pempi Vignette

Introduction

In epidemiological studies, an important quantity that needs to be estimated with great accuracy is the prevalence rate of a disease in order to learn more about central parameters such as case fatality rate and/or to plan and refine decisions about measures with regards to an epidemic or a pandemic. Traditionally, to measure the prevalence, a survey sample (of randomly chosen subjects from the population) is collected and the prevalence is estimated by the sample proportion. This process involves some financial and logistic efforts, which increase with the number of sampled participants, while at the same time, increasing the accuracy of the estimator. Having a sufficiently large sample is especially important when the (true) prevalence is very small, for example at the beginning of a new pandemic such as the one of the COVID-19. In this case, since the beginning of the outbreak, many measurements have been taken on the number of infected people, but only on the sub-population selected under some medical (and logistic) criteria. It is obvious that using these measurements as if they would be like a complete census, would lead to an underestimation of the prevalence.

In Guerrier, Kuzmics, and Victoria-Feser (2023), we propose to adequately use this information, together with data from a survey sample, in order to improve the accuracy of the prevalence estimation. In other terms, for a given (legal) statistical precision that might be required by authorities that finance a survey, the sample size can be a lot smaller if one adequately uses the information provided by the data collected in the sub-population. The possible misclassification errors of the (medical) testing devices used to collect the data, are also taken into account. The misclassification errors are actually induced by the sensitivity, i.e. the complement to the false positive rate, and by the specificity, i.e. the complement to the false negative rate, of the medical testing devices. The approach is a frequentist one, i.e. using cutoff values for the sensitivity and specificity, hence without the need to specify a (prior) distribution for these quantities.

Mathematical setup

As in Guerrier, Kuzmics, and Victoria-Feser (2023), we define the following unobserved random variable:

\[\begin{equation*} X_i= \left\{ \begin{array}{ll} 1 & \quad \mbox{if participant (in the survey sample) $i$ is positive,} \\ 0 & \quad \mbox{otherwise;} \end{array} \right. \end{equation*}\]

The objective is to provide an estimator for the unknown population proportion, i.e. the prevalence, given by

\[\begin{equation*} \pi = \Pr\left(X_i=1\right). \end{equation*}\]

Next, we define the following quantities.

\[\begin{eqnarray} Y_i&=& \left\{ \begin{array}{ll} 1 & \quad \mbox{if participant $i$ is tested positive in the survey sample,} \\ 0 & \quad \mbox{otherwise;} \end{array} \right.\nonumber \\ Z_i&=& \left\{ \begin{array}{ll} 1 & \quad \mbox{if participant $i$ was declared positive with the official procedure,} \\ 0 & \quad \mbox{otherwise.} \end{array} \right. \label{eqn:Y-Z} \end{eqnarray}\]

and

\[\begin{alignat}{4} &R_{11}&&=\sum_{i=1}^n Y_i Z_i, \quad\quad\quad &&R_{10}&&=\sum_{i=1}^n(1-Y_i)Z_i, \label{eqn:R-R4} \\ &R_{01}&&= \sum_{i=1}^nY_i(1-Z_i), \quad\quad \quad &&R_{00}&&= \sum_{i=1}^n(1-Y_i)(1-Z_i)=n-R_{11}-R_{10}+R_{01}. \nonumber \end{alignat}\]

In words, \(R_{11}\) is the number of participants in the survey sample that are tested positive and have also been declared positive through the official procedure; \(R_{10}\) is the number of participants in the survey sample that are tested negative but have been declared positive through the official procedure; \(R_{01}\) is the number of participants in the survey sample that are tested positive but have been declared negative through the official procedure; \(R_{00}\) is the number of participants in the survey sample that are tested negative and have been declared negative through the official procedure. We also make use of \(R_{\ast 1} = \sum_{i=1}^nY_i = R_{11} + R_{01}\), the number of participants that are tested positive in the survey sample. The \(R_{jk}\) are central quantities for computing the proposed estimators. In the case of stratified sampling, these quantities can be modified accordingly (see below in the section on stratified sampling).

Random sampling

For simplicity of exposition, we first consider the unweighted (or random) sampling case, the weighted one is treated below.

We also allow for the possibility that the outcome of (medical) tests can be subject to misclassification error. Hence, we define \(\alpha = \Pr(Y_i=1\vert X_i=0)\) and \(\beta = \Pr(Y_i=0\vert X_i=1)\). The probabilities \(\alpha\) and \(\beta\), are the (assumed known) FP rate (\(\alpha = 1-\mbox{specificity}\)) and FN rate (\(\beta = 1-\mbox{sensitivity}\)) of the particular medical test employed in the survey. Moreover, we will make use of the (known) prevalence \(\pi_0 =\Pr(Z_i=1)\) from the official procedure. As previously explained, \(\pi_0\) is the joint probability of being selected in the official procedure and declared positive, so that we have \(\pi_0\leq \pi\).

Since the objective is to take advantage of the information provided by the official procedure in estimating the prevalence \(\pi\), we also need to take into account the possible biases in the official data. Therefore, we define \(\alpha_0 = \Pr(Z_i=1\vert X_i=0)\), the FP rate of the official procedure and \(\beta_0 = \Pr(Z_i=0\vert X_i=1)\) the FN rate of the official procedure. It turns out (see Guerrier, Kuzmics, and Victoria-Feser (2023) for details) that \(\alpha_0\) is a negligible quantity and hence can be set to \(0\), and \(\beta_0\) can be deduced from the other available or estimable quantities as we have

\[ \beta_0 = 1- \frac{\pi_0-\alpha_0(1-\pi)}{\pi}. \]

The success probabilities (see Guerrier, Kuzmics, and Victoria-Feser (2023) for their derivation), denoted by \(\tau_{jk}(\pi)\) associated to each \(R_{jk}\), \(j,k \in \{0,1\}\) are given by % \[\begin{eqnarray} \begin{aligned} \tau_{11}(\pi) &= \Pr(Z_i=1, Y_i=1) =\pi\Delta\alpha_0+(\pi_0-\alpha_0)(1-\beta)+\alpha\alpha_0, \\ \tau_{10}(\pi) &= \Pr(Z_i=1,Y_i=0) = -\pi\Delta\alpha_0+(\pi_0-\alpha_0)\beta+(1-\alpha)\alpha_0, \\ \tau_{01}(\pi) &= \Pr(Z_i=0,Y_i=1) = \pi\Delta(1-\alpha_0)-(\pi_0-\alpha_0)(1-\beta)+\alpha(1-\alpha_0), \\ \tau_{00}(\pi) &= \Pr(Z_i=0,Y_i=0) = -\pi\Delta(1-\alpha_0)-(\pi_0-\alpha_0)\beta+(1-\alpha)(1-\alpha_0), \end{aligned} \label{eqn:tau-pi} \end{eqnarray}\] % where \(\Delta =1-(\alpha+\beta)\). Without misclassification error, we would have \(\tau_{11}(\pi)=\pi_0\), \(\tau_{10}(\pi)=0\), \(\tau_{01}(\pi)=\pi-\pi_0\), \(\tau_{00}(\pi)=1-\pi\).

Based on these definitions, we present a Conditional Maximum Likelihood Estimators (CMLE) and a Method of Moment Estimator (MME) estimator. The formal derivations and properties are provided in Guerrier, Kuzmics, and Victoria-Feser (2023). There, we also consider a Marginal MLE (MMLE) when some data is missing, and some Generalized Method of Moment (GMM) estimators. The likelihood function for \(\pi\) can be obtained from the multinomial distribution with categories provided by \(R_{11},R_{10},R_{01},R_{00}\) and their associated success probabilities \(\tau_{11}(\pi),\tau_{10}(\pi),\tau_{01}(\pi),\tau_{00}(\pi)\). The CMLE, conditional on the information provided by the official procedure, \(\widehat{\pi}\), generally, has no closed-form solution but can be computed numerically. However, in the case when \(\alpha_0 = 0\), we obtain a closed-form solution given by

\[\begin{equation} \widehat{\pi} = \frac{\pi_0 R_{00} + R_{01}}{\Delta \left(R_{01}+R_{00}\right)} - \frac{\pi_0 \beta}{\Delta} - \frac{\alpha}{\Delta}. \label{eq:cmle-closedform} \end{equation}\]

When \(\alpha_0=\alpha=\beta=0\), this further reduces to

\[\begin{equation} \widehat{\pi} = \pi_0 \frac{n - R_{\ast 1}}{n - R_{11}} + \frac{R_{01}}{ \left(n - R_{11}\right)}. \label{eqn:MLE} \end{equation}\]

Considering the data used in Guerrier, Kuzmics, and Victoria-Feser (2023), this estimator can be computed as follows:

# Load pempi
library(pempi)

# Austrian data (November 2020)
pi0 = 93914/7166167

# Load data
data("covid19_austria")

# Random sampling
n = nrow(covid19_austria)
R1 = sum(covid19_austria$Y == 1 & covid19_austria$Z == 1)
R2 = sum(covid19_austria$Y == 0 & covid19_austria$Z == 1)
R3 = sum(covid19_austria$Y == 1 & covid19_austria$Z == 0)
R4 = sum(covid19_austria$Y == 0 & covid19_austria$Z == 0)

# Compute CMLE
conditional_mle(R1 = R1, R2 = R2, R3 = R3, R4 = R4, pi0 = pi0)
## Method: Conditional MLE
## 
## Estimated proportion: 2.9317%
## Standard error      : 0.2639%
## 
## Confidence interval at the 95% level:
## Asymptotic Approach: 2.4145% - 3.4489%
## 
## Assumed measurement error: alpha  = 0%, beta = 0%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 55.30%
## CI at the 95% level: 47.41% - 63.18%
## 
## Estimated ascertainment rate: 
## pi0/pi = 44.70%
## CI at the 95% level: 36.82% - 52.59%
## 
## Sampling: Random

Note that the notation and conventions used in the text and in are slightly amended for convenience in this package. In particular, we use R1 for \(R_{11}\), R2 for \(R_{10}\), R3 for \(R_{01}\) and R4 for \(R_{00}\). Considering the following measurement error \(\alpha = 0.01\), \(\alpha_0 = 0\) and \(\beta = 0.1\), we obtain:

# Assumed measurement errors
alpha0 = 0
alpha  = 1/100
beta   = 10/100

# Compute CMLE with measurement error
conditional_mle(R1 = R1, R2 = R2, R3 = R3, R4 = R4, pi0 = pi0,
                alpha = alpha, alpha0 = alpha0, beta = beta)
## Method: Conditional MLE
## 
## Estimated proportion: 2.0200%
## Standard error      : 0.2962%
## 
## Confidence interval at the 95% level:
## Asymptotic Approach: 1.4394% - 2.6006%
## 
## Assumed measurement error: alpha  = 1%, beta = 10%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 35.12%
## CI at the 95% level: 16.48% - 53.77%
## 
## Estimated ascertainment rate: 
## pi0/pi = 64.88%
## CI at the 95% level: 46.23% - 83.52%
## 
## Sampling: Random

Alternatively, we can consider an estimator from the class of GMM estimators based on the random variable \(\mathbf{R} =[R_{11}/n, R_{10}/n, R_{01}/n, R_{00}/n]\) with expectation \(\mathbb{E}[\mathbf{R}] = {\tau}(\pi)=[\tau_{11}(\pi), \tau_{10}(\pi), \tau_{01}(\pi), \tau_{00}(\pi)]\). A particular case is given by an MME based on \(R_{01}\) (with expectation \(\tau_{01}(\pi)\)), which, again assuming an interior solution exists, is given by

\[\begin{equation} \widetilde{\pi} = \frac{1}{\Delta(1-\alpha_0)} \left(\frac{R_{01}}{n} + \pi_0 - \beta\pi_0 - \alpha_0 \Delta - \alpha\right). \label{eqn:MME-ME} \end{equation}\]

When \(\alpha_0=\alpha=\beta=0\), this reduces to

\[\begin{equation} \widetilde{\pi} = \pi_0 + \frac{R_{01}}{n}. \label{eqn:MME} \end{equation}\]

Since we have that \(\mathbb{E}[\widetilde{\pi} ] = \pi\), the MME is unbiased. This estimator can be computed as follows:

# Without measurement error
moment_estimator(R3 = R3, n = n, pi0 = pi0)
## Method: Moment Estimator
## 
## Estimated proportion: 2.9262%
## Standard error      : 0.2635%
## 
## Confidence intervals at the 95% level:
## Asymptotic Approach: 2.4099% - 3.4426%
## Clopper-Pearson    : 2.4506% - 3.5308%
## 
## Assumed measurement error: alpha  = 0%, beta = 0%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 55.21%
## CI at the 95% level: 47.31% - 63.12%
## 
## Estimated ascertainment rate: 
## pi0/pi = 44.79%
## CI at the 95% level: 36.88% - 52.69%
## 
## Sampling: Random
# With measurement error
moment_estimator(R3 = R3, n = n, pi0 = pi0, alpha = alpha,
                 alpha0 = alpha0, beta = beta)
## Method: Moment Estimator
## 
## Estimated proportion: 2.0171%
## Standard error      : 0.2960%
## 
## Confidence intervals at the 95% level:
## Asymptotic Approach: 1.4369% - 2.5973%
## Clopper-Pearson    : 1.4827% - 2.6963%
## 
## Assumed measurement error: alpha  = 1%, beta = 10%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 35.03%
## CI at the 95% level: 16.34% - 53.72%
## 
## Estimated ascertainment rate: 
## pi0/pi = 64.97%
## CI at the 95% level: 46.28% - 83.66%
## 
## Sampling: Random

In the pempi package the marginal MLE is also implemented and can be used as follows:

# Without measurement error
marginal_mle(R1 = R1, R3 = R3, n = n, pi0 = pi0)
## Method: Marginal MLE
## 
## Estimated proportion: 2.9317%
## Standard error      : 0.2639%
## 
## Confidence interval at the 95% level:
## Asymptotic Approach: 2.4145% - 3.4489%
## 
## Assumed measurement error: alpha  = 0%, beta = 0%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 55.30%
## CI at the 95% level: 47.41% - 63.18%
## 
## Estimated ascertainment rate: 
## pi0/pi = 44.70%
## CI at the 95% level: 36.82% - 52.59%
## 
## Sampling: Random
# With measurement error
marginal_mle(R1 = R1, R3 = R3, n = n, pi0 = pi0, 
             alpha = alpha, beta = beta, alpha0 = alpha0)
## Method: Marginal MLE
## 
## Estimated proportion: 2.0200%
## Standard error      : 0.2962%
## 
## Confidence interval at the 95% level:
## Asymptotic Approach: 1.4394% - 2.6006%
## 
## Assumed measurement error: alpha  = 1%, beta = 10%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 35.12%
## CI at the 95% level: 16.48% - 53.77%
## 
## Estimated ascertainment rate: 
## pi0/pi = 64.88%
## CI at the 95% level: 46.23% - 83.52%
## 
## Sampling: Random

These results can be compared to the standard survey MLE which can be computed as follows:

# Without measurement error
survey_mle(R = R1 + R3, n = n)
## Method: Survey MLE
## 
## Estimated proportion: 3.1441%
## Standard error      : 0.3647%
## 
## Confidence intervals at the 95% level:
## Asymptotic Approach: 2.4294% - 3.8588%
## Clopper-Pearson    : 2.4680% - 3.9433%
## 
## Assumed measurement error: alpha = 0%, beta = 0% 
## Sampling: Random
# With measurement error
survey_mle(R = R1 + R3, n = n, alpha = alpha, beta = beta)
## Method: Survey MLE
## 
## Estimated proportion: 2.4091%
## Standard error      : 0.4097%
## 
## Confidence intervals at the 95% level:
## Asymptotic Approach: 1.6060% - 3.2122%
## Clopper-Pearson    : 1.6495% - 3.3070%
## 
## Assumed measurement error: alpha = 1%, beta = 10% 
## Sampling: Random

Stratified sampling

In many applications sampling is not uniformly random, but stratified. This is also the case for the COVID-19 data from the Austrian survey sample that we apply our method to. For such cases, one can use different approaches. In this section, we present a method that considers a prevalence estimator formed as a weighted sum of prevalence estimators \(\widetilde{\pi}^k\) associated to each stratum \(k\), i.e., a generalization of the MME, as well as a Weighted M-Estimator (WME). In both cases, we have to rely on asymptotic theory for computing CIs.

It actually turns out that the resulting estimators are based on similar quantities provided previously, but weighted ones. Let \(\gamma_i\) denote the sampling weight associated to subject \(i=1,\ldots,n\), which is proportional to the reciprocal of the sampling probability for subject \(i\), and adjusted such that \(\sum_{i=1}^n \gamma_i = n\). Let also

\[\begin{eqnarray} \overline{R}_{11}&=&\sum_{i=1}^n\gamma_iY_i Z_i=\sum_{i=1}^n\gamma_iR_{i11}, \nonumber \\ \overline{R}_{10}&=&\sum_{i=1}^n\gamma_i(1-Y_i)Z_i=\sum_{i=1}^n\gamma_iR_{i10}, \nonumber \\ \overline{R}_{01}&=& \sum_{i=1}^n\gamma_iY_i(1-Z_i)= \sum_{i=1}^n\gamma_iR_{i01}, \nonumber\\ \overline{R}_{00}&=& \sum_{i=1}^n\gamma_i(1-Y_i)(1-Z_i)=\sum_{i=1}^n\gamma_iR_{i00}, \nonumber \\ \overline{R}_{*1}&=&\sum_{i=1}^n\gamma_iY_i=\overline{R}_{11}+\overline{R}_{01}. \nonumber \end{eqnarray}\]

With the MME approach, we consider the possibility that different groups of people (such as in different towns or provinces), or even each participant \(i\), are associated to different prevalence \(\pi_i\). We are, however, only interested in the overall prevalence \(\pi=(1/n)\sum_{i=1}^n \gamma_i \pi_i\), given the additional information provided by the official procedure. Moreover, the (known but biased) prevalence from the official procedure could also be different for each (groups of) subject(s) \(i\), with \(\pi_{i0}, i=1,\ldots,n\). Note, however, that \(\pi_0 =(1/n) \sum_{i=1}^n \gamma_i \pi_{i0}\) and that \(\pi_0\) is known. A general approach then consists in obtaining estimates for each \(\pi_i\) using some method and then take their weighted average as an estimator for \(\pi\). Using the MME approach based on the \(R_{i01}\) and assuming, as done before, that the parameters \(\alpha,\beta\) are specific to the medical test used and, thus, independent of subject \(i\), we obtain a weighted MME for \(\pi\) as

\[\begin{equation} \widetilde{\pi} = \frac{1}{\Delta(1-\alpha_0)} \left( \frac{\overline{R}_{01}}{n} + \pi_0 (1 - \beta) - \alpha_0 \Delta - \alpha\right). \label{eqn:CWMLE-alpha0} \end{equation}\]

We also have that \(\mathbb{E}\left[\widetilde{\pi}\right]= \pi\). When \(\alpha_0=\alpha=\beta=0\), we get

\[\begin{equation*} \widetilde{\pi} = \pi_0 + \frac{\overline{R}_{01}}{n}. \end{equation*}\]

An alternative approach is to consider a WME \(\widehat{\pi}\) as proposed for example by Wooldridge (2001). Generally, it has no closed-form solution but can be computed numerically. However, in the case when \(\alpha_0 = 0\), we obtain a closed-form solution given by

\[\begin{equation*} \widehat{\pi} = \frac{\pi_0 \overline{R}_{00} + \overline{R}_{01}}{\Delta \left(\overline{R}_{01}+\overline{R}_{00}\right)} - \frac{\pi_0 \beta}{\Delta} - \frac{\alpha}{\Delta}. \end{equation*}\]

When \(\alpha_0=\alpha=\beta=0\), this further reduces to

\[\begin{equation*} \widehat{\pi} = \pi_0 \frac{n - \overline{R}_{\ast 1}}{n - \overline{R}_{11}} + \frac{R_{01}}{ \left(n - R_{11}\right)}. \end{equation*}\]

In other words, in the case of stratified sampling, one replaces the \(R_{jk}\) by their weighted counterparts \(\overline{R}_{jk}\).

Considering the data used in Guerrier, Kuzmics, and Victoria-Feser (2023), these estimators can be computed as follows:

# Load pempi
library(pempi)

# Austrian data (November 2020)
pi0 = 93914/7166167

# Weighted sampling
R1w = sum(covid19_austria$weights[covid19_austria$Y == 1 & covid19_austria$Z == 1])
R2w = sum(covid19_austria$weights[covid19_austria$Y == 0 & covid19_austria$Z == 1])
R3w = sum(covid19_austria$weights[covid19_austria$Y == 1 & covid19_austria$Z == 0])
R4w = sum(covid19_austria$weights[covid19_austria$Y == 0 & covid19_austria$Z == 0])

# Average of squared weights 
V = mean(covid19_austria$weights^2)

# Compute CMLE
conditional_mle(R1 = R1w, R2 = R2w, R3 = R3w, R4 = R4w, 
                pi0 = pi0, V = V)
## Method: Conditional MLE
## 
## Estimated proportion: 2.9841%
## Standard error      : 0.3294%
## 
## Confidence interval at the 95% level:
## Asymptotic Approach: 2.3385% - 3.6297%
## 
## Assumed measurement error: alpha  = 0%, beta = 0%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 56.08%
## CI at the 95% level: 46.58% - 65.59%
## 
## Estimated ascertainment rate: 
## pi0/pi = 43.92%
## CI at the 95% level: 34.41% - 53.42%
## 
## Sampling: Stratified with V = 1.51
# Compute MME
moment_estimator(R3 = R3w, pi0 = pi0, n = n, V = V)
## Method: Moment Estimator
## 
## Estimated proportion: 2.9818%
## Standard error      : 0.3292%
## 
## Confidence interval at the 95% level:
## Asymptotic Approach: 2.3365% - 3.6270%
## 
## Assumed measurement error: alpha  = 0%, beta = 0%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 56.05%
## CI at the 95% level: 46.54% - 65.56%
## 
## Estimated ascertainment rate: 
## pi0/pi = 43.95%
## CI at the 95% level: 34.44% - 53.46%
## 
## Sampling: Stratified with V = 1.51
# Survey MLE
survey_mle(R = R1w + R3w, pi0 = pi0, n = n, V = V)
## Method: Survey MLE
## 
## Estimated proportion: 3.1280%
## Standard error      : 0.4471%
## 
## Confidence interval at the 95% level:
## Asymptotic Approach: 2.2517% - 4.0042%
## 
## Assumed measurement error: alpha = 0%, beta = 0% 
## Sampling: Stratified with V = 1.51

These estimators are also defined in the stratified case with measurement error. For example, for the MME:

# Compute MME
moment_estimator(R3 = R3w, pi0 = pi0, n = n, V = V, alpha = alpha,
                 alpha0 = alpha0, beta = beta)
## Method: Moment Estimator
## 
## Estimated proportion: 2.0794%
## Standard error      : 0.3699%
## 
## Confidence interval at the 95% level:
## Asymptotic Approach: 1.3544% - 2.8045%
## 
## Assumed measurement error: alpha  = 1%, beta = 10%,
##                            alpha0 = 0% 
## 
## Estimated false negative rate of the
## official procedure: beta0 = 36.98%
## CI at the 95% level: 15.00% - 58.95%
## 
## Estimated ascertainment rate: 
## pi0/pi = 63.02%
## CI at the 95% level: 41.05% - 85.00%
## 
## Sampling: Stratified with V = 1.51

References

Guerrier, Stéphane, Christoph Kuzmics, and Maria-Pia Victoria-Feser. 2023. “Assessing COVID-19 Prevalence in Austria with Infection Surveys and Case Count Data as Auxiliary Information.” https://arxiv.org/abs/2012.10745.
Wooldridge, J. M. 2001. “Asymptotic Properties of Weighted M-Estimators for Standard Stratified Samples.” Econometric Theory 17: 451–70.