Bayesian Estimation of Pneumonia Etiology: Epidemiologic Considerations and Applications to the Pneumonia Etiology Research for Child Health Study

Abstract

In pneumonia, specimens are rarely obtained directly from the infection site, the lung, so the pathogen causing infection is determined indirectly from multiple tests on peripheral clinical specimens, which may have imperfect and uncertain sensitivity and specificity, so inference about the cause is complex. Analytic approaches have included expert review of case-only results, case-control logistic regression, latent class analysis, and attributable fraction, but each has serious limitations and none naturally integrate multiple test results. The Pneumonia Etiology Research for Child Health (PERCH) study required an analytic solution appropriate for a case-control design that could incorporate evidence from multiple specimens from cases and controls and that accounted for measurement error. We describe a Bayesian integrated approach we developed that combined and extended elements of attributable fraction and latent class analyses to meet some of these challenges and illustrate the advantage it confers regarding the challenges identified for other methods.

Keywords: .; Bayes theorem; epidemiologic methods; etiologic estimations; pneumonia; statistical models.

Figure 1.

Alternative analytic approaches used for determining pneumonia etiology. Abbreviations: BCX+, positive blood culture; Cor, coronavirus; Hinf, Haemophilus influenzae; HMPV, human metapneumovirus A/B; NP/OP, nasopharyngeal/oropharyngeal; PCR, polymerase chain reaction; Rhino, rhinovirus; RSV, respiratory syncytial virus; S. aur, Staphylococcus aureus; Spn, Streptococcus pneumoniae.

Figure 2.

Estimating the etiologic fraction from a study with 2 types of measurements, 1 with control data, and accounting for imperfect sensitivity of both measurements: the PERCH integrated analysis method. The PERCH integrated analysis can combine multiple specimens (shown here for 2 but can integrate more specimen/test measurements, such as whole-blood polymerase chain reaction [PCR] and lung aspirate culture and PCR) and adjust each measurement for pathogen-specific sensitivity to estimate the etiologic fraction using all available evidence. Abbreviations: BCx, blood culture; NP/OP PCR, nasopharyngeal/oropharyngeal polymerase chain reaction; PIA, PERCH integrated analysis.

Figure 3.

Analysis of 500 simulated datasets for a study that had only 1 specimen (A) and resulting etiologic fraction estimates using attributable fraction and PERCH integrated analysis methods (B). A, Analyses performed on measurements from 600 cases and 600 controls for each of 500 simulated datasets. *Prevalence and odds ratios estimated by averaging across the 500 datasets that were created based on the true etiology, sensitivity, and specificity values. B, Description of boxplots: Bold black line, mean of the true value across the 500 datasets; Diamond, average etiologic estimate across the 500 datasets; Vertical line through diamond, confidence interval around the average etiologic estimate; Boxplot, distribution of the etiologic estimates across the 500 datasets. Numbers above boxplots indicate numeric value of the diamond. Abbreviations: NA, not applicable; NoA, none-of-the-above; PIA, PERCH integrated analysis.

Figure 4.

Etiologic fraction estimates using PERCH integrated analysis: distribution of results from analysis of 500 simulated datasets containing cases with known etiology (A) and results from 1 randomly selected dataset (B). A, Pathogens A through D represent true pneumonia-causing pathogens that were tested for, pathogen E represents a pathogen that was tested for but does not cause pneumonia, and NoA represents pathogens that cause pneumonia but were not tested for. Slashes in table indicate not applicable for the pathogen. Description of boxplots: Bold black line, mean of the true value across the 500 datasets; Boxplots display the distribution of etiologic fraction point estimates from 500 simulated datasets: Diamond, average etiologic estimate across the 500 datasets; Vertical line through diamond, confidence interval around the average etiologic estimate; Numbers above boxplots indicate the numeric value of the diamond; whiskers denote the 5th and 95th percentiles of the etiologic fraction point estimates. B, Bubble plot presenting the PERCH integrated analysis results from 1 randomly selected dataset. The area of the bubble is proportional to the estimated etiologic fraction (number above the bubbles) divided by its standard error (ie, the larger the bubble, the greater the degree of confidence in the estimate). Abbreviations: BrS, bronze-standard data (imperfect sensitivity and imperfect specificity; eg, nasopharyngeal polymerase chain reaction); NoA, none-of-the-above pathogens; PIA, PERCH integrated analysis; SS, silver-standard data (imperfect sensitivity and perfect specificity; eg, blood culture).

Figure 5.

Impact of inaccurate or imprecise sensitivity priors on etiologic estimation from the PERCH integrated analysis using 500 simulated datasets: a sensitivity analysis compared with the base scenario presented in Figure 4. A, Impact of changing true bronze-standard (BrS) sensitivity for pathogen A from 75% (white) to 90% (gray). B, Impact of changing true bronze-standard (BrS) sensitivity for pathogen B1, B2, and B3 from 75% (white) to 90% (gray). C, Impact of changing true silver-standard (SS) sensitivity for pathogens B1, B3, B4, C1, C3, D1, and D3 from 15% (white) to 5% (gray) and changing SS sensitivity prior from 5%–25% (white) to 10%–20% (gray). D, Impact of increasing width of SS sensitivity prior from 5%–25% (white) to 1%–50% (gray). Description of boxplots: Bold black line, mean of the true value across the 500 datasets; Boxplots display the distribution of etiologic fraction point estimates from 500 simulated datasets: Diamond, average etiologic estimate across the 500 datasets; Vertical line through diamond, confidence interval around the average etiologic estimate; Numbers above boxplots indicate the numeric value of the diamond; whiskers denote the 5th and 95th percentiles of the etiologic fraction point estimates. Abbreviation: BrS, bronze-standard data (imperfect sensitivity and imperfect specificity; eg, nasopharyngeal polymerase chain reaction); NoA, none-of-the-above (ie, pathogens not tested for); SS, silver-standard data (imperfect sensitivity and perfect specificity; eg, blood culture).