This repository has been archived by the owner on Feb 6, 2020. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 1
/
2018_ipwbootstrap_zhao.tex
516 lines (414 loc) · 17 KB
/
2018_ipwbootstrap_zhao.tex
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
\documentclass{beamer}
\usetheme{metropolis}
\usepackage[english]{babel}
\usepackage[utf8x]{inputenc}
\usepackage{amstext}
%\usepackage{coloremoji}
\usepackage{layout}
\usepackage{multirow}
\usepackage{graphicx}
\graphicspath{ {figs/} }
\setbeameroption{hide notes}
\setbeamertemplate{note page}[plain]
\usepackage{listings}
\usepackage{datetime}
\usepackage{url}
% specifications for presenter mode
%\beamerdefaultoverlayspecification{<+->}
%\setbeamercovered{transparent}
% math shorthand
\usepackage{bm}
\usepackage{amsmath}
\usepackage{mathtools}
\newcommand{\R}{\mathbb{R}}
\newcommand{\D}{\mathcal{D}}
\newcommand{\E}{\mathbb{E}}
\newcommand{\pr}{\mathbb{P}}
\newcommand{\F}{\mathcal{F}}
\newcommand{\X}{\mathcal{X}}
\newcommand{\lik}{\mathcal{L}}
\DeclarePairedDelimiterX{\infdivx}[2]{(}{)}{%
#1\;\delimsize\|\;#2%
}
\newcommand{\infdiv}{D\infdivx}
\DeclarePairedDelimiter{\norm}{\lVert}{\rVert}
\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator*{\argmax}{arg\,max}
% indepndence notation macro
\newcommand\independent{\protect\mathpalette{\protect\independenT}{\perp}}
\def\independenT#1#2{\mathrel{\rlap{$#1#2$}\mkern2mu{#1#2}}}
% Bibliography
\usepackage{natbib}
\bibpunct{(}{)}{,}{a}{}{;}
\usepackage{bibentry}
%\nobibliography*
\title[ipw-bootstrap]{Sensitivity Analysis for Inverse Probability Weighting
Estimators via the Percentile Bootstrap \small (Q.~Zhao, D.S.~Small, \&
B.B.~Bhattacharya, 2017)}
\subtitle{\vspace*{0.5em} \scriptsize for ``Observational Study Design and
Causal Inference'' (Statistics 260),\\ organized by S.~Pimentel, Spring 2018,
University of California, Berkeley}
\author{Nima Hejazi}
\institute{Group in Biostatistics,\\ University of California, Berkeley\\
\url{https://statistics.berkeley.edu/~nhejazi}
}
\date{12 April 2018}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\AtBeginSubsection[]{
\begin{frame}{Outline}
\tableofcontents[currentsection,currentsubsection]
\end{frame}
}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
\begin{frame}
\titlepage
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Introduction}
\subsection{Preliminaries}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Preliminaries: Notation}
\begin{itemize}
\itemsep12pt
\item Data: $(A_1, \bm{X_1}, Y_1), \ldots, (A_n, \bm{X_n}, Y_n)
\stackrel{\text{iid}}{\sim} F_0$, for a (binary) treatment $A_i$ and
measured confounders $\bm{X_i}$.
\item Outcome: $Y_i = A_iY_i(1) + (1 - A_i)Y_i(0)$, using potential outcome
notation.
\item Estimand (Parameter): $\Delta \coloneqq \E_0[Y(1)] - \E_0[Y(0)]$\\
(average treatment effect)
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Preliminaries: Identifiability and Assumptions}
\begin{itemize}
\itemsep12pt
\item Stable Unit Treatment Value Assumption (SUTVA)
\item No Unmeasured Confounders (NUC): $(Y(0), Y(1)) \independent A \mid
\bm{X}$\\(strong ignorability)
\item Overlap: $e_0(x) \coloneqq \pr_0(A = 1 \mid X = x) \in (0, 1)$\\
(bounded propensity score)
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Preliminaries: IPW Estimators}
\begin{itemize}
\itemsep12pt
\item Inverse Probability of Weighting (IPW) Estimators:
\[
\hat{\Delta}_{\text{IPW}} = \frac{1}{n} \sum_{i = 1}^n
\frac{A_iY_i}{\hat{e}(X_i)} - \frac{(1 - A_i)Y_i}{1 - \hat{e}(X_i)}
\]
\item $\hat{\Delta}$ consistently estimates $\Delta$ as long as
$\hat{e}(\bm{X})$ converges to $e_0$.
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Objectives}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{The Objective: Sensitivity Analysis}
\begin{itemize}
\itemsep12pt
\item Under violations of the NUC assumption, $\hat{\Delta}_{\text{IPW}}$ is
biased and has a confidence interval that doesn't cover $\Delta$ properly.
\item The goal of a sensitivity analysis is to gauge the degree to which a
statistical inference is incorrect under violations of the NUC assumption.
\begin{itemize}
\itemsep6pt
\item To what extent could the existence of potentially unmeasured
confounders invalidate our findings?
\item ...
\end{itemize}
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{The Objective: Sensitivity Analysis}
\begin{itemize}
\itemsep12pt
\item Under the NUC assumption, we have the following
\[
e_a(x,y) \coloneqq \pr_0(A = 1 \mid \bm{X} = x, Y(a) = y) = e_a(x),
\]
for $a \in \{0, 1\}$.
\begin{itemize}
\itemsep6pt
\item $e_a(x,y)$ --- ``complete data'' selection probability.
\item $e_a(x)$ --- ``observed data'' selection probability.
\end{itemize}
\item Thus, a sensitivity model might consider gauging whether $e_a(x,y) =
e_a(x)$ holds in order to assess violations of NUC.
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Methodology}
\subsection{Sensitivity: Parameter, Analysis, Inference}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Sensitivity Parameter}
\begin{itemize}
\itemsep12pt
\item Marginal Sensitivity Model: let
\[
\mathcal{E}(\Lambda) = \left\{e(x,y): \frac{1}{\Lambda} \leq
\text{OR}(e(x,y), e_0(x)) \leq \Lambda, x \in \mathcal{X}, y \in
\R \right\}
\]
\item Then, for observational causal inference, let us assume that
$e_a(x,y) \in \mathcal{E}(\Lambda),$ for $a \in \{0, 1\}$.
\item ...
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Sensitivity Parameter}
\begin{itemize}
\itemsep12pt
\item For convenience, we'll use a logistic representation of the sensitivity
model:
\[ h_0(x,y) = g_0(x) - g_0(x,y),\]
where
\begin{itemize}
\itemsep6pt
\item $g_0(x)=\text{logit}(e_0(x))=\text{log}\frac{e_0(x)}{1-e_0(x)}$
\item Similarly, let $g_0(x,y) = \text{logit}(e_0(x,y))$.
\end{itemize}
\item Then, we may express $\mathcal{E}(\Lambda)$ as
\[\mathcal{E}(\Lambda) = \{e^{(h)}(x,y): h \in \mathcal{H}(\lambda)\},\]
where $\mathcal{H}(\lambda) = \{h: \mathcal{X} \times \R \to \R$ and
$\lVert h \rVert_{\infty} \leq \lambda\}$.
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Parametric Model for Sensitivity Analysis}
\begin{itemize}
\itemsep12pt
\item Ideally, $e_0(x)$ would be estimated nonparametrically, but, in many
cases, we restrict ourselves to parametric models:
\[
\begin{aligned}
e_{\beta_0}(x) =& \argmin_{\beta \in \Theta} \text{KL}(\pr_0(A = 1 \mid
\bm{X} = x) \Vert \pr_{\beta}(A = 1 \mid \bm{X} = x))\\
=& \argmax_{\beta \in \Theta} \E_0[A \cdot \text{log}e_{\beta}(X) +
(1 - A) \cdot \text{log}(1 - e_{\beta}(x)) \mid \bm{X} = x]
\end{aligned}
\]
\item As before, now have $e_0(x,y) \in \mathcal{E}_{\beta_0}(\Lambda)$, where
\[\mathcal{E}_{\beta_0}(\Lambda) \coloneqq \left\{e(x,y): \frac{1}{\Lambda}
\leq \text{OR}(e(x,y), e_{\beta_0}(x,y)) \leq \Lambda, x \in \mathcal{X},
y \in \R \right\} \]
\end{itemize}
\note{
This parametric setup broadens the definition of sensitivity analysis:
\begin{itemize}
\item Model misspecification: $e_{\beta_0}(x) \neq e_0(x)$
\item No unmeasured confouding: $e_0(x) \neq e_0(x,y)$
\end{itemize}
}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Confidence Intervals in Sensitivity Analysis}
\begin{itemize}
\itemsep12pt
\item ($1 - \alpha$)--CI: $\pr_0(\Delta \in [\text{L}, \text{U}]) \geq 1 -
\alpha$ is true for any $F_0$ s.t.~$h_0 \in \mathcal{H}(\lambda)$, a
collection of sensitivity models.
\item Asymptotic ($1 - \alpha$)--CI if $\liminf\limits_{n \rightarrow \infty}
\pr_0(\Delta \in [\text{L}, \text{U}]) \geq 1 - \alpha$
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Inverse Probability Weighting Estimators}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{IPW Point Estimates}
\begin{itemize}
\itemsep12pt
\item Note that the ATE may be expressed $\hat{\Delta}^{(h_0, h_1)} \coloneqq
\hat{\mu}^{(h_1)}(1) - \hat{\mu}^{h_0}(0)$ for $h_0, h_1 \in
\mathcal{H}(\lambda)$.
\item Recall that the IPW estimator for a the mean $\mu$ in a missing data
problem may be expressed:
\[
\hat{\mu}^{(h)}_{\text{IPW}} = \frac{1}{n} \sum_{i = 1}^n \frac{A_i
Y_i}{\hat{e}^(h)(\bm{X_i}, Y_i)}
\]
\item The \textit{stabilized} IPW (SIPW) estimator is often used instead
\[
\hat{\mu}^{(h)}_{\text{SIPW}} = \left[\frac{1}{n} \sum_{i = 1}^n
\frac{A_i}{\hat{e}^(h)(\bm{X_i}, Y_i)}\right]^{-1} \left[\frac{1}{n}
\sum_{i = 1}^n \frac{A_iY_i}{\hat{e}^(h)(\bm{X_i}, Y_i)}\right]
\]
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Interval Construction for SIPW Estimators}
\begin{itemize}
\itemsep12pt
\item \textbf{Problem:} need to estimate variance computationally (e.g., via
bootstrap) while considering all violations of NUC in
$\mathcal{H}(\lambda)$.
\item Given ($1-\alpha$)--CIs for all $h \in \mathcal{H}(\lambda)$:
\[
\liminf\limits_{n \rightarrow \infty} \pr_0(\mu^{(h)} \in [L^{(h)},
U^{(h)}]) \geq 1 - \alpha
\]
\item Then, we have that the following is an asymptotic ($1-\alpha$)--CI under
the collection of sensitivity models $\mathcal{H}(\lambda)$:
\[
L = \inf_{h \in \mathcal{H}(\lambda)}L^{(h)}, U = \sup_{h \in
\mathcal{H}(\lambda)}U^{(h)}
\]
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Percentile Bootstrap and Interval Construction}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Percentile Bootstrap}
\begin{itemize}
\itemsep12pt
\item Finally, the centerpiece! But first, we need still more notation.
\item Let $\pr_n$ be the empirical measure on the sample $\bm{T_1}, \ldots,
\bm{T_n}$, where $\bm{T_i} = (A_i, \bm{X'_{i}}, A_i Y_i)$.
\item Further, let $\bm{\hat{T_1}}, \ldots, \bm{\hat{T_n}}$ be
i.i.d.~re-samples from $\pr_n$.
\item Then, the SIPW estimate $\hat{\hat{\mu}}^{(h)}$ may be computed over the
bootstrap re-samples $\{\bm{\hat{T_i}}\}_{i \in [n]}$.
\item Now, for $h \in \mathcal{H}(\lambda)$, percentile bootstrap CI:
\[
[L^{(h)}, U^{(h)}] = \left[Q_{\frac{\alpha}{2}}(\hat{\hat{\mu}}^{(h)}),
Q_{1 - \frac{\alpha}{2}}(\hat{\hat{\mu}}^{(h)})\right],
\]
where $Q_{\alpha}(\hat{\hat{\mu}}) \coloneqq \inf\{t:
\hat{\pr}_n(\hat{\hat{\mu}} \leq t) \geq \alpha\}$
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Percentile Bootstrap}
\begin{itemize}
\itemsep8pt
\item The percentile bootstrap interval $[L^{(h)}, U^{(h)}]$ is an
asymptotically valid CI of the target estimate for the parametric
sensitivity model $e^{(h)} \in \mathcal{E}_{\beta_0}(\Lambda)$.
\item The bootstrap is not valid if the missingness probability is modeled
nonparametrically (Abadie \& Imbens, 2008).
\item Percentile bootstrap CI under collection of sensitivity models:
\[
[L, U] = \left[\left(Q_{\frac{\alpha}{2}}(\inf_{h \in
\mathcal{H}(\lambda)} \hat{\hat{\mu}}^{(h)})\right), \left(Q_{1 -
\frac{\alpha}{2}}(\sup_{h \in \mathcal{H}(\lambda)}
\hat{\hat{\mu}}^{(h)})\right)\right]
\]
\begin{itemize}
\item \small Since infimum/supremum is inside the quantile function, the
problem is efficiently solved by linear programming.
\item \small Exchange of quantile and infimum/supremum is justified by a
generalized (von Neumann's) minimax/maximin inequality.
\end{itemize}
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Linear Fractional Programming of SIPW Point Estimates}
\begin{itemize}
\itemsep12pt
\item Bootstrap Intervals of the range of SIPW point estimates:
\[
[L_B, U_B] = \left[\left(Q_{\frac{\alpha}{2}}(\inf_{h \in
\mathcal{H}(\lambda)} \hat{\hat{\mu}}^{(h)})_{b \in [B]}\right),
\left(Q_{1 - \frac{\alpha}{2}}(\sup_{h \in \mathcal{H}(\lambda)}
\hat{\hat{\mu}}^{(h)})_{b \in [B]}\right)\right]
\]
\item In the linear programming problem, the optimization variable is merely
$z_i = e^{h(\bm{X}_i, Y_i)}$, as all other relevant quantities may be
readily estimated from the observed data.
\item Computation is extremely efficient, with complexity $O(nB + n
\text{log}n)$.
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{CI for the ATE in the Sensitivity Model}
\begin{itemize}
\itemsep12pt
\item Using the approach we've been discussing, we readily obtain an
asymptotically valid CI for the ATE via the percentile bootstrap:
\[
\left[Q_{\frac{\alpha}{2}}(\hat{\hat{\Delta}}^{(h_0, h_1)}),
Q_{1 - \frac{\alpha}{2}}(\hat{\hat{\Delta}}^{(h_0, h_1)})\right],
\]
where $Q_{\frac{\alpha}{2}}(\hat{\hat{\Delta}}^{(h_0, h_1)})$ is the
$\alpha$-th bootstrap quantile of the SIPW estimates.
\item From this, we obtain CIs
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Extensions: Augmented IPW Estimators}
\begin{itemize}
\itemsep12pt
\item AIPW estimators are \textit{``double robust''}, incorporating an extra
nuisance parameter: $f_0(\bm{x}) = \E_0[Y \mid A = 1, \bm{X} = \bm{x}]$.
\item The AIPW estimator $\hat{\Delta}_{\text{AIPW}}$ is consistent for
$\Delta$ as long as one of $\hat{e}(\bm{x})$ and $\hat{f}(\bm{x})$ is
consistent.
\item \textbf{Limitation:} In order to compute asymptotically valid confidence
intervals, the outcome regression model $\hat{f}(\bm{X}_i)$ must be
parametric.\\ \pause
\small{\textit{Perhaps this could be loosened.}}
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Discussion}
\subsection{Comparisons}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Comparison with Rosenbaum Sensitivity}
\begin{itemize}
\itemsep12pt
\item Rosenbaum's method obtains point estimates and CIs under the collection
of models $\mathcal{R}(\Gamma)$.
\begin{itemize}
\itemsep8pt
\item Here, classically, we assume that the causal effect is additive and
constant.
\item This means the Fisher null can be used to determine whether an
effect $\Delta$ ought to be included in the ($1-\alpha$)--CI.
\end{itemize}
\item The present approach is a hybrid of existing approaches in the sense
that it considers a range of differencs between $A\mid\bm{X}$ and
$A\mid\bm{X}, Y(a)$.
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Comparison with Rosenbaum Sensitivity}
\begin{itemize}
\itemsep8pt
\item Generally, Rosenbaum's method treats the sample as the population, the
present approach treats the observations as i.i.d.~samples from a
super-population.
\item The present method uses IPW-type estimators --- importantly, this makes
exact matching completely unnecessary.
\item The present approach is natural for IPW-type estimators while
Rosenbaum's is natural for matched designs.
\item Most methods use randomization inference based on Fisher's sharp null,
the present approach takes a point estimation perspective.
\item Heterogeneous treatment effects; applicability to missing data problems.
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection{Remarks}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Thoughts and Impressions}
\Huge What do I think?
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{I've Talked Enough$\ldots$}
\Huge What do you think?
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% don't want dimming with references
\setbeamercovered{}
\beamerdefaultoverlayspecification{}
\begin{frame}[c,allowframebreaks]{References}
\small
\bibliographystyle{plainnat}
\nocite{*}
\bibliography{refs}
\itemize
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\end{document}