1. Confidence intervals for linear combinations of variances

2016-03-19
Source

Let $x_i$’s be some observations of independent random variables $X_i \sim \theta_i \dfrac{\chi^2_{d_i}}{d_i}$. In this article we will take a look at some methods to get a confidence interval about a linear combination of the $\theta_i$’s. This situation occurs when we are interested in the variances of interest in an ANOVA model with random effects.

Satterthwaite method

In order to get a confidence interval about a linear combination $\sum a_i\theta_i$ with nonnegative coefficents $a_i$, the Satterthwaite approximation consists in doing as if $\sum a_i X_i \sim \left(\sum a_i\theta_i\right) \frac{\chi^2_\nu}{\nu} \quad \textrm{with }\; \nu = \frac{{\left(\sum a_ix_i\right)}^2}{\sum\dfrac{{(a_i x_i)}^2}{d_i}}.$ Thus, taking a $100(1-\alpha)\%$-dispersion interval $[b^-, b^+]$ of the $\chi^2_\nu$ distribution, one gets the approximate $100(1-\alpha)\%$-confidence interval about $\sum a_i\theta_i$: $\left[\nu\frac{\sum a_ix_i}{b^+}, \nu\frac{\sum a_ix_i}{b^-}\right].$ This interval is returned by the ciSatt function below, taking the quantiles $\chi^2_\nu\bigl(\frac{\alpha}{2}\bigr)$ and $\chi^2_\nu\bigl(1-\frac{\alpha}{2}\bigr)$ for $b^-$ and $b^+$ respectively.

ciSatt <- function(x, dofs, a, alpha=0.05){
s <- sum(a*x)
nu <- s^2/sum((a*x)^2/dofs)
lwr <- s*nu/qchisq(1-alpha/2, nu)
upr <- s*nu/qchisq(alpha/2,nu)
return(c("lwr"=lwr, "upr"=upr))
}

Graybill & Wang’s method

Graybill & Wang provided another method for nonnegative linear combinations. Their approximate $100(1-\alpha)\%$-confidence interval about $\sum a_i\theta_i$ is $\left[\sum a_i x_i - \sqrt{\sum{(G_ia_ix_i)}^2}, \sum a_i x_i + \sqrt{\sum{(H_ia_ix_i)}^2}\right]$ where $G_i = 1 - \frac{d_i}{\chi^2_{d_i}\bigl(1-\frac{\alpha}{2}\bigr)} \quad \text{and }\; H_i = \frac{d_i}{\chi^2_{d_i}\bigl(\frac{\alpha}{2}\bigr)} - 1.$

Ting & al’s generalization

Graybill & Wang’s confidence interval has been generalized to the case when some $a_i$ are negative by Ting & al (see also Burdick & al). It is returned by the function ciTing given below.

ciTing <- function(x, dofs, a, alpha=0.05){
G <- 1 - sapply(dofs, function(d){ d/qchisq(1-alpha/2,d) })
H <- sapply(dofs, function(d){ d/qchisq(alpha/2,d) }) - 1
s <- sum(a*x)
if(all(a>0)){ # this is Graybill & Wang's confidence interval
lwr <- s - sqrt(sum((G*a*x)^2))
upr <- s + sqrt(sum((H*a*x)^2))
return(c("lwr"=lwr, "upr"=upr))
}
pos <- which(a>0); neg <- which(a<0)
A12 <- sum(sapply(pos, function(q){
sapply(neg, function(r){
Fqr <- qf(1-alpha/2, dofs[q], dofs[r]) # alpha/2 in the article is an error
Lqr <- ((Fqr-1)^2 - G[q]^2*Fqr^2 - H[r]^2)/Fqr
return(Lqr*a[q]*x[q]*a[r]*x[r])
})
}))
B12 <- sum(sapply(pos, function(q){
sapply(neg, function(r){
Fqr <- qf(alpha/2, dofs[q], dofs[r]) # 1-alpha/2 in the article is an error
Lqr <- ((Fqr-1)^2 - H[q]^2*Fqr^2 - G[r]^2)/Fqr
return(Lqr*a[q]*x[q]*a[r]*x[r])
})
}))
lwr <- s - sqrt(sum((G[pos]*a[pos]*x[pos])^2) + sum((H[neg]*a[neg]*x[neg])^2) - A12)
upr <- s + sqrt(sum((H[pos]*a[pos]*x[pos])^2) + sum((G[neg]*a[neg]*x[neg])^2) - B12)
return(c("lwr"=lwr, "upr"=upr))
}

We will study the performance of this confidence interval on an example. Some improvements of this interval are given in Ting & al’s paper, but we do not provide them here.

Example

Consider a balanced three-way ANOVA model with one fixed factor and two random factors: $y_{hijk} = \underset{\mu_h}{\underbrace{\mu + A_h}} + B_i + C_j + {(AB)}_{hi} + {(AC)}_{hj} + {(BC)}_{ij} + {(ABC)}_{hij} + \epsilon_{hijk}, \\ \quad h = 1,\ldots, H, \quad i = 1,\ldots, I, \quad j = 1, \ldots, J, \quad k = 1, \ldots, K.$

$\scriptsize \begin{array}{lccc} \text{Source} & \text{dof} & \text{Mean square} & \text{Expected mean square} \\ B & I-1 & S^2_B & \theta_B = \sigma^2_E + K \sigma^2_{ABC} + HK \sigma^2_{BC} + JK \sigma^2_{AB} + HJK \sigma^2_B \\ C & J-1 & S^2_C & \theta_C = \sigma^2_E + K \sigma^2_{ABC} + HK \sigma^2_{BC} + IK \sigma^2_{AC} + HIK \sigma^2_C \\ A \times B & (H-1)(I-1) & S^2_{AB} & \theta_{AB}= \sigma^2_E + K \sigma^2_{ABC} + JK \sigma^2_{AB} \\ B \times C & (I-1)(J-1) & S^2_{BC} & \theta_{BC} = \sigma^2_E + K \sigma^2_{ABC} + HK \sigma^2_{BC} \\ A \times C & (H-1)(J-1) & S^2_{AC} & \theta_{AC} = \sigma^2_E + K \sigma^2_{ABC} + IK \sigma^2_{AC} \\ A \times B \times C & (H-1)(J-1)(K-1) & S^2_{ABC} & \theta_{ABC} = \sigma^2_E + K \sigma^2_{ABC} \\ E & n - HIJ & S^2_E & \theta_E = \sigma^2_E \end{array}$ In matricial form, the variance components and the expected mean squares are related by $\small \begin{pmatrix} \theta_B \\ \theta_C \\ \theta_{AB} \\ \theta_{BC} \\ \theta_{AC} \\ \theta_{ABC} \\ \theta_E \end{pmatrix} = \begin{pmatrix} HJK & 0 & JK & HK & 0 & K & 1 \\ 0 & HIK & 0 & HK & IK & K & 1 \\ 0 & 0 & JK & 0 & 0 & K & 1 \\ 0 & 0 & 0 & HK & 0 & K & 1 \\ 0 & 0 & 0 & 0 & IK & K & 1 \\ 0 & 0 & 0 & 0 & 0 & K & 1 \\ 0 & 0 & 0 & 0 & 0 & 0 & 1 \end{pmatrix} \begin{pmatrix} \sigma^2_B \\ \sigma^2_C \\ \sigma^2_{AB} \\ \sigma^2_{BC} \\ \sigma^2_{AC} \\ \sigma^2_{ABC} \\ \sigma^2_E \end{pmatrix}$ and conversely, $\small \begin{pmatrix} \sigma^2_B \\ \sigma^2_C \\ \sigma^2_{AB} \\ \sigma^2_{BC} \\ \sigma^2_{AC} \\ \sigma^2_{ABC} \\ \sigma^2_E \end{pmatrix} = \begin{pmatrix} \frac{1}{HJK} & 0 & -\frac{1}{HJK} & -\frac{1}{HJK} & 0 & \frac{1}{HJK} & 0 \\ 0 & \frac{1}{HIK} & 0 & -\frac{1}{HIK} & -\frac{1}{HIK} & \frac{1}{HIK} & 0 \\ 0 & 0 & \frac{1}{JK} & 0 & 0 & -\frac{1}{JK} & 0 \\ 0 & 0 & 0 & \frac{1}{HK} & 0 & -\frac{1}{HK} & 0 \\ 0 & 0 & 0 & 0 & \frac{1}{IK} & -\frac{1}{IK} & 0 \\ 0 & 0 & 0 & 0 & 0 & \frac{1}{K} & -\frac{1}{K} \\ 0 & 0 & 0 & 0 & 0 & 0 & 1 \end{pmatrix} \begin{pmatrix} \theta_B \\ \theta_C \\ \theta_{AB} \\ \theta_{BC} \\ \theta_{AC} \\ \theta_{ABC} \\ \theta_E \end{pmatrix}$

We look for a confidence interval about the reproductibily variance of factor $B$: $\sigma^2_{\textrm{repro}, B} = \sigma^2_B + \sigma^2_{AB} + \sigma^2_{BC} + \sigma^2_{ABC}$ which is the linear combination of the expected mean squares $\small \begin{multline} \frac{1}{HJK} \theta_B + \left(\frac{1}{JK}-\frac{1}{HJK}\right)\theta_{AB} + \left(\frac{1}{HK}-\frac{1}{HJK}\right)\theta_{BC} \\ + \left(\frac{1}{HJK}-\frac{1}{JK} + \frac{1}{K}-\frac{1}{HK}\right)\theta_{ABC} - \frac{1}{K} \theta_E \\ = \frac{1}{HJK}\left(\theta_B + (H-1)\theta_{AB} + (J-1)\theta_{BC} + (1-H-HJ-J)\theta_{ABC} - HJ\theta_E \right) \quad (\ast) \end{multline}$ and is estimated by $\frac{1}{HJK}\left(S^2_B + (H-1)S^2_{AB} + (J-1)S^2_{BC} + (1-H-HJ-J)S^2_{ABC} - HJS^2_E \right).$

Checking the coverage

Let us check the frequentist coverage of the confidence interval. We firstly write a function to sample the mean squares:

rMeanSquares <- function(nsims, H, I, J, K, sigma2B=1, sigma2C=1, sigma2AB=1, sigma2BC=1, sigma2AC=1, sigma2ABC=1, sigma2E=1){
VCnames <- c("B", "C", "AB", "BC", "AC", "ABC", "E")
VC <- setNames(c(sigma2B, sigma2C, sigma2AB, sigma2BC, sigma2AC, sigma2ABC, sigma2E), VCnames)
dofs <- setNames(c(I-1, J-1, (H-1)*(I-1), (I-1)*(J-1), (H-1)*(J-1), (H-1)*(J-1)*(K-1), H*I*J*(K-1)), VCnames)
M <- rbind(
c(H*J*K, 0, J*K, H*K, 0, K, 1),
c(0, H*I*K, 0, H*K, I*K, K, 1),
c(0, 0, J*K, 0, 0, K, 1),
c(0, 0, 0, H*K, 0, K, 1),
c(0, 0, 0, 0, I*K, K, 1),
c(0, 0, 0, 0, 0,K, 1),
c(0, 0, 0, 0, 0, 0, 1)
)
thetas <- setNames(as.vector(M %*% VC), VCnames)
sims <- matrix(numeric(1), nrow=nsims, ncol=7)
colnames(sims) <- VCnames
for(j in VCnames){
sims[,j] <- thetas[j]/dofs[j]*rchisq(nsims, dofs[j])
}
attr(sims, "VC") <- VC
attr(sims, "dofs") <- dofs
return(sims)
}

Here we simulate the mean squares using not too small values of the degress of freedom.

# simulations
H <- 10; I <- 15; J <- 10; K <- 5
nsims <- 10000
sims <- rMeanSquares(nsims, H=H, I=I, J=J, K=K)
dofs <- attr(sims, "dofs")
VC <- attr(sims, "VC")
sigma2Brepro <- sum(VC[c("B", "AB", "BC", "ABC")])
# linear combination
a <- 1/K*c(1/H/J, 1/J-1/H/J, 1/H-1/H/J, 1/H/J-1/J+1-1/H, -1)
# confidence intervals
Bounds <- matrix(numeric(1), nrow=nsims, ncol=2)
colnames(Bounds) <- c("lwr", "upr")
for(i in 1:nsims){
Bounds[i,] <- ciTing(sims[i,c("B", "AB", "BC", "ABC", "E")], dofs[c("B", "AB", "BC", "ABC", "E")], a=a)
}
# coverage of the upper one-sided interval:
mean(Bounds[,"lwr"] < sigma2Brepro)
##  0.9697
# coverage of the lower one-sided interval:
mean(Bounds[,"upr"] > sigma2Brepro)
##  0.981
# coverage of the two-sided interval:
mean(Bounds[,"lwr"] < sigma2Brepro & Bounds[,"upr"] > sigma2Brepro)
##  0.9507

As we observe, the difference between the coverage obtained from the simulations and the nominal coverage does not exceed $1\%$ for each of the three confidence intervals (upper one-sided, lower one-sided and two-sided).

A small degrees of freedom example

Now let us assess the frequentist coverage with $H=3$, $I=3$, $J=3$ and $K=5$.

# simulations
H <- 3; I <- 3; J <- 3; K <- 5
nsims <- 10000
set.seed(666)
sims <- rMeanSquares(nsims, H=H, I=I, J=J, K=K)
dofs <- attr(sims, "dofs")
VC <- attr(sims, "VC")
sigma2Brepro <- sum(VC[c("B", "AB", "BC", "ABC")])
# linear combination
a <- 1/K*c(1/H/J, 1/J-1/H/J, 1/H-1/H/J, 1/H/J-1/J+1-1/H, -1)
# confidence intervals
Bounds <- matrix(numeric(1), nrow=nsims, ncol=2)
colnames(Bounds) <- c("lwr", "upr")
for(i in 1:nsims){
Bounds[i,] <- ciTing(sims[i,c("B", "AB", "BC", "ABC", "E")], dofs[c("B", "AB", "BC", "ABC", "E")], a=a)
}
# coverage of the upper one-sided interval:
mean(Bounds[,"lwr"] < sigma2Brepro)
##  0.9496
# coverage of the lower one-sided interval:
mean(Bounds[,"upr"] > sigma2Brepro)
##  0.9995
# coverage of the two-sided interval:
mean(Bounds[,"lwr"] < sigma2Brepro & Bounds[,"upr"] > sigma2Brepro)
##  0.9491

This time, the coverage is not so close to the nominal level. The upper one-sided confidence interval ([lwr, Inf[) is too short, and the lower one-sided confidence interval (]-Inf, upr]) is too long.
In other words, the lower bound and the upper bound are higher than desired.
Let’s have a look to the bounds:

##             lwr       upr
##  [1,] 2.7472674  86.87765
##  [2,] 3.9919586 212.14470
##  [3,] 0.6890338  20.96861
##  [4,] 2.1775242  40.13811
##  [5,] 1.8697206  85.46661
##  [6,] 1.9688034 155.66416
##  [7,] 1.4805632 108.91109
##  [8,] 1.4817979  26.15695
##  [9,] 1.6689184  57.09501
## [10,] 1.3509618  18.37216

The upper bound is quite big ($\sigma_{\textrm{repro},B}=4$ here).

Shortening the intervals with the Satterthwaite approximation

Recall our linear combination of the mean squares:

\begin{align*} & a_1 S^2_B + a_2 S^2_{AB} + a_3 S^2_{BC} + a_4 S^2_{ABC} + a_5 S^2_E \\ = & a_1 x_1 + a_2 x_2 + a_3 x_3 + a_4 x_4 + a_5 x_5 \end{align*} with coefficients $a_1,a_2,a_3,a_4>0$, $a_5<0$, and degrees of freedom $2$, $4$, $4$, $16$ and $108$.

A degree of freedom of $2$ is pretty small, and it could be the cause of the large upper bound.
To circumvent this problem, we could try to replace $a_1x_1 + a_2x_2$ with its Satterthwaite approximation: $\underset{y}{\underbrace{a_1 x_1 + a_2 x_2}} + a_3 x_3 + a_4 x_4 + a_5 x_5$ and then apply the Ting & al interval to the new linear combination $y+a_3 x_3 + a_4 x_4 + a_5 x_5$. Let’s look what it gives for the second row of simulations:

x <- sims[2, c("B", "AB", "BC", "ABC", "E")]
dofs <- c(2, 4, 4, 16, 108)
y <- sum(a[1:2]*x[1:2])
nu <- y^2/sum((a[1:2]*x[1:2])^2/dofs[1:2])
x_new <- c(y, x, x, x)
dofs_new <- c(nu, dofs, dofs, dofs)
a_new <- c(1, a, a, a)
# original interval:
ciTing(x, dofs, a)
##        lwr        upr
##   3.991959 212.144698
# new interval:
ciTing(x_new, dofs_new, a_new)
##       lwr       upr
##  3.332372 79.301873

The upper bound is considerably smaller. Now let’s have a look at the coverage when we apply this method to the previous simulations:

# confidence intervals
Bounds_new <- matrix(numeric(1), nrow=nsims, ncol=2)
colnames(Bounds_new) <- c("lwr", "upr")
for(i in 1:nsims){
x <- sims[i, c("B", "AB", "BC", "ABC", "E")]
y <- sum(a[1:2]*x[1:2])
nu <- y^2/sum((a[1:2]*x[1:2])^2/dofs[1:2])
x_new <- c(y, x, x, x)
dofs_new <- c(nu, dofs, dofs, dofs)
Bounds_new[i,] <- ciTing(x_new, dofs_new, a_new)
}
# coverage of the upper one-sided interval:
mean(Bounds_new[,"lwr"] < sigma2Brepro)
##  0.9759
# coverage of the lower one-sided interval:
mean(Bounds_new[,"upr"] > sigma2Brepro)
##  0.9967
# coverage of the two-sided interval:
mean(Bounds_new[,"lwr"] < sigma2Brepro & Bounds_new[,"upr"] > sigma2Brepro)
##  0.9726

This time, the upper one-sided interval achieves a coverage close to the nominal value. The lower one-sided interval still have a too large coverage, but the upper bounds we get are generally pretty shorter:

##            lwr       upr
## [1,] 2.7472674  86.87765
## [2,] 3.9919586 212.14470
## [3,] 0.6890338  20.96861
## [4,] 2.1775242  40.13811
## [5,] 1.8697206  85.46661
## [6,] 1.9688034 155.66416
Note that the method we proposed here is not intended to be a general one. The only thing we propose to the user is to explore the performance of the confidence intervals with the help of simulations for a specific design (values of $H$, $I$, $J$ and $K$) and the expected values of the variance components. We also recall that Ting & al’s paper provides some improvements of the confidence intervals that we did not consider here.