-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsimulation_sampling_distribution.R
211 lines (189 loc) · 10.9 KB
/
simulation_sampling_distribution.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
library(data.table)
library(ggplot2)
library(grid)
library(latex2exp) # For ggplot2 xlab
library(parallel)
library(Rsolnp)
source("hmm_functions.R")
source("hmm_parameters.R")
source("ggplot_utils.R")
set_ggplot_theme()
set.seed(321321)
## True HMM parameters used in simulation
params0 <- get_params0()
get_estimates <- function(params0, n_panel_elements, n_random_starts) {
panel <- replicate(n_panel_elements, simulate_hmm(params0), simplify=FALSE)
estimates <- get_hmm_and_minimum_distance_estimates_random_initialization(params0, panel, n_random_starts)
return(estimates)
}
cluster <- makeCluster(detectCores())
n_replications <- 100
n_random_starts <- 6
outfile_format <- "simulation_sampling_distribution_with_%s_random_initial_parameters_panel_size_%s_%s_replications.rds"
clusterExport(cluster, c("baum_welch",
"eq_function_minimum_distance",
"get_estimates",
"get_expectation_maximization_estimates",
"get_hmm_and_minimum_distance_estimates_random_initialization",
"get_min_distance_estimates",
"get_random_initial_parameters",
"get_transition_probs_from_M_S_joint",
"n_random_starts",
"objfn_minimum_distance",
"params0",
"simulate_discrete_markov",
"simulate_hmm",
"valid_panel_element",
"valid_parameters"))
panel_sizes <- c(100, 200, 500, 1000, 5000, 10000) # Slow, panel size 5000 took 22 hours on my laptop (4 CPUs)
panel_sizes <- c(100, 200, 500, 1000)
for(panel_size in panel_sizes) {
outfile <- sprintf(outfile_format, n_random_starts, panel_size, n_replications)
if(file.exists(outfile)) {
message(outfile, " already exists; won't re-run simulation")
} else {
message("running replications with random initialization, time is ", Sys.time())
replications_random_initialization <- parLapply(cluster, rep(panel_size, n_replications), function(x) {
get_estimates(params0=params0, n_panel_elements=x, n_random_starts=n_random_starts) # Slow...
})
message("done, time is ", Sys.time())
message("saving ", outfile)
saveRDS(replications_random_initialization, outfile)
}
}
stopCluster(cluster)
infiles <- sprintf(outfile_format, n_random_starts, panel_sizes, n_replications)
## This is a list of lists: for each set of parameter values, we have multiple replications of the simulation
all_simulations <- lapply(infiles, readRDS)
stopifnot(all(sapply(all_simulations, length) == n_replications))
dataframes <- lapply(all_simulations, function(replications) {
df <- data.frame(panel_size=sapply(replications, function(replication) {
return(replication$em_params_hat_best_likelihood$panel_size)
}),
em_params_hat_n_iterations=sapply(replications, function(replication) {
return(replication$em_params_hat_best_likelihood$n_em_iterations)
}),
em_params_hat_pr_y_11=sapply(replications, function(replication) {
return(replication$em_params_hat_best_likelihood$pr_y[1, 1])
}),
em_params_hat_pr_y_22=sapply(replications, function(replication) {
return(replication$em_params_hat_best_likelihood$pr_y[2, 2])
}),
min_dist_params_hat_pr_y_11=sapply(replications, function(replication) {
return(replication$min_dist_params_hat_best_objfn$pr_y[1, 1])
}),
min_dist_params_hat_pr_y_22=sapply(replications, function(replication) {
return(replication$min_dist_params_hat_best_objfn$pr_y[2, 2])
}))
for(time_index in seq_along(params0$P_list)) {
varname <- sprintf("em_params_hat_P%s_11", time_index)
df[, varname] <- sapply(replications, function(replication) {
return(replication$em_params_hat_best_likelihood$P_list[[time_index]][1, 1])
})
varname <- sprintf("min_dist_params_hat_P%s_11", time_index)
df[, varname] <- sapply(replications, function(replication) {
return(replication$min_dist_params_hat_best_objfn$P_list[[time_index]][1, 1])
})
}
return(df)
})
df <- rbindlist(dataframes)
df$panel_size_label <- sprintf("%s%s", ifelse(df$panel_size %in% c(100, 1000), "sample size = ", ""), df$panel_size)
df$panel_size_label <- factor(df$panel_size_label,
levels=sprintf("%s%s", ifelse(sort(unique(df$panel_size)) %in% c(100, 1000), "sample size = ", ""), sort(unique(df$panel_size))))
estimators <- c("em", "min_dist")
## Transition probs
for (time_index in seq_along(params0$P_list)) {
for(estimator in estimators) {
variable <- sprintf("%s_params_hat_P%s_11", estimator, time_index)
p <- (ggplot(df, aes_string(x=variable)) +
geom_vline(xintercept=params0$P_list[[time_index]][1, 1], color="grey", linetype=2) +
geom_histogram(binwidth=0.005, color="black", fill="white") +
facet_wrap(~ panel_size_label, scale="free_x") +
geom_rug(aes(x=true_value), data=data.frame(true_value=params0$P_list[[time_index]][1, 1])) + # Correct parameter value
ylab("") + xlab(""))
outfile <- sprintf("simulation_methodology_%s_random_initialization_P_hat_%s_11_sampling_distribution.png", estimator, time_index)
message("saving ", outfile)
ggsave(outfile, p, width=12.5, height=8)
}
}
## Misclassification probs
for(matrix_index in c("11", "22")) {
for(estimator in estimators) {
variable <- sprintf("%s_params_hat_pr_y_%s", estimator, matrix_index)
true_value <- ifelse(matrix_index == "11", params0$pr_y[1, 1], params0$pr_y[2, 2])
p <- (ggplot(df, aes_string(x=variable)) +
geom_vline(xintercept=true_value, color="grey", linetype=2) +
geom_histogram(binwidth=0.005, color="black", fill="white") +
facet_wrap(~ panel_size_label, scale="free_x") +
geom_rug(aes(x=true_value), data=data.frame(true_value=true_value)) + # Correct parameter value
ylab("") + xlab(""))
outfile <- sprintf("simulation_methodology_%s_random_initialization_pr_y_%s_sampling_distribution.png", estimator, matrix_index)
message("saving ", outfile)
ggsave(outfile, p, width=12.5, height=8)
}
}
df_melted <- melt(df, id.vars=c("panel_size", "panel_size_label"))
p <- (ggplot(df_melted, aes(x=value)) +
geom_histogram(binwidth=0.01, fill="white", color="black") +
facet_wrap(~ variable, scale="free_x"))
p # Looks correct, compare to params0 -- careful, lots of estimates hit max_iter
## Do min dist estimates vary with initial parameters?
P_hat_range_min_dist <- c(lapply(all_simulations, function(list_of_replications) {
sapply(list_of_replications, function(replication) {
matrix_of_P_hat <- sapply(replication$min_dist_params_hat, function(x) {
return(c(x$P_list, recursive=TRUE))
})
P_hat_range <- max(apply(matrix_of_P_hat, 1, max) - apply(matrix_of_P_hat, 1, min))
return(P_hat_range)
})
}), recursive=TRUE)
sum(P_hat_range_min_dist > 0.01) # 60 simulations out of length(P_hat_range_min_dist)=1200 where P_hat differs by more than 0.01 across initializations
sum(P_hat_range_min_dist > 0.10) # 37 simulations out of length(P_hat_range_min_dist)=1200 where P_hat differs by more than 0.10 across initializations
P_hat_range_em <- c(lapply(all_simulations, function(list_of_replications) {
sapply(list_of_replications, function(replication) {
matrix_of_P_hat <- sapply(replication$em_params_hat_list, function(x) {
return(c(x$P_list, recursive=TRUE))
})
P_hat_range <- max(apply(matrix_of_P_hat, 1, max) - apply(matrix_of_P_hat, 1, min))
return(P_hat_range)
})
}), recursive=TRUE)
sum(P_hat_range_em > 0.01) # 1183 simulations out of length(P_hat_range_em)=1200
sum(P_hat_range_em > 0.10) # 394 simulations out of length(P_hat_range_em)=1200
## RMSE of each estimator as fn of sample size
dtable <- data.table(df)
setkey(dtable, panel_size)
with(subset(dtable, panel_size <= 5000), t.test(x=(em_params_hat_pr_y_11 - params0$pr_y[1, 1])^2,
y=(min_dist_params_hat_pr_y_11 - params0$pr_y[1, 1])^2)) # P-value 0.02
with(subset(dtable, panel_size <= 5000), t.test(x=(em_params_hat_P1_11 - params0$P_list[[1]][1, 1])^2,
y=(min_dist_params_hat_P1_11 - params0$P_list[[1]][1, 1])^2)) # P-value 0.11
with(subset(dtable, panel_size <= 5000), t.test(x=(em_params_hat_P2_11 - params0$P_list[[2]][1, 1])^2,
y=(min_dist_params_hat_P2_11 - params0$P_list[[2]][1, 1])^2)) # P-value 0.09
with(subset(dtable, panel_size <= 5000), t.test(x=(em_params_hat_P3_11 - params0$P_list[[3]][1, 1])^2,
y=(min_dist_params_hat_P3_11 - params0$P_list[[3]][1, 1])^2)) # P-value 0.004
dtable_rmses <- dtable[, list(rmse_em_pr_y_11=sqrt(mean((em_params_hat_pr_y_11 - params0$pr_y[1, 1])^2)),
rmse_min_dist_pr_y_11=sqrt(mean((min_dist_params_hat_pr_y_11 - params0$pr_y[1, 1])^2, na.rm=TRUE)),
rmse_em_P1_11=sqrt(mean((em_params_hat_P1_11 - params0$P_list[[1]][1, 1])^2)),
rmse_min_dist_P1_11=sqrt(mean((min_dist_params_hat_P1_11 - params0$P_list[[1]][1, 1])^2, na.rm=TRUE)),
rmse_em_P2_11=sqrt(mean((em_params_hat_P2_11 - params0$P_list[[2]][1, 1])^2)),
rmse_min_dist_P2_11=sqrt(mean((min_dist_params_hat_P2_11 - params0$P_list[[2]][1, 1])^2, na.rm=TRUE)),
rmse_em_P3_11=sqrt(mean((em_params_hat_P3_11 - params0$P_list[[3]][1, 1])^2)),
rmse_min_dist_P3_11=sqrt(mean((min_dist_params_hat_P3_11 - params0$P_list[[3]][1, 1])^2, na.rm=TRUE))),
by="panel_size"]
dtable_melted <- melt(dtable_rmses, id.vars="panel_size")
dtable_melted[, estimator := ifelse(grepl("_em_", variable), "EM", "MinDist")]
dtable_melted[, parameter := ifelse(grepl("pr_y_11", variable), "first entry of misclassification matrix",
sprintf("first entry of %s", gsub("rmse_min_dist_|rmse_em_|_11", "", variable)))]
p <- (ggplot(dtable_melted, aes(x=panel_size, y=value, color=estimator)) +
geom_point() + geom_line(size=1.2) +
scale_color_manual("estimator", values=c("black", "grey")) +
scale_x_continuous("panel size", breaks=c(100, 1000, 5000, 10000)) +
scale_y_continuous("root mean squared error") +
geom_hline(yintercept=0, linetype=2, color="grey") +
facet_wrap(~ parameter, scales="free_y"))
p
outfile <- "simulation_methodology_random_initialization_rmse_em_and_min_dist.png"
message("saving ", outfile)
ggsave(outfile, p, width=11, height=8)
message("done with simulation, time is ", Sys.time())