intensity-code-and-data/ipe_model_code.R at master · r-clements/intensity-code-and-data · GitHub

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
#AUTHOR: Robert Clements
#DATE: February 18, 2013
#SUMMARY: Fit an ordinal regression model, support vector machine, and a gam with a spatial term,
#to the cleaned up DBMI database.
require(mgcv)
require(ordinal)
require(e1071)
require(ggplot2) #for some very nice plotting
require(ggmap) #for including maps
require(RColorBrewer)


#############
#load data and explore it

ipe.data <- read.table("Raw data/all_clean_ordinal_data.txt", header = TRUE)
head(ipe.data)

p <- ggplot(ipe.data, aes(factor(MCS), dist))
p + geom_boxplot()

p <- ggplot(ipe.data, aes(factor(MCS), Mw))
p + geom_boxplot()

p <- ggplot(ipe.data, aes(dist, MCS, alpha = .1))
p + geom_point() + theme(legend.position="none")

brewer.div <- colorRampPalette(brewer.pal(9, "RdBu"), interpolate = "spline")
cols <- (brewer.div(10))

p <- ggplot(ipe.data, aes(log(dist), Mw, color = MCS))
p + geom_point() +
  scale_colour_gradientn(colours = cols,
                         breaks = 2:11,
                         guide = "legend")


################
#visualize the locations

locations <- unique(ipe.data[,c("lon", "lat")])
#ity = get_map(location = "italy", zoom = 5, source = "osm", color = "bw")
#open street map's server is often unavailable, so I'd suggest saving this map
#save(ity, file = "ity")
load("ity")
locs <- ggmap(ity)
locs <- locs +
  geom_point(data = locations, mapping = aes(x = lon, y = lat))
locs


################
#start modeling
#ordered logistic model first

#make MCS a factor, which it is
ipe.data$MCS <- as.factor(ipe.data$MCS)

#start with ordinal regression with cumulative link
formula1 <- MCS ~ Mw + dist

model.ord <- clm(formula1, data = ipe.data, link = "logit", threshold = "flexible")
summary(model.ord)


################
#svm next

#support vector machine with defaults
formula1 <- MCS ~ Mw + dist

model.svm <- svm(formula1, data = ipe.data, type = "C-classification", kernel = "radial")

#circles are support vectors, triangles are the rest of the data
#terrain colors are our different classes, rainbow colors are the classes that
#each point belongs to
plot(model.svm, ipe.data, Mw~dist, dataSymbol = 2, svSymbol = 1, symbolPalette = rainbow(10),
     color.palette = terrain.colors)


################
#gam last

#gam with smooth terms and spatial term
gam.data <- ipe.data
gam.data$MCSgam <- as.numeric(gam.data$MCS)

formula.gam <- MCSgam ~ te(dist, Mw) + s(dist) + s(Mw) + s(lon, lat, bs="tp", k=10)

model.gam <- gam(formula.gam, data = gam.data)

summary(model.gam)


################
#load testing data and get our IPE model predictions for comparison

#our testing data and IPE predictions
test.data <- read.table("ipe_test_data1.txt", h=TRUE)
IPE1 <- test.data$Ipre
temp.data <- read.table("ipe_test_data2.txt", h=TRUE)
IPE2 <- temp.data$Ipre

test.data.internet <- read.table("ipe_test_data_internet1.txt", h=TRUE)
IPE.internet1 <- test.data.internet$Ipre
temp.data <- read.table("ipe_test_data_internet2.txt", h=TRUE)
IPE.internet2 <- temp.data$Ipre

rm(temp.data)

pred.ord <- predict(model.ord, newdata=test.data, type="class")[[1]]
pred.svm <- predict(model.svm, newdata=test.data)
pred.gam <- round(predict(model.gam, newdata=test.data))

pred.ord.internet <- predict(model.ord, newdata=test.data.internet, type="class")[[1]]
pred.svm.internet <- predict(model.svm, newdata=test.data.internet)
pred.gam.internet <- round(predict(model.gam, newdata=test.data.internet))


################
#compare using mae

#use mae, which may not make a lot of sense
mae.ord <- mean(abs(test.data$MCSgam-as.numeric(pred.ord)))
mae.svm <- mean(abs(test.data$MCSgam-as.numeric(pred.svm)))
mae.gam <- mean(abs(test.data$MCSgam-(pred.gam)))

mae.ord.internet <- mean(abs(test.data.internet$MCSgam-as.numeric(pred.ord.internet)))
mae.svm.internet <- mean(abs(test.data.internet$MCSgam-as.numeric(pred.svm.internet)))
mae.gam.internet <- mean(abs(test.data.internet$MCSgam-(pred.gam.internet)))

mae.all <- c(mae.ord, mae.svm, mae.gam)
mae.all.internet <- c(mae.ord.internet, mae.svm.internet, mae.gam.internet)

#compare with our two Italy IPEs
IPE1.mcs <- round(IPE1)-1
IPE2.mcs <- round(IPE2)-1
IPE.internet1.mcs <- round(IPE.internet1)-1
IPE.internet2.mcs <- round(IPE.internet2)-1

mae.IPE1 <- mean(abs(test.data$MCSgam-IPE1.mcs))
mae.IPE2 <- mean(abs(test.data$MCSgam-IPE2.mcs))
mae.IPE.internet1 <- mean(abs(test.data.internet$MCSgam-IPE.internet1.mcs))
mae.IPE.internet2 <- mean(abs(test.data.internet$MCSgam-IPE.internet2.mcs))

mae.IPE <- c(mae.IPE1, mae.IPE2)
mae.IPE.internet <- c(mae.IPE.internet1, mae.IPE.internet2)

mae.all
mae.IPE

mae.all.internet
mae.IPE.internet