Still 'dairy farm fever'? A Bayesian model for leptospirosis notification data in New Zealand

Abstract

Routinely collected public health surveillance data are often partially complete, yet remain a useful source by which to monitor incidence and track progress during disease intervention. In the 1970s, leptospirosis in New Zealand (NZ) was known as 'dairy farm fever' and the disease was frequently associated with serovars Hardjo and Pomona. To reduce infection, interventions such as vaccination of dairy cattle with these two serovars was implemented. These interventions have been associated with significant reduction in leptospirosis incidence, however, livestock-based occupations continue to predominate notifications. In recent years, diagnosis is increasingly made by nucleic acid detection which currently does not provide serovar information. Serovar information can assist in linking the recognized maintenance host, such as livestock and wildlife, to infecting serovars in human cases which can feed back into the design of intervention strategies. In this study, confirmed and probable leptospirosis notification data from 1 January 1999 to 31 December 2016 were used to build a model to impute the number of cases from different occupational groups based on serovar and month of occurrence. We imputed missing occupation and serovar data within a Bayesian framework assuming a Poisson process for the occurrence of notified cases. The dataset contained 1430 notified cases, of which 927 had a specific occupation (181 dairy farmers, 45 dry stock farmers, 454 meatworkers, 247 other) while the remaining 503 had non-specified occupations. Of the 1430 cases, 1036 had specified serovars (231 Ballum, 460 Hardjo, 249 Pomona, 96 Tarassovi) while the remaining 394 had an unknown serovar. Thus, 47% (674/1430) of observations had both a serovar and a specific occupation. The results show that although all occupations have some degree of under-reporting, dry stock farmers were most strongly affected and were inferred to contribute as many cases as dairy farmers to the burden of disease, despite dairy farmer being recorded much more frequently. Rather than discard records with some missingness, we have illustrated how mathematical modelling can be used to leverage information from these partially complete cases. Our finding provides important evidence for reassessing the current minimal use of animal vaccinations in dry stock. Improving the capture of specific farming type in case report forms is an important next step.

Keywords: Markov chain Monte Carlo; data imputation; epidemiology; leptospirosis.

Figure 1.

Posterior median and 90% CI for the expected number of notified leptospirosis cases from each occupation in each year. Crosses indicate the number of cases reported for each occupation.

Figure 2.

Posterior median and 90% CI for the proportion of notified leptospirosis cases from each month by occupation. Crosses indicate the proportions in the observed data. Horizontal line shows equal proportions.

Figure 3.

Posterior median and 90% CI for the proportion of notified leptospirosis cases from each serovar by occupation (circles). Crosses indicate the proportions in the observed data with 90% Goodman multinomial CIs.

Figure 4.

Posterior median and 90% CI for the inferred occupation of notified leptospirosis cases recorded as farmer (left plot) or unknown (right plot).

Figure 5.

Posterior median and 90% CI for the proportion of notified leptospirosis cases with each recorded occupation, split by inferred occupation in the four panels. This presentation is the reverse of figure 4.

Figure 6.

Posterior histogram for the probability that a notified leptospirosis case has a serovar recorded. Cross and horizontal lines indicate the proportion observed in the data and 90% binomial CI.