3 ;;; Copyright (c) 2008--, by A.J. Rossini <blindglobe@gmail.com>
4 ;;; See COPYRIGHT file for any additional restrictions (BSD license).
5 ;;; Since 1991, ANSI was finally finished. Modified to match ANSI
9 ;;;; regression.lsp XLISP-STAT regression model proto and methods
10 ;;;; XLISP-STAT 2.1 Copyright (c) 1990, by Luke Tierney
11 ;;;; Additions to Xlisp 2.1, Copyright (c) 1989 by David Michael Betz
12 ;;;; You may give out copies of this software; for conditions see the file
13 ;;;; COPYING included with this distribution.
15 ;;;; Incorporates modifications suggested by Sandy Weisberg.
17 ;;; This version uses lisp-matrix for underlying numerics.
19 (in-package :lisp-stat-regression-linear
)
21 ;;; Regresion Model Prototype
23 ;; The general strategy behind the fitting of models using prototypes
24 ;; is that we need to think about want the actual fits are, and then
25 ;; the fits can be used to recompute as components are changes. One
26 ;; catch here is that we'd like some notion of trace-ability, in
27 ;; particular, there is not necessarily a fixed way to take care of the
28 ;; audit trail. save-and-die might be a means of recording the final
29 ;; approach, but we are challenged by the problem of using advice and
30 ;; other such features to capture stages and steps that are considered
31 ;; along the goals of estimating a model.
33 ;; Note that the above is a stream-of-conscience response to the
34 ;; challenge of reproducibility in the setting of prototype "on-line"
37 (defvar regression-model-proto nil
38 "Prototype for all regression model instances.")
40 (defproto regression-model-proto
41 '(x y intercept sweep-matrix basis weights
44 residual-sum-of-squares
51 "Normal Linear Regression Model")
54 (defun regression-model
59 (included (repeat t
(length y
)))
63 (doc "Undocumented Regression Model Instance")
65 "Args: (x y &key (intercept T) (print T) (weights nil)
66 included predictor-names response-name case-labels)
67 X - list of independent variables or X matrix
68 Y - dependent variable.
69 INTERCEPT - T to include (default), NIL for no intercept
70 PRINT - if not NIL print summary information
71 WEIGHTS - if supplied should be the same length as Y; error
73 assumed to be inversely proportional to WEIGHTS
74 PREDICTOR-NAMES, RESPONSE-NAME, CASE-LABELS
75 - sequences of strings or symbols.
76 INCLUDED - if supplied should be the same length as Y, with
77 elements nil to skip a in computing estimates (but not
78 in residual analysis).
79 Returns a regression model object. To examine the model further assign the
80 result to a variable and send it messages.
81 Example (data are in file absorbtion.lsp in the sample data directory):
82 (def m (regression-model (list iron aluminum) absorbtion))
83 (send m :help) (send m :plot-residuals)"
85 ((typep x
'matrix-like
) x
)
86 ((or (typep x
'sequence
)
89 (make-vector (length x
) :initial-contents x
)))
90 (t x
))) ;; actually, might should barf.
92 ((typep y
'vector-like
) y
)
94 (numberp (car x
))) (make-vector (length y
) :initial-contents y
))
95 (t y
))) ;; actually, might should barf.
96 (m (send regression-model-proto
:new
)))
101 (send m
:intercept intercept
)
102 (send m
:weights weights
)
103 (send m
:included included
)
104 (send m
:predictor-names predictor-names
)
105 (send m
:response-name response-name
)
106 (send m
:case-labels case-labels
)
110 (format t
"~S~%" (send m
:doc
))
111 (format t
"X: ~S~%" (send m
:x
))
112 (format t
"Y: ~S~%" (send m
:y
))))
113 (if print
(send m
:display
))
117 ;; regression-model is the old API, but regression as a generic will
118 ;; be the new API. We need to distinguish between APIs which enable
119 ;; the user to do clear activities, and APIs which enable developers
120 ;; to do clear extensions and development, and underlying
121 ;; infrastructure to keep everything straight and enabled.
123 ;; There are conflicting theories for how to structure the
124 ;; specification of mathematical models, along with the statistical
125 ;; inference, along with the data which is instantiating the model.
127 ;; i.e.: mathematical model for the relationships between components,
128 ;; between a component and a summarizing parameter, and between
131 ;; statistical inference describes the general approach for
132 ;; aggregating into a decision and has impliciations for the scale up
133 ;; from the model on a single instance to the generalization.
135 ;; The data represents the particular substantive context that is
136 ;; driving the model/inference combination, and about which we hope to
137 ;; generate knowledge.
139 ;; numerical analysis selects appropriate algorithms/implementations
140 ;; for combining the above 3.
142 ;; the end result is input on the decision being made (which could be
143 ;; specific (decision analysis/testing), risk-analysis (interval
144 ;; estimation) , most likely/appropriate selection (point estimation)
151 (defgeneric regression
;; assumes x/y from lisp-matrix -- start of a set of generics.
153 "Args: (x y &key (intercept T) (print T) (weights nil)
154 included predictor-names response-name case-labels)
155 X - list of independent variables or X matrix
156 Y - dependent variable.
157 INTERCEPT - T to include (default), NIL for no intercept
158 PRINT - if not NIL print summary information
159 WEIGHTS - if supplied should be the same length as Y; error
161 assumed to be inversely proportional to WEIGHTS
162 PREDICTOR-NAMES, RESPONSE-NAME, CASE-LABELS
163 - sequences of strings or symbols.
164 INCLUDED - if supplied should be the same length as Y, with
165 elements nil to skip a in computing estimates (but not
166 in residual analysis).
167 Returns a regression model object. To examine the model further assign the
168 result to a variable and send it messages.
169 Example (data are in file absorbtion.lsp in the sample data directory):
170 (def m (regression-model (list iron aluminum) absorbtion))
171 (send m :help) (send m :plot-residuals)"
172 (let ((m (send regression-model-proto
:new
)))
177 (send m
:intercept intercept
)
178 (send m
:weights weights
)
179 (send m
:included included
)
180 (send m
:predictor-names predictor-names
)
181 (send m
:response-name response-name
)
182 (send m
:case-labels case-labels
)
186 (format t
"~S~%" (send m
:doc
))
187 (format t
"X: ~S~%" (send m
:x
))
188 (format t
"Y: ~S~%" (send m
:y
))))
189 (if print
(send m
:display
))
194 (defmeth regression-model-proto
:isnew
()
195 (send self
:needs-computing t
))
197 (defmeth regression-model-proto
:save
()
199 Returns an expression that will reconstruct the regression model."
200 `(regression-model ',(send self
:x
)
202 :intercept
',(send self
:intercept
)
203 :weights
',(send self
:weights
)
204 :included
',(send self
:included
)
205 :predictor-names
',(send self
:predictor-names
)
206 :response-name
',(send self
:response-name
)
207 :case-labels
',(send self
:case-labels
)))
209 ;;; Computing and Display Methods
211 (defmeth regression-model-proto
:compute
()
213 Recomputes the estimates. For internal use by other messages"
214 (let* ((included (if-else (send self
:included
) 1 0))
217 (intercept (send self
:intercept
))
218 (weights (send self
:weights
))
219 (w (if weights
(* included weights
) included
))
220 (m (make-sweep-matrix x y w
)) ;;; ERROR HERE
221 (n (matrix-dimension x
1))
222 (p (- (matrix-dimension m
0) 1))
225 (reduce #'* (mapcar #'standard-deviation
(list-of-columns x
)))))
228 (sweep-operator m
(iseq 1 n
) tol
)
229 (sweep-operator m
(iseq 0 n
) (cons 0.0 tol
)))))
231 "~%REMOVEME: regr-mdl-prto :compute~%Sweep= ~A~%x= ~A~%y= ~A~%m= ~A~%tss= ~A~%"
232 sweep-result x y m tss
)
233 (setf (slot-value 'sweep-matrix
) (first sweep-result
))
234 (setf (slot-value 'total-sum-of-squares
) tss
)
235 (setf (slot-value 'residual-sum-of-squares
)
236 (mref (first sweep-result
) p p
))
237 ;; SOMETHING WRONG HERE! FIX-ME
238 (setf (slot-value 'basis
)
239 (let ((b (remove 0 (second sweep-result
))))
240 (if b
(- (reduce #'-
(reverse b
)) 1)
241 (error "no columns could be swept"))))))
243 (defmeth regression-model-proto
:needs-computing
(&optional set
)
244 "Message args: ( &optional set )
246 If value given, sets the flag for whether (re)computation is needed to
247 update the model fits."
249 (if set
(setf (slot-value 'sweep-matrix
) nil
))
250 (null (slot-value 'sweep-matrix
)))
252 (defmeth regression-model-proto
:display
()
255 Prints the least squares regression summary. Variables not used in the fit
256 are marked as aliased."
257 (let ((coefs (coerce (send self
:coef-estimates
) 'list
))
258 (se-s (send self
:coef-standard-errors
))
260 (p-names (send self
:predictor-names
)))
261 (if (send self
:weights
)
262 (format t
"~%Weighted Least Squares Estimates:~2%")
263 (format t
"~%Least Squares Estimates:~2%"))
264 (when (send self
:intercept
)
265 (format t
"Constant ~10f ~A~%"
266 (car coefs
) (list (car se-s
)))
267 (setf coefs
(cdr coefs
))
268 (setf se-s
(cdr se-s
)))
269 (dotimes (i (array-dimension x
1))
271 ((member i
(send self
:basis
))
272 (format t
"~22a ~10f ~A~%"
273 (select p-names i
) (car coefs
) (list (car se-s
)))
274 (setf coefs
(cdr coefs
) se-s
(cdr se-s
)))
275 (t (format t
"~22a aliased~%" (select p-names i
)))))
277 (format t
"R Squared: ~10f~%" (send self
:r-squared
))
278 (format t
"Sigma hat: ~10f~%" (send self
:sigma-hat
))
279 (format t
"Number of cases: ~10d~%" (send self
:num-cases
))
280 (if (/= (send self
:num-cases
) (send self
:num-included
))
281 (format t
"Number of cases used: ~10d~%" (send self
:num-included
)))
282 (format t
"Degrees of freedom: ~10d~%" (send self
:df
))
285 ;;; Slot accessors and mutators
287 (defmeth regression-model-proto
:doc
(&optional new-doc append
)
288 "Message args: (&optional new-doc)
290 Returns the DOC-STRING as supplied to m.
291 Additionally, with an argument NEW-DOC, sets the DOC-STRING to
292 NEW-DOC. In this setting, when APPEND is T, append NEW-DOC to DOC
293 rather than doing replacement."
295 (when (and new-doc
(stringp new-doc
))
296 (setf (slot-value 'doc
)
305 (defmeth regression-model-proto
:x
(&optional new-x
)
306 "Message args: (&optional new-x)
308 With no argument returns the x matrix-like as supplied to m. With an
309 argument, NEW-X sets the x matrix-like to NEW-X and recomputes the
311 (when (and new-x
(typep new-x
'matrix-like
))
312 (setf (slot-value 'x
) new-x
)
313 (send self
:needs-computing t
))
316 (defmeth regression-model-proto
:y
(&optional new-y
)
317 "Message args: (&optional new-y)
319 With no argument returns the y vector-like as supplied to m. With an
320 argument, NEW-Y sets the y vector-like to NEW-Y and recomputes the
323 (typep new-y
'vector-like
))
324 (setf (slot-value 'y
) new-y
) ;; fixme -- pls set slot value to a vector-like!
325 (send self
:needs-computing t
))
328 (defmeth regression-model-proto
:intercept
(&optional
(val nil set
))
329 "Message args: (&optional new-intercept)
331 With no argument returns T if the model includes an intercept term,
332 nil if not. With an argument NEW-INTERCEPT the model is changed to
333 include or exclude an intercept, according to the value of
336 (setf (slot-value 'intercept
) val
)
337 (send self
:needs-computing t
))
338 (slot-value 'intercept
))
340 (defmeth regression-model-proto
:weights
(&optional
(new-w nil set
))
341 "Message args: (&optional new-w)
343 With no argument returns the weight vector-like as supplied to m; NIL
344 means an unweighted model. NEW-W sets the weights vector-like to NEW-W
345 and recomputes the estimates."
347 (typep new-w
'vector-like
))
348 (setf (slot-value 'weights
) new-w
)
349 (send self
:needs-computing t
))
350 (slot-value 'weights
))
352 (defmeth regression-model-proto
:total-sum-of-squares
()
355 Returns the total sum of squares around the mean.
356 This is recomputed if an update is needed."
357 (if (send self
:needs-computing
)
358 (send self
:compute
))
359 (slot-value 'total-sum-of-squares
))
361 (defmeth regression-model-proto
:residual-sum-of-squares
()
364 Returns the residual sum of squares for the model.
365 This is recomputed if an update is needed."
366 (if (send self
:needs-computing
)
367 (send self
:compute
))
368 (slot-value 'residual-sum-of-squares
))
370 (defmeth regression-model-proto
:basis
()
373 Returns the indices of the variables used in fitting the model, in a
375 This is recomputed if an update is needed."
376 (if (send self
:needs-computing
)
377 (send self
:compute
))
381 (defmeth regression-model-proto
:sweep-matrix
()
384 Returns the swept sweep matrix. For internal use"
385 (if (send self
:needs-computing
)
386 (send self
:compute
))
387 (slot-value 'sweep-matrix
))
389 (defmeth regression-model-proto
:included
(&optional new-included
)
390 "Message args: (&optional new-included)
392 With no argument, NIL means a case is not used in calculating
393 estimates, and non-nil means it is used. NEW-INCLUDED is a sequence
394 of length of y of nil and t to select cases. Estimates are
396 (when (and new-included
397 (= (nelts new-included
) (send self
:num-cases
)))
398 (setf (slot-value 'included
) (copy-seq new-included
))
399 (send self
:needs-computing t
))
400 (if (slot-value 'included
)
401 (slot-value 'included
)
402 (repeat t
(send self
:num-cases
))))
404 (defmeth regression-model-proto
:predictor-names
(&optional
(names nil set
))
405 "Message args: (&optional (names nil set))
407 With no argument returns the predictor names. NAMES sets the names."
408 (if set
(setf (slot-value 'predictor-names
) (mapcar #'string names
)))
409 (let ((p (matrix-dimension (send self
:x
) 1))
410 (p-names (slot-value 'predictor-names
)))
411 (if (not (and p-names
(= (length p-names
) p
)))
412 (setf (slot-value 'predictor-names
)
413 (mapcar #'(lambda (a) (format nil
"Variable ~a" a
))
415 (slot-value 'predictor-names
))
417 (defmeth regression-model-proto
:response-name
(&optional
(name "Y" set
))
418 "Message args: (&optional name)
420 With no argument returns the response name. NAME sets the name."
422 (if set
(setf (slot-value 'response-name
) (if name
(string name
) "Y")))
423 (slot-value 'response-name
))
425 (defmeth regression-model-proto
:case-labels
(&optional
(labels nil set
))
426 "Message args: (&optional labels)
427 With no argument returns the case-labels. LABELS sets the labels."
428 (if set
(setf (slot-value 'case-labels
)
430 (mapcar #'string labels
)
431 (mapcar #'(lambda (x) (format nil
"~d" x
))
432 (iseq 0 (- (send self
:num-cases
) 1))))))
433 (slot-value 'case-labels
))
437 ;;; None of these methods access any slots directly.
440 (defmeth regression-model-proto
:num-cases
()
442 Returns the number of cases in the model."
443 (nelts (send self
:y
)))
445 (defmeth regression-model-proto
:num-included
()
447 Returns the number of cases used in the computations."
448 (sum (if-else (send self
:included
) 1 0)))
450 (defmeth regression-model-proto
:num-coefs
()
452 Returns the number of coefficients in the fit model (including the
453 intercept if the model includes one)."
454 (if (send self
:intercept
)
455 (+ 1 (nelts (send self
:basis
)))
456 (nelts (send self
:basis
))))
458 (defmeth regression-model-proto
:df
()
460 Returns the number of degrees of freedom in the model."
461 (- (send self
:num-included
) (send self
:num-coefs
)))
463 (defmeth regression-model-proto
:x-matrix
()
465 Returns the X matrix for the model, including a column of 1's, if
466 appropriate. Columns of X matrix correspond to entries in basis."
467 (let ((m (select (send self
:x
)
468 (iseq 0 (- (send self
:num-cases
) 1))
469 (send self
:basis
))))
470 (if (send self
:intercept
)
471 (bind2 (repeat 1 (send self
:num-cases
)) m
)
474 (defmeth regression-model-proto
:leverages
()
476 Returns the diagonal elements of the hat matrix."
477 (let* ((weights (send self
:weights
))
478 (x (send self
:x-matrix
))
480 (m* (* (m* x
(send self
:xtxinv
)) x
)
481 (repeat 1 (send self
:num-coefs
)))))
482 (if weights
(* weights raw-levs
) raw-levs
)))
484 (defmeth regression-model-proto
:fit-values
()
486 Returns the fitted values for the model."
487 (m* (send self
:x-matrix
) (send self
:coef-estimates
)))
489 (defmeth regression-model-proto
:raw-residuals
()
491 Returns the raw residuals for a model."
492 (- (send self
:y
) (send self
:fit-values
)))
494 (defmeth regression-model-proto
:residuals
()
496 Returns the raw residuals for a model without weights. If the model
497 includes weights the raw residuals times the square roots of the weights
499 (let ((raw-residuals (send self
:raw-residuals
))
500 (weights (send self
:weights
)))
501 (if weights
(* (sqrt weights
) raw-residuals
) raw-residuals
)))
503 (defmeth regression-model-proto
:sum-of-squares
()
505 Returns the error sum of squares for the model."
506 (send self
:residual-sum-of-squares
))
508 (defmeth regression-model-proto
:sigma-hat
()
510 Returns the estimated standard deviation of the deviations about the
512 (let ((ss (send self
:sum-of-squares
))
513 (df (send self
:df
)))
514 (if (/= df
0) (sqrt (/ ss df
)))))
516 ;; for models without an intercept the 'usual' formula for R^2 can give
517 ;; negative results; hence the max.
518 (defmeth regression-model-proto
:r-squared
()
520 Returns the sample squared multiple correlation coefficient, R squared, for
522 (max (- 1 (/ (send self
:sum-of-squares
) (send self
:total-sum-of-squares
)))
525 (defmeth regression-model-proto
:coef-estimates
()
528 Returns the OLS (ordinary least squares) estimates of the regression
529 coefficients. Entries beyond the intercept correspond to entries in
531 (let ((n (array-dimension (send self
:x
) 1))
532 (indices (flatten-list
533 (if (send self
:intercept
)
534 (cons 0 (+ 1 (send self
:basis
)))
535 (list (+ 1 (send self
:basis
))))))
536 (m (send self
:sweep-matrix
)))
537 (format t
"~%REMOVEME2: Coef-ests: ~% Sweep Matrix: ~A ~% array dim 1: ~A ~% Swept indices: ~A ~% basis: ~A"
538 m n indices
(send self
:basis
))
539 (coerce (compound-data-seq (select m
(+ 1 n
) indices
)) 'list
))) ;; ERROR
541 (defmeth regression-model-proto
:xtxinv
()
543 Returns ((X^T) X)^(-1) or ((X^T) W X)^(-1)."
544 (let ((indices (if (send self
:intercept
)
545 (cons 0 (1+ (send self
:basis
)))
546 (1+ (send self
:basis
)))))
547 (select (send self
:sweep-matrix
) indices indices
)))
549 (defmeth regression-model-proto
:coef-standard-errors
()
551 Returns estimated standard errors of coefficients. Entries beyond the
552 intercept correspond to entries in basis."
553 (let ((s (send self
:sigma-hat
)))
554 (if s
(* (send self
:sigma-hat
) (sqrt (diagonalf (send self
:xtxinv
)))))))
556 (defmeth regression-model-proto
:studentized-residuals
()
558 Computes the internally studentized residuals for included cases and externally studentized residuals for excluded cases."
559 (let ((res (send self
:residuals
))
560 (lev (send self
:leverages
))
561 (sig (send self
:sigma-hat
))
562 (inc (send self
:included
)))
564 (/ res
(* sig
(sqrt (pmax .00001 (- 1 lev
)))))
565 (/ res
(* sig
(sqrt (+ 1 lev
)))))))
567 (defmeth regression-model-proto
:externally-studentized-residuals
()
569 Computes the externally studentized residuals."
570 (let* ((res (send self
:studentized-residuals
))
571 (df (send self
:df
)))
572 (if-else (send self
:included
)
573 (* res
(sqrt (/ (- df
1) (- df
(^ res
2)))))
576 (defmeth regression-model-proto
:cooks-distances
()
578 Computes Cook's distances."
579 (let ((lev (send self
:leverages
))
580 (res (/ (^
(send self
:studentized-residuals
) 2)
581 (send self
:num-coefs
))))
582 (if-else (send self
:included
) (* res
(/ lev
(- 1 lev
) )) (* res lev
))))
585 (defun plot-points (x y
&rest args
)
587 (declare (ignore x y args
))
588 (error "Graphics not implemented yet."))
590 ;; Can not plot points yet!!
591 (defmeth regression-model-proto
:plot-residuals
(&optional x-values
)
592 "Message args: (&optional x-values)
593 Opens a window with a plot of the residuals. If X-VALUES are not supplied
594 the fitted values are used. The plot can be linked to other plots with the
595 link-views function. Returns a plot object."
596 (plot-points (if x-values x-values
(send self
:fit-values
))
597 (send self
:residuals
)
598 :title
"Residual Plot"
599 :point-labels
(send self
:case-labels
)))
601 (defmeth regression-model-proto
:plot-bayes-residuals
603 "Message args: (&optional x-values)
605 Opens a window with a plot of the standardized residuals and two
606 standard error bars for the posterior distribution of the actual
607 deviations from the line. See Chaloner and Brant. If X-VALUES are not
608 supplied the fitted values are used. The plot can be linked to other
609 plots with the link-views function. Returns a plot object."
611 (let* ((r (/ (send self
:residuals
)
612 (send self
:sigma-hat
)))
613 (d (* 2 (sqrt (send self
:leverages
))))
616 (x-values (if x-values x-values
(send self
:fit-values
)))
617 (p (plot-points x-values r
618 :title
"Bayes Residual Plot"
619 :point-labels
(send self
:case-labels
))))
620 (map 'list
#'(lambda (a b c d
) (send p
:plotline a b c d nil
))
621 x-values low x-values high
)
622 (send p
:adjust-to-data
)