ACTEX Learning

R Formula & Review Sheet

(updated 10/11/2024)

DISTRIBUTIONS
Functions related to distributions belong to the (automatically loaded) stats package and have the form:
d<distribution>(x, <parameters>) returns the value of the PDF (f (x) or P (X = x)
p<distribution>(x, <parameters>) returns the value of the CDF (F (x)) or P (X ≤ x))
p<distribution>(x, <parameters>, lower=FALSE) returns 1 − F (x) or P (X > x)
q<distribution>(p, <parameters>) returns the min value of x for which P (X ≤ x) ≥ p (p-th quantile)
q<distribution>(p, <parameters>, lower=FALSE) returns the min value of x for which P (X > x) ≥ p
r<distribution>(n, <parameters>) returns an iid random sample of size n from the distribution

Continuous Distributions:
The parameters for Continuous Distributions are:
(a) Uniform Distribution (U (a, b)) unif(<>,a,b), dunif(<>, a, b), punif(<>, a, b), …
(b) Exponential Distribution (Exp(λ)) exp(<>, lambda)
(c) Gamma Distribution (Gamma(α, λ)) gamma(<>, alpha, lambda)
(d) Chi-square Distribution (χ2df ) chisq(<>,df)
(e) Beta Distribution (Beta(α, β)) beta(<>,alpha,beta)
(f) Normal Distribution (N (µ, σ )) 2
norm(<>,mu,sigma)
(g) Log-normal Distribution (Lognormal(µ, σ 2 )) lnorm(<>,mu,sigma)
(h) Standard Normal Distribution (N(0,1)) norm(<>)
(i) Student’s t-distribution (t(ν) ) t(<>, nu)
(j) F-Distribution (F (d1 , d2 )) f(<>, d1, d2)

Discrete Distributions:
The parameters for Discrete Distributions are:
(a) Binomial Distribution (Bin(n, p)) binom(<>, n, p)
(b) Poisson Distribution (P(λ)) pois(<>, lambda)
(c) Geometric Distribution (Geom(p)) geom(<>, p)
(d) Negative Binomial Distribution (NB(k, p)) nbinom(<>, k, p)
(e) Hypergeometric Distribution (HG(N, K, n)) hyper(<>, N, K, n)
(f) Uniform Distribution:
R does not provide built-in functions to deal with the discrete uniform distribution. Here are some alternatives:

x <- seq(a, b, i) # generates a sequence a, a+i, a+2i, ... a+ki <= b

length(x[x<=t]) / length(x) # yields P[X<=t]
length(x[x>=p | x>=q]) / length(x) # yields P[X>=p or X>=q]
length(x[x>=p & x>=q]) / length(x[x >= q]) # yields P[X=>p | X>=q] (conditional probability)

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 2

TABLES AND PLOTS

1. Frequency Table:
Example: Frequency table for an iid sample X with Xi ∼ Bin(100, 0.2). The bottom row is the number of occurrences of
the corresponding value in the top row in the sample.

table(rbinom(100, 100, 0.1))

##
## 4 5 6 7 8 9 10 11 12 13 14 15 16 18
## 2 6 3 13 11 13 15 13 12 8 1 1 1 1

2. Bar chart:
Example: Bar chart for the PDF of Bin(50, 0.4):
x <- 0:50
barplot(dbinom(x, 50, 0.4), names = x)
0.06
0.00

0 4 8 13 18 23 28 33 38 43 48

3. Plots:
Example: Plot of the PDF of N (10, 22 ):

x <- seq(0, 30, 0.01)

plot(x, dnorm(x, 10, 2), # x and y
type = "l", # type (p = points, l = lines, o = both overlaid)
main = "Normal Distribution Graph", # plot title
xlab = "x-values", ylab = "probabilities", # axis labels
xlim = c(0,30), ylim = c(0,0.25), # axis limits
col = "red", lwd = 2, # color and line width
lty = 2) # line type (1 = solid, 2 = dashed, 3 = dotted)

Normal Distribution Graph

probabilities

0.15
0.00

0 5 10 15 20 25 30

x−values

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 3

Plots over existing plots

These function calls are ran after a call to plot(). They all support the argument col=<> to assign colors.
(a) lines(x, y): Adds a line plot.
(b) points(x, y): Adds a scatter plot.
(c) abline(h = y0): Adds a horizontal line at y0
(d) abline(v = x0): Adds a vertical line at x0 .
(e) abline(a, b): Adds the straight line y = a + b · x.

Example: Plot of the PDFs of N (10, 22 ) and N (20, 42 ), highlighting the µs:

x <- seq(0, 30, 0.01)

plot(x, dnorm(x, 10, 2), type = "l", xlab = "x-values", ylab="probabilities")
lines(x, dnorm(x, 20, 4))
abline(v = 10, col = 'red')
abline(v = 20, col = 'red')
0.20
probabilities

0.10
0.00

0 5 10 15 20 25 30

x−values

Q-Q Plots
(a) qqnorm(x): Plots quantiles of sample data x (vertical axis) vs quantiles of a normal distribution (horizontal
axis).
(b) qqplot(sim,x): The same as above, but with generic distribution quantiles sim
(e.g. generated by qexp(seq(0, 1, 0.05), lambda))
(c) qqline(x): Adds a reference line on a Q-Q plot created with qqnorm().
Example: Q-Q plot of normal data in two different ways:
y <- rnorm(20, 50, 1) # generate a random sample
qqnorm(y) # approach 1: qqnorm
qqline(y)
sim <- qnorm(seq(0, 1, 0.01), 0, 1) # generate quantiles
qqplot(sim, y) # approach 2: qqplot

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 4

Normal Q−Q Plot

Sample Quantiles

52
50
48
−2 −1 0 1 2

Theoretical Quantiles
52
50
y

48

−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

sim

CENTRAL LIMIT THEOREM (CLT)

Example: Assume we have an iid sample X̄ of size 50 from Bin(30, 0.6). By the CLT

X̄ ∼ N (18, 0.144)

N <- 30
p <- 0.6
set.seed(30) # setting a seed makes the code results repeatable!
xbar <- rep(0,1000) # initialize a sequence [1, 2, 3, ..., 1000]
for (i in 1:1000) {
x <- rbinom(50, N, p) # generate a sample of size 50 from Bin(30, 0.6)
xbar[i] <- mean(x) # replace each value of the sequence by the sample mean
}
hist(xbar, prob = TRUE, xlab = "Sample means")
rng <- range(xbar) # returns a vector with [min(xbar), max(xbar)]
xvals <- seq(rng[0], rng[1], 0.001)
lines(xvals, dnorm(xvals, N * p, sqrt((N * p * (1-p)) / 50)), col = "red")

Histogram of xbar
1.2
Density

0.6
0.0

16.5 17.0 17.5 18.0 18.5 19.0

Sample means

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 5

DATA ANALYSIS
Scatter Plots:

plot(<x>, <y>) # same function as before

plot(<dataframe>) # for multivariate data and bivariate data

Correlation:

cor(<x>, <y>, method = "<>")

cor(<dataframe>, method = "<>")
# method can be "pearson" (by default), "spearman" or "kendall"

Correlation Test:

cor.test(<x>, <y>, method = "pearson", alt = ”two.sided”, conf = 0.95)

# method can be "pearson" (by default), "spearman" or "kendall"

CONFIDENCE INTERVALS AND HYPOTHESIS TESTING

Functions to compute confidence intervals (CI) or conduct hypothesis testing (HT), have the form:

<name>.test(<data>,
alternative = <>, # can be "two.sided", "less" or "greater"
conf.level = 0.95,
mu = <>) # depends on the parameter in study;
# when set, this it becomes the Null Hypothesis

Test function attributes:

$method: type of test
$statistic: test statistic
$p.value: p-value
$parameter: DoF
$observed: observed counts
$expected: expected counts under the Null Hypothesis

Example:

test <- t.test(x, alternative = "two.sided", conf.level = 0.95, mu = 0)

test$p.value

## [1] 1.113842e-43

One Sample CI/HT:

1. Mean (σ unknown):

t.test(x, alternative = "two.sided", conf.level = 0.95) # CI

t.test(x, alternative = "two.sided", conf.level = 0.95, mu = 0) # HT
# x is a vector of observations

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 6

2. Probability p of a Bernoulli trial

binom.test(x, n, alternative = "two.sided", conf.level = 0.95) # CI

binom.test(x, n, alternative = "two.sided", conf.level = 0.95, p=0.5) # HT
# x is number of successes, n is number of trials

3. The rate parameter λ (i.e. mean) of a Poisson distribution:

poisson.test(x, alternative = "two.sided", conf.level = 0.95) # CI

poisson.test(x, alternative = "two.sided", conf.level = 0.95, r = 1) # HT
# x is number of events in one time base, r is the event rate.

Example: Output from calling the t.test function on a normal sample with a probable µ = 2

x <- rnorm(50, 20, 3)

t.test(x, alternative = "two.sided", conf.level = 0.95, mu = 2)

##
## One Sample t-test
##
## data: x
## t = 44.174, df = 49, p-value < 2.2e-16
## alternative hypothesis: true mean is not equal to 2
## 95 percent confidence interval:
## 19.06296 20.68939
## sample estimates:
## mean of x
## 19.87618

Two Sample CI/HT:

1. Difference of means (σ unknown):

t.test(x1, x2, alternative = "two.sided", var.equal = F, conf = 0.95) # CI

t.test(x1, x2, alternative = "two.sided", var.equal = F, conf = 0.95, mu = 0) # HT
# mu is the difference of the means; x1 and x2 are vectors with data observations

2. Difference of means (paired data; e.g. before-and-after):

t.test(x1, x0, alternative = "two.sided", conf = 0.95, paired = T) # CI

t.test(x1, x0, alternative = "two.sided", conf = 0.95, paired = T, mu = 0) # HT

3. Variance Ratio:

var.test(x1, x2, alternative = "two.sided", conf = 0.95) # CI

var.test(x1, x2, alternative = "two.sided", conf = 0.95, ratio = 1) # HT
# ratio is the ratio of the population variances; x1 and x2 are vectors with data observations

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 7

4. Difference of proportions (Bernoulli trials):

prop.test(x, n, alternative = "two.sided", conf.level = 0.95, correct = T) # CI

prop.test(x, n, alternative = "two.sided", conf.level = 0.95, correct = T, p) # HT
# x is a vector c(x1, x2) of the number of successes
# n is a vector c(n1, n2) of the number of trials
# p is a vector c(p1, p2) of the proportions to be HT

Chi-squared tests:
1. Goodness-of-fit (only HT makes sense; H0 is the hypothesis that the data is generated from a multinomial distribution
with the ps as probabilities):

chisq.test(x, p) # HT
# x is a vector of integers (the counts of each class)
# p is the probabilities of each class

Example: A man sitting in his balcony has spotted a number of birds belonging to one of six species.
The exact classification is given below:

Species 1 2 3 4 5 6
Frequency 8 10 13 9 11 14

Test at 5% level of significance whether or not the data are compatible with the assumption that the particular area is
visited by the same proportion of birds belonging to these six species in the population. (Goodness-of-fit test)
# Using the chi-squared test:
obs <- c(8, 10, 13, 9, 11, 14)
n <- length(obs)
ep <- rep(1/n, n) # a vector c(1/n, 1/n, ..., 1/n)
pval1 <- chisq.test(obs, p = ep)$p.value
pval1

## [1] 0.7799664

# Using First Principles:

ef <- ep * sum(obs)
ts <- sum((obs-ef)^2 / ef)
pval2 <- pchisq(ts, n-1, lower.tail = FALSE)
pval2

## [1] 0.7799664

Since the p-value is greater than 0.05, we have no reason to reject the hypothesis that the six species occur with the same
frequency.

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 8

LINEAR REGRESSION
Linear Regression Models:

X <- rnorm(50, 20, 2) # generate pseudorandom normal data for X and Y

Y <- 20 * X + 42 + rnorm(50, 20, 2)

model1 <- lm(Y ~ X) # Y = a + Bx

model2 <- lm(Y ~ X - 1) # Y = Bx
model3 <- lm(Y ~ 1) # Y = a

summary(model1)

##
## Call:
## lm(formula = Y ~ X)
##
## Residuals:
## Min 1Q Median 3Q Max
## -4.5506 -0.8937 -0.1579 0.9268 3.1274
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 59.5070 2.1692 27.43 <2e-16 ***
## X 20.1292 0.1045 192.53 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.639 on 48 degrees of freedom
## Multiple R-squared: 0.9987,Adjusted R-squared: 0.9987
## F-statistic: 3.707e+04 on 1 and 48 DF, p-value: < 2.2e-16

Since the p-value is less than 5%, we should reject H0 (that all coefficients except the intercept are zero).
The coefficients could have also been obtained via model1$coef or coef(model1).

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 9

Using models for prediction and Confidence Interval for Mean and Individual Response:

newdata <- data.frame(X=30)

predict(model1, newdata) # point estimate of the response

## 1
## 663.3824

predict(model1, newdata, interval = "confidence", level = 0.95) # mean response

## fit lwr upr

## 1 663.3824 661.3582 665.4067

predict(model1, newdata, interval = "predict", level = 0.95) # individual response

## fit lwr upr

## 1 663.3824 659.5145 667.2504

Checking the residuals:

layout(matrix(c(1, 2), 1, 2))

plot(model1, 1)
plot(model1, 2)
Standardized residuals

Residuals vs Fitted Normal Q−Q

0 2 4

20 20
Residuals

1
−1

47 47
−4

19
−3

19

400 450 500 550 −2 −1 0 1 2

Fitted values Theoretical Quantiles

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 10

ANOVA Tables and Confidence intervals for parameters:

size <- c(3, 4, 5, 6, 4, 5, 6, 7, 7, 8, 9, 10)

pop <- c("A", "A", "A", "A", "B", "B", "B", "B", "C", "C", "C", "C")

model <- lm(size ~ pop)

anv <- anova(model)
anv

## Analysis of Variance Table

##
## Response: size
## Df Sum Sq Mean Sq F value Pr(>F)
## pop 2 34.667 17.3333 10.4 0.004572 **
## Residuals 9 15.000 1.6667
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

anv[1, 2] # individual table entries are available using this syntax:

## [1] 34.66667

confint(model, level = 0.95) # confidence intervals for the parameters

## 2.5 % 97.5 %
## (Intercept) 3.039784 5.960216
## popB -1.065058 3.065058
## popC 1.934942 6.065058

MULTIPLE LINEAR REGRESSION

Regression Models:

Form Model
model <- lm(Y ~ X1 + X2 + X3) Y = a + b1 · X1 + b2 · X2 + b3 · X3
model <- lm(Y ~ X1 * X2) Y = a + b1 · X1 + b2 · X2 + b12 · X1 · X2
model <- lm(Y ~ X1 + X2 + X1:X2) Y = a + b1 · X1 + b2 · X2 + b12 · X1 · X2
model <- lm(Y ~ I(X1^2)) Y = a + b1 · X12 ; the I(<arg>) makes <arg> “literal”

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 11

# generating a fake dataset

set.seed(10)
X1 <- rnorm(100, 20, 2)
X2 <- rnorm(100, 10, 4)
Y <- 42 + 20 * X1 + rnorm(100, 0, 4)

model1 <- lm(Y ~ X1 + X2)

summary(model1)

##
## Call:
## lm(formula = Y ~ X1 + X2)
##
## Residuals:
## Min 1Q Median 3Q Max
## -10.6062 -2.2655 0.2998 2.6385 9.4370
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 43.4326 4.2955 10.11 <2e-16 ***
## X1 19.9993 0.2083 96.00 <2e-16 ***
## X2 -0.1355 0.1011 -1.34 0.183
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 3.895 on 97 degrees of freedom
## Multiple R-squared: 0.9896,Adjusted R-squared: 0.9894
## F-statistic: 4632 on 2 and 97 DF, p-value: < 2.2e-16

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 12

ANOVA Table:

anova(model1)

## Analysis of Variance Table

##
## Response: Y
## Df Sum Sq Mean Sq F value Pr(>F)
## X1 1 140548 140548 9262.0579 <2e-16 ***
## X2 1 27 27 1.7951 0.1834
## Residuals 97 1472 15
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Confidence Interval of parameters:

confint(model1, c(1, 2, 3), level = 0.95)

## 2.5 % 97.5 %
## (Intercept) 34.9071989 51.95802047
## X1 19.5858137 20.41275538
## X2 -0.3362104 0.06521903

Individual Response and Mean Response:

newdata <- data.frame(X1 = 0.03, X2 = 0.02)

predict(model1, newdata, interval = "confidence") # For mean response

## fit lwr upr

## 1 44.02988 35.51763 52.54213

predict(model1, newdata, interval = "predict") # For individual response

## fit lwr upr

## 1 44.02988 32.53062 55.52914

Checking the fit of the model (Residuals):

# resid(model1) # this function yields the residuals

layout(matrix(c(1, 2), 1, 2))
plot(model1, 1) # residuals vs fitted values
plot(model1, 2) # Q-Q plot of residuals

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 13

Standardized residuals
Residuals vs Fitted Normal Q−Q

10
75 80 75 80

2
Residuals

0
−10

−2
68 68

350 400 450 500 −2 −1 0 1 2

Fitted values Theoretical Quantiles

GENERALISED LINEAR MODELS (GLMS)

GLM: glm(<formula>, family = <family>(link = <link_function>))

Family Default link function Other link functions

gaussian identity log, inverse
Gamma inverse log, identity
binomial logit log, probit
poisson log identity, sqrt

# generate fake date

set.seed(10)
n <- 1000
age <- runif(n, min = 18, max = 70) # Age between 18 and 70
vehicle_type <- rbinom(n, size = 1, prob = 0.4) # 40% SUVs
b <- list(X0 = -2, X1 = 0.05, X2 = 0.7)
lambda <- exp(b[['X0']] + b[['X1']] * age + b[['X2']] * vehicle_type)
Y <- rpois(n, lambda)

# create a GLM model

DF <- data.frame(Y, age, vehicle_type)
g <- glm(formula = Y ~ age + vehicle_type, data = DF, family = poisson())
g

##
## Call: glm(formula = Y ~ age + vehicle_type, family = poisson(), data = DF)
##
## Coefficients:
## (Intercept) age vehicle_type
## -2.00875 0.05106 0.68335
##
## Degrees of Freedom: 999 Total (i.e. Null); 997 Residual
## Null Deviance: 2583
## Residual Deviance: 1115 AIC: 3228

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 14

ANOVA Table and Confidence Intervals of Parameters and Model Predictions:

# g is the GLM above!
anova(g)

## Analysis of Deviance Table

##
## Model: poisson, link: log
##
## Response: Y
##
## Terms added sequentially (first to last)
##
##
## Df Deviance Resid. Df Resid. Dev
## NULL 999 2582.6
## age 1 1193.09 998 1389.5
## vehicle_type 1 274.39 997 1115.1

confint(g, par = c(1, 2, 3), level = 0.90) # 90% confidence interval for all 3 parameters

## Waiting for profiling to be done...

## 5 % 95 %
## (Intercept) -2.16265206 -1.85726127
## age 0.04844901 0.05368993
## vehicle_type 0.61529167 0.75160250

new.data <- data.frame(age = c(30, 30), vehicle_type = c(1, 0))

predict(g, new.data)

## 1 2
## 0.2062803 -0.4770718

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 15

Checking the Model Fit:

layout(matrix(c(1, 2), 1, 2))

plot(g, 1) #residuals vs fitted values
plot(g, 2) #Q-Q plot of residuals

Residuals vs Fitted Normal Q−Q

Std. Pearson resid.

6
2 4 6

819 819
Residuals

514 514
259

4
259

2
−2

−2
−1.0 0.0 0.5 1.0 1.5 2.0 −3 −2 −1 0 1 2 3

Predicted values Theoretical Quantiles

Model Comparison:

g1 <- glm(formula = Y ~ age, data = DF, family = poisson())

g1$aic

## [1] 3500.063

g$aic

## [1] 3227.675

# model g is better because AIC(g) < AIC(g1)!

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 16

BAYESIAN STATISTICS AND CREDIBILITY THEORY

Poisson-Gamma Model: If X ∼ Poisson(λ) is the distribution of the number of annual claims, then under the Gamma-
Poisson model – having a λ ∼ Gamma(α, β) prior – yields a posterior λ|X ∼ Gamma( xi + α, n + β) on λ. In this case,
P

the posterior mean is: P

xi + α
λ̂B =
n+β
set.seed(100)
pm <- rep(0, 1000)
for (i in 1:1000) {
lambda <- rgamma(1, shape = 9, rate = 4)
x <- rpois(n = 20, lambda)
pm[i] <- (sum(x) + 9) / (20 + 4)
}
mean(pm)

## [1] 2.208667

Credibility Theory: In this case,

λ̂Credibility = Z · λ̂Individual + (1 − Z) · λ̂Prior

where: λ̂Prior = β, the expected value of the prior, as in the example above; λ̂Individual = and Z = is the
P
α xi n
n n+β

Credibility Factor.

set.seed(100)
n <- 20
Z <- n / (n + 4)

cp <- rep(0, 1000)

for (i in 1:1000) {
lambda <- rgamma(1,shape = 9, rate = 4)
x <- rpois(n = n, lambda)
cp[i] <- Z * mean(x) + (1-Z) * 9 / 4
}
mean(cp)

## [1] 2.208667

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 17

EMPIRICAL BAYES CREDIBILITY THEORY (EBCT)

Model 1: For the claims dataset below (5 entities, 6 years), determine the future premium.

data <- data.frame(

X2015 = c(1350, 1000, 2050, 4670, 7150),
X2016 = c(1080, 3080, 4550, 4270, 3480),
X2017 = c(1800, 1700, 1900, 4800, 4900),
X2018 = c(2040, 2820, 1600, 9070, 4810),
X2019 = c(1000, 5760, 4200, 3770, 8800),
X2020 = c(1180, 3670, 2650, 5250, 7000)
)
rownames(data) <- c("A", "B", "C", "D", "E")

n <- ncol(data) # Number of time periods (e.g., years)

entity_mean_claims <- rowMeans(data) # Mean claim for each entity
entity_claim_variances <- apply(data, 1, var) # Variance of claims for each entity
overall_mean_claim <- mean(entity_mean_claims) # Overall mean claim amount
variance_of_entity_means <- var(entity_mean_claims) # Variance of entity mean claims
average_entity_variance <- mean(entity_claim_variances) # Average variance across entities
# Adjusting for within-entity variance
var.theta <- variance_of_entity_means - (average_entity_variance / n)
Z <- n / (n + average_entity_variance / var.theta)
# Credibility-adjusted estimate: Z * m_i + (1 - Z) * m
premiums <- Z * entity_mean_claims + (1 - Z) * overall_mean_claim
premiums

## A B C D E
## 1662.853 3083.215 2923.090 5129.247 5768.262
5000

Entity Mean
Future Premium
Value

2000

A B C D E

Entity

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 18

Model 2: For this model, include a second dataframe consisting of policy volume.

volume <- data.frame(

X2015 = c(7, 9, 4, 6, 4),
X2016 = c(7, 11, 5, 6, 8),
X2017 = c(8, 3, 6, 7, 8),
X2018 = c(7, 11, 3, 8, 8),
X2019 = c(13, 8, 3, 6, 9),
X2020 = c(7, 12, 2, 13, 10)
)
rownames(volume) <- c("A", "B", "C", "D", "E")

n_i <- rowSums(volume) # Total exposure for each group (n_i)

m_i <- rowSums(data) / n_i # Observed claim frequency per group (m_i)
total_claims <- sum(data) # Total claims across all groups
total_exposure <- sum(volume) # Total exposures across all groups
m <- total_claims / total_exposure # Overall average claim frequency (m)
P_ij <- data / volume # Claim frequencies per period and group
s2i <- numeric(nrow(data)) # Initialize vector for s_i^2
for (i in 1:nrow(data)) {
exposures_i <- as.numeric(volume[i, ])
P_ij_i <- as.numeric(P_ij[i, ])
squared_deviations <- exposures_i * (P_ij_i - m_i[i])^2
s2i[i] <- sum(squared_deviations) / n_i[i]
}

Es <- mean(s2i) # Average within-group variance (E[s(theta)])

V_m_theta <- sum(n_i * (m_i - m)^2) / sum(n_i) # Variance of m_i's (V[m(theta)])
K <- Es / V_m_theta # K = E[s(theta)] / V[m(theta)]
Z_i <- n_i / (n_i + K) # Credibility factor per group (Z_i)
hat_P_i <- Z_i * m_i + (1 - Z_i) * m # Credibility-adjusted claim frequency
hat_P_i

## A B C D E
## 180.9165 337.8925 725.0477 687.0478 762.1102

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 19

PRINCIPAL COMPONENT ANALYSIS (PCA)

Computing PCA:
Example: From First Principles.

set.seed(10)
df <- rnorm(1000, c(2, 10), c(1, 2))
data <- matrix(df, 500, 2)

X <- scale(data, center = T, scale = F) # scale the data

A <- t(X) %*% X # compute the unnormalized covariance matrix
W <- eigen(A)$vectors # the principal components of X
V <- eigen(A)$values # the squares of the singular values of X

explained.var <- V / sum(V) # the % explained variance of each variable

Example: Using the built-in prcomp(<>) function.

# data and df are the variables in the example above
pca <- prcomp(data)

pca$rotation # the principal components of X

vars <- pca$sdev^2

explained.var <- vars / sum(vars) # the % explained variance of each variable

plot(data) # doesn't plot the principal components!

data[,2]

10
5
0

0 5 10 15

data[,1]

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 20

BOOTSTRAP INFERENCE
Parametric bootstrap:
Example: X is a sample of size 300 from a normal distribution N (10, 52 ). Let’s compute the bootstrap empirical
distribution of X̄:

x <- rnorm(300, 10, 5)

N <- 300
# the replicate(<n>, <expr>) function generates a vector by evaluating <expr> <n> times.
bm <- replicate(10000, mean(rnorm(N, mean = mean(x), sd = sd(x))))

# The 95% CI of the sample mean is:

quantile(bm, c(0.025, 0.975))

## 2.5% 97.5%
## 9.744772 10.855255

Non-parametric bootstrap:
Example: X is now a sample of size 300 from an unknown distribution; let’s compute the bootstrap empirical distribution
of X̄:

bm <- replicate(10000, mean(sample(x, replace = T)))

# The 95% CI of the sample mean is:

quantile(bm, c(0.025, 0.975))

## 2.5% 97.5%
## 9.731208 10.866606

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 21

NON-PARAMETRIC PERMUTATION TESTS

Test whether two sample means are equal using a Random permutation test:
Example: X1 and X2 are samples: we’re interested in finding out whether or not the they come from the same population
via Random Permutation Testing the sample means differences. In this case, we don’t look at all permutations of indices.

X1 <- rnorm(100, 10, 5) # generate a normal sample X1

X2 <- rnorm(200, 5, 5) # generate a normal sample X2

n1 <- length(X1) # n1 is the size of X1

result <- c(X1, X2) # concatenate the two samples together
index <- 1:length(result) # generate an index 1, 2, ..., n1 + n2
diff <- rep(0, 10000) # initialize 10k permutations

for (i in 1:10000) {
p <- sample(index, n1, replace = F) # sample n1 indexes out of the n1 + n2
diff[i] <- mean(result[p]) - mean(result[-p]) # compute this sample means differences
}
obs <- abs(mean(X1) - mean(X2)) # sample means differences
length(diff[abs(diff) >= obs]) / length(diff) # probability that a permutation sample has
# greater diff than the observed one(i.e. a p-value!)

Permutation test: Example: X1 and X2 are samples: we’re interested in finding out whether or not the they come
from the same population via Permutation Testing the sample means differences.
X1 <- rnorm(10, 10, 5) # generate a normal sample X1; needs to be very small
X2 <- rnorm(10, 2, 5) # generate a normal sample X2; needs to be very small

n1 <- length(X1) # n1 is the size of X1

result <- c(X1, X2) # concatenate the two samples together
index <- 1:length(result) # generate an index 1, 2, ..., n1 + n2

p <- combn(index, n1) # generate all combinations; explodes with n1 and n2!
n <- choose(n1 + n2, n1) # combinations n1 + n2 choose n1
diff <- rep(0, n) # initialize all combinations
for (i in 1:n) {
diff[i] <- mean(result[p[,i]]) - mean(result[-p[,i]]) # compute this sample means differences
}
#P-value
obs <- abs(mean(X1) - mean(X2)) # sample means differences
length(diff[abs(diff) >= obs]) / length(diff) # probability that a permutation sample has
# greater diff than the observed one(i.e. a p-value!)

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 22

CS2 - RISK MODELLING AND SURVIVAL ANALYSIS

CREATING FUNCTIONS
Syntax:

function.name <- function(arg1, arg2, ...) {

<expressions>
<last expression> # this one is returned by the function
}

Example: Creating pdf, cdf and random sampling function for Pareto distribution X ∼ Pareto(a, l).

dpareto <- function(x, a, l) { # pdf

(a * l^a) / ((l+x)^(a+1))
}

ppareto <- function(x, a, l) { # cdf

1 - (l / (l+x))^a
}

rpareto <- function(n, a, l) { # random sampler

u <- runif(n)
l * ((1-u)^(-1/a) - 1)
}

DIFFERENTIATION AND INTEGRATION

Symbolic Differentiation:
Example: Differentiate the function f (x) = x2 + 3x + 5.

expr <- expression(x^2 + 3*x + 5) # Define the function

D(expr, "x") # Differentiate the expression with respect to x

## 2 * x + 3

Numerical Integration:
Example: Estimate 0 2x + 3dx
R1

f <- function(x) { 2*x + 3 } # Define the function

integrate(f, lower = 0, upper = 1) # Integrate the function from 0 to 1

## 4 with absolute error < 4.4e-14

integrate(f, lower = 0, upper = 1)$value # Yields the value only

## [1] 4

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 23

LOSS DISTRIBUTION
Generating Maximum Likelihood Estimates (MLE) and analysing the fit:
Example: Use MLE to estimate the parameters of a normally distributed sample:

LogL <- function(x, params) {

if (params[2] <= 0) { # sigma can never be negative;
Inf # make the log-likelihood infinite if it is!
} else {
-sum(log(dnorm(x, params[1], params[2]))) # the log-likelihood function for the sample X
} # and params (mu, sigma)
}

sample <- rnorm(200, 10, 5)

initial.params = c(1, 100)
MLE <- nlm(LogL, initial.params, x = sample)
MLE$estimate # the estimate should be ~ the true parameters!

## [1] 10.353933 4.844735

hist(sample, freq = F, breaks = 10)

x <- seq(-20, 20, 0.05)
y <- dnorm(x, MLE$estimate[1], MLE$estimate[2])
lines(x, y, col = "darkblue")

Histogram of sample
0.08
Density

0.04
0.00

−5 0 5 10 15 20 25

sample

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 24

MLE using the MASS package:

Example: Use the fitdistr(<data>,<distribution>) function to compute the MLE of parameters of a Gamma-
distributed sample.

library(MASS) # install.packages("MASS") if needed!

x <- rgamma(2000, 20, 0.5) # generate a sample from a Gamma distribution

fitdistr(x, "gamma")

## shape rate
## 19.98918983 0.50248312
## ( 0.62686916) ( 0.01595715)

# fitdistr(x, "gamma", list(shape = 20, rate = 0.5)) # if an initial candidate exists!

Using Q-Q plots:

Example: Generate estimates for the parameters using the Method of Moments and then check with a Q-Q plot.

ozone <- airquality$Ozone # airquality is a built-in dataset included in R

ozone <- ozone[!is.na(ozone)]

a <- mean(ozone)^2 / sd(ozone)^2 # Estimating a and l using method of moments

l <- a / mean(ozone)

X <- rgamma(2000, a, l)
qqplot(X, ozone, ylab = "Sample quantile", ylim = c(0,200))
abline(0, 1, col = "blue")
200
Sample quantile

100
0

0 50 100 150 200 250 300

sx

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 25

REINSURANCE
Proportional Reinsurance:
Example: Suppose Insurance claims follow a distribution LogN(5, 2) and the value of α is 75% (that is, the reinsurer pays
25% of the claims). How does this affect the payments?

set.seed(50)
claim <- rlnorm(1000, 5, 2)
a <- 0.75

paid.insurance <- a * claim

paid.reinsurer <- claim - paid.insurance
(mean(paid.insurance) - mean(claim)) / mean(claim) * 100 # % decrease in mean

## [1] -25

(sd(paid.insurance) - sd(claim)) / sd(claim) * 100 # % decrease in std. deviation

## [1] -25

Excess of Loss Reinsurance:

Example: Assume the claims are distributed as Pareto(3, 600) and the retention limit M = 1000. What do the insurer
payments look like?

set.seed(10)
rpareto <- function(n, a, l) {
l * ((1 - runif(n))^(-1/a) - 1)
}
claim <- rpareto(10000, 3, 600)

m <- 1000
paid.insurance <- pmin(claim, m)
paid.reinsurer <- pmax(0, claim - m)
(mean(paid.insurance) - mean(claim)) / mean(claim) * 100

## [1] -15.14916

(sd(paid.insurance) - sd(claim)) / sd(claim) * 100

## [1] -48.57061

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 26

Inflation Case:
Example: Assume an inflation rate of 12% per annum. Assess the impact of inflation on the mean and standard deviation
of claims paid by insurance and reinsurer, given that the retention limit M = 1000.

set.seed(10)
rpareto <- function(n, a, l) {
l * ((1 - runif(n))^(-1/a) - 1)
}
claim <- rpareto(10000, 3, 600)

m <- 1000
paid.insurance <- pmin(claim, m)
paid.reinsurer <- pmax(claim - m, 0)

i <- 0.12
k <- (1+i)
paid.insurance.i <- pmin(k * claim, m)
paid.reinsurer.i <- pmax(k*claim - m, 0)

# insurer plays under inflation rate!

(mean(paid.insurance.i) - mean(paid.insurance)) / mean(paid.insurance) * 100

## [1] 9.299076

# reinsurer plays above inflation rate!

(mean(paid.reinsurer.i) - mean(paid.reinsurer)) / mean(paid.reinsurer) * 100

## [1] 27.12794

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 27

RISK MODELS
Collective Risk Models:
Example: Estimate the distribution of

S = X1 + X2 + X3 + · · · + XN , N ∼ Poisson(900), Xi ∼ Gamma(680, 0.23)

set.seed(10)

s <- rep(0, 10000)

for (i in 1:10000) {
n <- rpois(1, lambda = 900)
s[i] <- sum(rgamma(n, shape = 680, rate = 0.23))
}
cat("Mean:", mean(s), "| Variance:", var(s))

## Mean: 2661070 | Variance: 7768288482

length(s[s > 2500000]) / length(s)

## [1] 0.967

Example: Estimate the distribution of S, but with a retention limit of 2 100.

si <- rep(0, 10000)

sr <- rep(0, 10000)
m <- 2100
for (i in 1:10000) {
n <- rpois(1, lambda = 900)
x <- (rgamma(n, shape = 600, rate = 0.23))
si[i] <- sum(pmin(x, m))
sr[i] <- sum(pmax(0, x - m))
}
cat("Mean:", mean(si), "| Variance:", var(si))

## Mean: 1891784 | Variance: 4046519806

cat("Mean:", mean(sr), "| Variance:", var(sr))

## Mean: 458234.3 | Variance: 248071912

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 28

Individual Risk Model:

Example: Company XYZ has sold 1200 policies; 300 to smokers (qS = 0.002) and 900 to non-smokers (qN S = 0.001).

S = Y1 + Y2 + · · · + Y300 + Z1 + · · · + Z900 , Yi ∼ Gamma(220, 0.75), Zi ∼ Gamma(310, 0.45)

set.seed(10)
S <- rep(0, 10000)
for (j in 1:10000) {
Y <- rgamma(300, shape = 220, rate = 0.75)
Z <- rgamma(900, shape = 310, rate = 0.45)
T <- c(Y, Z)
DY <- rbinom(300, 1, 0.002)
DZ <- rbinom(900, 1, 0.001)
D <- c(DY, DZ)
S[j] <- sum(T * D)
}

p.700 <- length(S[S > 700]) / length(S)

cat("Mean:", mean(S), "| Variance:", var(S), "| P(S > 700)", p.700)

## Mean: 795.0162 | Variance: 475206.3

hist(S)

Histogram of S
3500
Frequency

1500
0

0 1000 2000 3000 4000 5000

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 29

SURVIVAL MODELS
Creating a function for Makeham’s Law:
Example: Assuming Makeham’s Law, µ(x) = A + B ∗ C x , we will compute the survival
 Z t 
t px = Sx (t) = P [Tx > t] = exp − µ(x + s)ds
0

first using first principles, then numerically. After that we compute the expectation e̊x and the curtate expectation ex .

p <- list(A = 0.00004, B = 0.0000025, C = 1.1) # store the model parameters in a list
mu <- function(x) {
p$A + p$B * p$C^x
}

# Let's compute 10p35

tpx1 <- function(x, t) {
int <- p$A * t + p$B * (p$C^t - 1) * (p$C^x) / log(p$C)
exp(-int)
}

tpx2 <- function(x, t) {

i <- integrate(mu, x, x + t)$value
exp(-i)
}

tpx1(35, 10)

## [1] 0.9984264

tpx2(35, 10)

## [1] 0.9984264

integrate(tpx1, 0, Inf, x = 50)$value # E[mu]

## [1] 54.75777

sum(tpx1(50, 1:1000)) # the curtate expectation at 50 years old

## [1] 54.2578

The flexsurv library: The (automatically loaded) stats package does not include functions to deal with the Gompertz
distribution. That’s where the flexsurv function comes in.

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 30

library(flexsurv)

rate <- 0.0000025 # beta (rate parameter)

C <- 1.23 # C = exp(alpha)
shape <- log(C) # alpha (shape parameter)

p <- list(A = 0, B = rate, C = C)

x <- 30
t <- 25
Sxt <- pgompertz(x+t, shape, rate, lower.tail = FALSE) # S(x + t)
Sx <- pgompertz(x, shape, rate, lower.tail = FALSE) # S(x)
tpx <- Sxt / Sx
tpx

## [1] 0.3473415

# Simulate 10 lifetimes of lives currently aged 50

set.seed(42)
rgompertz(10, shape, rate * C^50)

## [1] 9.745937 10.235528 3.080972 8.405690 6.342800 5.206165 7.299388

## [8] 1.566020 6.490136 6.969379

# Hazard rate at age 60

hgompertz(30, shape, rate) # the same as rate * exp(shape * 30)

## [1] 0.001244782

# Plot the hazard rate over the age range 30 to 100

x <- seq(30, 100)
y <- hgompertz(x - 30, shape, rate)
plot(x, y, type = "l", xlab = "Ages", ylab = "Hazard rate")
Hazard rate

4
2
0

30 40 50 60 70 80 90 100

Ages

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 31

1 - exp(-Hgompertz(50, shape, rate))

## [1] 0.3145845

pgompertz(50, shape, rate)

## [1] 0.3145845

mu <- function(x) {
p$B * p$C^x
}

mu(55)

## [1] 0.2201512

hgompertz(55, shape, rate)

## [1] 0.2201512

integrate(mu, 0, 45)$value

## [1] 0.1341624

Hgompertz(45, shape, rate)

## [1] 0.1341624

Survival Distributions:
1. Exponential distribution: Tx ∼ Exp(0.075)
# P(Tx <= 30)
lambda <- 0.075
pexp(30, lambda)

## [1] 0.8946008

2. Weibull Distribution: Tx ∼ Weibull(0.0002, 2.5)

shape <- 2.5
scale <- 0.0002^(-1 / shape)

# P(Tx <= 30)

pweibull(30, shape, scale)

## [1] 0.6268969

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 32

PROPORTIONAL HAZARDS MODELS

Example: Estimating β from First Principles – according to the Cox Proportional Hazards model.
Consider a study where patients either die within 5 years after surgery or survive beyond 5 years (censored). Surgeries
took place at Hospital A or Hospital B; we want to determine if the hospital affects survival – in particular the relative
hazard of surgery at Hospital A compared to Hospital B (in this case it’s exp(β)).

data <- data.frame(X = c(0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1), # Covariate: 1 if Hospital A, 0 if B

T = c(60, 32, 60, 60, 60, 4, 18, 60, 9, 31, 53, 17), # Time of death in months
C = c(0, 1, 0, 0, 0, 1, 1, 0, 1, 1, 1, 1)) # 0 if censored, 1 if death

neg_log_partial_likelihood <- function(beta, df) {

events <- which(df$C == 1)
event_times <- T[events]
sorted_event_indices <- events[order(event_times)]

log_PL <- 0
for (i in sorted_event_indices) {
risk_set <- which(df$T >= df$T[i])
new.term <- beta * df$X[i] - log(sum(exp(beta * df$X[risk_set])))
log_PL <- log_PL + new.term
}

-log_PL
}

mle <- nlm(neg_log_partial_likelihood, 0, df = data)

beta_hat <- mle$estimate

cat("Estimated beta (regression coefficient):", beta_hat, "\n")

## Estimated beta (regression coefficient): 2.118368

cat("Hazard ratio (exp(beta)):", exp(beta_hat), "\n")

## Hazard ratio (exp(beta)): 8.317556

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 33

Using the survival Library:

Example: Estimating β from the Cox PH model using the survival library.

library(survival)
data <- data.frame(X = c(0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1), # Covariate: 1 if Hospital A, 0 if B
T = c(60, 32, 60, 60, 60, 4, 18, 60, 9, 31, 53, 17), # Time of death in months
C = c(0, 1, 0, 0, 0, 1, 1, 0, 1, 1, 1, 1)) # 0 if censored, 1 if death

surv_object <- Surv(time = data$T, event = data$C)

cox_model <- coxph(surv_object ~ X, data = data)

summary(cox_model)

## Call:
## coxph(formula = surv_object ~ X, data = data)
##
## n= 12, number of events= 7
##
## coef exp(coef) se(coef) z Pr(>|z|)
## X 2.118 8.318 1.092 1.939 0.0525 .
## ---
## Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
##
## exp(coef) exp(-coef) lower .95 upper .95
## X 8.318 0.1202 0.9776 70.77
##
## Concordance= 0.741 (se = 0.068 )
## Likelihood ratio test= 5.54 on 1 df, p=0.02
## Wald test = 3.76 on 1 df, p=0.05

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 34

ESTIMATING THE LIFETIME DISTRIBUTION FUNCTION

Kaplan-Meier:

data <- data.frame(tj = c( 4, 5, 8, 10, 11, 15),

dj = c( 2, 2, 1, 3, 1, 0),
nj = c(100, 98, 96, 90, 87, 86))

survival.km <- cumprod(1 - data$dj / data$nj)

plot(data$tj, survival.km, type = "s",

xlim = c(0, 25), ylim = c(0.85, 1),
xlab = "Time (t)", ylab = "Survival Probabilities",
main = "Kaplan-Meier Estimator")

Nelson-Aalen:
# data is the same as in the previous example!

lambda <- cumsum(data$dj / data$nj)

survival.na <- exp(-lambda)

plot(data$tj, survival.na, type = "s",

xlim = c(0, 25), ylim = c(0.85, 1),
xlab = "Time (t)", ylab = "Survival Probabilities",
main = "Nelson-Aalen Estimator")

cbind(data$tj, survival.km, survival.na) # Kaplan-Meyers < Nelson-Aalen!

Kaplan−Meier Estimator Nelson−Aalen Estimator

Survival Probabilities

Survival Probabilities
0.95

0.95
0.85

0.85

0 5 10 15 20 25 0 5 10 15 20 25

Time (t) Time (t)

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 35

Using the Survival library:

library(survival)

# generate a synthetic dataset

set.seed(10)
n <- 200
event_time <- rweibull(n, shape = 2, scale = 10)
censor_time <- rexp(n, rate = 0.5)
observed_time <- pmin(event_time, censor_time)
status <- ifelse(event_time <= censor_time, 1, 0)

# Create the survival data frame

surv.data <- data.frame(Patient = 1:n, event.time = observed_time, event.code = status)

#Creating a Survival Object

surv.obj <- Surv(surv.data$event.time, surv.data$event.code)

#Generating estimates
survival.km <- survfit(surv.obj ~ 1, conf.type = "plain")
survival.na <- survfit(surv.obj ~ 1, conf.type = "plain", stype = 2)

plot(survival.km, main = "Kaplan-Meier and Nelson-Aalen Survival functions",

xlab = "time", ylab = "S(t)", col = "red")
lines(survival.na, col = "blue")
legend("topright", legend = c("KM", "NA"), cex = 0.4, col = c("red", "blue"), lwd = 1)

Kaplan−Meier and Nelson−Aalen Survival functions

KM
NA
0.8
S(t)

0.4
0.0

0 2 4 6 8

time

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 36

MORTALITY RATES
Example: Compute µ70 for the insured population in this example, between 2013 and 2017. In this case, each person is
exposed while they are

• alive

• between 70 and 71 years of age

• within the study period (2013 and 2017)

• insured

# Creating a synthetic dataset

set.seed(10)
n <- 1000
data <- data.frame(ID = 1:n, BIRTH = runif(n, 1920, 1955),
DEATH = NA, ENTRY = runif(n, 2010, 2015))
data$DEATH <- ifelse(runif(n) < 0.8, data$BIRTH + runif(n, 70, 90), NA)

d.70 <- data$BIRTH + 70 # 70th birthday

d.71 <- data$BIRTH + 71 # 71st birthday
d.0 <- 2013 # investigation start date
d.1 <- 2017 # investigation end date
data$DEATH[is.na(data$DEATH)] <- d.1 # if someone dies after d1, consider d1

d.A <- pmax(d.70, d.0, data$ENTRY)

d.B <- pmin(d.71, d.1, data$DEATH)

exposure <- pmax(0, d.B - d.A) # ignore negative exposures!

ecx <- sum(exposure)

pop <- data$DEATH > d.0 & data$DEATH < d.1 # consider only deaths in the study period
age <- data$DEATH[pop] - data$BIRTH[pop] # age at death
d <- age[age >= 70 & age < 71] # age of death is 70 years old
mu70 <- length(d) / ecx
mu70 # deaths per person.year

## [1] 0.06773931

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 37

MARKOV CHAINS
Markov Chain library: Install (using install.packages(<>)) and load (using library(<>) the package markovchain
before running this code.

P <- matrix(c(0.25, 0.75, 0.2, 0.8), 2, 2, byrow = T) # transition probability matrix

states <- c("S1", "S2")
mc <- new("markovchain", transitionMatrix = P, states = states)

rowSums(mc@transitionMatrix) # notice the @ when accessing properties of mc!

is.irreducible(mc)

## [1] TRUE

period(mc)

## [1] 1

Example: Computing steady-state behaviour:

exp.dist <- function(start, P, n){

pos <- start
for (i in 1:n) {
pos <- pos %*% P # the operator %*% is matrix multiplication
}
pos
}

X0 <- c(0.1, 0.9)

exp.dist(X0, mc@transitionMatrix, 10) # equivalently, this package allows for the syntax: X0 * mc^10

## S1 S2
## [1,] 0.2105263 0.7894737

steadyStates(mc)

## S1 S2
## [1,] 0.2105263 0.7894737

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 38

Premium Calculations:
Example: Computing premiums using first principles.

discount <- c(0, 0.15)

prem <- 100 * (1 - discount)
P <- matrix(c(0.1, 0.9, 0.05, 0.95), 2, 2, byrow = T)
P

## [,1] [,2]
## [1,] 0.10 0.90
## [2,] 0.05 0.95

X0 <- c(0.5, 0.5)

prem.inc <- 0
for (i in 1:4) {
prem.inc <- prem.inc + sum(X0 * prem)
X0 <- X0 %*% P
}
prem.inc

## [1] 350.2216

Example: Computing premiums using the markovchain package.

discount <- c(0, 0.15)

prem <- 100 * (1 - discount)
P <- matrix(c(0.1, 0.9, 0.05, 0.95), 2, 2, byrow = T)

states <- c("0%", "25%")

mc <- new("markovchain", transitionMatrix = P, states = states)

X0 <- c(0.5, 0.5)

X0 %*% expectedRewards(mc, 4 - 1, prem)

## [,1]
## [1,] 350.2216

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 39

Sampling: Example: Sampling from a Markov Chain and fitting a Markov Chain from a sequence of states, using the
markovchain library.

P <- matrix(c(0.35, 0.65, 0.8, 0.2), 2, 2, byrow = T)

P

## [,1] [,2]
## [1,] 0.35 0.65
## [2,] 0.80 0.20

states <- c("0%", "20%")

mc <- new("markovchain", transitionMatrix = P, states = states)

#Sampling
#rmarkovchain(<n>,<mc>,t0="start state",include.t0=TRUE)
set.seed(10)
data <- rmarkovchain(5, mc, t0 = "20%", include.t0 = T) # generates a vector/sequence of states
data

## [1] "20%" "0%" "20%" "0%" "20%" "0%"

#Fitting the markov chain

data <- rmarkovchain(20, mc, t0 = "20%", include.t0 = T)
fit <- markovchainFit(data)
fit$estimate

## MLE Fit
## A 2 - dimensional discrete Markov Chain defined by the following states:
## 0%, 20%
## The transition matrix (by rows) is defined as follows:
## 0% 20%
## 0% 0.5384615 0.4615385
## 20% 0.8571429 0.1428571

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 40

Example: Simulating a time-inhomogeneous Markov chain

 
1 − 0.015t 0.015t
P (t) =  
1 − 0.02t 0.02t

prob_01 <- function(x){

0.015 * x
}

prob_11 <- function(x){

0.02 * x
}

px <- function(x) {
matrix(c((1 - prob_01(x)), prob_01(x), 1 - prob_11(x), prob_11(x)), 2, 2, byrow = T)
}

fx <- function(age, years, pos){

for (i in age:(age + years - 1)) {
pos <- pos %*% px(i)
}
pos
}

# starting from the first state at 35 years old, these are the state probabilities 10 years later:
fx(35, 10, c(1, 0))

## [,1] [,2]
## [1,] 0.1606768 0.8393232

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 41

MARKOV JUMP PROCESSES

Time-homogeneous case:
Example: In a continuous-time Markov Chain with generator matrix (λi s in year−1 )
 
−1 0.3 0.7
 
G = 0.8 −1.5
 
0.7 
 
0.6 1.2 −1.8

what’s the probability of going from state 2 to state 3 in a month?

states <- paste("State", 1:3)

GM <- matrix(c(-1, 0.3, 0.7, 0.8, -1.5, 0.7, 0.6, 1.2, -1.8),
3, 3, byrow = T, dimnames = list(states, states))

# only works for small time steps!!

h <- 1 / 12 # a month is 1 / 12 of a year
ph <- diag(3) + h * GM # diag(3) is the three-dimensional identity matrix
ph[2, 3]

## [1] 0.05833333

Example: Now we’re interested in the same probability, but over 30 years (we already computed it for one month!)
# ph and states as in the example above
library("markovchain")
markov_chain <- new("markovchain", transitionMatrix = ph, states = states)
markov_chain ^ (12 * 30)

## Unnamed Markov chain^360

## A 3 - dimensional discrete Markov Chain defined by the following states:
## State 1, State 2, State 3
## The transition matrix (by rows) is defined as follows:
## State 1 State 2 State 3
## State 1 0.4133333 0.3066667 0.28
## State 2 0.4133333 0.3066667 0.28
## State 3 0.4133333 0.3066667 0.28

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 42

Simulating a Markov jump process:

Example: Let’s use the approximation P ≈ In + h · G, valid for small h.

states <- paste("state",1:3)

GM <- matrix(c(-1, 0.3, 0.7, 0.8, -1.5, 0.7, 0.6, 1.2, -1.8),
3, 3, byrow = T, dimnames = list(states, states))

ph <- function(h, GM){

diag(nrow(GM)) + h * GM
}

library(markovchain)
set.seed(10)
sims <- rmarkovchain(1000, markov, include.t0 = T)
head(sims)

## [1] "state 3" "state 3" "state 3" "state 3" "state 3" "state 3"

Example: Now for an exact method:

# GM is as above
library(markovchain)
set.seed(10)

TPM <- generatorToTransitionMatrix(GM)

markov <- new("markovchain", transitionMatrix = TPM)
sims <- rmarkovchain(1000, markov, include.t0 = T)
head(sims)

## [1] "state 3" "state 2" "state 1" "state 3" "state 2" "state 1"

Example: Let’s simulate the waiting times for sims.

ljs = -diag(GM) # the diagonal values of the generator matrix are the
waiting_times <- rexp(1000, ljs[sims2]) # rate parameters of the waiting times (exp.) distributions.
head(waiting_time)

## [1] 0.2335325 0.8978444 0.4619815 0.3163581 0.2385942 0.9672260

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 43

Time-Inhomogeneous: Example: Simulating a time-inhomogenous Markov jump process. We use the approximation

P (t, t + h) ≈ I + h · GM(t)

and iterate it:

P (t, s) ≈ P (t, t + h) · · · P (t + h, t + 2h) · · · P (s − h, s)

states <- paste("State", 1:3)

GM <- function(t){
matrix(c(-t, 0.3*t, 0.7*t,
0.8*t, -1.5*t, 0.7*t,
0.6*t, 1.2*t, -1.8*t), 3, 3, byrow = T, dimnames = list(states, states))
}

Ph <- function(h, GM) {

diag(nrow(GM)) + h*GM
}

Pts <- function(t, s, h){

M <- diag(length(states))
dimnames(M) <- list(states, states)
n <- (s-t) / h - 1
for (i in 0:n) {
M <- M %*% Ph(h, GM(t + h*i))
}
M
}

Pts(1, 2, 1/10000)

## State 1 State 2 State 3

## State 1 0.4562438 0.2703372 0.2734190
## State 2 0.3890577 0.3375233 0.2734190
## State 3 0.3765769 0.3265005 0.2969225

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 44

TIME SERIES
Simulating a Stationary Time Series

yt = 0.6yt−1 − 0.2yt−2 + et + 0.3et−1 (ARMA(2, 1) et ∼ N (0, 52 )

| {z } | {z }
AR MA

set.seed(10)
yt <- arima.sim(n = 10, list(ar = c(0.6, -0.2), ma = 0.3), sd = 5)
yt

## Time Series:
## Start = 1
## End = 10
## Frequency = 1
## [1] 8.3867410 4.6225682 5.6760659 7.6692436 5.0251545 -3.1594544
## [7] -5.3088712 1.7814489 5.9338181 0.9469157

Simulating a non-stationary Time Series:

xt = xt−1 + et et ∼ N (0, 52 ) a random walk!

set.seed(50)
N <- 10
et <- rnorm(N, mean = 0, sd = 5)
xt <- cumsum(et)

Using the ACF to determine if the series needs differencing to become stationary:
N <- 100
et <- rnorm(n, mean = 0, sd = 5)
xt <- cumsum(et)
xt

acf(xt, main = "Sample ACF of data", ylab = "Sample ACF")

# Differencing is needed, since the decay in the ACF is slow.

Sample ACF of data

Sample ACF

0.8
0.4
0.0

0 5 10 15 20 25 30

Lag

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 45

Phillips-Perron test:

PP.test(xt) # H0 is the hypothesis that the series has a unit root (and therefore is not stationary).

##
## Phillips-Perron Unit Root Test
##
## data: xt
## Dickey-Fuller = -1.9202, Truncation lag parameter = 7, p-value = 0.6121

# since the p-value is greater than the significance level of 0.05, we don't reject H0.

# take differences:
zt <- diff(xt, lag = 1, differences = 1)
acf(zt, main = "Sample ACF of differenced series")

Sample ACF of differenced series

0.8
ACF

0.2
−0.4

0 5 10 15

Lag

PP.test(zt)

##
## Phillips-Perron Unit Root Test
##
## data: zt
## Dickey-Fuller = -7.2067, Truncation lag parameter = 3, p-value = 0.01

This time, H0 should be rejected; no further differencing seems to be needed. This is also visible in the ACF plot – the
decay of autocorrelations is very swift.
A heuristic to pick the order of differencing is to take the point where the variance of the series goes up!

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 46

ut <- xt
for (i in 0:4) {
print(paste0("order = ", i, ", var = ", var(ut))) # paste0 concatenates strings!
ut <- diff(ut, lag = 1, differences = 1)
}

## [1] "order = 0, var = 21044.8336811315"

## [1] "order = 1, var = 53.5795254513324"
## [1] "order = 2, var = 79.279157622838"
## [1] "order = 3, var = 173.982472108447"
## [1] "order = 4, var = 473.573641010363"

Identifying and removing trends using linear least-squares: Example: Identifying the trend
N <- 100
yt <- arima.sim(n = N, list(ar = c(0.6, -0.2), ma = 0.3), sd = 5)
xt <- 10 + 5 * (1:N) + yt
ts.plot(xt)
t <- seq(1:N)
trend <- lm(xt ~ t)
lines(t, trend$fitted.values, type = "l", col = "blue")
200 400
xt

0 20 40 60 80 100

Time

Example: Removing the trend

ts.plot(xt - trend$fitted.values)
xt − trend$fitted.values

10
0
−15

0 20 40 60 80 100

Time

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 47

Identifying and Removing Seasonality:

# ldeaths is a built-in time series data set
plot(ldeaths)
acf(ldeaths)
spectrum(ldeaths, main = "Periodogram", xlab = "Frequency")
# the 1-year seasonality is clear by visual inspection of all plots!

3000
ldeaths

1500

1974 1975 1976 1977 1978 1979 1980

Time

Series ldeaths Periodogram

5e+02 5e+04
spectrum
0.5
ACF

−0.5

0.0 0.5 1.0 1.5 0 1 2 3 4 5 6

Lag Frequency
bandwidth = 0.0481

# removing the seasonality

dif <- diff(ldeaths, lag = 12, differences = 1)
plot(ldeaths, main = "Original time series")
acf(ldeaths, main = "Sample ACF of original time series")
plot(dif, main = "Differenced Time Series")
acf(dif, main = "Sample ACF of Diff time series")

Original time series Sample ACF of original time series

3000

0.5
ldeaths

ACF

−0.5
1500

1974 1976 1978 1980 0.0 0.5 1.0 1.5

Time Lag

Differenced Time Series Sample ACF of Diff time series

1000

0.5
0

ACF
dif

−1500

−0.5

1975 1976 1977 1978 1979 1980 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4

Time Lag

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 48

Identifying and removing seasonality using built-in functions:

Example: Using decompose(<>)

d <- decompose(ldeaths, type = "additive")

plot(ldeaths)
lines(d$trend, col = "blue")
lines(d$seasonal, col = "green", lwd = 1.3)
lines(d$seasonal + d$trend, col = "red", lwd = 1.3)

Example: Using stl(<>)

d <- stl(ldeaths, s.window = "periodic")

plot(ldeaths)
lines(d$time.series[, "trend"], col = "blue")
lines(d$time.series[, "seasonal"] + d$time.series[, "trend"], col = "red", lwd = 1.3)

using decompose using stl

3000

3000
ldeaths

ldeaths
1500

1500

1974 1976 1978 1980 1974 1976 1978 1980

Time Time

Fitting a time series model: For MA(q) models, we determine the order by looking at the ACF and counting how many
lags it takes to ”cut off”, and confirming that the ACF ”tails off” (decays exponentially).
For AR(p) models, we determine the order by looking at the PACF and counting how many lags it takes to cut off, and
confirming that the ACF tails off.

xt <- arima.sim(list(order = c(3, 0, 0), # simulating an AR(3) process

ar = c(1, -1, 1/2)), n = 100)

plot(xt, main = "Time series")

acf(xt, main = "sample ACF") # the ACF tails off
pacf(xt, main = "sample PACF") # the PACF cuts off at lag 3

What model to fit sample ACF sample PACF

0.4
4

0.8

0.2
2

Partial ACF

0.0
0.4
ACF
0
xt

−0.2
0.0
−2

−0.4
−0.6
−4

−0.4

0 20 40 60 80 100 0 5 10 15 20 5 10 15 20

Time Lag Lag

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 49

Example: Fitting an ARIMA model (i.e. determining the coefficients): the lower the AIC the better!

xt <- arima.sim(list(order = c(0, 0, 3), ma = c(1, -1, 1/2)), n = 100)

arima(xt, order = c(0, 0, 3))

##
## Call:
## arima(x = xt, order = c(0, 0, 3))
##
## Coefficients:
## ma1 ma2 ma3 intercept
## -0.1770 -0.2840 0.1213 0.2499
## s.e. 0.1008 0.1186 0.1215 0.1402
##
## sigma^2 estimated as 4.453: log likelihood = -216.68, aic = 443.37

arima(xt, order = c(0, 0, 4))$aic

## [1] 440.5388

Time Series Forecasting:

Example: Forecasting a time series model

xt <- arima.sim(list(order = c(3, 0, 0), ar = c(1, -1, 1/2)), n = 100)

model <- ar(xt)
preds <- predict(model, n.ahead = 30)

plot(xt, xlim = c(0, 130), main = "Forecasting Example", ylab = "Data")

lines(preds$pred, col = "red")

Forecasting Example
4
2
Data

−2

0 20 40 60 80 100 120

Time

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 50

Exponential Smoothing:

set.seed(50)
N <- 100
etx <- rnorm(N, mean = 0, sd = 5)
xt <- cumsum(etx)
xt <- ts(xt, start=c(2020, 1), frequency = 12)
HW <- HoltWinters(xt, alpha = 0.7, beta = F, gamma = F)
HW

## Holt-Winters exponential smoothing without trend and without seasonal component.

##
## Call:
## HoltWinters(x = xt, alpha = 0.7, beta = F, gamma = F)
##
## Smoothing parameters:
## alpha: 0.7
## beta : FALSE
## gamma: FALSE
##
## Coefficients:
## [,1]
## a -58.94712

predict(HW, n.ahead = 1, level = 0.95, prediction.interval = T) # yields a confidence interval

## fit upr lwr

## May 2028 -58.94712 -48.66337 -69.23087

HW2 <- HoltWinters(xt, beta = F, gamma = F) # uses the optimal alpha

predict(HW2, level = 0.95, prediction.interval = T)

## fit upr lwr

## May 2028 -56.76395 -47.00935 -66.51856

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 51

Multivariate Time Series:

Example: Co-integrated time series (two series Xt and Yt , both I(1), such that there is a stationary linear combination
aXt + bYt ).

# Generate two random walks (they are I(1)).

set.seed(50)
N <- 10
etx <- rnorm(N, mean = 0, sd = 5)
xt <- cumsum(etx)

ety <- rnorm(N, mean = 0, sd = 10)

yt <- cumsum(ety)

find.cointegrated <- function(x) { # the lowest the p-value the most likely
comb <- xt*x[1] + yt*x[2] # for the linear combination to be stationary
PP.test(comb)$p.value
}

v <- c(1, 1) # initial search parameters are a = 1, b = 1

X <- nlm(find.cointegrated, v) # nonlinear minimization (Newton-like algorithm)

X$estimate

## [1] 1.924800 0.470916

# Confirm stationarity.
zt <- X$estimate[1]*xt + X$estimate[2]*yt
PP.test(zt)

##
## Phillips-Perron Unit Root Test
##
## data: zt
## Dickey-Fuller = -4.0434, Truncation lag parameter = 2, p-value =
## 0.02174

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved
 ACTEX Learning R Formula & Review Sheet 52

Example: Stationarity of multivariate time series


Xt = 0.3Xt−1 + 0.45Yt−1 + et,x


Yt = 0.2Xt−1 + 0.4Yt−1 + et,y



A <- matrix(c(0.3, 0.2, 0.45, 0.4), 2, 2)

eigen(A)$values

## [1] 0.65413813 0.04586187

# Both values have absolute value less than 1, so the process is stationary.

Copyright © ArchiMedia Advantage Inc.

www.ACTEXLearning.com Need Help? Email [email protected] All Rights Reserved