Chapter 19 Rosenbaum-Style p-Values for Matched Observational Studies with Unmeasured Confounding

Rosenbaum, 1987b () introduced a sensitivity analysis technique for matched observational studies. Although it works for general matched studies (Rosenbaum, 2002b, ), the theory is most elegant for one-to-one matching. Different from Chapters 17 and 18, Rosenbaum-type sensitivity analysis works best for matched observational studies for testing the sharp null hypothesis of no individual treatment effect.

96 A model for sensitivity analysis with matched data

Consider a matched observational study, with (i,j) indexing unit j in pair i (i=1,\ldots,n;j=1,2). With exactly matched pairs, unit (i,1) and unit (i,2) have the same covariates X_{i}. Assume IID sampling, and extend Definition 11.1 to define the propensity score as

e_{ij}=\textup{pr}\{Z_{ij}=1\mid X_{i},Y_{ij}(1),Y_{ij}(0)\}.

Let \mathbb{S}_{i}=\{Y_{i1}(1),Y_{i1}(0),Y_{i2}(1),Y_{i2}(0)\} denote the set of all potential outcomes within pair i. Conditioning on the event that Z_{i1}+Z_{i2}=1, we have

\displaystyle\pi_{i1} \displaystyle= \displaystyle\textup{pr}\{Z_{i1}=1\mid X_{i},\mathbb{S}_{i},Z_{i1}+Z_{i2}=1\}
\displaystyle= \displaystyle\frac{\textup{pr}\{Z_{i1}=1,Z_{i2}=0\mid X_{i},\mathbb{S}_{i}\}}{\textup{pr}\{Z_{i1}+Z_{i2}=1\mid X_{i},\mathbb{S}_{i}\}}
\displaystyle= \displaystyle\frac{\textup{pr}\{Z_{i1}=1,Z_{i2}=0\mid X_{i},\mathbb{S}_{i}\}}{\textup{pr}\{Z_{i1}=1,Z_{i2}=0\mid X_{i},\mathbb{S}_{i}\}+\textup{pr}\{Z_{i1}=0,Z_{i2}=1\mid X_{i},\mathbb{S}_{i}\}}
\displaystyle= \displaystyle\frac{e_{i1}(1-e_{i2})}{e_{i1}(1-e_{i2})+(1-e_{i1})e_{i2}}

Define o_{ij}=e_{ij}/(1-e_{ij}) as the odds of the treatment for unit (i,j), and we have

\pi_{i1}=\frac{o_{i1}}{o_{i1}+o_{i2}}.

Under ignorability, e_{ij} is only a function of X_{i}, and therefore, e_{i1}=e_{i2} and \pi_{i1}=1/2. Thus the treatment assignment mechanism conditional on covariates and potential outcomes is equivalent to that from an MPE with equal treatment and control probabilities. This is a strategy to analyze matched observational studies we discussed in Chapter 75.

In general, e_{ij} is also a function of the unobserved potential outcomes, and it can range from 0 to 1. Rosenbaum, 1987b ()’s model for sensitivity analysis imposes bounds on the odds ratio o_{i1}/o_{i2}.

Assumption 19.1 (Rosenbaum’s sensitivity analysis model)

The odds ratios are bounded by

o_{i1}/o_{i2}\leq\Gamma,\quad o_{i2}/o_{i1}\leq\Gamma,\quad(i=1,\ldots,n)

for some pre-specified \Gamma\geq 1. Equivalently,

\frac{1}{1+\Gamma}\leq\pi_{i1}\leq\frac{\Gamma}{1+\Gamma}\quad(i=1,\ldots,n)

for some pre-specified \Gamma\geq 1.

Under Assumption 19.1, we have a biased MPE with unequal and varying treatment and control probabilities across pairs. When \Gamma=1, we have \pi_{i1}=1/2 and thus a standard MPE. Therefore, \Gamma>1 measures the deviation from the ideal MPE due to the omitted variables in matching.

97 Worst-case p-values under Rosenbaum’s sensitivity analysis model

Consider testing the sharp null hypothesis

H_{0\textsc{f}}:Y_{ij}(1)=Y_{ij}(0)\text{ for }i=1,\ldots,n\text{ and }j=1,2

based on the within-pair differences \hat{\tau}_{i}=(2Z_{i1}-1)(Y_{i1}-Y_{i2})\ (i=1,\ldots,n). Under H_{0\textsc{f}}, |\hat{\tau}_{i}| is fixed but S_{i}=I(\hat{\tau}_{i}>0) is random if \hat{\tau}_{i}\neq 0. Consider the following class of test statistics:

T=\sum_{i=1}^{n}S_{i}q_{i},

where q_{i}\geq 0 is a function of (|\hat{\tau}_{1}|,\ldots,|\hat{\tau}_{n}|). Special cases include the sign statistic, the pair t statistic (up to some constant shift), and the Wilcoxon signed-rank statistic:

T=\sum_{i=1}^{n}S_{i},\quad T=\sum_{i=1}^{n}S_{i}|\hat{\tau}_{i}|,\quad T=\sum_{i=1}^{n}S_{i}R_{i},

where (R_{1},\ldots,R_{n}) are the ranks of (|\hat{\tau}_{1}|,\ldots,|\hat{\tau}_{n}|).

What is the null distribution of the test statistic T under Assumption 19.1 with a general \Gamma? It can be quite complicated because Assumption 19.1 does not fully specify the exact values of the \pi_{i1}’s. Fortunately, we do not need to know the exact distribution of T but rather the worst-case distribution of T that yields the largest p-value under Assumption 19.1. This worst-case distribution corresponds to

S_{i}\stackrel{{\scriptstyle\textup{IID}}}{{\sim}}\text{Bernoulli}\left(\frac{\Gamma}{1+\Gamma}\right).

The corresponding distribution of T has mean

E_{\Gamma}(T)=\frac{\Gamma}{1+\Gamma}\sum_{i=1}^{n}q_{i},

and variance

\textup{var}_{\Gamma}(T)=\frac{\Gamma}{(1+\Gamma)^{2}}\sum_{i=1}^{n}q_{i}^{2},

with a Normal approximation

\frac{T-\frac{\Gamma}{1+\Gamma}\sum_{i=1}^{n}q_{i}}{\sqrt{\frac{\Gamma}{(1+\Gamma)^{2}}\sum_{i=1}^{n}q_{i}^{2}}}\rightarrow\textsc{N}(0,1)

in distribution. In practice, we can report a sequence of p-values as a function of \Gamma.

98 Examples

98.1 Revisiting the LaLonde data

We conduct Rosenbaum-style sensitivity analysis in the matched LaLonde data. Using the Matching package, we can construct the matched dataset.

library("Matching")
library("sensitivitymv")
library("sensitivitymw")
dat <- read.table("cps1re74.csv",header=T)
dat$u74 <- as.numeric(dat$re74==0)
dat$u75 <- as.numeric(dat$re75==0)
y = dat$re78
z = dat$treat
x = as.matrix(dat[, c("age", "educ", "black",
"hispan", "married", "nodegree",
"re74", "re75", "u74", "u75")])
matchest = Match(Y = y, Tr = z, X = x)
ytreated = y[matchest$index.treated]
ycontrol = y[matchest$index.control]
datamatched = cbind(ytreated, ycontrol)

We consider using the test statistic T=\sum_{i=1}^{n}S_{i}|\hat{\tau}_{i}|. Under the ideal MPE with \Gamma=1, we can simulate the distribution of T and obtain the p-value 0.002, as shown in the first subfigure in Figure 23. With a slightly larger \Gamma=1.1, the worst-case distribution of T shifts to the right, and the p-value increases to 0.011. If we further increase \Gamma to 1.3, then the worst-case distribution of T shifts further and the p-value exceeds 0.05. Figure 24 shows the histogram of the \hat{\tau}_{i}’s and the p-value as a function of \Gamma; \Gamma=1.233 measures the maximum confounding that we can still reject the null hypothesis at level 0.05.

We can also use the senmw function in the sensitivitymw package to obtain a sequence of p-values against \Gamma.

Gamma = seq(1, 1.4, 0.001)
Pvalue = Gamma
for(i in 1:length(Gamma))
{
Pvalue[i] = senmw(datamatched, gamma = Gamma[i],
method = "t")$pval
}

Figure 24 shows the plot of the p-value against \Gamma.

插图
Figure 23: The worst-case distributions of T=\sum_{i=1}^{n}S_{i}|\hat{\tau}_{i}| with S_{i} IID Bernoulli(\Gamma/(1+\Gamma)), based on the matched LaLonde data.
插图
Figure 24: p-value as a function of \Gamma, based on the matched LaLonde data.
98.2 Two examples from Rosenbaum’s packages

The erpcp dataset is from the R package sensitivitymw. It contains n=39 matched pairs of a welder and a control, based on observed covariates age and smoking. The outcome is the DNA elution rate. Figure 25(a) shows the histogram of the within-pair differences of the outcomes and the p-value against \Gamma based on the pair t-statistic. The following R code generates Figure 25(a).

par(mfrow = c(1, 2), mai = c(0.8, 0.8, 0.3, 0.3))
data(erpcp)
hist(erpcp[, 1] - erpcp[, 2], main = "erpcp",
xlab = expression(hat(tau)[i]),
freq = FALSE)
Gamma = seq(1, 5, 0.005)
Pvalue = Gamma
for(i in 1:length(Gamma))
{
Pvalue[i] = senmw(erpcp, gamma = Gamma[i], method = "t")$pval
}
gammastar = Gamma[which(Pvalue >= 0.05)[1]]
gammastar
plot(Pvalue ~ Gamma, type = "l",
xlab = expression(Gamma),
ylab = "p-value")
abline(h = 0.05, lty = 2)
abline(v = gammastar, lty = 2)

The lead250 dataset is from the R package sensitivitymw. It contains n=250 matched pairs of a daily smoker and a control of nonsmoker, based on observed covariates gender, age, race education level, and household income from NHANES. The outcome is the blood lead level in \mu g/l. Figure 25(b) shows the histogram of the within-pair differences of the outcomes and the p-value against \Gamma based on the pair t-statistic. The following R code generates Figure 25(b).

par(mfrow = c(1, 2), mai = c(0.8, 0.8, 0.3, 0.3))
data(lead250)
hist(lead250[, 1] - lead250[, 2],
main = "lead250",
xlab = expression(hat(tau)[i]),
freq = FALSE)
Gamma = seq(1, 2.5, 0.001)
Pvalue = Gamma
for(i in 1:length(Gamma))
{
Pvalue[i] = senmw(lead250, gamma = Gamma[i], method = "t")$pval
}
gammastar = Gamma[which(Pvalue >= 0.05)[1]]
gammastar
plot(Pvalue ~ Gamma, type = "l",
xlab = expression(Gamma), ylab = "p-value")
abline(h = 0.05, lty = 2)
abline(v = gammastar, lty = 2)
插图
(a) erpcp data
插图
(b) ead250 data
Figure 25: Two examples: histogram of the \hat{\tau}_{i}’s and p-value as a function of \Gamma

99 Homework Problems

19.1 A model for Assumption 19.1

Assumption 19.1 has the following equivalent form.

Assumption 19.2

The propensity score satisfies the following model

\log\frac{\pi_{ij}}{1-\pi_{ij}}=g(X_{i})+\gamma U_{ij},\quad(i=1,\ldots,n;j=1,2)

where g(\cdot) is an unknown function and U_{ij}\in[0,1] is a a bounded unobserved covariate for unit (i,j).

Show that if Assumption 19.2 holds, then Assumption 19.1 must hold with \Gamma=e^{\gamma}.

Remark: Rosenbaum, 2002b (, page 108) also shows that if Assumption 19.1 holds, then Assumption 19.2 must hold for some U_{ij}’s. Assumption 19.2 may be easier to interpret because \gamma measures the log of the conditional odds ratio of U on the treatment; see Chapter 31.F for the interpretation of the coefficients in the logistic regression.

19.2 Application of Rosenbaum’s approach

Re-analyze Example 10.3 using Rosenbaum’s approach based on matching.

19.3 Recommended reading

Rosenbaum, (2015) provides a tutorial for his two R packages for sensitivity analysis with matched observational studies.

本书:A First Course in Causal Inference,作者 Peng Ding。原文:arXiv:2305.18793v2,按 CC BY 4.0 许可发布。本站(yuanens.com)把 arXiv 网页版改成分章的纯 HTML,把公式渲染成图片,方便在 Kindle 等电子书阅读器上看,内容没有改动。转换难免有错,以原文为准。本站与作者和出版社无关。