09:41 petered wrote:On March 28 2014 09:13 GolemMadness wrote:On March 28 2014 09:01 xes wrote:Sufficiency should open up his own thread on this so we can put our cancer in there. On March 28 2014 09:01 Gahlo wrote:
I don't know which is a worse discussion, this or warwick support. This is basically the equivalent of a Warwick support discussion the context of categorical data analysis. Yes, please make this into a blog or something. I know, right? Can we please just get back to anecdotal evidence for backing up our theories? I mean, as evidenced by nearly every facet of modern day research, numbers are basically useless for analyzing anything. "Every facet of modern research" demands numbers because they are necessary for the precision and rigor required in their respective fields. Not to mention that they all went through hundreds of years of qualitative analysis and general theory before they reached the point where numbers could be practically applied. We're talking about a game that has only existed for less than 10 years. Even when applied to a game like baseball there was at least some qualitative understanding of the statistics being mined (and like 100+ years of baseball theory) before something like Moneyball could happen. Nobody has the qualitative understanding of the game necessary to draw proper conclusions from data like this and build meaningful models because the qualitative understanding of what a lot of numbers actually mean isn't there yet. The statistics we have are the analogues for stuff like batting averages that were proven to be useless. The complex aggregate statistics which are actually meaningful don't even exist yet.
I think you should posting too. I think I have lost about all the respect I had for you.
On March 28 2014 10:02 Goumindong wrote:Show nested quote +On March 28 2014 09:02 Sufficiency wrote:On March 28 2014 08:48 Goumindong wrote:
All those values are binary though so it's identical to a linear system. I hate to say this. but I think you should stop posting about this and read some books before you further embarrass yourself. Try this one, it's the classic: http://www.amazon.com/Categorical-Data-Analysis-Alan-Agresti/dp/0470463635Pay strong attention to 3-way contingency tables and inference on such tables. Nothing against you in particular, but judging from your feedback so far, it seems that you have some degrees of training in basic applied statistics. But all you have been doing so far was throwing jargon at me, and it sounds to me that you have no idea what is actually happening beyond trying to impress the ordinary reader with a bunch of verbal diarrhea. I chose to avoid jargon as much as possible so it's easy to grasp for any reader. I also chose to not disclose my educational background, training, experience, and publication records because I feel it's not useful and will only make me sound condescending. The issue is that everything you're saying sounds the same to me. If your independent variables take two possible values it doesn't matter if you take the log or not. The interpretation of your coefficients changes slightly but it's the same system. Because you have only two values a log transformation is indistinguishable from a linear transformation.
Sigh.
EDIT: let me give you a more serious response. Words are that you are an economist. What you just said to me is like I just ask you "WTF WHY DO PEOPLE RESPOND TO INCENTIVES?!?!?!".
I am serious. The matter of taking the log is really fundamental. It's like the economic principle that people respond to incentives and everything has a cost.
|
On March 28 2014 09:51 TheYango wrote:Show nested quote +On March 28 2014 09:41 petered wrote:On March 28 2014 09:13 GolemMadness wrote:On March 28 2014 09:01 xes wrote:Sufficiency should open up his own thread on this so we can put our cancer in there. On March 28 2014 09:01 Gahlo wrote:
I don't know which is a worse discussion, this or warwick support. This is basically the equivalent of a Warwick support discussion the context of categorical data analysis. Yes, please make this into a blog or something. I know, right? Can we please just get back to anecdotal evidence for backing up our theories? I mean, as evidenced by nearly every facet of modern day research, numbers are basically useless for analyzing anything. "Every facet of modern research" demands numbers because they are necessary for the precision and rigor required in their respective fields. Not to mention that they all went through hundreds of years of qualitative analysis and general theory before they reached the point where numbers could be practically applied. We're talking about a game that has only existed for less than 10 years. Even when applied to a game like baseball there was at least some qualitative understanding of the statistics being mined (and like 100+ years of baseball theory) before something like Moneyball could happen.
So you need 100+ years of history to be able to understand a game through numbers? Yango you disappoint, that is nonsensical. The math behind data analysis is not that drastically different from one area of research to the next, making it possible to draw from the rich experience of other fields of research. The sheer volume of LoL games played and the numerous quantitative measures that can be drawn from each game make it a fantastic target for this type of analysis.
Now, I'll be honest, I didn't even look very much at what Sufficiency posted. I am just so confused when people flip their shit about someone trying to use numbers to better understand the game. Sure there are challenges and caveats with some studies, but that does not mean that he is going down the wrong path or that what he has presented is useless. Disagree with his study/conclusions fine, but this will inevitably be the future of understanding the game, since that is what is happening is just about every other field.
|
|
|
|
|
|
EPT
