DE LA IGLESIA EN ARGENTINA
4. Criterios pastorales comunes
clude the International Classification of Diseases (Anonymous 1977) and the Diagnostic
and Statistical Manual of Mental Disorders (DSM-4).9 The DSM-4 lists categories of
psychiatric disorders and their associated diagnostic criteria, as defined by the Ameri- can Psychiatric Association. It is used globally by clinicians and researchers as well as pharmaceutical companies and policy makers. An example of syndromic groupings used in a veterinary setting would be the disease classification system for dairy cattle used by Livestock Improvement Corporation (New Zealand, Livestock Improvement Corpora- tion, 2008). Roberts et al. (2006) provide details of a system where equine diseases are classified by body system.
Data providers for veterinary surveillance systems should include farm managers, veteri- narians, drug companies, abattoirs, and diagnostic laboratories. Each of these collect a range of information in a format which is often difficult (if not impossible) to be com- bined into a format that is able to be used for meaningful epidemiological analysis. I propose that a set of standards need to be devised for recording of health and produc- tion information in domestic animals. This would provide two benefits. Firstly, it would allow information from a range of different sources to be assimilated, enhancing the over- all sensitivity of the system to detect outbreaks and emerging disease conditions. The second benefit is that, once developed, system components (e.g. animal-farm databases, animal movement databases) could be sold on to other countries. This would offset costs for the country developing the technology and dramatically reduce deployment costs for the country purchasing the technology, effectively eliminating the need to ‘reinvent the wheel’.
3.5
Methods for incomplete data
An important aspect of veterinary surveillance is the provision of valid estimates of dis- ease frequencies which may be used to inform policy decisions on control measures to be implemented. Valid estimates of disease frequency are possible when all case events have been captured by the system and updated data on the population at risk are available. The reality in animal health is that data obtained from the many sources may be subject to bias (selection and misclassification), which in turn decreases the validity of the correspond-
9
ing estimates of disease frequency. As a result, decisions informed by these estimates may be flawed. Sources of bias include underreporting and inconsistent reporting, which may vary with the source of the data and over space and time. Recognising that data gathered from surveillance systems generally reflect a proportion of the actual number of cases that occur (under ascertainment) a number of approaches, pioneered for AIDS surveillance and wildlife management, provide a means by which one can estimate the number of unrecognised cases. These methods include back-calculation or back extrap- olation (Brookmeyer & Gail 1986, 1988), capture-mark-recapture approaches (Hook & Regal 1995), and epidemic transmission models (Wang & Ruan 2004).
Back-calculation methods were developed during the early years of the HIV/AIDS epi- demic to estimate the incidence of infection on the basis of reported clinical cases and the distribution of incubation periods (Brookmeyer & Gail 1986, 1988). The method is based on the principle that the number of clinical cases of disease in a population will be dependent on the number of infected individuals present and the length of the incubation period (Donnelly et al. 2003, Brookmeyer 2004). This method has been used to estimate the incidence of BSE in the United Kingdom (Anderson et al. 1996, Donnelly et al. 2002, 2003) and France (Supervie & Costagliola 2004, 2007). Supervie & Costagliola (2004) used a modification of the back-calculation method to estimate the age and year-specific incidence risk of BSE in French cattle between 1990 and 2001 based on a previous study which showed that 20% of cases were identified by passive surveillance. These authors estimated that 301,200 (95% CI 27,600 – 837,600) cattle were infected with BSE during the study period, a number many times greater than the 103 cases that had been iden- tified at the time by passive surveillance. This method is limited by its dependence on the observed number of cases which are themselves prone to underreporting. In addition, knowledge of the distributional form of the incubation period is critical. Reviews of these methods and their application to AIDS and BSE are provided by Donnelly et al. (2003) and Brookmeyer (2004).
Capture-mark-recapture methods have been developed to estimate the size of wild animal populations on the basis of capturing, marking, releasing and recapturing animals over a period of time (Seber 1982). These methods have since been used in a number of non-ecological settings, including studies to estimate the level of under-counting in a population census (Darroch et al. 1993) and to estimate the level of underreporting of
3.5 Methods for incomplete data 73
health events to provide adjusted estimates of disease frequency (Hook & Regal 1995, LaPorte et al. 1995). Examples of situations where capture-mark-recapture methods have been used in human epidemiology include the assessment of the most accurate data source for estimating the frequency of adolescent injuries (LaPorte et al. 1995), assessment of the surveillance sensitivity for sexually transmitted disease in The Netherlands (Reintjes et al. 1999), estimation of the prevalence of malaria in The Netherlands (Hest et al. 2002) and estimation of the incidence of stroke in the United Kingdom (Tilling et al. 2001). These examples are based on data aggregated over a single time period. Capture-mark-recapture methods use statistical models to aggregate data captured by multiple data sources (health registries, surveillance databases, birth or death registries) that contain incomplete and partially overlapping data as well as sources that may not be independent. The total number of cases of disease is then computed as the sum of the observed cases and the estimated number of unobserved cases from the capture-mark-recapture model. The use of these models rely on the following assumptions (Hook & Regal 1995):
• the population of interest should be constant or closed during the study period;
• information recorded on each unit of interest in the separate data sources must have a common, unique identifier to facilitate matching of information from different data sources;
• each unit of interest should have the same ‘catchability’ (that is, an equal probability of being monitored); and
• data sources must be independent (the probability of a unit being captured by one source does not depend on the remaining sources).
A number of models are available to conduct capture-mark-recapture analyses and their use will depend on whether the data sources are considered to be independent (two source models) or dependent (log-linear models) (International Working Group for Disease Mon- itoring and 1995). Examples of the use of capture-mark-recapture methods in veterinary science are rare. To the best of my knowledge, only two veterinary examples have been published: del Rio Vilas et al. (2005) and B¨ohning & del Rio Vilas (2008). In Great Britain del Rio Vilas et al. (2005) used capture-mark-recapture methods to estimate the number of holdings infected with scrapie as well as to estimate the sensitivity of three
scrapie surveillance systems. The authors used three capture-mark-recapture methods (two source models, log-linear models, and a sample coverage method) to aggregate data from three scrapie surveillance data sources: (1) statutory notifications of scrapie-positive holdings recorded within the Scrapie Notification Database (SND) (n= 141), (2) positive holdings (n= 67) from an abattoir survey (AS) of sheep greater than 18 months of age and, (3) scrapie-positive holdings (n = 12) from a fallen stock (FS) survey between January 2001 and April 2002. Using the two source model approach, the data sources were treated as independent and the estimated number of missing scrapie-positive holdings obtained from pairwise combinations of the data sets (i.e. SND-AS, SND-FS, and AS-FS), ignor- ing the third. The number of scrapie-positive holdings missed by each data source was 936 for the SND-AS comparison, 170 for the SND-FS comparison, and 336 for the AS- FS comparison. The estimated number of missed cases from the SND-AS comparison was 5.5 times greater than the number missed by the SND-FS comparison. This illus- trates one of the major limitations of using capture-mark-recapture methods when data sources show some level of dependence. In the second approach, a series of log-linear models were fitted under various assumptions of independence or dependence between sources. The most significant model estimated a total of 1,653 (95% CI 354 – 6,434) missed scrapie-positive holdings under the assumption that data recorded in the SND and AS were related. The prevalence of scrapie-positive holdings in Great Britain was esti- mated to be 0.82%. The Rcapture package (Baillargeon & Rivest 2007) implemented in the statistical software package R (R Development Core Team 2008) provides a compre- hensive and accessible set of tools for analysing capture-mark-recapture data.