9 September 2014

18 minute read

The objectives of this paper are to

1) Describe the requirements to the epidemiological tools for herd diagnosis

2) Provide examples of tools we have developed during recent years.

Monitoring and risk factor analysis are the key components, writes Carsten Enevoldsen of the Royal Veterinary and Agricultural University of Denmark.

The major challenge for the user is to select the best balance between sensitivity and specificity. The quality and utility of epidemiological tools may be characterized by four criteria: Correctness, validity, precision, and transparency.

Graphical tools and options for displaying recordings for individual animals are effective tools to ensure correctness of recordings. A series of available tools are described. Several types of smoothing techniques are implemented.

The advantages of an Internet platform for development and implementation of epidemiological tools for herd diagnoses, barriers to implementation of epidemiological tools and options for promoting implementation in cattle practice are discussed.

Introduction

A diagnosis in an individual animal can be described as a complex pattern of deviations from a more or less well defined normal state. The disease process in individuals is usually of a relatively short duration. Consequently, it is reasonably easy for a veterinarian or a skilled farmer to judge whether the condition is changing (monitoring). E.g., it is quite easy to judge whether a treatment of a lame cow has the expected effect.

The herd shall be seen as an organism like the individual cow. The animals, the housing system, the management, the farmer, and the interactions between these components together form a production system. This system is also dynamic.

Changes in the input to the system (e.g. the feed ration or management routine) may cause responses very quickly but changes may also take months or even years to take effects because of the long calving intervals, climatic effects etc. Such long term changes are very difficult to detect without access to correct (systematic), valid and precise data records that are transformed to appropriate information. If the relations between changes in input and responses are not revealed (e.g., the effect of a new drug for mastitis treatment or a management change are not assessed correctly), the decision maker may use the available (scarce) resources inefficiently. Consequently, the veterinarian working with herd health needs a set of tools to describe this dynamic system efficiently and to estimate the relations between input and output with sufficient validity and precision.

The objectives of this paper are to 1) describe the requirements to the epidemiological tools for herd diagnosis and 2) provide examples of tools we have developed during recent years.

Requirements to the Epidemiological Tools for Herd Diagnosis

Epidemiology and statistical analysis are not the same

When the clinician works with an individual animal she or he evaluates changes over time, compares the observations with norms (reference values) or maybe simply compares the individual of interest with its (comparable) herd mates.

The same process applied to a population (e.g. a herd) essentially is epidemiology, which (often) is defined as "the occurrence and spread (distribution), the determinants (risk-factors), and (biological) consequences of health disorders in populations". This definition also indicates that epidemiology is not merely statistical analysis.

Epidemiological analyses require a solid insight into the "biology" of the population of interest. E.g., the biological and pathobiological responses of individuals and management processes are factors that determine population structure and dynamics. However, statistical analysis is necessary to perform proper epidemiological analyses.

Monitoring and risk factor anaysis

Epidemiology serves two major purposes in herd (health) management: Monitoring and risk factor analysis. Key indicators of health, fertility, production efficiency etc. are calculated and used to monitor performance. Usually the values of the indicators are compared to results obtained in other herds or compared to historical results in the same herd (“targets”). This proces is often called bench-marking.

In a risk factor analysis the response of interest (e.g. milk production, SCC or ketosis) is estimated for each of several sub-populations characterized by (risk) factors, which we hypothesize are affecting the response of interest. The statistical analyses will primarily provide estimates of the magnitude of the relations (e.g. average milk loss associated with ketosis) and the uncertainty associated with the estimate (usually some P-value or confidence intervals).

The epidemiologist will then formulate a herd-level diagnosis by means of a synthesis of these statistical estimates together with general and specific knowledge about biological, technical and managerial aspects related to the system, issues related to data quality, and common sense. Sometimes this process is called a “causal analysis” (Markusfeld, 1993). However, the distinction between monitoring and risk factor analysis is not clear because monitoring can also be seen as a mono-factorial analysis with calendar time as the “risk factor”.

Sensitivity and specificity

For obvious reasons all the information derived from data are describing historical events or developments. The decision taker will usually feel more confident with recent data compared to older data because he or she usually will prefer a tool to monitor the production that is highly sensitive to recent change.

However, if only the most recent data are used, the uncertainty associated with the estimates derived from the statistical analyses will increase because of the direct link between uncertainty and the number of observations. Including more historical data, in contrast, will reduce the risk of getting false positive results (increase specificity).

False positives may be very important if it is costly to react to an alarm (e.g. it is costly to go and find 10 cows without mastitis for every one with mastitis). The search for an optimal balance between sensitivity and specificity of the monitoring tools is a major issue in rational herd management. In addition to these statistical aspects, historical events may also be caused by completely different factors.

Criteria to assess the quality and utility of epidemiological tools

The quality and utility of epidemiological tools may be characterized by four criteria: Correctness, validity, precision, and transparency.

Correctness: It may seem obvious that data and information used for management must be correct. However, there are numerous possibilities to encounters errors. Data recording may be incorrect, data transfer to data files may fail sporadically, and computer programs may have bugs. It is not easy to avoid such errors completely.

If the user knows only little about data management and data is agregated, even serious errors may remain undetected for quite a while. Efficient logical checks that reduce recording errors can be incoorporated in the recording systems but such facilities requires much effort to design efficiently. If the raw data (e.g. calving dates) are used directly for operational herd management the correctness of data will be maximized.

The validity of a measurement tool is defined as the average ability of the tool to provide the correct estimate of the condition of interest. For example, the rate of abortions calculated from the available data may be higher or lower than the true rate due to use of an incorrect method for estimation of the rates. The term bias may also be used to describe poor validity.

Precision is defined as the degree of repeatability (or reproducibility) of the measurements on the same observational units. It is very desirable to have a high degree of precision because it will allow us to adjust our results if validity is poor. Poor precision can only be handled by increasing the number of measurements and only if validity is good or the bias is known.

Transparency means that the users have possibilities to see how data and information is derived, handled and used. Graphical display of raw data and systems that allows the user to change assumptions (sensitivity analysis) are tools to increase transparency. It is a key characteristic of these tools that they seek to present the identification numbers and, if possible, major characteristics of the animals included in the analyses.

This facility gives the users better options for evaluating the data quality, which obviously is important for the validity of the results. In this way we hope to maximize the users’ trust in the data analyses. When the number of observations is low, one or few observations can be very influential if they have very strange characteristics. Such influential observations are more likely to be identified when individual animals are visualized.

Examples of Epidemiological Tools for Herd Diagnosis

Monitoring – general issues

Numerous tools are available to utilize data for monitoring. So-called smoothing techniques are particularly well suited. The best known probably are rolling averages. E.g., if the average of the 12 latest monthly estimates of pregnancy rates is calculated every month, occurrence of erratic changes will be reduced (increased specificity).

The major problem with this approach is that a change in the latest rolling average can be caused by the exclusion of the pregnancy rate 13 months ago. Thysen & Enevoldsen (1994) has suggested an alternative approach that puts more weight on the most recent observations and includes tools to handle the uncertainty in a consistent way. Similar methods have been developed to handle continuous data like SCC and electrical conductivity (Thysen, 1993).

The near future probably will bring several new techniques to perform "smoothing" of more complex data with the purpose to explore time trends or patterns in the data. For example, it will be possible to include several (time) factors in the analysis. In that case monitoring in fact becomes similar to riusk factor analysis (se below).

In larger herds with more than approximately 150 dairy cows virtually all the epidemiological and statistical techniques available to research can be applied. The limiting factors are insufficient data quality and lack of users with sufficient experience with such analyses and application of the results. Application of rather advanced methods on a national scale has been in place in Israel for several years (Markusfeld, 1993). In smaller herds a number of problems arises due to the smaller number of cows.

However, with proper modifications, most of the principles described in this paper can be applied even down to herd sizes as low as 40-50 cows.

Risk factor analysis – general issues

Several computer programs for centralized herd management or stand-alone PC-programs provide some risk factor analyses. Usually responses to one factor at a time are estimated. Some – probably rather few - practicing consultants also download data and conduct ad hoc analyses by means of spreadsheets or similar rather simple tools. Such mono-factorial analyses do provide some insight into the causal effects of various factors in the herds. However, often more factors must be taken account to see a more complete picture.

An example where a multi-factorial approach obviously is needed is a comparison of the risk of milk fever (MF) in a group of summer calvings with the risk among winter calvings. If we knew that there were more old cows in the winter group than in the summer group, we clearly would not accept a simple (mono-factorial) comparison because we know that the MF-risk increases with parity.

Then we could calculate the MF-risk separately by parity, perform significance tests, and present the results separately. That would be correct, but we would have a complicated result with many numbers and it is likely that the statistical tests would indicate weak and inconclusive associations because there would be very few observations in most parities.

To solve these problems, we could calculate some sort of weighted average MF-risk for all parities to give each group (summer or winter) an equal parity frequency. Such standardization has been widely used and is correct although statistical tests become rather complicated to perform. Often herd results are standardized to, for instance, 40% 1st, 30% 2nd, and 30% later lactation cows. We could instead handle this problem by using a multi-factorial model that effectively takes a factor like parity into account. Details are beyond the scope of this paper. Such multi-factorial analyses were first implemented on a larger scale in practice by the cooperative veterinary service (the Hachaklait) in Israel (Markusfeld, 1993).

HerdView – a tool for monitoring and monofactorial risk factor analysis

HerdView is a freeware program that uses reproductive and disease events during a 60-week period (http://web.agrsci.dk/jbs/ith/Www-herdview/index.html - includes a maual). It presents individual cows by lifelines with event occurrences shown by symbols and herd performance by graphs of weekly event counts. A robust change-point procedure reveals changes over time in herd reproductive efficiency, and age-dependencies are visualized by survival graphs. Analyses can be stratified by parity and selected subsets of cows for separate mono-factorial risk factor analyses. The user has access to all details about the records and the program thus provides a high degree of transparency.

Veterinary Production Analysis (VPA) for monitoring and multifactorial risk factor analysis

The main aspects of the Hachaklait system (Markusfeld, 1993) has been implemented in a system adapted to the generally smaller herd sizes in Denmark (www.dhd-vpa.dk). The VPA contains both monitoring of numerous performance indicators and analyses of several risk factors based on linear and logistic multivariable models. Problems related to the smaller herd size (currently about a 100 cows on average) have been handled as described by Enevoldsen (1997).

The VPR System – a web platform for practicing cattle veterinarians

A steadily increasing number of dairy veterinarians are recording clinical findings in well defined groups of high risk cows in Danish dairy herds. These recordings make it possible to detect disease in the individual cow earlier and intervene effectively. This will reduce production loss at cow level. Systematically recorded clinical findings can also be used to obtain valuable information about the herd. The number of clinical registrations from one herd increases rapidly.

To become useful, these registrations have to be merged with information about calvings, milk production etc. It is almost impossible to do this by hand and achieve information about the development over time in the specific herd. We have developed an Internet portal that helps the practicing dairy veterinarians to solve these problems. We call it The VPR System. The letters VPR are taken from the full Danish name: “Veterinær ProduktionsRådgivning”, which can be translated to Veterinary Production Consultancy.

Compared to stand-alone PC-programs, a web system allows easier implementation and updating of advanced statistical models because all changes in software and hardware are made centrally at the server. Consequently, the end-users do not have to know much about the technical issues. It is absolutely necessary to have a fast Internet connection (broad band) because several output files are 1-2 MB. You get access to the login page of The VPR System at this link: http://130.226.13.142/. A short introductory text (in Danish) describes the terms of use.

The VPR System is non-commercial. The system aims at giving dairy practice a platform for development and utilization of clinical observations recorded from systematic dairy herd health programs. Data files stored in the system will be used for research and development purposes. Strict confidentiality will be observed. Any use of the output files for consultancy or management puposes is completely at the responsibility of the users.

The use of The VPR System is independent of other people, which makes is possible to use the entire system and all the features whenever you want to. The data collection is supported with error checking and error correcting facilities. It is always possible to get your clinical registrations out of the system as an Excel-spredsheet file. The VPR System can create the data files needed to use the HerdView program.

Tools to record and utilize systematic clinical recordings

The VPR System contains a facility for data entry and creates an action plan for farmers or veterinarians. The plan contains cows that have to be examined or handled (e.g. drying-off and start of insemenation). Information from clinical registration is included and it is possible for the user to define herd-specific time intervals or criteria that are necessary to create a completely herd specific plan. This facility promotes correctness of both data recording and data management.

There is a series of graphical analyses that describes the development in clinical registrations over time (monitoring). The menu currently includes scores of Body Condition (BCS), CMT, vaginal discharge, clinical condition of the udder, ketone reactions in milk and urine, hock condition and body measurements of heifers. Most plots contain output from rather complex and advanced statistical models to describe trends in the average status in subsets of the herds (smoothing).

BCS is usually recorded at drying off and at calving. We have developed a plot that shows the distribution of BCS in relation to drying off and calving. Scores for each cow are connected by lines. Because it is a generally accepted goal to avoid loss of BCS during the dry period, cows with BCS loss are marked with red lines and their cow number is given at the x-axis. Cows without BCS loss are marked with blue.  

Consequently, a plot like this allows quite intense monitoring both at cow and herd level. At herd level, an ideal situation would be a plot with parallel blue “cow lines”. The plot will be very a very sensitive indicator of emerging trends over time. The plot also shows the stage of lactation where the BCS was measured after calving (the left Y-axis). This facility allows the user to monitor how well the data recording follows the recording schedule. Similar plots are currently available for scores like CMT, vaginal discharge, ketone reactions, and contusions.

Lactation curve analysis

The VPR System contains a concept to provide a visual analysis of parity specific lactation curves during a two-year period. Lactation curves from individual cows are shown. The curves are predicted by means of a so-called mixed model that takes into account correlations between milk recordings. Other curves represent average energy corrected milk yield at 10, 60 and 305 days in milk (DIM).

Finally, there is an average lactation curve for the entire observation period. Consequently, this plot presents the following parameters visually: Average parity-specific yield at various stages of lactation, changes in the averages during the observation period and variability of these parameters (e.g. parallel lines will indicate a minimum variability in persistency).

Amount and contents (fat and protein) of milk is the essential output from dairy herds. Changes in the profiles of milk production and causes of these changes during time provide very important information about production efficiency and health. The principles used for analysis of yield are applied to fat and protein as well. The estimates also allow us to benchmark against other similar herds. Currently we are exploring the potentials and barriers towards specifying the model as a multi-herd model with built-in benchmarking derived from variance components. For research purposes we have used the model to estimate effects of potential determinants of milk production loss (Bennedsgaard et al., 2003).

We are currently working on supplementing the baseline model with options for interactive estimation of such effects within herd. We used a multilevel random effects model to assess the effects of various types of clinical herd-level intervention on the average milk yield 9-92 dpp. In one prototype analysis the main effect of intervention was highly significant (p<0.001). The average estimate of the intervention was 0.8 kg milk but with huge differences between herds. It seems straight forward to adapt our models to other types of outcomes and interventions. We are also working on including pedigree information in our models. This probably will improve the precision of the herd specific estimates derived from the models because we can utilize the information from sires with offspring in multiple herds.

Barriers to implementation of epidemiological tools for herd diagnosis and options for promoting implementation

Despite the availability of numerous rather advanced tools for proper epidemiological analyses of herd data, rather few practicing consultants apparently use those tools. Why is that so? One obvious explanation is that the consultants have not received a proper training at the university. This is probably true. Most training in statistics and epidemiology until now seems to have been detached from real life examples. It is our experience that courses based on real herd data do improve the analytical skills among veterinary students considerably (Enevoldsen & Houe, 2003). However, post graduate training may be more efficient because the consultants at that time have practical experiences that allow them to appreciate the need for quantitative analyses based on a scientific approach (specification and test of hypotheses).

Another explanation to the consultants’ still limited use of proper analyses could be that they do not see a need that justifies the additional efforts. At a recent seminar for Danish cattle consultants there were some indications (unpublished) that many consultants were satisfied with qualitative assessments of the status in the herds. Consequently, there seems to be a need for demonstration of the need for epidemiology in practice.

In the most advanced version of monitoring, actual results are compared with results predicted by means of some prognostic (simulation) model (Enevoldsen et al., 1995; Østergaard et al., 2004). Apparently, very few such prognostic models are used in practice. However, simulation models may have a considerable potential for promoting the use of efficient epidemiological tools because simulation demonstrates the need for estimation of herd specific parameters and provides an efficient tool to combine these estimates into whole herd evaluations. If the use of epidemiological tools and synthesis of information by means of simulation is combined with an ongoing dialogue with the farmer, values and attitudes of the farmer may be revealed (Andersen & Enevoldsen, 2006). In that way we may create a whole-herd oriented framework for computer-assisted learning.

Literature