-
Notifications
You must be signed in to change notification settings - Fork 2
/
5software.Rmd
416 lines (335 loc) · 13.5 KB
/
5software.Rmd
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
```{r setup01ss, include=FALSE}
require(knitr)
require(glmnet)
require(kableExtra)
require(dplyr)
require(xgboost)
require(SuperLearner)
require(sl3)
require(Rsolnp)
require(ltmle)
require(AIPW)
require(tmle3)
require(sl3)
options(knitr.kable.NA = '')
cachex=TRUE
cachexy=FALSE
```
```{r dataload_01, cache=cachex, echo = TRUE}
# Read the data saved at the last chapter
ObsData <- readRDS(file = "data/rhcAnalytic.RDS")
dim(ObsData)
```
# Pre-packaged software
## tmle
- The _tmle_ package can handle
- both binary and
- continuous outcomes, and
- uses the _SuperLearner_ package to construct both models just like we did in the steps above.
- The default SuperLearner library for estimating the outcome includes [@tmlePkgDocs]
- `SL.glm`: generalized linear models (GLMs)
- `SL.glmnet`: LASSO
- `tmle.SL.dbarts2`: modeling and prediction using BART
- The default library for estimating the propensity scores includes
- `SL.glm`: generalized linear models (GLMs)
- `tmle.SL.dbarts.k.5`: SL wrappers for modeling and prediction using BART
- `SL.gam`: generalized additive models: (GAMs)
- It is certainly possible to use different set of learners
- More methods can be added by
- specifying lists of models in the _Q.SL.library_ (for the outcome model) and
- _g.SL.library_ (for the propensity score model) arguments.
- Note also that the outcome $Y$ is required to be within the range of $[0,1]$ for this method as well,
- so we need to pass in the transformed data, then transform back the estimate.
```{r tmlepkg, cache=cachex, message=FALSE, warning=FALSE}
set.seed(1444)
# transform the outcome to fall within the range [0,1]
min.Y <- min(ObsData$Y)
max.Y <- max(ObsData$Y)
ObsData$Y_transf <- (ObsData$Y-min.Y)/(max.Y-min.Y)
# run tmle from the tmle package
ObsData.noYA <- dplyr::select(ObsData,
!c(Y_transf, Y, A))
SL.library = c("SL.glm",
"SL.glmnet",
"SL.xgboost")
```
```{r tmlepkg33, cache=cachex, message=FALSE, warning=FALSE}
tmle.fit <- tmle::tmle(Y = ObsData$Y_transf,
A = ObsData$A,
W = ObsData.noYA,
family = "gaussian",
V = 3,
Q.SL.library = SL.library,
g.SL.library = SL.library)
tmle.fit
```
```{r tmlepkgtr2, cache=cachex, message=FALSE, warning=FALSE}
summary(tmle.fit)
```
```{r tmlepkgtr, cache=cachex, message=FALSE, warning=FALSE}
tmle_est_tr <- tmle.fit$estimates$ATE$psi
tmle_est_tr
# transform back the ATE estimate
tmle_est <- (max.Y-min.Y)*tmle_est_tr
tmle_est
```
```{r, cache=TRUE, echo = TRUE}
saveRDS(tmle_est, file = "data/tmle.RDS")
```
```{r tmlepkg2, cache=cachex, results='hide', message=FALSE, warning=FALSE}
tmle_ci <- paste("(",
round((max.Y-min.Y)*tmle.fit$estimates$ATE$CI[1], 3), ", ",
round((max.Y-min.Y)*tmle.fit$estimates$ATE$CI[2], 3), ")", sep = "")
```
```{r, cache=TRUE, echo = TRUE}
tmle.ci <- (max.Y-min.Y)*tmle.fit$estimates$ATE$CI
saveRDS(tmle.ci, file = "data/tmleci.RDS")
```
```{r, cache=cachex, echo=FALSE}
cat("ATE from tmle package: ", tmle_est, tmle_ci, sep = "")
```
Notes about the _tmle_ package:
* does not scale the outcome for you
* can give some error messages when dealing with variable types it is not expecting
* practically all steps are nicely packed up in one function, very easy to use but need to dig a little to truly understand what it does
Most helpful resources:
* [CRAN docs](https://cran.r-project.org/web/packages/tmle/tmle.pdf)
* [tmle package paper](https://www.jstatsoft.org/article/view/v051i13)
## tmle (reduced computation)
We can use the previously calculated propensity score predictions from SL (calculated using `WeightIt` package) in the `tmle` to reduce some computing time.
```{r tmlepkg33b, cache=cachex, message=FALSE, warning=FALSE}
ps.obj <- readRDS(file = "data/ipwslps.RDS")
ps.SL <- ps.obj$weights
tmle.fit2 <- tmle::tmle(Y = ObsData$Y_transf,
A = ObsData$A,
W = ObsData.noYA,
family = "gaussian",
V = 3,
Q.SL.library = SL.library,
g1W = ps.SL)
tmle.fit2
```
```{r tmlepkgtrb, cache=cachex, message=FALSE, warning=FALSE}
# transform back ATE estimate
(max.Y-min.Y)*tmle.fit2$estimates$ATE$psi
```
## sl3 (optional)
```{r}
# install sl3 if not done so
# remotes::install_github("tlverse/sl3")
```
The _sl3_ package is a newer package, that implements two types of Super Learning:
- **discrete Super Learning**,
- in which the best prediction algorithm (based on cross-validation) from a specified library is returned, and
- **ensemble Super Learning**,
- in which the best linear combination of the specified algorithms is returned (@coyle2021sl3).
The first step is to create a sl3 task which keeps track of the roles of the variables in our problem (@coyle2021tlverse).
```{r sl301, cache=cachexy}
require(sl3)
# create sl3 task, specifying outcome and covariates
rhc_task <- make_sl3_Task(
data = ObsData,
covariates = colnames(ObsData)[-which(names(ObsData) == "Y")],
outcome = "Y"
)
```
```{r sl30156, cache=cachexy}
rhc_task
```
Next, we create our SuperLearner. To do this,
- we need to specify a **selection of machine learning algorithms** we want to include as candidates, as well as
- a **metalearner** that the SuperLearner will use to combine or choose from the machine learning algorithms provided (@coyle2021tlverse).
```{r sl302, cache=cachexy}
# see what algorithms are available for a continuous outcome
# (similar can be done for a binary outcome)
sl3_list_learners("continuous")
```
The chosen candidate algorithms can be created and collected in a Stack.
```{r sl303, cache=cachexy, results='hide', message=FALSE, warning=FALSE}
# initialize candidate learners
lrn_glm <- make_learner(Lrnr_glm)
lrn_lasso <- make_learner(Lrnr_glmnet) # alpha default is 1
xgb_5 <- Lrnr_xgboost$new(nrounds = 5)
# collect learners in stack
stack <- make_learner(
Stack, lrn_glm, lrn_lasso, xgb_5
)
```
The stack is then given to the SuperLearner.
```{r sl304, cache=cachexy, results='hide', message=FALSE, warning=FALSE}
# to make an ensemble SuperLearner
sl_meta <- Lrnr_nnls$new()
sl <- Lrnr_sl$new(
learners = stack,
metalearner = sl_meta)
# or a discrete SuperLearner
sl_disc_meta <- Lrnr_cv_selector$new()
sl_disc <- Lrnr_sl$new(
learners = stack,
metalearner = sl_disc_meta
)
```
The SuperLearner is then trained on the sl3 task we created at the start and then it can be used to make predictions.
```{r sl305, cache=cachexy, message=FALSE, warning=FALSE}
set.seed(1444)
# train SL
sl_fit <- sl$train(rhc_task)
# or for discrete SL
# sl_fit <- sl_disc$train(rhc_task)
# make predictions
sl3_data <- ObsData
sl3_data$sl_preds <- sl_fit$predict()
sl3_est <- mean(sl3_data$sl_preds[sl3_data$A == 1]) -
mean(sl3_data$sl_preds[sl3_data$A == 0])
sl3_est
```
```{r, cache=TRUE, echo = TRUE}
saveRDS(sl3_est, file = "data/sl3.RDS")
```
Notes about the _sl3_ package:
* fairly easy to implement & understand structure
* large selection of candidate algorithms provided
* unsure why result is so different
* very different structure from _SuperLearner_ library, but very customizable
* could use more explanations of when to use what metalearner and what exactly the structure of the metalearner construction means
Most helpful resources:
* [tlverse sl3 page](https://tlverse.org/sl3/)
* [sl3 GitHub repository](https://github.com/tlverse/sl3/)
* [tlverse handbook chapter 6](https://tlverse.org/tlverse-handbook/tmle3.html)
* Vignettes in R
<!---
## ltmle
Similarly to the _tmle_ package, the _ltmle_ package gives the direct TMLE result with the call of one function.
```{r ltmlepkg, cache=cachex, message=FALSE, warning=FALSE}
# exclude Y_transf since ltmle scales automatically
ltmle_data <- dplyr::select(ObsData, !Y_transf)
```
```{r ltmlepkgxd, cache=cachex, message=FALSE, warning=FALSE}
# run ltmle
ltmle_est <- ltmle(ltmle_data,
Anodes = "A",
Ynodes = "Y",
abar = list(1,0),
SL.cvControl=list(V=3),
SL.library = SL.library,
estimate.time = FALSE)
```
```{r ltmlepkgpr, cache=cachex}
summary(ltmle_est)
```
```{r, cache=TRUE, echo = TRUE}
saveRDS(ltmle_est, file = "data/ltmle.RDS")
```
```{r ltmlepkg2, cache=cachex, include = FALSE}
# # print result & confidence intervals
# ltmle_ci <- paste("(",
# round(summary(ltmle_est)[["treatment"]][["CI"]][,"2.5%"], 3),
# ", ", round(summary(ltmle_est)[["treatment"]][["CI"]][,"97.5%"], 3),
# ")", sep = "")
# cat("ATE from ltmle package: ",
# ltmle_est$estimates[["tmle"]], ltmle_ci, sep = "")
```
- The main difference between the _tmle_ and the _ltmle_ package is that the _ltmle_ package is designed to handle longitudinal data, with measurements data recorded for each subject at multiple timepoints.
- More information on the use of the _ltmle_ package in these settings can be found [here](https://cran.r-project.org/web/packages/ltmle/vignettes/ltmle-intro.html).
## AIPW
- The _aipw_ package implements augmented inverse probability weighting (another type of DR method).
- It has similar parameters as the _tmle_ package.
- It is typically used with the _SuperLearner_ library.
```{r aipwpkg, cache=cachex, results='hide', message=FALSE, warning=FALSE, include = FALSE}
set.seed(1444)
# construct AIPW estimator
aipw <- AIPW$new(Y=ObsData$Y,
A=ObsData$A,
W=ObsData[colnames(ObsData)[-which(names(ObsData) == "Y")]],
Q.SL.library = c("SL.glm",
"SL.glmnet",
"SL.xgboost"),
g.SL.library = c(c("SL.glm",
"SL.glmnet",
"SL.xgboost")),
k_split=3,
verbose=FALSE)
```
```{r aipwpkg2, cache=cachex, results='hide', message=FALSE, warning=FALSE, include = FALSE}
# fit AIPW object
aipw$fit()
```
```{r aipwpkg3, cache=cachex, results='hide', message=FALSE, warning=FALSE, include = FALSE}
# calculate ATE
aipw$summary()
print(aipw$estimates)
aipw_est <- aipw$estimates$RD[["Estimate"]]
# 95% CI
aipw_ci <- paste(" (",
round(aipw$estimates$RD["95% LCL"], 3), ", ",
round(aipw$estimates$RD["95% UCL"], 3), ")", sep = "")
```
```{r, cache=TRUE, echo = TRUE, include = FALSE}
saveRDS(aipw, file = "data/aipw.RDS")
```
```{r asg, echo=FALSE, include = FALSE}
cat("ATE from aipw package: ", aipw_est, aipw_ci, sep = "")
```
--->
## RHC results
Gathering previously saved results:
```{r summarytable0, cache=cachex, echo=TRUE, results='hold', warning=FALSE, message=FALSE}
fit.reg <- readRDS(file = "data/adjreg.RDS")
TEr <- fit.reg$coefficients[2]
CIr <- as.numeric(confint(fit.reg, 'A'))
fit.matched <- readRDS(file = "data/match.RDS")
TEm <- fit.matched$coefficients[2]
CIm <- as.numeric(confint(fit.matched, 'A'))
TEg <- readRDS(file = "data/gcomp.RDS")
CIg <- readRDS(file = "data/gcompci.RDS")
CIgc <- CIg$percent[4:5]
TE1g <- readRDS(file = "data/gcompxg.RDS")
TE2g <- readRDS(file = "data/gcompls.RDS")
TE3g <- readRDS(file = "data/gcompsl.RDS")
ipw <- readRDS(file = "data/ipw.RDS")
TEi <- ipw$coefficients[2]
CIi <- as.numeric(confint(ipw, 'A'))
ipwsl <- readRDS(file = "data/ipwsl.RDS")
TEsli <- ipwsl$coefficients[2]
CIsli <- as.numeric(confint(ipwsl, 'A'))
tmleh <- readRDS(file = "data/tmlepointh.RDS")
tmlecih <- readRDS(file = "data/tmlecih.RDS")
tmlesl <- readRDS(file = "data/tmle.RDS")
tmlecisl <- readRDS(file = "data/tmleci.RDS")
slp <- readRDS(file = "data/sl3.RDS")
ci.b <- rep(NA,2)
ks <- 2.01
ci.ks <- c(0.6,3.41)
point <- as.numeric(c(TEr, TEm, TEg, TE1g, TE2g,
TE3g, TEi, TEsli, tmleh,
tmlesl, slp, ks))
CIs <- cbind(CIr, CIm, CIgc, ci.b, ci.b, ci.b,
CIi, CIsli, tmlecih, tmlecisl,
ci.b, ci.ks)
```
```{r summarytable, cache=cachex, echo=TRUE}
method.list <- c("Adj. Reg","PS match",
"G-comp (linear reg.)","G-comp (xgboost)",
"G-comp (lasso)", "G-comp (SL)",
"IPW (logistic)", "IPW (SL)",
"TMLE (9 steps)", "TMLE (package)",
"sl3 (package)", "Keele and Small (2021) paper")
results <- data.frame(method.list)
results$Estimate <- round(point,2)
results$`2.5 %` <- CIs[1,]
results$`97.5 %` <- CIs[2,]
kable(results,digits = 2)%>%
row_spec(10, bold = TRUE, color = "white", background = "#D7261E")
```
```{block, type='rmdcomment'}
@keele2021comparing used TMLE-SL based on an ensemble of 3 different learners: (1) GLM, (2) random forests, and (3) LASSO.
```
## Other packages
Other packages that may be useful:
| Package | Resources | Notes |
|---|---|---|
| ltmle | [CRAN vignette](https://cran.r-project.org/web/packages/ltmle/vignettes/ltmle-intro.html) | Longitudinal |
| tmle3 | [GitHub](https://github.com/tlverse/tmle3), [framework overview](https://tlverse.org/tmle3/articles/framework.html), [tlverse handbook](https://tlverse.org/tlverse-handbook/tmle3.html) | tmle3 is still under development |
| aipw | [GitHub](https://github.com/yqzhong7/AIPW), [CRAN vignette](https://cran.r-project.org/web/packages/AIPW/vignettes/AIPW.html) | Newer package for AIPW (another DR method) |
| Others | [van der Laan research group](https://www.stat.berkeley.edu/users/laan/Software/) | |
You can find many other related packages on [CRAN](https://cran.r-project.org/search.html) or GitHub.