Statistics
September 27, 2011
These are all straight forward:
(define (mean xs) (/ (sum xs) (length xs)))
(define (std-dev xs)
(let* ((n (length xs)) (x-bar (/ (sum xs) n)))
(define (diff x) (- x x-bar))
(sqrt (/ (sum (map square (map diff xs))) (- n 1)))))
(define (linear-regression xs ys)
(let* ((n (length xs))
(x (sum xs)) (y (sum ys))
(xx (sum (map square xs)))
(xy (sum (map * xs ys)))
(yy (sum (map square ys)))
(slope (/ (- (* n xy) (* x y)) (- (* n xx) (* x x))))
(intercept (- (/ y n) (* slope (/ n) x))))
(values slope intercept)))
(define (correlation xs ys)
(let* ((n (length xs)) (x-bar (mean xs)) (y-bar (mean ys)))
(define (x-diff x) (- x x-bar)) (define (y-diff y) (- y y-bar))
(/ (sum (map * (map x-diff xs) (map y-diff ys)))
(- n 1) (std-dev xs) (std-dev ys))))
We used sum and square from the Standard Prelude. Here is a simple example:
> (define xs '(1.47 1.50 1.52 1.55 1.57 1.60 1.63 1.65 1.68 1.70 1.73 1.75 1.78 1.80 1.83))
> (define ys '(52.21 53.12 54.48 55.84 57.20 58.57 59.93 61.29 63.11 64.47 66.28 68.10 69.92 72.19 74.46))
> (mean xs)
1.6506666666666665
> (mean ys)
62.078
> (std-dev xs)
0.11423451233985206
> (std-dev ys)
7.037514983490772
> (linear-regression xs ys)
61.272186542107434
-39.06195591883866
> (correlation xs ys)
0.9945837935768895
In 1973 statistician F. J. Anscombe constructed four datasets that have remarkably different shapes, shown at right, but that share common mean x=9 and y=7.5, standard deviation x=3.32 and y=2.03, slope 0.5, intercept 3.0 and correlation 0.816:
> (define xs-1 '(10 8 13 9 11 14 6 4 12 7 5))
> (define ys-1 '(8.04 6.95 7.58 8.81 8.33 9.96 7.24 4.26 10.84 4.82 5.68))
> (define xs-2 '(10 8 13 9 11 14 6 4 12 7 5))
> (define ys-2 '(9.14 8.14 8.74 8.77 9.26 8.10 6.13 3.10 9.13 7.26 4.74))
> (define xs-3 '(10 8 13 9 11 14 6 4 12 7 5))
> (define ys-3 '(7.46 6.77 12.74 7.11 7.81 8.84 6.08 5.39 8.15 6.42 5.73))
> (define xs-4 '(8 8 8 8 8 8 8 19 8 8 8))
> (define ys-4 '(6.58 5.76 7.71 8.84 8.47 7.04 5.25 12.50 5.56 7.91 6.89))
You can run the program at http://programmingpraxis.codepad.org/8WJpBVc9.
The implementation of your standard deviation (and thus correlation) is wrong, given the definitions on page 1. Your definition says to divide by N, you divide by N – 1…
My implementation in Go:
package main import ( "fmt" "math" ) func mean(data []float64) float64 { var sum float64 for _, x := range data {sum += x} return sum / float64(len(data)) } func sd(data []float64) float64 { mn := mean(data) var sd float64 for _, x := range data {diff := x - mn; sd += diff * diff} return math.Sqrt(sd / float64(len(data))) } func regress(xs, ys []float64) (slope, intercept float64) { var sum_xs, sum_xs2, sum_ys, sum_xs_ys, n float64 for i, _ := range xs { sum_xs += xs[i] sum_xs2 += xs[i] * xs[i] sum_ys += ys[i] sum_xs_ys += xs[i]*ys[i] } n = float64(len(xs)) slope = (n*sum_xs_ys - sum_xs*sum_ys) / (n*sum_xs2 - sum_xs*sum_xs) intercept = sum_ys/n - slope*sum_xs/n return } func correlation(xs, ys []float64) float64 { x_bar := mean(xs) y_bar := mean(ys) var cor float64 for i, _ := range xs {cor += (xs[i] - x_bar) * (ys[i] - y_bar)} return cor / (float64(len(xs) - 1)*sd(xs)*sd(ys)) } func main() { xs := []float64{1.47, 1.50, 1.52, 1.55, 1.57, 1.60, 1.63, 1.65, 1.68, 1.70, 1.73, 1.75, 1.78, 1.80, 1.83} ys := []float64{52.21, 53.12, 54.48, 55.84, 57.20, 58.57, 59.93, 61.29, 63.11, 64.47, 66.28, 68.10, 69.92, 72.19, 74.46} fmt.Println(mean(xs), mean(ys)) fmt.Println(sd(xs), sd(ys)) fmt.Println(regress(xs, ys)) fmt.Println(correlation(xs, ys)) return }I think I was taught to divide by n – 1 when the deviation from the (unknown) population mean is wanted but the (known) sample mean is used instead in the formula. The sample values are said to lose one “degree of freedom” because they can not all deviate freely from their own mean.
See: http://en.wikipedia.org/wiki/Standard_deviation
If you divide by n, the standard deviation is biased. Dividing by n-1 gives an unbiased standard deviation.
Python solution
http://pastebin.com/vrV9J4vN
By way of conversation, here is an approach I find much fun. I lift constants to be vecs (indexed sequences) so that everything is uniform, and then I map binary or unary operations on these vecs. Like in the language of R but more rigidly and in Scheme. The goal is a special language that allows to explore descriptions like “the mean square deviation from the mean” in the code itself. Someone should write The Structure and Interpretation of Statistical, er, Something.
Ok, I get carried away. A variation on the theme anyway. I’ve included one of Anscombe’s cases.
(define (vec . args) (case-lambda ((k) (list-ref args k)) (() (length args)))) (define fun (case-lambda ((f u w) (case-lambda ((k) (f (u k) (w k))) (() (u)))) ((f v) (case-lambda ((k) (f (v k))) (() (v)))))) I (define (con v c) (case-lambda ((k) c) (() (v)))) (define (sum v) (con v (do ((k 0 (+ k 1)) (s 0 (+ s (v k)))) ((= k (v)) s)))) (define (mean v) (fun / (sum v) (con v (v)))) (define (mean1 v) (fun / (sum v) (con v (- (v) 1)))) (define (var1 v) (let* ((d (fun - v (mean v)))) (mean1 (fun * d d)))) (define (stddev1 v) (fun sqrt (var1 v))) (define (std1 v) (fun / (fun - v (mean v)) (stddev1 v))) (define (cor1 u w) (mean1 (fun * (std1 u) (std1 w)))) (define xs-1 (vec 10 8 13 9 11 14 6 4 12 7 5)) (define ys-1 (vec 8.04 6.95 7.58 8.81 8.33 9.96 7.24 4.26 10.84 4.82 5.68)) (define (test) `((means ,((mean xs-1) 0) ,((mean ys-1) 0)) (devs ,((stddev1 xs-1) 0) ,((stddev1 ys-1) 0)) (corr ,((cor1 xs-1 ys-1) 0)))) ;((means 9 7.500909090909093) ; (devs 3.3166247903554 2.031568135925815) ; (corr 0.8164205163448399))@Jussi Piitulainen, Paul Hofstra:
Yes, but that’s not how he defined standard deviation on page 1. Thus the confusion..
@DGel: Yes. It may be better to deviate from the definition on page 1, especially when even the model implementation does so.
How ugly :)
let sum = List.fold_left (+.) 0.;; let mean = function | [] -> 0. | l -> sum l /. (float_of_int (List.length l));; let std = function | [] -> 0. | l -> let mu = mean l in let le = float_of_int (List.length l) in sqrt ((List.fold_left (fun r e -> r +. ((e -. mu) *. (e -. mu))) 0. l) /. le);; let slope xs ys = let le = float_of_int (List.length xs) in let sumxs = sum xs in let sumys = sum ys in ((le *. List.fold_left2 (fun r xi yi -> r +. (xi *. yi)) 0. xs ys) -. (sumxs *. sumys)) /. ((le *. (List.fold_left (fun r xi -> r +. (xi *. xi)) 0. xs)) -. (sumxs *. sumxs));; let intercept xs ys = let le = float_of_int (List.length xs) in ((sum ys) /. le) -. (((slope xs ys) *. (sum xs)) /. le);; let lin_reg xs ys = ((slope xs ys), (intercept xs ys));; let correlation xs ys = let le = float_of_int (List.length xs) in let mx = mean xs in let my = mean ys in let sx = std xs in let sy = std ys in (List.fold_left2 (fun r xi yi -> r +. ((xi -. mx) *.(yi -. my))) 0. xs ys) /. ((le -. 1.) *. sx *. sy) ;; let xs = [1.47;1.50;1.52;1.55;1.57;1.60;1.63;1.65;1.68;1.70;1.73;1.75;1.78;1.80;1.83];; let ys = [52.21;53.12;54.48;55.84;57.20;58.57;59.93;61.29;63.11;64.47;66.28;68.10;69.92;72.19;74.46];; let wrap point v = let eps = 0.001 in point -. eps < v && v < point +. eps;; assert (wrap 1.6506666666666665 (mean xs));; assert (wrap 62.078 (mean ys));; assert (wrap 0.110 (std xs));; assert (wrap 6.798 (std ys));; assert (wrap 61.272186542107434 (slope xs ys));; assert (wrap (-39.06195591883866) (intercept xs ys));; assert (wrap 1.06562549311809596 (correlation xs ys));;