Huge rec and a reminder to myself to look at this again when not in bed.

My last stats class was in ’75. All I can say without the aid of more coffee is: Wow! Nice work.

So the consensus is that Peralta hits poorly in cold weather, right? I majored in Philosophy.

Nice. The high BB/K rate at the extreme temperatures is intriguing.

Did you control for number of PAs for each temperature? That shouldremove some of the outliers in Jhonny’s charts – perhaps you can get a tighter result on the R-squared. For instance, the 100-degree data point for Jhonny has a .000 OPS, but no value for BB/K – that leads me to assume he has no K’s for 100-degree ABs, so very few ABs.

Interesting analysis. If you’ve got the raw data in a spreadsheet, I’d love to take at look at it; I think I have an idea that would help remove those small sample outliers at the ends of the graph.

These are all extremely weak correlations, and as we know correlation does not imply causation. This apparent trend, even if it is real, could simply be the result of players’ tendency to start the year a little rusty. Look at the first graph, much wider variance inthe beginning of the year.

This is definitely a cool idea, thanks. So we can expect him to put up huge numbers in another month. Problem fixed.

I like this, but I think what it suggests is that Peralta is pretty normal when it comes to temperature/performance. The one chart that makes Peralta look difference is the BB/K chart, but with that chart, the effect is minimal and the explanatory power represented by Peralta’s curve is also not far removed from zero. It’s great to see it laid out, though.

I take it you used equally sized temperature bins all across the spectrum? It might clean things up a bit if, instead of constraining the bins to be equally sized in temperature, make them equally sized in terms of number of plate appearances. So, in other words, sort the plate appearances by temperature, and divide them evenly into however many groups. That way you’ll be less likely to have the bloated variability at the extreme temperatures, where there’s less data (and if it’s still there it might tell you something substantive, instead of simply reflecting a smaller number of PAs per data point).

I think the uneven data issue might have something to do with the shape of the BB/K curves also. In Jhonny’s case, the small samples at the ends give you a bunch of zeros, which pulls the curve down. On the league graph, reducing the sample size gives you more variability, but because the measure is a ratio of two positive quantities, the variability is asymmetrical. That is, there’s more room to skew the ratio up than down, which will tend to pull the curve up when you’re fitting the curve by minimizing squared errors.

Ideally you probably want to use a log scale for a ratio measure, which will make things more symmetrical. It also will give you the same curve for BB/K and K/BB (just upside-down), which using a raw ratio scale doesn’t do. Once you’ve fit the regression line on the log scale, you can always transform back to raw ratios, to make reading the graph more intuitive.

If you wanted to send me your raw data, I might play around with it a little…

I disagree with your “2.5x great effect on OPS” conclusion because that regression is an extremely weak predictor (as a single variable model should be in this instance). The model is so weak that it’s basically useless – as evidenced by the stochastic, non-trending data points.

Perhaps you should retitle this “Johny Peralta is Zero Degrees Kelvin”

De La put in the freezer today.