Picciolo

Please not Girardi. Did you see Josh Johnson is going under the knife later this week & will miss a second season thanks to ol Joe? His confrontational style won't fit well with Melvin/Ash either.

When he gets canned in the next couple of months we can talk about how great of a pitching coach he is. If he's such a great pitching coach, show me one good thing that has happened to the Rays under his watch other than Shields. If he's so great, why is he overhauling the mechanics of a late inning reliever so that he can get to first a little quicker? Orvella's stuff is better than everyone you mention except for maybe Parra, but they play different positions (Balfour? Are you kidding me?). Last I checked, we have piles of crap like Shouse and Spurling on our staff. If our AAA insurance is so great, how are they not on this frickin roster? Why would you ever overlook an opportunity to grab a high ceiling player for very cheap just because we're better and we have some overrated depth?

As we approach the trade deadline, is there a Blue Light special out there who you'd like to see in a Brewers uni for the stretch run (or even as an insurance policy in the minors)? A year ago, I had Oliver Perez on my wish list. This year it's Chad Orvella of the Devil Rays. His story is somewhat similar to Ollie's -- he's in a bad organization, good stuff, bad pitching coach, mechanics a bit of a mess thanks to some unnecessary tinkering. He's got late inning stuff but just needs a good pitching coach & a better environment, IMO. A trade would do him as good as it did for Ollie, IMO. He has struggled in his MLB stints, but he has absolutely dominated AAA (and he's on a tear since his latest demotion). I don't think he's a 4A pitcher though -- his stuff is too good. I think the Brewers could get him very cheaply as I don't think Joe Maddon likes Orvella & Chad has had a couple failed stints in the bigs. Any other bargains out there or any thoughts about Chad?

I don't have either, actually. I was just going to write something quick.

No. Again, the scale effect, among other things. When interpretting a linear regression, we say that "holding all else constant, a one point increase in X increases Y by ___." The "all else constant" phrase means variables in the equation and other variables not in the equation. Going back to my earlier example, someone who adds a pound of weight will be related to a smaller increase in homeruns holding balls in play constant, than a one point change in balls in play holding weight constant because of the sheer size of weight relative to a percentage. Another reason we can't make the statement is multicollinearity. As you know, OBP and SLG will be highly correlated with each other. The goal in designing a model is to have ind. variables that are highly correlated with the Y variable, not each other. Multicollinearity biases the coefficient estimates. As an example, I would be willing to bet any amount of money that I could do a study that shows that smoking decreases the chances of developing lung cancer. How? I'd just add a bunch of variables that are related to smoking, such as: Number of lighters owned Cases of Milwaukee's Best Light drank per week Number of bowling and dart leagues they belong to Social status Lives in a trailer park (Y or N) Parents smoke (Y or N) Number of trips to the local smoke shop etc. Now if I add a variable such as "number of cigs smoked per day" to this equation...I have no idea what will show up. It would be just as likely that smoking would show as as negatively related to cancer as positively related to cancer because the coefficients will all be screwed up. I'm not familiar with the other studies you mention. I'd have to look at those sometime. Off the top of my head, the only way to compare OPS and SLG would be an F-test which would tell you if one variable is significantly different from the other in terms of explanatory power. It wouldn't give relative weights though. In statistics, even if we have two variables with different coefficients (assuming the exact same scale & no multicollinearity), it doesn't necessarily mean that they are statistically different. An F-test would show whether OBP is statistically different from SLG in explaining run production. I have to work on updating my SAS license, or I'd run the test right now. As a side note, in multiple regression, the only way to compare two independent variable coefficients is if they are of the same scale, there is zero multicollinearity, and the model has an R-squared of 1.

Yeah, you can generally find me over on BN (who are you?). Hopefully our conversations over there have been pleasant. http://forum.brewerfan.net/images/smilies/smile.gif I understand the stats frustration. It's not for everyone. I guess it makes it all the more sick that stats is more of a hobby for me then, huh? I've taken a few PhD stats courses and have taken my share of lumps at conferences. I discovered that I hated the PhD, but loved the stats.

Forget about multiple regression for a second. Let's look at an equation with just one X variable. Hopefully this won't get too confusing...this is tough to do in this format, but I'll give it a shot. Let's say my Y variable is homeruns and player's body weight is my X variable. HR=X*BodyWeight Let's say I'm looking at 5 players. So I have 5 homerun numbers and the body weight for the 5 players. For example, my five pairs of numbers are (X,Y): (220,10), (230,15), 235,12), (215,27) and (220,28 ) . We can calculate the intercept and the X coefficients using these numbers. To do so, we need to calculate some additional numbers. 1) I need the mean of the X numbers, which is (220+230+235+215+220)/5=224. 2) I also need the mean of the Y, which is 18.4. 3) Next, I need to subtract the mean of the X from each individual X, and the mean of Y from each individual Y. 4) I need to square the difference between the X's and the X mean (this is where the term least "squares" comes from). 5) I need to multiply each (X-224) by its corresponding (Y-18.4). Hopefully the formatting don't get too screwed up... Y........X........ X-224........ (X-224)^2........Y-18.4........(X-224)*(Y-18.4) 10........220........-4............ 16............ -8.4.............. 33.6 15........230........ 6............ 36............ -3.4..............-20.4 12........235........11...........121............ -6.4..............-70.4 27........215........-9............ 81............. 8.6.............. -77.4 28........220........ -4............16............. 9.6.............. -38.4 Total.................................270.................................-173 To calculate the X in the regression equation divide the total of the last column by the sum of the X-squares: -173/270= -.64 I won't go into the details of how to calculate the intercept, as it's not really important here, but the calculations are similar. So from my observations, I get the following equation: HR=162 - .64*BodyWeight So a couple of observations here: 1. As you can see, we're not assigning weights to anything. We are calculating the actual coefficient from the numbers in our sample of 5. 2. Coefficients can be negative. 3. We can use the equation to predict homeruns for someone who weighs 185 pounds. We take 162-.64*185lbs and we get 43.6 HR's. 4. The coefficients are really slopes. For each 1 pound increase in body weight, the model predicts that homeruns will decrease by .64. That's what least squares regression does, it finds the best slope for the data we have. 5. The OPS analysis was trying to predict runs scored I believe. So people were saying that each point of OBP is 1.8 times better than SLG. They calculated the coefficients for the two numbers the same way we just did in the example above and divided a couple of the coefficients -- which I've already shown is absolutely, positively wrong. 6. All of the calculations can be done in Excel. 7. Multiple regression uses the same type of calculations to get as many coefficients as is necessary (if you thought the above calcs were long, the work increases geometrically with each add'l variable). Yes, I know. http://forum.brewerfan.net/images/smilies/frown.gif People can make sense without cutting corners in their analysis (I may have failed here, but I'm winging it). It's something I find incredibly frustrating. Not so much on this board as for most of us, this is a hobby. But for people that do this for a living, I pull my hair out over a lot of the analysis -- especially when the analysis violates very basic regression rules. Many of the conclusions I see out there cannot be drawn from the methods used. Moneyball was an incredibly frustrating book to get thru for me.

Coefficients aren't relative weights. Basically, without going into too much statistical jargon, they are sample means. I'll throw out an example...see if this clears it up: Suppose I wanted to do a homerun study, and I decided to regress the number of homeruns a player hits on the players' weight, and the percentage of balls he puts in play so that: HR = X1*weight + X2*Percent Before I even run my regression, I pretty much know which coefficient will be smaller. It will be the players weight because each individual player weight is going to be very large compared to a percentage. Say I run my regression on my sample and suppose I get the following numbers... HR = 4 + .1*Weight + 16*Percent So that someone who weighed 200 lbs and put 50% of his balls in play would work out to 32 HR's (just plugging the numbers in). Now, if coefficients were relative weights, we would say that each percentage point of balls in play is 160 times more important than another pound of weight. Now say that instead of entering the balls in play number as a percentage, I decide to enter them in as whole numbers. In other words, I multiply the percentages by 100. Guess what happens to the coefficients. The intercept (4) stays the same as does X1. The only number that changes is X2 so that I'd get something that looked like: HR = 4 + .1*Weight + .16*Percent Someone that weighs 200 lbs and puts 50% in play still works out to 32 HR's. But now, each percentage point increase is only worth 1.6 times a pound of weight. All I did was change the scale. This is why you can't use coefficients as relative weights. It's apples & oranges.

The problem with the moneyball analysis, and the one linked by rluz below (I mention this in the linked discussion) is that OBP is constrained between 0 and 1. SLG is constrained b/w 0 and 4. You can't divide the coefficients of two variables to get relative weights when they are of different scales.

Absolutely true. The Perry's & Niekro's of the world put some spit on the ball before the current age of legal spitballs. Come on, you're going to compare licking a baseball to injecting yourself with an illegal drug? What illegal drug are you using?

I'm not sure when the baseball HOF succumbed to the "innocent until proven guilty" clause. Anyone who thinks he wasn't a cheat in his career should take a test of their own. Those who are grounded in reality & know he's a cheat should be subject to other tests if they think cheaters should be rewarded with the highest honors in their profession. Mac should be scorned for being a cheat and a liar and a druggie, not immortalized.

LOL! Actually, I was thinking that there is a lot more that goes into run-scoring than OBP and SLG, such as speed, weather, ballpark, opposing pitcher, league, manager style, and yes, battleness. Actually, I think the PC term is grindingness, but now we're splitting hairs. OK. I did miss that one. I'll have to think about that one some more when I don't have a few pale ales in my system. http://forum.brewerfan.net/images/smilies/smile.gif

R-squared's run from 0 to 1. A 1 would indicate that runs could be explained entirely by OBP and SLG. .79 is actually very high considering you're only looking at two variables. When you are running multivariate regression models, each variable will have a "p-value". Some software packages may report only a t-score. A very low p-value (<.05) would indicate that a variable is significantly related to runs. A high t-score (>1.96) indicates the same thing. We can start to compare relative importance to scoring runs by looking at the relative p-value (t-value) of OBP and SLG. There are other issues to consider, but this is at least a start. The reason you can't compare two coefficients of different scales is for the following reason: Example (forget the intercept term): Runs=2*OBP + 1.5*SLG If you're looking at a team that has OBP of 300 and SLG of 400 you get 2*300+1.5*400=1200. In other words, they contribute equally to runs. When I use Excel for regression, I use tools>internet options>regression.

Finally! One of them hangers got deposited!

You can't take the ratio of two coefficients of different scales like this. What were the p-values on each?