Noninvasive Tuberculosis Screening in Free-Living Primate Populations in Gombe National Park, Tanzania

Abstract

Recent advances in noninvasive detection methods for mycobacterial infection in primates create new opportunities for exploring the epidemiology of tuberculosis in free-living species. Chimpanzees (Pan troglodytes schweinfurthii) and baboons (Papio anubis) in Gombe National Park, Tanzania, were screened for infection with pathogens of the Mycobacterium tuberculosis Complex using Fecal IS6110 PCR; none was positive. This study demonstrates the feasibility of large-scale mycobacterial screening in wild primates.

Technical Appendix 1

Advanced Statistics for Prevalence Estimation of Rare Disease

MTC prevalence among chimpanzee and baboons populations was estimated from laboratory results. Wilson score confidence intervals were chosen as an exact method for small n and rare disease (Brown et al. 2001; Newcombe 2012). This method for estimating confidence intervals accounts for the limits of a proportion, being bound by 0 and 1, and also for interval asymmetry that occurs when estimating a low proportion, providing a more precise estimate for the upper limit. As MTC infection was expected to be rare, if present, in these primate populations, advanced statistical methods for rare diseases, where prevalence is expected to be close to 0, were also employed. Agresti–Coull prevalence shrinkage estimators and respective confidence intervals were compared with the maximum likelihood estimator. For comparison, the maximum likelihood prevalence estimator (P) is calculated as:

$$P = \frac{m}{n}$$

(1)

where m is the number positive and n is the number sampled. In contrast, the prevalence shrinkage estimator (P _Ψ ) for rare disease is estimated as:

$$P_{\psi } = \frac{{\left( {m + \psi } \right)}}{{\left( {n + 2\psi } \right)}}$$

(2)

where Ψ is a pseudo-frequency greater than 0, commonly set as z ²/2 (Newcombe 2012). When α is set to 0.05, z = 1.96 and Ψ = 1.92. Thus, by setting Ψ = 2, as done for the Agresti–Coull confidence interval (Newcombe 2012), the equation is reduced to:

$$P_{{\text{AC}}} = \frac{{\left( {m + 2} \right)}}{{\left( {n + 4} \right)}}$$

(3)

This method essentially adds two positive and two negative samples to the calculation of the prevalence estimate. The Agresti–Coull estimate is appropriate for situations where the true prevalence is expected to be close to 0, thus where disease may more easily be missed in surveillance; however, the shrinkage estimator approaches 0.5 where n is small (Newcombe 2012). Therefore, this comparison was made for the population-level prevalence estimates, as the Agresti–Coull prevalence estimator is expected to perform better for n > 40 (Newcombe 2012).

True Prevalence Estimation

True prevalence (TP) was calculated as:

$${\text{TP}} = \frac{{{\text{AP}} + {\text{Sp}}{-}1}}{{{\text{Se}} + {\text{Sp}} - 1}}$$

(4)

where AP is the apparent or maximum likelihood prevalence, P (Dohoo et al. 2009; Rogan and Gladen 1978). Sensitivity (Se) and specificity (Sp) estimates from our previous validation study were used as inputs for the calculation (Wolf et al. 2015). Ninety-five percent Wilson score confidence intervals were calculated in Epitools and were based on variance estimates that incorporate additional uncertainty associated with the sample sizes used in Se and Sp estimation (Rogan and Gladen 1978; Sergeant 2015; Wolf et al. 2015).