-
Notifications
You must be signed in to change notification settings - Fork 3
/
SWSamp.R
1329 lines (1171 loc) · 61.4 KB
/
SWSamp.R
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
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
make.swt <- function(I=NULL,J=NULL,H=NULL,K,design="cross-sec",mu=NULL,b.trt,
b.time=NULL,sigma.y=NULL,sigma.e=NULL,rho,sigma.a=NULL,rho.ind=NULL,
sigma.v=NULL,X=NULL,family="gaussian",natural.scale=TRUE) {
## Normal outcome
## inputs:
# I = number of clusters
# J = number of randomisation time points (excluding baseline)
# K = average sample size in each cluster
# H = number of units randomised at each time point
# design = type of design. Can be "cross-sec" (default) or "cohort" (repeated measurements)
# mu = baseline mean
# b.trt = intervention effect
# b.time = (optional) time effect (on linear time trend scale!)
# sigma.y = *total* individual variability
# sigma.e = *residual* individual variability
# rho = ICC at the cluster level
# sigma.a = the sd of the cluster random effect (default at NULL)
# rho.ind = ICC at the individual level (for cohort design, default at NULL)
# sigma.v = the sd of the treatment random effect (default at NULL)
# X = SW design matrix (default at NULL and will be computed automatically)
# family = type of outcome to be simulated (options are "gaussian", "binomial" and "poisson")
# natural.scale = whether the input values for the trt effect etc are passed on the natural scale or not
#
# GB + Rosalind Leach (November 2015)
# CHECK ON VALID FAMILY TYPE
flag=1
if(family=="gaussian" | family=="binomial" | family=="poisson"){flag=0}
if(flag==1){ stop("Available options for the argument 'family' are: gaussian, binomial or poisson")}
# CREATE DESIGN MATRIX FROM sw.design.mat IF X HAS NOT BEEN PROVIDED
if(is.null(X)) {X <- sw.design.mat(I=I,J=J,H=NULL)} else {row.names(X) <- sample(1:I,I)}
if(family == "gaussian"){
# SETS DEFAULT FOR mu AND b.time IF THEY HAVE NOT BEEN PORVIDED
if(is.null(mu)) {mu=0} # Baseline (set to 0 if not specified)
if(is.null(b.time)) { # Time effect (if not specified, set to)
b.time <- .5*b.trt } # some pre-defined proportion of the
# SIMULATES VARIABLES FROM THE INPUTS GIVEN
treatment <- rep(t(X),each=K,1) # Treatment allocation at each time point
time <- rep(seq(0,J),each=K,I) # K measurements during the trial
person <- seq(1,K) # individuals IDs
cluster.order <- rep(sample(1:I)) # order at which clusters switch
cluster <- rep(cluster.order, each=(J+1)*K) # cluster IDs (randomised)
id <- NULL
# Calculate the hyperparameters
mu.a <- 0 # Within cluster (random intercept) mean
# CROSS-SECTIONAL DESIGN -------------------------------------------------
if(design=="cross-sec") {
# CHECK THE RIGHT COMBINATION OF sigma.e, sigma.y, sigma.a, rho HAS BEEN GIVEN
if((sum(sapply(list(sigma.y,sigma.e),is.null)) !=1) & (sum(sapply(list(sigma.a,rho),is.null)) !=1)) { # Need to pass 1 of 'sigma.y' or 'sigma.e' and 1 of 'sigma.a' or 'rho
stop("Please provide either 'sigma.y' (*total* individual variability) or 'sigma.e' (*residual* variability)
and either 'sigma.a' (*within cluster* variability) or 'rho' (ICC)")}
# IF STATEMENTS TO CALCULATE ALL RELEVANT PARAMETERS FROM THE INPUTS GIVEN
if(is.null(sigma.e) & is.null(sigma.a)) {
sigma.a <- rho*sigma.y
sigma.e <- sqrt(sigma.y^2-sigma.a^2)
}
if(is.null(sigma.e) & is.null(rho)) {
sigma.e <- sqrt(sigma.y^2-sigma.a^2)
rho <- sigma.a^2/sigma.y^2
}
if(is.null(sigma.y) & is.null(sigma.a)) {
sigma.a <- sqrt(sigma.e^2*rho/(1-rho))
sigma.y <- sqrt(sigma.a^2+sigma.e^2)
}
if(is.null(sigma.y) & is.null(rho)) {
sigma.y <- sqrt(sigma.e^2+sigma.a^2)
rho <- sigma.a^2/sigma.y^2
}
# SIMULATE REMAINING VARIABLES
a <- rnorm(I, mu.a, sigma.a) # cluster-level intercept
# CALCULATE LINEAR PREDICTOR
linpred <- mu + a[cluster] + b.trt*treatment + b.time*time # Calculates the linear predictor
}
# COHORT DESIGN -------------------------------------------------
if (design=="cohort") {
# CHECK THE RIGHT COMBINATION OF sigma.e, sigma.y, sigma.a, rho HAS BEEN GIVEN
if((sum(sapply(list(sigma.y,sigma.e),is.null)) !=1) || (sum(sapply(list(sigma.a,rho),is.null)) !=1) ||
(sum(sapply(list(sigma.v,rho.ind),is.null)) !=1)) {
stop("Please provide either 'sigma.y' (*total* individual variability) or 'sigma.e' (*residual* variability)
AND either 'sigma.a' (*within cluster* variability) or 'rho' (cluster-level ICC)
AND either 'sigma.v' (*within cluster-individual* variability) or 'rho.ind (individual-level ICC)") }
# IF STATEMENTS TO CALCULATE ALL RELEVANT PARAMETERS FROM THE INPUTS GIVEN
if(is.null(sigma.y) & is.null(sigma.a) & is.null(sigma.v)) {
r1 <- rho.ind/(1-rho.ind)
r2 <- rho/(1-rho)
sigma.v <- sqrt((r1*(sigma.e^2/(1-rho)))/(1-r1*r2))
sigma.a <- sqrt(r2*(sigma.e^2+sigma.v^2))
sigma.y <- sqrt(sigma.e^2+sigma.a^2+sigma.v^2)
}
if(is.null(sigma.y) & is.null(rho) & is.null(sigma.v)) {
r <- rho.ind/(1-rho.ind)
sigma.v <- sqrt(r*(sigma.a^2+sigma.e^2))
rho <- sigma.a^2/(sigma.a^2+sigma.e^2+sigma.v^2)
sigma.y <- sqrt(sigma.a^2+sigma.e^2+sigma.v^2)
}
if(is.null(sigma.y) & is.null(sigma.a) & is.null(rho.ind)) {
r <- rho/(1-rho)
sigma.a <- sqrt(r*(sigma.v^2+sigma.e^2))
sigma.y <- sqrt(sigma.a^2+sigma.e^2+sigma.v^2)
rho.ind <- sigma.v^2/(sigma.a^2+sigma.e^2+sigma.v^2)
}
if(is.null(sigma.y) & is.null(rho) & is.null(rho.ind)) {
rho <- sigma.a^2/(sigma.a^2+sigma.e^2+sigma.v^2)
rho.ind <- sigma.v^2/(sigma.a^2+sigma.e^2+sigma.v^2)
sigma.y <- sqrt(sigma.a^2+sigma.e^2+sigma.v^2)
}
if(is.null(sigma.e) & is.null(sigma.a) & is.null(sigma.v)) {
sigma.a <- sqrt(sigma.y^2*rho)
sigma.v <- sqrt(sigma.y^2*rho.ind)
sigma.e <- sqrt(sigma.y^2-sigma.a^2-sigma.v^2)
}
if(is.null(sigma.e) & is.null(rho) & is.null(sigma.v)) {
rho <- sigma.a^2/sigma.y^2
sigma.v <- rho.ind/sigma.y^2
sigma.e <- sqrt(sigma.y^2-sigma.a^2-sigma.v^2)
}
if(is.null(sigma.e) & is.null(sigma.a) & is.null(rho.ind)) {
sigma.a <- sqrt(rho*sigma.y^2)
rho.ind <- sigma.v^2/sigma.y^2
sigma.e <- sqrt(sigma.y^2-sigma.a^2-sigma.v^2)
}
if(is.null(sigma.e) & is.null(rho) & is.null(rho.ind)) {
rho <- sigma.a^2/sigma.y^2
rho.ind <- sigma.v^2/sigma.y^2
sigma.e <- sqrt(sigma.y^2-sigma.a^2-sigma.v^2)
}
# SIMULATE REMAINING VARIABLES
start <- seq(1,I*K,by=K)
id <- numeric # Individual ID (for repeated measurements)
id <- rep(seq(start[cluster.order[1]],start[cluster.order[1]]+K-1),(J+1))
for (i in 2:I) {
id <- c(id,rep(seq(start[cluster.order[i]],start[cluster.order[i]]+K-1),(J+1)))
}
mu.v <- 0 # within individuals (random intercept) mean
v <- rnorm(K*I,mu.v,sigma.v) # individual-level intercept
sigma.a <- sqrt(sigma.y^2*rho) # Within cluster (random intercept) sd
a <- rnorm (I, mu.a, sigma.a) # cluster-level intercept
# CALCULATE LINEAR PREDICTOR
linpred <- mu + a[cluster] + v[id] + b.trt*treatment + b.time*time
}
# SIMULATES DATA
n <- I*(J+1)*K
y <- rnorm(I*(J+1)*K, linpred, sigma.e)
}
if(family=="binomial"){
# SETS DEFAULT FOR mu AND b.time IF THEY HAVE NOT BEEN PORVIDED
if(is.null(mu)) {
if (natural.scale==TRUE) {
mu=.5 } # Default baseline probability on [0-1]
else if (natural.scale==FALSE) {
mu=log(.5/.5) } # Default baseline probability on logit scale
}
# SIMULATES VARIABLES FROM THE INPUTS GIVEN
treatment <- rep(t(X),each=K,1) # Treatment allocation at each time point
time <- rep(seq(0,J),each=K,I) # K measurements during the trial
person <- seq(1,K) # individuals IDs
cluster.order <- rep(sample(1:I)) # order at which clusters switch
cluster <- rep(cluster.order, each=(J+1)*K) # cluster IDs (randomised)
id <- NULL
# If natural.scale = T, then assume that the user is giving values for p0 (mu) and OR directly (b.trt)
if (natural.scale==TRUE) {
OR <- b.trt # for simplicity defines the OR
b.trt <- log(b.trt) # logOR to be used in the linear predictor
p0 <- mu # for simplicity defines the baseline probability
mu <- log(mu/(1-mu)) # Converts baseline probability to log odds to be used in the linear predictor
}
# But if natural.scale = F, then the user has passed data on the logit scale for p0 (mu) and on the logOR (b.trt)
if (natural.scale==FALSE) {
OR <- exp(b.trt) # Defines the OR on the natural scale
p0 <- exp(mu)/(1+exp(mu)) # Rescales the baseline to the [0-1] range
}
# Now defines p1 and sigma.e
p1 <- OR*(p0/(1-p0))/(1+OR*(p0/(1-p0))) # Estimates the outcome probability for intervention
sigma.e <- sqrt(((p0*(1-p0))+(p1*(1-p1)))/2) # Pooled estimate of the within cluster sd
if(is.null(b.time)) { # Time effect (if not specified, set to)
b.time <- .5*b.trt } # some pre-defined proportion of the treatment effect
# CALCULATE THE HYPERPARAMETERS
mu.a <- 0 # Within cluster (random intercept) mean
# CROSS-SECTIONAL DESIGN -------------------------------------------------
if(design=="cross-sec") {
# CHECK THE RIGHT COMBINATION OF sigma.a, rho HAS BEEN GIVEN
if (sum(sapply(list(rho,sigma.a),is.null)) !=1) { #cannot provide both sigma.a and rho
stop("exactly one of 'rho' and 'sigma.a' must be null") }
# IF STATEMENTS TO CALCULATE ALL RELEVANT PARAMETERS FROM THE INPUTS GIVEN
if(is.null(sigma.a)){ # if sigma.a is not given, calculate it as a function of rho
sigma.a <- sqrt(sigma.e^2*rho/(1-rho)) } # Within cluster (random intercept) sd
if(is.null(rho)){ # if sigma.a is not given, calculate it as a function of rho
rho <- (sigma.a^2/(sigma.e^2 + sigma.a^2)) } # if rho is not given, calculate it as a function of sigma.a
# SIMULATE REMAINING VARIABLES
a <- rnorm(I, mu.a, sigma.a) # cluster-level intercept
# CALCULATE LINEAR PREDICTOR ON THE LOGIT SCALE!
linpred <- mu + a[cluster] + b.trt*treatment + b.time*time
}
# COHORT DESIGN -------------------------------------------------
if (design=="cohort") {
# CHECK THE RIGHT COMBINATION OF sigma.a, sigma.v, rho, rho.ind HAS BEEN GIVEN
if((sum(sapply(list(rho, sigma.a),is.null)) !=1) || (sum(sapply(list(sigma.a,rho),is.null)) !=1) ||
(sum(sapply(list(sigma.v,rho.ind),is.null)) !=1)) {
stop("Please provide either 'sigma.a' (*within cluster* variability) or 'rho' (cluster-level ICC)
AND either 'sigma.v' (*within cluster-individual* variability) or 'rho.ind (individual-level ICC)") }
# IF STATEMENTS TO CALCULATE ALL RELEVANT PARAMETERS FROM THE INPUTS GIVEN
if(is.null(sigma.v) & is.null(sigma.a)) {
r1 <- rho.ind/(1-rho.ind)
r2 <- rho/(1-rho)
sigma.v <- sqrt((r1*(sigma.e^2/(1-rho)))/(1-r1*r2))
sigma.a <- sqrt(r2*(sigma.e^2+sigma.v^2))
}
if(is.null(rho.ind) & is.null(sigma.a)) {
r <- rho/(1-rho)
sigma.a <- sqrt(r*(sigma.v^2+sigma.e^2))
#rho.ind <-
}
if(is.null(rho.ind) & is.null(rho)) {
rho <- sigma.a^2/(sigma.a^2+sigma.e^2+sigma.v^2)
rho.ind <- sigma.v^2/(sigma.a^2+sigma.e^2+sigma.v^2)
}
if(is.null(rho) & is.null(sigma.v)) {
r <- rho.ind/(1-rho.ind)
sigma.v <- sqrt(r*(sigma.a^2+sigma.e^2))
rho <- sigma.a^2/(sigma.a^2+sigma.e^2+sigma.v^2)
}
# SIMULATE REMAINING VARIABLES
start <- seq(1,I*K,by=K)
id <- numeric # Individual ID (for repeated measurements)
id <- rep(seq(start[cluster.order[1]],start[cluster.order[1]]+K-1),(J+1))
for (i in 2:I) {
id <- c(id,rep(seq(start[cluster.order[i]],start[cluster.order[i]]+K-1),(J+1)))
}
mu.v <- 0 # within individuals (random intercept) mean
v <- rnorm(K*I,mu.v,sigma.v) # individual-level intercept
### sigma.a <- sqrt(sigma.y^2*rho) # Within cluster (random intercept) sd
a <- rnorm (I, mu.a, sigma.a) # cluster-level intercept
# CALCULATE LINEAR PREDICTOR ON THE LOGIT SCALE!
linpred <- mu + a[cluster] + v[id] + b.trt*treatment + b.time*time
}
# SIMULATES DATA
n <- I*(J+1)*K
# Converts log odds back to probability
p = (exp(linpred))/(1+exp(linpred))
y <- rbinom(n, 1, p)
}
if(family == "poisson"){
# SETS DEFAULT FOR mu AND b.time IF THEY HAVE NOT BEEN PROVIDED
if(is.null(mu)) {
if (natural.scale==TRUE) {
mu=1} # Default baseline rate set to 1
else if (natural.scale==FALSE) {
mu=log(1)} # Default baseline rate on log scale
}
# SIMULATES VARIABLES FROM THE INPUTS GIVEN
treatment <- rep(t(X),each=K,1) # Treatment allocation at each time point
time <- rep(seq(0,J),each=K,I) # K measurements during the trial
person <- seq(1,K) # individuals IDs
cluster.order <- rep(sample(1:I)) # order at which clusters switch
cluster <- rep(cluster.order, each=(J+1)*K) # cluster IDs (randomised)
id <- NULL # initialise ID vector
# If natural.scale = T, then assume that the user is giving values for p0 and OR directly
if (natural.scale==TRUE) {
RR <- b.trt # For convenience defines the RR
b.trt <- log(b.trt) # logRR to be used in the linear predictor
lambda0 <- mu # For convenience defines the baseline rate
mu <- log(mu) # Converts baseline probability to log rate to be used in the linear predictor
}
# But if natural.scale = F, then the user has passed data on the logit scale for p0 and on the logOR
if (natural.scale==FALSE) {
RR <- exp(b.trt) # Defines the RR on the natural scale
lambda0 <- exp(mu) # Rescales the baseline rate to the [0-\infty] range
}
lambda1 <- lambda0*RR # rate for intervention
sigma.e <- sqrt((lambda0+lambda1)/2) # Within cluster sd (estimated using pooled value)
if(is.null(b.time)) { # Time effect (if not specified, set to)
b.time <- .5*b.trt } # some pre-defined proportion of the
# CALCULATE THE HYPERPARAMETERS
mu.a <- 0 # Within cluster (random intercept) mean
# CROSS-SECTIONAL DESIGN -------------------------------------------------
if(design=="cross-sec") {
# CHECK THE RIGHT COMBINATION OF sigma.a, rho HAS BEEN GIVEN
if (sum(sapply(list(rho,sigma.a),is.null)) !=1) { #cannot provide both sigma.a and rho
stop("exactly one of 'rho' and 'sigma.a' must be null") }
# IF STATEMENTS TO CALCULATE ALL RELEVANT PARAMETERS FROM THE INPUTS GIVEN
if(is.null(sigma.a)){ # if sigma.a is not given, calculate it as a function of rho
sigma.a <- sqrt(sigma.e^2*rho/(1-rho)) } # Within cluster (random intercept) sd
if(is.null(rho)){ # if sigma.a is not given, calculate it as a function of rho
rho <- (sigma.a^2/(sigma.e^2 + sigma.a^2)) } # if rho is not given, calculate it as a function of sigma.a
# SIMULATE REMAINING VARIABLES
a <- rnorm(I, mu.a, sigma.a) # cluster-level intercept
# CALCULATE LINEAR PREDICTOR ON THE LOG SCALE
linpred <- mu + a[cluster] + b.trt*treatment + b.time*time
}
# COHORT DESIGN -------------------------------------------------
if (design=="cohort") {
# CHECK THE RIGHT COMBINATION OF sigma.a, rho, rho.ind HAS BEEN GIVEN
if((sum(sapply(list(rho, sigma.a),is.null)) !=1) || (sum(sapply(list(sigma.a,rho),is.null)) !=1) ||
(sum(sapply(list(sigma.v,rho.ind),is.null)) !=1)) {
stop("Please provide either 'sigma.a' (*within cluster* variability) or 'rho' (cluster-level ICC)
AND either 'sigma.v' (*within cluster-individual* variability) or 'rho.ind (individual-level ICC)") }
# IF STATEMENTS TO CALCULATE ALL RELEVANT PARAMETERS FROM THE INPUTS GIVEN
if(is.null(sigma.v) & is.null(sigma.a)) {
r1 <- rho.ind/(1-rho.ind)
r2 <- rho/(1-rho)
sigma.v <- sqrt((r1*(sigma.e^2/(1-rho)))/(1-r1*r2))
sigma.a <- sqrt(r2*(sigma.e^2+sigma.v^2))
}
if(is.null(rho.ind) & is.null(sigma.a)) {
r <- rho/(1-rho)
sigma.a <- sqrt(r*(sigma.v^2+sigma.e^2))
#rho.ind <-
}
if(is.null(rho.ind) & is.null(rho)) {
rho <- sigma.a^2/(sigma.a^2+sigma.e^2+sigma.v^2)
rho.ind <- sigma.v^2/(sigma.a^2+sigma.e^2+sigma.v^2)
}
if(is.null(rho) & is.null(sigma.v)) {
r <- rho.ind/(1-rho.ind)
sigma.v <- sqrt(r*(sigma.a^2+sigma.e^2))
rho <- sigma.a^2/(sigma.a^2+sigma.e^2+sigma.v^2)
}
# SIMULATE REMAINING VARIABLES
start <- seq(1,I*K,by=K)
id <- numeric # Individual ID (for repeated measurements)
id <- rep(seq(start[cluster.order[1]],start[cluster.order[1]]+K-1),(J+1))
for (i in 2:I) {
id <- c(id,rep(seq(start[cluster.order[i]],start[cluster.order[i]]+K-1),(J+1)))
}
mu.v <- 0 # within individuals (random intercept) mean
v <- rnorm(K*I,mu.v,sigma.v) # individual-level intercept
### sigma.a <- sqrt(sigma.y^2*rho) # Within cluster (random intercept) sd
a <- rnorm (I, mu.a, sigma.a) # cluster-level intercept
# CALCULATE LINEAR PREDICTOR
linpred <- mu + a[cluster] + v[id] + b.trt*treatment + b.time*time
}
# SIMULATES DATA
n <- I*(J+1)*K
y <- rpois(I*(J+1)*K, lambda=exp(linpred))
}
# RETURNS DATA FRAME
data <- data.frame(y, person, time, cluster, treatment, linpred, b.trt)
if (design=="cohort") {data <- data.frame(data,id)}
return(data)
}
sim.power <- function (I,J,H=NULL,K,design="cross-sec",mu=0,b.trt,b.time=NULL,
sigma.y=NULL,sigma.e=NULL,rho=NULL,sigma.a=NULL,
rho.ind=NULL,sigma.v=NULL,n.sims=1000,formula=NULL,
family="gaussian",natural.scale=TRUE,sig.level=0.05,n.cores=NULL,
method="lme",plot=FALSE,...) {
## Power analysis for repeated cohort design (cross-sectional) - continuous outcome
## The optional arguments include the specification of a user-defined function data
## which is used to simulate the data, according to any design the user has.
## So including data=data, where data=function(){...} specifies the simulation
## of a suitable dataset allows the user to select *any* kind of model for data
## generation. If data has inputs, then these must be specified as a list
## inpts, to be passed as an extra argument to sim.sw.cont.
## In addition to this, the user should then select a suitable formula
## which is in line with the model used for the simulation of the virtual trial.
## Also, the name of the "treatment" variable can be specified as a string in the
## extra argument treat.name. If not present then it is assumed that the name of the
## treatment variable is, unsurprisingly, "treatment"
## method is default at 'lme' indicating that a linear mixed model using lmer will be
## used for the analysis. The user can also specify 'INLA' or 'inla' which performs
## the analysis in Bayesian framework with INLA
exArgs <- list(...)
requireNamespace("lme4",quietly=TRUE)
requireNamespace("foreach",quietly=TRUE)
requireNamespace("doParallel",quietly=TRUE)
requireNamespace("parallel",quietly=TRUE)
requireNamespace("iterators",quietly=TRUE)
# If not specified, uses all cores available to the machine
if(is.null(n.cores)) {n.cores <- parallel::detectCores()}
# Usually a good idea to leave 1 core free for the OS to use
doParallel::registerDoParallel(cores=n.cores-1)
# Defines basic formulation for the model
if(is.null(formula)){
formula=y~treatment+factor(time)+(1|cluster)
if(design=="cohort") {formula <- update.formula(formula,.~.+(1|id))}
# Now updates the formula so it's given in INLA notation if method==INLA
if (method=="INLA" || method=="inla") {
# Checks the the package INLA is installed
if (!isTRUE(requireNamespace("INLA", quietly = TRUE))) {
stop("You need to install the packages 'INLA'. Please run in your R terminal:\n install.packages('INLA', repos='http://www.math.ntnu.no/inla/R/stable')")
} else {
if (!is.element("INLA", (.packages()))) {
attachNamespace("INLA")
}
}
# This substitutes the notation '(1 | var)' to the notation 'f(var)'. Notice that INLA will assume that
# this actually means 'f(var,model="iid")' --- if specific prior are required, they need to be passed
# directly within the formula (which can't be left as NULL), using INLA notation!
formula <- as.formula(gsub("\\(1 \\| ","f(",deparse(formula)))
}
}
# Defines the name of the treatment variable (for which the main analysis is performed)
if(exists("treat.name",where=exArgs)) {treatment <- exArgs$treat.name} else {treatment <- "treatment"}
res <- list()
tic <- proc.time()
# For user-specified data generating processes
if(exists("data",where=exArgs)) {
func <- exArgs$data
if(exists("inpts",where=exArgs)) {inpts=exArgs$inpts} else {inpts=list()}
if (method=="INLA" || method=="inla") {
# INLA-specific options --- can be used to set priors & stuff
# 'control.fixed' can be used to set the priors for the fixed effects in the model
if(exists("control.fixed",where=exArgs)) {
control.fixed <- exArgs$control.fixed
} else {
control.fixed <- INLA::inla.set.control.fixed.default()
# prior mean = 0 for *all* fixed effects
# prior var = 1000 for *all* fixed effects
# prior mean = 0 for the intercept
# prior prec -> 0 for the intercept
}
# 'offset' can be used with Poisson model. This needs to be a string with the name of the relevant variable
if(exists("offset",where=exArgs)) {offset <- exArgs$offset} else {offset <- NULL}
# 'Ntrials' is the Binomial denominator. This needs to be a string with the name of the relevant variable
if(exists("Ntrials",where=exArgs)) {Ntrials <- exArgs$Ntrials} else {Ntrials <- NULL}
res <- foreach::foreach(iterators::icount(n.sims), .combine=cbind, .packages="INLA",
.export=c("make.swt","sw.design.mat")) %dopar% {
fake.data <- do.call(what=func,args=inpts)
inla.args <- list(formula,data=fake.data,family=family,control.fixed=control.fixed)
if (!is.null(offset)) {inla.args$offset <- fake.data[[offset]]}
if (!is.null(Ntrials)) {inla.args$Ntrials <- fake.data[[Ntrials]]}
m <- do.call(INLA::inla,args=inla.args)
theta <- m$summary.fixed[treatment,1]
sd.theta <- m$summary.fixed[treatment,2]
## This assumes that alpha=0.05 and so can consider the 95% CrI
ci.fixed <- m$summary.fixed[treatment,c(3,5)]
lower.lim.check <- ci.fixed[,1]>0
upper.lim.check <- ci.fixed[,2]<0
## Needs to simulate values from the posterior of theta and then get the quantiles if alpha neq 0.05
signif <- as.numeric(((sign(theta)==1 & lower.lim.check==TRUE) | (sign(theta)==-1 & upper.lim.check==TRUE)))
tval <- theta / sd.theta
pvalue <- 2*pnorm(abs(tval), lower.tail = FALSE)
# Random effects sds
mat <- INLA::inla.contrib.sd(m)$hyper
rnd.eff.sd <- as.matrix(mat[,1])
names(rnd.eff.sd) <- rownames(mat)
list(power=signif,theta=theta,sd.theta=sd.theta,
rnd.eff.sd=rnd.eff.sd,method=method,pvalue=pvalue)
}
}
if (method=="lme") {
# Each 10% (default) of the iterations are stored in the results list, labelled res_1, res_2, etc.
res <- foreach::foreach(iterators::icount(n.sims), .combine=cbind, .packages="lme4",
.export=c("make.swt","sw.design.mat")) %dopar% {
fake.data <- do.call(what=func,args=inpts)
# If the formula contains random effects, run lmer
check.random.effect <- !is.null(findbars(formula))
if(check.random.effect==TRUE) {
if(family=="gaussian") {
m <- lme4::lmer(formula, data=fake.data)
} else {
m <- lme4::glmer(formula, data=fake.data,family=family)
}
which.treat <- which(names(fixef(m))==treatment)
theta <- lme4::fixef(m)[which.treat]
sd.theta <- summary(m)$coefficients[which.treat,2]
Vcov <- vcov(m, useScale = FALSE)
se <- sqrt(diag(Vcov))
betas <- lme4::fixef(m)
tval <- betas / se
### NB: issue with p-values in random effect models (cannot determine df easily)
### An alternative (as recommended by Bates et al) is to use CIs instead of p-values
ci.fixed <- confint(m,names(fixef(m)),method="Wald",level=1-sig.level)
lower.lim.check <- ci.fixed[,1]>0
upper.lim.check <- ci.fixed[,2]<0
signif <- as.numeric(((sign(betas)==1 & lower.lim.check==TRUE) | (sign(betas)==-1 & upper.lim.check==TRUE))[2])
### This estimates the p-value based on Normal approximation --- may not be too correct...
### See: http://mindingthebrain.blogspot.co.uk/2014/02/three-ways-to-get-parameter-specific-p.html
pval <- 2*pnorm(abs(tval), lower.tail = FALSE)
pvalue <- pval[which.treat]
###signif <- pvalue < alpha
VC <- as.data.frame(lme4::VarCorr(m))
rnd.eff.sd <- VC[,which(colnames(VC)=="sdcor")]
# Finds the rows that needs to be reported
to.report <- c(which(is.na(VC[,3]==T), which(VC[,1]=="Residual")))
rnd.eff.sd <- VC[to.report,which(colnames(VC)=="sdcor")]
names(rnd.eff.sd)[which(is.na(VC[to.report,3])==T)] <- paste(VC[to.report,"grp"],VC[to.report,"var1"])
names(rnd.eff.sd) <- gsub(" NA","",names(rnd.eff.sd))
if(family=="gaussian") {method <- "lmer"} else {method <- "glmer"}
}
# If it doesn't then do lm
if (!check.random.effect) {
if (family=="gaussian") {
m <- lm(formula,data=fake.data)
} else {
m <- glm(formula, data=fake.data,family=family)
}
which.treat <- which(names(m$coefficients)==treatment)
theta <- m$coefficients[which.treat]
sd.theta <- summary(m)$coefficients[which.treat,2]
se <- summary(m)$coefficients[,2]
betas <- summary(m)$coefficients[,1]
tval <- summary(m)$coefficients[,3]
df <- summary(m)$df[2]
pval <- summary(m)$coefficients[,4]
pvalue <- pval[which.treat]
signif <- pvalue < sig.level
if(family=="gaussian") {method <- "lm"} else {method <- "glm"}
rnd.eff.sd=NULL
}
list(power=signif,theta=theta,sd.theta=sd.theta,
rnd.eff.sd=rnd.eff.sd,method=method,pvalue=pvalue)
}
}
}
# For standard SWT data generating processes (uses make.swt)
if(!exists("data",where=exArgs)) {
if(exists("X",where=exArgs)) {
X=exArgs$X
row.names(X) <- sample(1:I,I)
colnames(X) <- c("Baseline",paste0("Time ",1:J))
} else {
X=NULL
}
if (method=="INLA" || method=="inla") {
# INLA-specific options --- can be used to set priors & stuff
# 'control.fixed' can be used to set the priors for the fixed effects in the model
if(exists("control.fixed",where=exArgs)) {
control.fixed <- exArgs$control.fixed
} else {
control.fixed <- INLA::inla.set.control.fixed.default()
# prior mean = 0 for *all* fixed effects
# prior var = 1000 for *all* fixed effects
# prior mean = 0 for the intercept
# prior prec -> 0 for the intercept
}
# 'offset' can be used with Poisson model. This needs to be a string with the name of the relevant variable
if(exists("offset",where=exArgs)) {offset <- exArgs$offset} else {offset <- NULL}
# 'Ntrials' is the Binomial denominator. This needs to be a string with the name of the relevant variable
if(exists("Ntrials",where=exArgs)) {Ntrials <- exArgs$Ntrials} else {Ntrials <- NULL}
res <- foreach::foreach(iterators::icount(n.sims), .combine=cbind, .packages="INLA",
.export=c("make.swt","sw.design.mat")) %dopar% {
fake.data <- fake.data <- make.swt(I=I,J=J,H=H,K=K,design=design,mu=mu,b.trt=b.trt,
b.time=b.time,sigma.y=sigma.y,sigma.e=sigma.e,
rho=rho,sigma.a=sigma.a,rho.ind=rho.ind,sigma.v=sigma.v,
X=X,family=family,natural.scale=natural.scale)
inla.args <- list(formula,data=fake.data,family=family,control.fixed=control.fixed)
if (!is.null(offset)) {inla.args$offset <- fake.data[[offset]]}
if (!is.null(Ntrials)) {inla.args$Ntrials <- fake.data[[Ntrials]]}
m <- do.call(INLA::inla,args=inla.args)
theta <- m$summary.fixed[treatment,1]
sd.theta <- m$summary.fixed[treatment,2]
## This assumes that alpha=0.05 and so can consider the 95% CrI
ci.fixed <- m$summary.fixed[treatment,c(3,5)]
lower.lim.check <- ci.fixed[,1]>0
upper.lim.check <- ci.fixed[,2]<0
## Needs to simulate values from the posterior of theta and then get the quantiles if alpha neq 0.05
signif <- as.numeric(((sign(theta)==1 & lower.lim.check==TRUE) | (sign(theta)==-1 & upper.lim.check==TRUE)))
tval <- theta / sd.theta
pvalue <- 2*pnorm(abs(tval), lower.tail = FALSE)
# Random effects sds
mat <- INLA::inla.contrib.sd(m)$hyper
rnd.eff.sd <- as.matrix(mat[,1])
names(rnd.eff.sd) <- rownames(mat)
list(power=signif,theta=theta,sd.theta=sd.theta,
rnd.eff.sd=rnd.eff.sd,method=method,pvalue=pvalue)
}
}
if (method=="lme") {
res <- foreach::foreach(iterators::icount(n.sims), .combine=cbind, .packages="lme4",
.export=c("make.swt","sw.design.mat")) %dopar% {
fake.data <- make.swt(I=I,J=J,H=H,K=K,design=design,mu=mu,b.trt=b.trt,
b.time=b.time,sigma.y=sigma.y,sigma.e=sigma.e,
rho=rho,sigma.a=sigma.a,rho.ind=rho.ind,sigma.v=sigma.v,
X=X,family=family,natural.scale=natural.scale)
# If the formula contains random effects, run lmer
check.random.effect <- !is.null(findbars(formula))
if(check.random.effect==TRUE) {
if(family=="gaussian") {
m <- lme4::lmer(formula, data=fake.data)
} else {
m <- lme4::glmer(formula, data=fake.data,family=family)
}
which.treat <- which(names(fixef(m))==treatment)
theta <- lme4::fixef(m)[which.treat]
sd.theta <- summary(m)$coefficients[which.treat,2]
Vcov <- vcov(m, useScale = FALSE)
se <- sqrt(diag(Vcov))
betas <- lme4::fixef(m)
tval <- betas / se
### NB: issue with p-values in random effect models (cannot determine df easily)
### An alternative (as recommended by Bates et al) is to use CIs instead of p-values
ci.fixed <- confint(m,names(fixef(m)),method="Wald",level=1-sig.level)
lower.lim.check <- ci.fixed[,1]>0
upper.lim.check <- ci.fixed[,2]<0
signif <- as.numeric(((sign(betas)==1 & lower.lim.check==TRUE) | (sign(betas)==-1 & upper.lim.check==TRUE))[2])
### This estimates the p-value based on Normal approximation --- may not be too correct...
### See: http://mindingthebrain.blogspot.co.uk/2014/02/three-ways-to-get-parameter-specific-p.html
pval <- 2*pnorm(abs(tval), lower.tail = FALSE)
pvalue <- pval[which.treat]
###signif <- pvalue < alpha
VC <- as.data.frame(lme4::VarCorr(m))
rnd.eff.sd <- VC[,which(colnames(VC)=="sdcor")]
# Finds the rows that needs to be reported
to.report <- c(which(is.na(VC[,3]==T), which(VC[,1]=="Residual")))
rnd.eff.sd <- VC[to.report,which(colnames(VC)=="sdcor")]
names(rnd.eff.sd)[which(is.na(VC[to.report,3])==T)] <- paste(VC[to.report,"grp"],VC[to.report,"var1"])
names(rnd.eff.sd) <- gsub(" NA","",names(rnd.eff.sd))
if(family=="gaussian") {method <- "lmer"} else {method <- "glmer"}
}
# If it doesn't then do lm
if (!check.random.effect) {
if (family=="gaussian") {
m <- lm(formula,data=fake.data)
} else {
m <- glm(formula, data=fake.data,family=family)
}
which.treat <- which(names(m$coefficients)==treatment)
theta <- m$coefficients[which.treat]
sd.theta <- summary(m)$coefficients[which.treat,2]
se <- summary(m)$coefficients[,2]
betas <- summary(m)$coefficients[,1]
tval <- summary(m)$coefficients[,3]
df <- summary(m)$df[2]
pval <- summary(m)$coefficients[,4]
pvalue <- pval[which.treat]
signif <- pvalue < sig.level
if(family=="gaussian") {method <- "lm"} else {method <- "glm"}
rnd.eff.sd=NULL
}
list(power=signif,theta=theta,sd.theta=sd.theta,
rnd.eff.sd=rnd.eff.sd,method=method,pvalue=pvalue)
}
}
}
toc <- proc.time(); time2run <- (toc-tic)[3]; names(time2run) <- "Time to run (secs)"
theta=mean(unlist(res[2,]),na.rm=T); names(theta)=NULL
sd.theta <- mean(unlist(res[3,]),na.rm=T); names(sd.theta)=NULL
theta=c(theta,sd.theta); names(theta)=c("Estimate","Standard Error")
pvalue=unlist(res[6,]); names(pvalue)=NULL
method=unlist(res[5,1]); names(method)=NULL
power=mean(unlist(res[1,]),na.rm=T)
ci <- power+c(-qnorm(1-sig.level/2),qnorm(1-sig.level/2))*sqrt(var(pvalue<sig.level,na.rm=T)/n.sims)
ci <- power+c(-qnorm(1-sig.level/2),qnorm(1-sig.level/2))*sqrt(var(unlist(res[1,]),na.rm=T)/n.sims)
if(ci[1]<0) {ci[1] <- 0}; if(ci[2]>1) {ci[2] <- 1} # Constrains ci in [0;1]
if (!exists("data",where=exArgs)) {
setting <- list(n.clusters=I,n.time.points=J,avg.cluster.size=K,
design=design,formula=formula,method=method,family=family)
} else {setting=list(formula=formula,method=method,family=family)}
closeAllConnections()
# If the option 'plot=T' creates a moving average plot of the power
if (plot==TRUE) {
# Defines some cutoff points in the simulations to plot the moving average
percs <- seq(.1,1,by=.1)
# If the first %tile of the simulations is 0, then sets to 1
steps <- pmax(round(n.sims*percs),1)
# Computes the power moving average
mov_av <- numeric()
for (s in 1:length(steps)) {
mov_av[s] <- mean(unlist(res[1,c(1:steps[s])]))
}
# plot the moving average of power in a graph
plot(mov_av, type="o", ylim=c(0,1), ylab="Power", xlab="Percentage of Iterations", xaxt = 'n', main="Moving Average of Power")
axis(1, at=1:10, labels = c(paste((1:10)*10, "%", sep="")))
#### NEED TO CHECK THIS
abline(h=power,col="red",lwd=2)
abline(h=ci[1],col="red",lwd=2,lty=2)
abline(h=ci[2],col="red",lwd=2,lty=2)
}
if (!is.null(unlist(res[4,]))) {
rnd.eff.sd=apply(do.call(cbind.data.frame,res[4,]),1,mean,na.rm=T)
} else {
rnd.eff.sd=NULL
}
list(power=power,time2run=time2run,ci.power=ci,theta=theta,
rnd.eff.sd=rnd.eff.sd,setting=setting) # pvalue=pvalue,
}
#
## OLD VERSION WHICH DOESN'T WORK ON WINDOWS
# sim.swt <- function (I,J,H=NULL,K,design="cross-sec",mu=0,b.trt,b.time=NULL,
# sigma.y=NULL,sigma.e=NULL,rho=NULL,sigma.a=NULL,
# rho.ind=NULL,sigma.v=NULL,n.sims=1000,formula=NULL,
# alpha=0.05,n.cores=NULL,...) {
#
# ## Power analysis for repeated cohort design (cross-sectional) - continuous outcome
# ## The optional arguments include the specification of a user-defined function data
# ## which is used to simulate the data, according to any design the user has.
# ## So including data=data, where data=function(){...} specifies the simulation
# ## of a suitable dataset allows the user to select *any* kind of model for data
# ## generation. If data has inputs, then these must be specified as a list
# ## inpts, to be passed as an extra argument to sim.sw.cont.
# ## In addition to this, the user should then select a suitable formula
# ## which is in line with the model used for the simulation of the virtual trial.
# ## Also, the name of the "treatment" variable can be specified as a string in the
# ## extra argument treat.name. If not present then it is assumed that the name of the
# ## treatment variable is, unsurprisingly, "treatment"
#
# exArgs <- list(...)
#
# ## Uses parallel computation to speed up task
# # required <- c("lme4","foreach","doParallel","parallel","iterators")
# # for (i in 1:length(required)) {
# # txt1 <- paste0("if(!isTRUE(requireNamespace('",required[i],"',quietly=TRUE))) {")
# # txt2 <- paste0("stop('You need to install the package ",required[i],". Please run in your R terminal:
# # install.packages(",required[i],")')")
# # txt3 <- paste0("}")
# # txt <- paste(txt1,txt2,txt3,sep="\n",collapse=" ")
# # eval(parse(text=txt))
# # }
# requireNamespace("lme4",quietly=TRUE)
# requireNamespace("foreach",quietly=TRUE)
# requireNamespace("doParallel",quietly=TRUE)
# requireNamespace("parallel",quietly=TRUE)
# requireNamespace("iterators",quietly=TRUE)
#
# # If not specified, uses all cores available to the machine
# if(is.null(n.cores)) {n.cores <- parallel::detectCores()}
# # Usually a good idea to leave 1 core free for the OS to use
# doParallel::registerDoParallel(cores=n.cores-1)
#
# # Defines basic formulation for the model
# if(is.null(formula)){
# formula=y~treatment+factor(time)+(1|cluster)
# if(design=="cohort") {formula <- update.formula(formula,.~.+(1|id))}
# }
# # Defines the name of the treatment variable (for which the main analysis is performed)
# if(exists("treat.name",where=exArgs)) {treatment <- exArgs$treat.name} else {treatment <- "treatment"}
#
# # Simulates the datasets & analysis
# tic <- proc.time()
# ### NB Needs to use the .export argument to the function foreach, to include
# ### variables that are passed outside of the scope (eg b.trt) --- need to figure
# ### this out a bit better. But basically, linux/mac will make stuff available,
# ### while windows won't and so throw an error because some variable that is input
# ### the function in foreach is not loaded in memory...
# res <- foreach::foreach(iterators::icount(n.sims), .combine=cbind, .packages="lme4",
# .export=c("make.swt","sw.design.mat")) %dopar% {
# if(exists("data",where=exArgs)) {
# func <- exArgs$data
# if(exists("inpts",where=exArgs)) {inpts=exArgs$inpts} else {inpts=list()}
# fake.data <- do.call(what=func,args=inpts)
# }
# if(!exists("data",where=exArgs)) {
# if(exists("X",where=exArgs)) {
# X=exArgs$X
# row.names(X) <- sample(1:I,I)
# colnames(X) <- c("Baseline",paste0("Time ",1:J))
# } else {
# X=NULL
# }
# fake.data <- make.swt(I=I,J=J,H=H,K=K,design=design,mu=mu,b.trt=b.trt,
# b.time=b.time,sigma.y=sigma.y,sigma.e=sigma.e,
# rho=rho,sigma.a=sigma.a,rho.ind=rho.ind,
# sigma.v=sigma.v,X=X)
# }
#
# # If the formula contains random effects, run lmer
# check.random.effect <- !is.null(findbars(formula))
# if(check.random.effect==TRUE) {
# m <- lme4::lmer(formula, data=fake.data)
# which.treat <- which(names(fixef(m))==treatment)
# theta <- lme4::fixef(m)[which.treat]
# sd.theta <- summary(m)$coefficients[which.treat,2]
# Vcov <- vcov(m, useScale = FALSE)
# se <- sqrt(diag(Vcov))
# betas <- lme4::fixef(m)
# tval <- betas / se
# ### NB: issue with p-values in random effect models (cannot determine df easily)
# ### An alternative (as recommended by Bates et al) is to use CIs instead of p-values
# ci.fixed <- confint(m,names(fixef(m)),method="Wald",level=1-alpha)
# lower.lim.check <- ci.fixed[,1]>0
# upper.lim.check <- ci.fixed[,2]<0
# signif <- as.numeric(((sign(betas)==1 & lower.lim.check==TRUE) | (sign(betas)==-1 & upper.lim.check==TRUE))[2])
# ### This estimates the p-value based on Normal approximation --- may not be too correct...
# ### See: http://mindingthebrain.blogspot.co.uk/2014/02/three-ways-to-get-parameter-specific-p.html
# pval <- 2*pnorm(abs(tval), lower.tail = FALSE)
# pvalue <- pval[which.treat]
# ###signif <- pvalue < alpha
# VC <- as.data.frame(lme4::VarCorr(m))
# rnd.eff.sd <- VC[,which(colnames(VC)=="sdcor")]
# # Finds the rows that needs to be reported
# to.report <- c(which(is.na(VC[,3]==T), which(VC[,1]=="Residual")))
# rnd.eff.sd <- VC[to.report,which(colnames(VC)=="sdcor")]
# names(rnd.eff.sd)[which(is.na(VC[to.report,3])==T)] <- paste(VC[to.report,"grp"],VC[to.report,"var1"])
# names(rnd.eff.sd) <- gsub(" NA","",names(rnd.eff.sd))
# method <- "lmer"
# }
# # If it doesn't then do lm
# if (!check.random.effect) {
# m <- lm(formula,data=fake.data)
# which.treat <- which(names(m$coefficients)==treatment)
# theta <- m$coefficients[which.treat]
# sd.theta <- summary(m)$coefficients[which.treat,2]
# se <- summary(m)$coefficients[,2]
# betas <- summary(m)$coefficients[,1]
# tval <- summary(m)$coefficients[,3]
# df <- summary(m)$df[2]
# ## pval <- 2*pt(abs(tval),df=df,lower.tail = FALSE)
# pval <- summary(m)$coefficients[,4]
# pvalue <- pval[which.treat]
# signif <- pvalue < alpha
# method <- "lm"
# rnd.eff.sd=NULL
# }
# list(power=signif,theta=theta,sd.theta=sd.theta,rnd.eff.sd=rnd.eff.sd,method=method,pvalue=pvalue)
# }
# toc <- proc.time(); time2run <- (toc-tic)[3]; names(time2run) <- "Time to run (secs)"
# theta=mean(unlist(res[2,]),na.rm=T); names(theta)=NULL
# sd.theta <- mean(unlist(res[3,]),na.rm=T); names(sd.theta)=NULL
# theta=c(theta,sd.theta); names(theta)=c("Estimate","Standard Error")
# pvalue=unlist(res[6,]); names(pvalue)=NULL
# power=mean(unlist(res[1,]),na.rm=T)
# ci <- power+c(-qnorm(1-alpha/2),qnorm(1-alpha/2))*sqrt(var(pvalue<alpha,na.rm=T)/n.sims)
# if(ci[1]<0) {ci[1] <- 0}; if(ci[2]>1) {ci[2] <- 1} # Constrains ci in [0;1]
# if (!exists("data",where=exArgs)) {
# setting <- list(n.clusters=I,n.time.points=J,avg.cluster.size=K,
# design=design,formula=formula)
# } else {setting=list(formula=formula)}
#
# list(power=power,time2run=time2run,ci.power=ci,theta=theta,
# rnd.eff.sd=unlist(res[4,1]),setting=setting) # pvalue=pvalue,
# }
HH.binary <- function(p1,OR,I,J,K,rho=0,sig.level=0.05,which.var="within",X=NULL) {
# HH sample size calculations for binary outcome
if(is.null(X)) {
X <- sw.design.mat(I=I,J=J,H=NULL)
} else {
row.names(X) <- sample(1:I,I)
colnames(X) <- c("Baseline",paste0("Time ",1:J))
}
U <- sum(X)
W <- sum(apply(X,2,sum)^2)
V <- sum(apply(X,1,sum)^2)
# Data
p2 <- OR*(p1/(1-p1))/(1+OR*(p1/(1-p1)))
theta=abs(p1-p2)
if (which.var=="within") {
sigma.e=sqrt((p1*(1-p1)+p2*(1-p2))/2)
####sigma.e=sqrt((p1*(1-p1))) # Consistent with Hemming - stata
sigma.a <- sqrt(rho*sigma.e^2/(1-rho))
sigma.y <- sqrt(sigma.e^2+sigma.a^2)
}
if (which.var=="total") {
sigma.y=sqrt((p1*(1-p1)+p2*(1-p2))/2)
####sigma.y=sqrt((p1*(1-p1)))
sigma.a <- sqrt(sigma.y^2*rho)
sigma.e <- sqrt(sigma.y^2-sigma.a^2)
}
sigma <- sqrt(sigma.e^2/K)
# Power calculations
v <- (I*sigma^2*(sigma^2+((J+1)*sigma.a^2)))/((I*U-W)*sigma^2+(U^2+I*(J+1)*U-(J+1)*W-I*V)*sigma.a^2)
power <- pnorm(theta/sqrt(v)-qnorm(1-sig.level/2))
setting <- list(n.clusters=I,n.time.points=J,avg.cluster.size=K,design.matrix=X)
list(power=power,p1=p1,p2=p2,sigma.y=sigma.y,sigma.e=sigma.e,sigma.a=sigma.a,setting=setting)
}
HH.normal <- function(mu,b.trt,sigma,I,J,K,rho=0,sig.level=0.05,which.var="within",X=NULL) {
# HH sample size calculations for continuous (normal) outcome
# Stepped wedge design matrix
if(is.null(X)) {
X <- sw.design.mat(I=I,J=J,H=NULL)
} else {
row.names(X) <- sample(1:I,I)
colnames(X) <- c("Baseline",paste0("Time ",1:J))
}
U <- sum(X)
W <- sum(apply(X,2,sum)^2)
V <- sum(apply(X,1,sum)^2)
# Data
mu1 <- mu+b.trt
theta <- abs(mu1-mu)
# Assumes that the input sigma.y is in fact the within cluster sd
if (which.var=="within") {
sigma.e <- sigma
sigma.a <- sqrt(rho*sigma.e^2/(1-rho))
sigma.y <- sqrt(sigma.e^2+sigma.a^2)
}
# Assumes that the input sigma.y is total sd
if (which.var=="total") {
sigma.a <- sqrt(sigma^2*rho)
sigma.e <- sqrt(sigma^2-sigma.a^2)
sigma.y <- sigma
}
sigma <- sqrt(sigma.e^2/K)
# Power calculations
v <- (I*sigma^2*(sigma^2+((J+1)*sigma.a^2)))/((I*U-W)*sigma^2+(U^2+I*(J+1)*U-(J+1)*W-I*V)*sigma.a^2)
power <- pnorm(theta/sqrt(v)-qnorm(1-sig.level/2))
setting <- list(n.clusters=I,n.time.points=J,avg.cluster.size=K,design.matrix=X)
list(power=power,sigma.y=sigma.y,sigma.e=sigma.e,sigma.a=sigma.a,setting=setting)
}
HH.count <- function(lambda1,RR,I,J,K,rho=0,sig.level=0.05,which.var="within",X=NULL) {
# HH sample size calculations for continuous (normal) outcome
# Stepped wedge design matrix
if(is.null(X)) {
X <- sw.design.mat(I=I,J=J,H=NULL)
} else {
row.names(X) <- sample(1:I,I)
colnames(X) <- c("Baseline",paste0("Time ",1:J))
}
U <- sum(X)
W <- sum(apply(X,2,sum)^2)
V <- sum(apply(X,1,sum)^2)
# Data
lambda2 <- RR*lambda1
theta <- abs(lambda1-lambda2)
if (which.var=="within") {
sigma.e=(sqrt(lambda1)+sqrt(lambda2))/2 # Consistent with Hemming - stata
sigma.a <- sqrt(rho*sigma.e^2/(1-rho))
sigma.y <- sqrt(sigma.e^2+sigma.a^2)
}
if (which.var=="total") {
sigma.y <- (sqrt(lambda1)+sqrt(lambda2))/2 # Consisten with Hemming - stata
sigma.a <- sqrt(sigma.y^2*rho)
sigma.e <- sqrt(sigma.y^2-sigma.a^2)
}
sigma <- sqrt(sigma.e^2/K)
# Power calculations
v <- (I*sigma^2*(sigma^2+((J+1)*sigma.a^2)))/((I*U-W)*sigma^2+(U^2+I*(J+1)*U-(J+1)*W-I*V)*sigma.a^2)
power <- pnorm(theta/sqrt(v)-qnorm(1-sig.level/2))
setting <- list(n.clusters=I,n.time.points=J,avg.cluster.size=K,design.matrix=X)
list(power=power,lambda1=lambda1,lambda2=lambda2,sigma.y=sigma.y,sigma.e=sigma.e,sigma.a=sigma.a,setting=setting)
}
sw.design.mat <- function(I,J,H=NULL) {
## Creates the design matrix for a stepped wedge design
## Checks to see that data are consistent with SW design
if(sum(sapply(list(I,J,H), is.null)) != 1) {
warning("exactly one of 'I', 'J' and 'H' must be NULL")
}
if (is.null(I)) {