-
Notifications
You must be signed in to change notification settings - Fork 5
/
lec06.rnw
975 lines (848 loc) · 27.8 KB
/
lec06.rnw
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
\documentclass[table,10pt]{beamer}
\mode<presentation>{
%\usetheme{Goettingen}
\usetheme{Boadilla}
\usefonttheme[onlymath]{serif}
\usecolortheme{whale}
}
\usepackage{CJK}
\usepackage{graphicx}
\usepackage{amsmath, amsopn}
\usepackage{xcolor}
\usepackage[english]{babel}
\usepackage[T1]{fontenc}
\usepackage[latin1]{inputenc}
\usepackage{enumerate}
\usepackage{multirow}
\usepackage{url}
\ifx\hypersetup\undefined
\AtbBeginDocument{%
\hypersetup{unicode=true,pdfusetitle,
bookmarks=true,bookmarksnumbered=false,bookmarksopen=false,
breaklinks=false,pdfborder={0 0 0},pdfborderstyle={},backref=false,colorlinks=false}
}
\else
\hypersetup{unicode=true,pdfusetitle,
bookmarks=true,bookmarksnumbered=false,bookmarksopen=false,
breaklinks=false,pdfborder={0 0 0},pdfborderstyle={},backref=false,colorlinks=false}
\fi
\usepackage{breakurl}
\usepackage{color}
\usepackage{times}
\usepackage{xcolor}
\usepackage{listings}
\lstset{
language=R,
keywordstyle=\color{blue!70}\bfseries,
basicstyle=\ttfamily,
commentstyle=\ttfamily,
showspaces=false,
showtabs=false,
frame=shadowbox,
rulesepcolor=\color{red!20!green!20!blue!20},
breaklines=true}
\usepackage{algorithm,algorithmic}
\usepackage{float}
\DeclareMathOperator*{\argmax}{arg\,max}
%\newtheorem{theorem}{Theorem}
%\newtheorem{example}{Example}
%\newtheorem{definition}{Definition}
%\newtheorem{property}{Property}
\setlength{\parskip}{.5em}
\title[BI476]{BI476: Biostatistics - Case Studies}
\subtitle[anova]{Lec06: Linear Models and Generalizations}
\author[Maoying Wu]{Maoying,Wu\\{\small ricket.woo@gmail.com}}
\institute[CBB] % (optional, but mostly needed)
{
\inst{}
Dept. of Bioinformatics \& Biostatistics\\
Shanghai Jiao Tong University
}
\date{Spring, 2018}
\AtBeginSection[]
{
\begin{frame}<beamer>{Next Section ...}
\tableofcontents[currentsection]
\end{frame}
}
\begin{document}
\begin{CJK*}{UTF8}{gbsn}
<<setup, include=FALSE>>=
library(knitr)
opts_chunk$set(fig.path='figure/beamer-', fig.align='center',
fig.show='hold',size='footnotesize', out.width='0.60\\textwidth')
@
\begin{frame}
\titlepage
\end{frame}
\begin{frame}
\frametitle{Outline}
\tableofcontents
\end{frame}
\section{General Likelihood Theory}
\begin{frame}[t,containsverbatim]
\frametitle{General Likelihood Theory}
Consider a random variable $Y \sim f(Y;\theta)$ and $n$ i.i.d. observations $y=\{y_1, \dots, y_n\}$
\begin{itemize}
\item The \textbf{likelihood function}
$$
L(y;\theta) = \prod_{i=1}^n f(y_i;\theta)
$$
\item The \textbf{log-likelihood function}
$$
\ell(y;\theta) = \sum_{i=1}^n \log f(y_i;\theta)
$$
\item \textbf{Maximum likelihood estimator}
$$
\widehat{\theta} = \argmax_{\theta} L(y;\theta) = \argmax_{\theta} \ell(y;\theta)
$$
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{General Likelihood Theory}
\begin{itemize}
\item \textbf{Fisher's score function}:
$$
u(\theta) = \partial \ell(y;\theta)/\partial \theta
$$
\item $\hat{\theta}_{\text{MLE}}$ can be obtained by solving
$$
u(\theta)=0
$$
\item \textbf{Fisher's information matrix}
$$
I(\theta) = Var[u(\theta)] = E[u(\theta)u^T(\theta)] = -E\left[ \frac{\partial^2 \ell(y;\theta)}{\partial \theta \partial \theta^T}\right]
$$
\item \textbf{Asymptotic normal distribution}:
$$
\sqrt{n} (\hat{\theta} - \theta) \sim N(0_p, I^{-1}(\theta))
$$
\end{itemize}
\end{frame}
\begin{frame}[containsverbatim]
\frametitle{Newton-Raphson Method for MLE}
Generally MLE has no closed-form solution, numerical methods needed.
\begin{itemize}
\item First-order Taylor's extension for $u(\theta)$:
$$
u(\theta) \approx u(\theta^{(0)}) - I(\theta^{(0)})(\theta - \theta^{(0)})
$$
\end{itemize}
\begin{algorithm}[H]
\caption{Newton's method for MLE}
\begin{algorithmic}
\STATE Initialize with $\theta^{(0)}$
\WHILE {$\lVert \theta^{(k)} - \theta^{(k-1)} \rVert > \epsilon$}
\STATE Solve for $\theta^{(k)}$
$$
I(\theta^{(k-1)})(\theta^{(k)} - \theta^{(k-1)}) = u(\theta^{(k-1)})
$$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Likelihood Theory: Hypothesis Testing}
\framesubtitle{Wald Test}
$$
\sqrt{n}(\hat{\theta} - \theta) \sim N\left(0, I^{-1}(\theta)\right), \theta \in \mathbb{R}^p
$$
\begin{itemize}
\item $H_0: \theta = \theta_0$
\item \textbf{Test statistic}:
$$
W = (\hat{\theta}-\theta_0)^T I^{-1}(\hat{\theta})(\hat{\theta} - \theta_0)
$$
\item Asymptotically $\chi^2$ distributed:
$$
W \sim \chi_p^2
$$
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Likelihood Theory: Hypothesis Testing}
\framesubtitle{Score Test}
$$
u(\theta) \sim N\left(0_p,I(\theta)\right), \theta \in \mathbb{R}^p
$$
\begin{itemize}
\item $H_0: \theta = \theta_0$
\item \textbf{Test statistic}:
$$
Q = u^T(\theta_0) I^{-1}(\theta_0) u(\theta_0)
$$
\item Asymptotically $\chi^2$ distributed:
$$
Q \sim \chi_p^2
$$
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Likelihood Theory: Hypothesis Testing}
\framesubtitle{Likelihood Ratio Test (LRT) for Comparing Nested Models}
Two models with $p_1$ and $p_2$ ($p1 > p2$) parameters:
$$
\begin{array}{lcl}
\hat{\theta}_1 &=& \argmax_{\theta_1 \in \mathbb{R}^{p_1}} l(y;\theta_1)\\
\hat{\theta}_2 &=& \argmax_{\theta_2 \in \mathbb{R}^{p_2}} l(y;\theta_2)
\end{array}
$$
\begin{itemize}
\item Two models have no significant difference.
\item \textbf{Test statistic}:
$$
D = 2\left(l(y;\hat{\theta}_1) - l(y;\hat{\theta}_2)\right)
$$
\item Asymptotically $\chi^2$ distributed:
$$
D \sim \chi_p^2, p = p_1 - p_2
$$
\end{itemize}
\end{frame}
\section{Generalized Linear Models (GLMs)}
\begin{frame}[t,containsverbatim]
\frametitle{Generalized Linear Models (GLMs)}
\begin{itemize}
\item Extension to linear models.
\item The response $Y$ follows a conditional exponential-\textbf{family} distribution.
\item No assumption on the predictors, $X$
\item $Y$ and $X$ are connected through a linear prediction model:
$$
\eta = X\beta
$$
\begin{itemize}
\item $\eta = g(\mu)$, where $g(\cdot)$ is a \textbf{link function} of the
conditional expected value of $Y$ given $X$.
\item $\mu = E(Y|X)$ is the expected value of $Y$ given $X$.
\end{itemize}
\end{itemize}
\begin{lstlisting}
glm(formula = Y ~ X, family=gaussian(link="identity"))
\end{lstlisting}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Three Components of GLMs}
\framesubtitle{Exponential Family Distributions}
An exponential family is a family of distributions with pdf of the form:
$$
f(Y;\theta,\phi) = \exp\left( \frac{Y\theta - b(\theta)}{a(\phi)} + c(\phi, Y)\right)g(Y)
$$
where
\begin{itemize}
\item $g(Y)$ does not depend on $\theta$ or $\phi$;
\item $\theta$: canonical parameter; $\phi$: dispersion parameter;
\item $a(\cdot), b(\cdot), c(\cdot)$ are known functions.
\end{itemize}
\begin{example}[Gaussian Distribution]
$$
Y \sim \frac{1}{\sqrt{2\pi\sigma^2}} \exp\left(-\frac{(Y-\mu)^2}{2\sigma^2}\right)
$$
\begin{itemize}
\item $\theta = \mu$; $b(\theta) = \mu^2/2$
\item $\phi = \sigma; a(\phi) = \sigma^2$
\item $c(\phi,Y) = -Y^2/(2\sigma^2) - \log(2\pi)$
\end{itemize}
\end{example}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Three Components of GLMs}
\framesubtitle{Exponential Families: Properties}
\begin{itemize}
\item Canonical statistic $t(y) = y$;
\item Canonical parameter $\theta^* = \theta/a(\phi)$
\item The cumulant generating function
$$
\kappa\{\theta^*\} = \kappa\{\theta/a(\phi)\} = b(\theta)/a(\phi)
$$
\item The expected value of $Y$:
$$
E(Y) = \frac{\partial}{\partial \theta^*} \kappa\{\theta/a(\phi)\} = a(\phi) \frac{\partial}{\partial \theta} \kappa\{\theta/a(\phi)\} = b'(\theta)
$$
\item The variance of $Y$:
$$
V(Y) = \frac{\partial^2}{\partial \theta^{*2}}\kappa\{\theta/a(\phi)\} = a(\phi)^2 \frac{\partial^2}{\partial \theta^2} \kappa\{\theta/a(\phi)\} = b''(\theta)a(\phi)
$$
\item If $Y_i \stackrel{\text{i.i.d}}{\sim} f(Y_i;\theta, \phi)$ for $i=1,\dots,n$,
then the joint distribution of $Y_i$'s also follows an exponential-family distribution. And
$\sum Y_i$ is the \textbf{sufficient and complete statistics} for the canonical parameter.
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Three Components of GLMs}
\framesubtitle{Linear Predictors}
$$
\eta = X \beta = \beta_0 + X_1\beta_1 + \dots + X_p\beta_p
$$
is the dot product between the design matrix ($X = (\mathbf{1}, X_1, \dots, X_p)$) and the coefficient ($\beta = (\beta_0, \dots, \beta_p)^T$).
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Three Components of GLMs}
\framesubtitle{Link Function}
$$
\begin{aligned}
E(Y) = \mu\\
g(\mu) = \eta = \beta_0 + X_1\beta_1 + \dots + X_p\beta_p
\end{aligned}
$$
\begin{itemize}
\item $\eta = g(\mu)$ must be 1-1 mapping.
\item Usually $g(\cdot)$ is chosen to be a simple, continuously differentiable
function.
\item If possible, $g(\cdot)$ should be chosen based on data.
\end{itemize}
\begin{example}[Gaussian Distribution]
Since $b(\theta) = \mu^2/2 = \theta^2/2$, $E(Y) = b'(\theta) = \theta = \mu$
The canonical link function is
$$
\eta = E(Y) = \mu
$$
\end{example}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Iteratively Reweighted Least Square (IRWLS)}
\framesubtitle{MLE Solution for GLMs}
\begin{itemize}
\item $Y_i \sim \exp\left( \frac{y_i \theta_i - b(\theta_i)}{a(\phi)} + c(\phi,y_i)\right)$
\item $\theta_i = \theta(\mu_i) = \theta(h(\eta_i)) = \theta(h(x_i^T \beta)), h=g^{-1}$
\item The log-likelihood
$$
\begin{array}{lcl}
l(\beta) &=& \sum_{i=1}^n l_i(\beta)\\
l_i(\beta) &=& \frac{1}{a(\phi)}\left[ y_i\theta(h(x_i^T \beta)) - b(\theta(h(x_i^T \beta)))\right]
\end{array}
$$
\item $\beta$ can be obtained through Newton-Raphson's method with Fisher scoring
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Score function}
$$
u_r(\beta) = \frac{\partial l(\beta)}{\partial \beta_r} =
\sum_{i=1}^n u_{ir} = \sum_{i=1}^n \frac{\partial l_i(\beta)}
{\partial \beta_r}
$$
We can Compute $u_{ir}$:
$$
\frac{\partial l_i(\beta)}{\partial \beta_r} = \frac{\partial l_i}{\partial \theta_i}
\frac{\partial \theta_i}{\partial \mu_i}
\frac{\partial \mu_i}{\partial \eta_i}
\frac{\partial \eta_i}{\partial \beta_r}
$$
where
\begin{itemize}
\item $\frac{\partial l_i}{\partial \theta_i} = \frac{y_i - b'(\theta_i)}{a(\phi)} = \frac{y_i - \mu_i}{a(\phi)}$
\item $\frac{\partial \theta_i}{\partial \mu_i} = \left( \frac{\partial \mu_i}{\partial \theta_i}\right)^{-1} = v_i^{-1}$ since
\begin{itemize}
\item $\mu_i = b'(\theta_i)$
\item $\frac{\partial \mu_i}{\partial \theta_i} = b''(\theta_i) = Var(\mu_i) = v_i$
\end{itemize}
\item $\frac{\partial \eta_i}{\partial \beta_r} = x_{ir}$
\end{itemize}
$$
\frac{\partial l_i}{\partial \beta_r} = \frac{1}{a(\phi)} x_{ir} w_i \frac{\partial \eta_i}{\partial \mu_i} (y_i - \mu_i)
$$
where $w_i = v_i^{-1}(\partial \eta_i / \partial \mu_i)^{-2}$
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Score Function}
$$
\frac{\partial l}{\partial \beta} = \frac{1}{a(\phi)} X^T W \left[ \frac{\partial \eta}{\partial \mu} (Y - \mu)\right]
$$
where $W = \text{diag}(w_i)$.
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Fisher's Information Matrix}
Similarly, we can get
$$
A = \frac{1}{a(\phi)} X^T W X
$$
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Newton-Raphson Method for Fisher Scoring}
$$
X^T W(\beta^{(t-1)}) X (\beta^{(t)} - \beta^{(t-1)}) = X^T W(\beta^{(t-1)}) \left[ \frac{\partial \eta}{\partial \mu} (Y - \mu(\beta^{(t-1)}))\right]
$$
Then we can get:
$$
\begin{array}{lcl}
X^T W(\beta^{(t-1)}) X \beta^{(t)} &=& X^T W(\beta^{(t-1)}) \left\{ X\beta^{(t-1)} \left[\frac{\partial \eta}{\partial \mu} (Y - \mu(\beta^{(t-1)}))\right] \right\}\\
&=& X^T W(\beta^{(t-1)})z(\beta^{(t-1)})
\end{array}
$$
That is the reason why it is called \textbf{Iteratively Reweighted Least Squares}.
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{A Clinical Trial on Diastolic Blood Pressure (DBP)}
Here we present a data set of diastolic blood pressure measured in
small clinical trials in hypertension from the mid-to-late 1960s and
for approximately a decade thereafter. Diastolic blood pressure (DBP)
was measured (mmHg) in the supine position at baseline (i.e., ``DBP1'')
before randomization and monthly thereafter up to 4 months as indicated
by \texttt{DBP2}, \texttt{DBP3}, \texttt{DBP4} and \texttt{DBP5}.
Patients' age and sex were recorded at baseline and represent potential
covariates.
The primary objective in the analysis of this dataset is to test
whether treatment A (new drug) may be effective in lowering DBP as
compared to B (placebo) and to describe changes in DBP across the times
at which it was measured.
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Importing the dataset}
<<eval=TRUE>>=
dbp <- read.table("data/dbp.txt", header=T)
@
\end{frame}
\section{General Linear Regression}
\begin{frame}[t,containsverbatim]
\frametitle{Linear regression models}
$$
y = X\beta + \epsilon
$$
\begin{itemize}
\item The quantitative ``predictors'', $X$, contains the
treatments in the clinical trials as well as other
factors.
\item $X$ is the so-called design matrix with one column for
treatment.
\item $y$ is the measured clinical endpoint (e.g., DBP)
\item $\beta$ is the vector of regression parameters
\item $\epsilon \sim N(0, \sigma^2)$ is the error term
\item Treatment can be incorporated into the regression using
a \emph{dummy variable}, $X_1$ as
$$
X_1 = \begin{cases}
0 & TRT = A\\
1 & TRT = B
\end{cases}
$$
\item $X_2 = Age, X_3 = DBP1$
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Estimation of the Linear Models}
\begin{itemize}
\item Least-squares estimation of $\beta$ by minimizing
$$
\begin{array}{lcl}
\sum_{i=1}^n \epsilon_i^2 & = & (y-X\beta)^T (y-X\beta)\\
& = & y^T y - 2\beta X^T y + \beta^T X^T X \beta
\end{array}
$$
\item Taking the partial derivative of the error sum of squares
with respect to the component of $\beta$ and setting to
zero leads to
$$
X^T X \beta = X^T \hat{\beta}
$$
\item When $X^T X$ is invertible
$$
\hat{\beta} = (X^T X)^{-1} X^T y
$$
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Least-squares estimation}
\begin{description}
\item[Predicted values]\hfill\\
$\hat{y} = X\hat{\beta} = X(X^T X)^{-1} X^T y = Hy$,
where $H=X(X^T X)^{-1} X^T$ is called the \emph{hat matrix}
\item[Residuals for diagnostics]\hfill\\
$\hat{\epsilon} = y - \hat{y} = y - X\hat{\beta} = (I-H)y$
\item[Residual sum of squares]\hfill\\
$RSS = \hat{\epsilon}^T \hat{\epsilon} = y^T(I-H)^T(I-H)y = y^T (I-H) y$
\item[Unbiasedness]\hfill\\
$\hat{\beta}$ is unbiased with variance $var(\hat{\beta}) = (X^T X)^{-1} \sigma^2$ if $var(\epsilon) = \sigma^2 I$
\item[Variance estimate]\hfill\\
$\hat{\sigma}^2 = \frac{\hat{\epsilon}^T \hat{\epsilon}}{n-p} = \frac{RSS}{n-p}$,
where $n-p$ is the \emph{degrees of freedom} of the model.
\item[Coefficient of determination]\hfill\\
$R^2 = \frac{(\hat{y} - y)^T (\hat{y} - y)}{(y-\bar{y})^T (y-\bar{y})}$
\end{description}
\end{frame}
\section{Logistic Regression}
\begin{frame}[t,containsverbatim]
\frametitle{Clinical Trials for Beta Blockers}
This is a multi-center clinical trials in 22 centers to evaluate the
efficacy of beta-blockers in reducing the mortality after myocardial
infarction.
\begin{itemize}
\item \texttt{Deaths} is the number of deaths and \texttt
{Total} is the total number of patients enrolled at
each clincal \texttt{Center}.
\item \texttt{Treatment} represents whether patients at each
center were randomzied to \texttt{Control} or
\texttt{Treated} (Beta-blocker)
\item The outcome (death or non-death) are binomial and are
used to demonstrate the application of logistic
regression as well as remedies for over-dispersion
using a \texttt{quasi-likelihood} approach.
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Importing the data}
<<eval=TRUE>>=
betablocker <- read.table("data/betablocker.txt", header=TRUE)
head(betablocker)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Logistic regression for binary/binomial outcomes}
\begin{itemize}
\item When the response is Bernoulli (e.g., death/alive or
cured/uncured) or binomial (e.g. number of deaths among
a fixed total number of patients)
\item Linear regression method may yield biased estimates of
the covariate parameters.
\item $Y_i \sim \text{Bin}(n_i, p_i)$ can be written as
$$
P(Y_i = y_i) = \binom{n_i}{y_i} p_i^{y_i} (1-p_i)^{n_i - y_i}
$$
\item In the generalized linear model (GLM) framework, a \textbf
{linear predictor} is used to model the linear
relationship and a \textbf{link function} to link the
linear predictor to the binomial probability $p_i$
\begin{itemize}
\item Linear predictor
$$
\eta_i = \beta_0 + \beta_1 x_{i1} + \dots + \beta_q x_{iq} = X_i \beta
$$
\item The most commonly used link function for binomial
response is the \textbf{logit} function
$$
\eta = \log\left( \frac{p}{1-p}\right)
$$
\item Other link functions: \textbf{probit} and \textbf{complementary log-log}
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{MLE for Binomial Data}
\begin{itemize}
\item The likelihood function
$$
L(\beta | y) = \prod_{i=1}^r \binom{n_i}{y_i} p_i^{y_i}(1-p_i)^{n_i - y_i}
$$
\item The log-likelihood function
$$
\ell(\beta | y) = \sum_{i=1}^r \left[ \binom{n_i}{y_i} + y_i \log(p_i) + (n_i - y_i) \log(1-p_i)\right]
$$
\item For \textbf{logit} link function, $\eta = \log(\frac{p}{1-p})$, then $p=\frac{e^{\eta}}{1+e^{\eta}}$. The loglikelihood function
becomes
$$
\begin{array}{lcl}
\ell(\beta | y) & = & \sum_{i=1}^r \left[ \log \binom{n_i}{y_i} + y_i \log(p_i) + (n_i - y_i) \log(1-p_i) \right]\\
& = & \sum_{i=1}^r \left[ y_i \eta_i - n_i \log(1+e^{\eta_i}) + \log \binom{n_i}{y_i} \right]
\end{array}
$$
\item This can be solved using numerical iterative approach.
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Fitting a logistic regression model}
<<eval=TRUE>>=
# fit a logistic regression using glm
beta.glm <- glm(cbind(Deaths, Total-Deaths) ~ Treatment + Center,
family=binomial(link="logit"), data=betablocker)
# print the model fitting
summary(beta.glm)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Over-dispersion issues}
\begin{itemize}
\item \textbf{Over-dispersion} is a common phenomenon in
GLM including logistic and Poisson regression.
\item Over-dispersion is evident if the deviance from the
fitted model is too large.
\item In the MLE theory, redidual deviance is asymptotically
distributed as $\chi^2$ with appropriate degrees of
freedom.
\item For $\chi^2$ distributed deviance, its value should be
close to its degrees of freedom.
\begin{itemize}
\item When the deviance is greater than the df,
\textbf{over-dispersion} occurs.
\item When the deviance is smaller than the df,
\textbf{under-dispersion} occurs.
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Estimate and Adjust the Dispersion parameter}
Two-stage approach
\begin{itemize}
\item Estimate the dispersion parameter using Pearson's
$\chi^2$ as
$$
\hat{\phi} = \frac{\chi^2}{n-p}
$$
\item Use the estimated dispersion parameter to adjust the
model fit for further statistical inference.
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Estimating and fitting the dispersion parameter}
<<eval=TRUE>>=
est.dp <- sum(resid(beta.glm, type="pearson")^2)/beta.glm$df.res
summary(beta.glm, dispersion=est.dp)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Over-dispersion: Quasi-likelihood approach}
Combined approach.
\begin{itemize}
\item Permits estimating the dispersion parameter along with
the model parameters simultaneously - without assuming
an error distribution.
\item This approach requires only the mean and variance
function to be specified.
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Quasi-likelihood for binomial data}
<<eval=TRUE>>=
# fit a quasi-likelihood for binomial data
beta.glm2 <- glm(cbind(Deaths, Total-Deaths)~Treatment + Center,
family=quasibinomial, data=betablocker)
# print the model fitting
summary(beta.glm2)
@
\end{frame}
\section{Poisson Regression}
\begin{frame}[t,containsverbatim]
\frametitle{A Clinical Trial on Familial Andenomatous Polyposis}
A placebo-controlled clinical trial of a non-sterioidal anti-
inflammatory drug (NSAID) in treating FAP.
A planned interim analysis of the number of polyps to reveal
significant evidence favoring NSAID treatment to warrant termination of
the trial.
\begin{itemize}
\item \texttt{number}: Number of colonic polyps after 12 months
of treatment
\item \texttt{age}: Age at the baseline for patient.
\item \texttt{treat}: Treatment assignment allocation,
``drug'' or ``placebo''
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Importing the data}
<<eval=TRUE>>=
polyps <- read.table("data/polyps.txt", header=TRUE)
head(polyps)
str(polyps)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Poisson regression for clinical outcome with counts}
If $Y$ is Poisson ditributed with mean $\lambda \ge 0$,
$$
P(Y = y) = \frac{e^{\mu} \mu^y}{y!}
$$
For a Poisson random variable $Y$, $E(Y) = Var(Y) = \mu$.
Suppose $Y_i \sim Pois(\mu)$ represents count data and that we want to
model $Y_i$ as a function of a set of clinical covariates of $(x_1,
\dots, x_q)$.
A link function is needed to link $\mu_i$ to $(x_1, \dots, x_q)$ with
the linear predictor $\eta_i = X_i \beta$. The natural canonical log
link function can be written as:
$$
\log(\mu_i) = \eta_i = X_i \beta
$$
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Poisson regression}
The log-likelihood function is
$$
\begin{array}{lcl}
\ell{\beta} & = & \sum_{i=1}^n \log P(Y = y) = \sum_{i=1}^n \log \left[ \frac{e^{\mu_i} \mu_i^{y_i}}{y_i!} \right]\\
& = & \sum_{i=1}^n \left[ y_i x_i^T \beta - \exp(x_i^T \beta) - \log(y_i!)\right]
\end{array}
$$
Differetiating with respect to $\beta$ gives the MLE for $\hat{\beta}$ as the solution to
$$
\sum_{i=1}^n \left[ y_i - \exp(x_i^T \hat{\beta})\right] = 0
$$
There is no closed-form analytic solution. Numerical search algorithms
are needed.
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Fitting a Poisson Model to the count data}
<<eval=TRUE>>=
# Poisson regression
m0.polyps <- glm(number ~ treat*age, polyps,
family=poisson(link="log"))
# print the model fit
summary(m0.polyps)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Estimating and applying the dispersion parameter}
<<eval=TRUE>>=
est.dp <- sum(resid(m0.polyps, type="pearson")^2)/m0.polyps$df.res
summary(m0.polyps, dispersion=est.dp)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Refitting Poisson model without interaction}
<<eval=TRUE>>=
# Poisson regression
m1.polyps <- glm(number ~ treat + age, polyps,
family=poisson(link="log"))
# print the model fit
summary(m1.polyps)
est.dp <- sum(resid(m1.polyps, type="pearson")^2)/m1.polyps$df.res
summary(m1.polyps, dispersion=est.dp)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Quasi-Poisson Regression Model}
<<eval=TRUE>>=
# fit the quasi-Poisson
m2.polyps <- glm(number ~ treat + age, polyps,
family=quasipoisson())
# print the model fit
summary(m2.polyps)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Negative Binomial Approach}
Another remedy to overdispersion is to use a more general distribution
to relax the dependence of the mean and variance function, such as the
negative binomial (NB) or gamma distribution to model the over-
dispersion.
For a series of independent trials with $P(\text{success})=p$, the
random variable $N$ for the number of trials until the $k^{th}$ success
is observed has a \textbf{negative-binomial} distribution with mdf:
$$
P(N=n) = \binom{n-1}{k-1}p^k (1-p)^{n-k}
$$
The NB is an extension of the Poisson distribution from a Bayesian
perspective. It can be thought of as the Gamma-Poisson mixture. That
is, NB is a $\text{Poisson}(\lambda)$, where $\lambda \sim \Gamma(k, \frac{p}{1-p})$.
Mathematically, the NB is reparameterized for convenience in model
fitting by defining $Y=N-k$ and $p=\frac{1}{1+\alpha}$ as
$$
P(Y=y) = \binom{y+k-1}{k-1}\frac{\alpha^k}{(1+\alpha)^{y+k}}
$$
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Negative Binomial Approach}
Therefore,
$$
\begin{aligned}
E(Y) = \mu = k\alpha\\
Var(Y) = k\alpha + k\alpha^2 = \mu + \mu^2/k
\end{aligned}
$$
where an extra term $k\alpha^2$ can be used to model over-dispersion.
The log-likelihood function can be written as
$$
\ell(\alpha, k) = \sum_{i=1}^n \left( y_i \log\frac{\alpha}{1+\alpha} - k\log(1+\alpha) + \sum_{j=0}^{y_j-1}\log(j+k) - \log(y!)\right)
$$
The \emph{link function} to link mean response $\mu$ to a linear
combination of the clinical covariates $X$ is
$$
\eta = X\beta = \log\frac{\alpha}{1+\alpha} = \log \frac{\mu}{\mu+k}
$$
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Fitting a negative-binomial regression model}
<<eval=TRUE>>=
library(MASS)
# fit the negative binomial model
m3.polyps <- glm.nb(number ~ treat + age, polyps)
# print the model fit
summary(m3.polyps)
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Gamma-Poisson Model}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Relationship of Exponential, Poisson and Gamma Distributions}
Poisson and exponential distributions are very strongly related but they're fundamentally
different because the Poisson is discrete (a count variable) and the exponential is continuous
(a waiting time).
\begin{itemize}
\item If the time between a certain type of event is exponentially distributed with
rate $\lambda$, then the number of events in a given time period of length $t$
follows a Poisson distribution with parameter $\lambda t$.
\item The time to observe exacly $n$ events is a sum of independent exponentially
distributed random variables, and it follows a $\text{Gamma}(\lambda,n)$ distribution
(a.k.a \textbf{Erlang} distribution, to distinguish it from the general gamma
distribution where $n$ is allowed to be a non-integer).
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Beta-Binomial Model - Priors}
\begin{itemize}
\item If $X \sim \text{Bernoulli}(\theta)$, then for $D=\{X_1, X_2, \dots, X_n\}$, the likelihood
$$
p(D \big| \theta) = \theta^{n_1}(1-\theta)^{n_0}
$$
where
\begin{itemize}
\item $n_1 = \sum_{i=1}^n I(X_i = 1)$
\item $n_0 = \sum_{i=1}^n I(X_i=0)$
\end{itemize}
\item $n_1 \sim \text{Bin}(n, \theta)$
\item For the sake of convenience, the prior for the parameter
$\theta$ has the same form as the likelihood function
(conjugate prior), then
$$
p(\theta) = \theta^{\gamma_1}(1-\theta)^{\gamma_0}
$$
\item In a binomial model, the conjugate prior is a Beta
distribution:
$$
\text{Beta}(\theta|a,b) \propto \theta^{a-1}(1-\theta)^{b-1}
$$
where $a,b$ are the hyperparameters.
\end{itemize}
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Beta-Binomial Distribution}
<<eval=FALSE>>=
R <- 10000
T <- 20000
x <- matrix(NA, R+T, 2)
n <- 10
alpha <- 7
beta <- 19
x[1,] <- c(1, .5)
for (i in 2:(R+T)) {
x[i,2] <- rbeta(1, alpha, beta)
x[i,1] <- rbinom(1, n, x[i,2])
}
x <- x[(R+1):(R+T), ]
plot(table(x[,1])/T, col="sienna4", type="p", pch=20)
betabinomialdensity <- function(x, n, alpha, beta)
choose(n, x)*beta(alpha+x, beta+n-x)/beta(alpha,beta)
points(0:n, betabinomialdensity(0:n, n, alpha, beta),
type="o", pch=22, lty=2, col="red")
@
\end{frame}
\begin{frame}[t,containsverbatim]
\frametitle{Beta-Binomial Distribution}
<<eval=TRUE, echo=FALSE>>=
R <- 10000
T <- 20000
x <- matrix(NA, R+T, 2)
n <- 10
alpha <- 7
beta <- 19
x[1,] <- c(1, .5)
for (i in 2:(R+T)) {
x[i,2] <- rbeta(1, alpha, beta)
x[i,1] <- rbinom(1, n, x[i,2])
}
x <- x[(R+1):(R+T), ]
plot(table(x[,1])/T, col="sienna4", type="p", pch=20)
betabinomialdensity <- function(x, n, alpha, beta)
choose(n, x)*beta(alpha+x, beta+n-x)/beta(alpha,beta)
points(0:n, betabinomialdensity(0:n, n, alpha, beta),
type="o", pch=22, lty=2, col="red")
@
\end{frame}
\end{CJK*}
\end{document}