Cost implications of improving food safety in the Dutch dairy chain

1. Introduction

The increase in demand for safer food has triggered an integrated approach for improving food safety throughout the entire food chain (‘farm to table’), which emphasises responsibility for food safety at all stages of the food chain, including animal feed production, primary production, food processing, storage, transport and retail sale. To support this approach legally, the General Food Law Regulation (EU Directive 178/2002) was adopted in 2002 ( European Commission, 2002 ). This regulation lays down principles and obligations covering all stages of food production and distribution that must be gradually adopted by January 2007. Besides regulatory incentives, each chain participant has private incentives to prevent food safety incidents, because their impact can be devastating (e.g. reputation loss, product price reductions, temporary or permanent shut down of production facilities, lawsuits, increase of premium for product liability and decrease of product demand) ( Unnevehr and Jensen, 1999 ; Buzby et al. , 2001 ). Consequently, driven by both regulatory and private incentives, chain participants are encouraged to ensure food safety by adopting various control measures, and they need to find cost-effective ways of doing so.

The potential costs of food safety improvement and methods for evaluating them are discussed in the literature (e.g. Antle, 1999 ; Valeeva et al. , 2004 ; Valeeva, 2005 ). Existing studies on the costs of improving food safety focus mainly on the processor ( Gould et al. , 2000 ; Jensen and Unnevehr, 2000 ) and retailer ( Mortlock et al. , 2000 ) segments of the chain. The few studies that are available for the whole chain explore the cost-effectiveness of particular packages of control measures (e.g. van der Gaag et al. , 2004 ) in which cost-effectiveness is expressed as a ratio of change in impact (such as reduction in the prevalence of Salmonella at the end of the chain) relative to change in costs associated with pre-determined (fixed) packages of interventions. Little information is available to guide agents along the chain in choosing the cost-effective set of control measures to achieve a certain food safety level. Such information may help prioritise opportunities for allocating resources to improving food safety within the chain ( Valeeva et al. , 2004 ). The objective of this research is to identify cost-effective strategies for attaining various levels of food safety improvement and to examine how, for each level, the total effort is best allocated among the chain participants. In particular, the paper analyses (i) the contribution of chain participants to improving food safety, (ii) cost-effective combinations of control measures along the chain and (iii) extra costs associated with improving food safety and their distribution among chain participants.

Various alternative strategies could be adopted to achieve any given food safety level, and they may differ substantially in terms of cost. In this study, we quantify the cost of alternative strategies associated with achieving given levels of food safety improvement. The least-cost strategy for achieving a given level of food safety improvement is identified as the cost-effective strategy for that level.

The cost-effective strategies for each level of food safety improvement in the chain are solutions to an integer linear programming (ILP) model. Although chain participants are in fact independent of each other, the model assumes that they behave as a single actor looking for a cost-effective way to improve the chain's food safety. Thus, the model developed is for the limiting case of a chain regulator (leader) who is able to influence chain participants' decisions to achieve a desired food safety level. In reality, regulations, societal pressure, market requirements or own conviction might be incentives for some chain participants to implement additional control measures earlier than other participants. It follows that some participants may implement many control measures, whereas others adopt few additional measures for improving food safety. A more realistic approach should then account for the incentive structure within the chain. However, as shown by Hennessy et al . (2001) , incentives are far from efficient for optimal food safety provision along the chain: in fact, leadership by a firm or group of firms can enhance regulation by motivating other chain participants to improve food safety.

This paper analyses the dairy production chain for fluid pasteurised milk as a case study. Although dairy products are deemed one of the safest food groups, safety is a prime concern since hazards originating from dairy products could affect a large number of consumers. The paper is structured as follows. Section 2 elaborates on the general framework, assumptions, data and model for evaluating the cost-effectiveness of different strategies for improving food safety. Section 3 presents the results. Finally, Section 4 reports conclusions and discusses the implications of the results and contexts in which our approach can provide insights.

2. Methodology

2.1. General outline

Each strategy for improving food safety over the chain involves prevention, elimination or reduction of food safety hazards by means of various control measures at each chain stage. These measures include both specific interventions and new methods of managing process control ( FAO, 1997 ). The challenge is to find which set of control measures allows the chain as a whole to achieve a required level of food safety at the lowest cost.

An ILP model was developed to identify cost-effective strategies for attaining different food safety levels over the entire dairy chain. Linear programming (LP) is a mathematical technique designed to find the maximum or minimum of the linear function of decision variables subject to a set of constraints defined by linear inequalities ( Hazell and Norton, 1986 ). ILP is a special case of LP in which all decision variables are restricted to the integer values.

The study started with defining strategies for improving food safety as a set of control measures that could be implemented at different action points within the chain. The impact of the measures on food safety improvement was assessed by experts and formalised using adaptive conjoint analysis (ACA). The extra costs of implementing and maintaining these measures were calculated using partial budgeting. This information on alternatives, impacts and costs was used to build the ILP, which minimised the costs of attaining different food safety levels in the dairy chain.

Consumed fluid pasteurised milk is taken to be the end product of the whole chain. Food safety improvements are focused on two main groups of food safety hazards that are essential to the production of safe fluid pasteurised milk: chemical (antibiotics and dioxin) hazards and microbiological ( Salmonella , Escherichia coli , Mycobacterium paratuberculosis and Staphylococcus aureus ) hazards. ¹

The chain includes three blocks that can affect the food safety level of the end product: feed, farm and dairy-processing blocks. Strategies for improving food safety in the dairy chain include specific action points where control measures can be implemented. Action points and corresponding control measures refer to purchase, transport and storage of compound feed ingredients, identification and traceability of compound feed and its ingredients, the design of the production facilities at the feed plant, production practices and hygiene conditions for compound feed production and transport (feed block); purchase or production of feed, grazing of pastures, cattle movement and traceability, herd health and treatment, dairy cattle housing, calving and feeding of calves, water management, hygiene conditions on the farm (farm block); transport of raw milk to the processing factory, the design of production facilities at the dairy plant, production practices and hygiene conditions for raw milk processing and delivery of pasteurised milk to the sale unit (dairy-processing block).

Key action points (75 in total) and control measures relevant to them (186 in total) are presented in Table  1 . Within each action point, chain participants must decide which measure among those available to implement. These measures serve as main decision variables in the model in which their integer values are required to be binary, depending on whether or not a measure is selected for implementation within a certain action point.

Table  1 shows that control measures contain specific interventions, or actions, in the production process—such as ‘culling cows in salmonellosis and M. paratuberculosis cases’—as well as the new methods of managing process control—such as ‘compliance by feed ingredient manufacturers with Good Manufacturing Practice ⁺ (GMP ⁺ ) ² or another certified quality assurance (QA) system equivalent to GMP ⁺ '. Some measures are relevant for the control of both chemical and microbiological hazards. Only measures that are currently available for implementation and that comply with the European law are considered.

2.2. Food safety input

The impact of control measures on food safety tends to be quantified in different units that often cannot be easily compared. For example, pathogen reduction in pork can be measured in terms of colony-forming units per square centimetre ( Jensen et al. , 1998 ) or of pathogen prevalence expressed as a percentage ( van der Gaag et al. , 2004 ). In addition, for measures that require a lot of administration (e.g. ‘monitoring’), it is difficult to obtain an absolute measure of effectiveness. Therefore, this study uses experts' perceptions to obtain consistent quantitative information on the impact of control measures for improving food safety. Data on the perceptions of 22 experts from research, industry, regulatory authorities, extension and farmers' organisations were collected via two workshops and personal interviews (four experts represented the feed block, 13 the farm block, and five the processing block). ³ Initially, the impact of control measures was assessed per block, and separately for chemical and microbiological improvements, using ACA ( Green et al. , 1991 ; Johnson, 1991 ).

ACA is a widely used analytical tool from the group of conjoint measurement techniques. Conjoint measurement assumes that an individual's preference for a multi-attribute product (say, an apple) can be represented by a subjective utility function, which is a monotonic function of interval scale utilities, which are simple additive functions of what are often called partworths , or subjective utility values, associated with different levels of the discrete product attributes (e.g. ‘sweet’ and ‘bitter’ are different levels of the attribute ‘taste’) ( Churchill, 1999 ; Lattin et al ., 2003 ). Following this approach, an expert's preference for a particular strategy for improving food safety can be expressed as a function of the partworths of control measures that can be implemented within action points associated with improving food safety (where action points are equivalent to attributes and control measures to attribute levels) ( Valeeva et al ., 2005a , b ). For examples of action points and control measures, see the first column of Table  1 .

In brief, the current application of conjoint measurement uses ratings obtained from each expert who scores the overall preferences for a set of hypothetical strategies for improving food safety. Each strategy represents a systematically predetermined combination of action points, each of which is represented by only one of the control measures corresponding to the action point. On the basis of these ratings, the conjoint measurement procedures identify a mathematical function, which is interval-scaled and which corresponds best to the set of subjective evaluations (ordinal judgments) of the strategies made by the expert.

Experts' choices between different strategies reveal actual trade-offs between control measures at different action points (rather than dealing with measures within single action points), similar to choices that often have to be made in practice.

Conjoint analysis methods can differ according to the form of utility function, data collection method, construction and presentation of strategies for improving food safety, estimation techniques. However, they all attempt to find a set of partworth utilities (reflected by regression coefficients) in a function that relates the control measures being considered within the strategy (represented by vector of dummy variables) to overall preference or utility (expert's rating) of that strategy. A set of partworths for the separate control measures is defined so that they sum to the total utility of each strategy. Most conjoint analysis methods can study only a limited number of action points (often six or seven at most). ACA, by contrast, combines ‘self-explicated’ and conjoint measurement approaches to allow up to 30 action points to be included ( Green and Srinivasan, 1990 ).

Six computerised ACA questionnaires, one each for chemical and microbiological food safety for each of the three chain blocks, were analysed using the ACA software ( ACA User Manual System, 2002 ). Each questionnaire included four sections. In the first section, for each action point, the expert rated the control measures according to their perceived contribution to improving food safety (1 = smallest contribution, 7 = largest contribution). In the second section, for each action point, the expert indicated the extent of the difference (1 = very small, 7 = very large) between the control measures rated ‘smallest contribution’ and ‘largest contribution’ in the first section. These first two sections represent the self-explicated part of the ACA model, which used to estimate intermediate partworths. For each pairwise comparison in the third section, the expert was shown two hypothetical partial strategies for food safety improvement differing by control measures corresponding to two or three action points. The expert expressed his preference and its strength for one of the proposed strategies (1 = strong preference for strategy 1; 9 = strong preference for strategy 2). The third section is the conjoint part of ACA model, which refines the partworths obtained in the previous two sections.

The three sections yielded data for each control measure within each action point for each expert. Experts' preferences were analysed using an additive, main-effects partworth model. For each expert, the ACA procedure updates the paired-comparison questions, using information on intermediate estimates of the utilities. The procedure is adaptive in the sense that each paired comparison is constructed so as to take advantage of the information obtained about the experts' partworths in preceding sections ( Green et al ., 1991 ). This allows ACA to analyse the responses of each expert individually. These intermediate estimates are based on an OLS regression using the pooled self-explicated and conjoint (paired-comparison) data, which ensures that the overall utilities of the two partial food safety strategies are nearly equal. Details of the ACA process, regression layout and updating utility procedure are provided by Green et al . (1991) . The final ACA section investigated the (internal) predictive validity of the estimated utilities ( Valeeva et al ., 2005b , p. 1608).

The partworth utilities estimated by ACA and measured in synthetic units indicate the relative contribution of each control measure to the expert's preference for the strategy for improving food safety per block. The partworths and their units are referred to here as ‘food safety coefficients’ and ‘points’, respectively. Thus, each food safety coefficient within a block may be interpreted as the relative effectiveness of a particular control measure, in terms of points, for improving food safety in that block when the measure is implemented.

Finally, the experts were asked to allocate a total of 100 points over the three blocks according to the need for additional control measures against chemical and microbiological hazards in order to achieve the highest food safety level for consumed fluid pasteurised milk. The average points per block were used as weights to convert average food safety coefficients per block into chain food safety coefficients. These chain coefficients were scaled so as to sum to a total of 100 points for all alternative control measures among which the choice should be performed at each action point along the chain. These 100 points do not correspond to any strategy for improving food safety. As specified in the previous section of this paper, each strategy implies the choice of only one alternative at each action point in the chain, i.e. implementation of one measure at each action point excludes others.

The second and third columns of Table  1 show the chemical and microbiological food safety coefficients obtained for the control measures corresponding to the key action points along the dairy chain. Each chain coefficient indicates the relative effectiveness of a control measure for improving food safety in the chain when the measure is implemented. The sum of control measure coefficients that compose a certain chain strategy represents the total effect of this strategy on food safety in the chain compared with the minimum food safety level. The ‘minimum food safety level’ refers to the strategy when at each action point along the chain the control measure with the lowest coefficient is chosen. Note that this minimum level does not represent any legal minimum food safety level. Rather, it indicates the lowest possible level given the measures available in this study. Table  1 indicates that the measures with the lowest coefficients at each action point tend to represent measures that are not adopted in practice. For example, ‘ingredients are not traceable to their transporter/manufacturer or farm of origin’ is the measure with the lowest effect for the action point ‘feed ingredient identification and traceability’ in the feed block. By contrast, the maximum food safety level refers to the strategy in which, at each action point along the chain, a control measure with the highest coefficient is selected for implementation, assuming that each measure is being correctly implemented. The sum of the highest food safety coefficients at each action point shows the maximum food safety level achievable within this study, corresponding to 67 microbiological and 69 chemical points.

2.3. Economic input

Costs of control measures, in Euros per ton of milk, were calculated throughout the chain, using partial budgeting. ⁴ This method estimates the incremental change in cost, relative to a base situation, because of a change in production activities ( Boehlje and Eidman, 1984 ; Huirne and Dijkhuizen, 1997 ). In this study, for each chain participant separately, partial budgets were used to estimate the extra costs resulting from the change in control measures within a certain action point relative to the measure representing the minimum level of improving food safety (the base situation), assuming that implementation of all changes is feasible.

Data on costs of control measures for the feed and processing blocks were collected in 2003 directly through interviews with representatives of compound feed and dairy-processing companies and the regulatory authority for the animal feed sector. The cost estimates are considered representative of average Dutch compound feed and dairy factories producing fresh dairy products.

For the dairy-processing block, cost data are needed for a representative factory. During the period 2003–4, there were 55 dairy factories in the Netherlands, of which 13 factories produced fresh dairy products. ⁵ About 11.7 million tons of milk were processed annually, of which 1.5 million tons were used to produce fresh dairy products ( Dutch Dairy Board, 2005 ). The average manufacturer of fresh dairy products processes about 140,000 tons of processed milk a year. However, factory equipment has particular production capacity, which is sometimes unused. Furthermore, for many control measures, representatives interviewed in this study could only provide data relating to a specific factory. Thus, to represent the average dairy factory in this study, we chose a real factory with 160,000 tons of processed milk, which is the closest to the average. For this factory, 53 per cent of its raw milk goes for fluid pasteurised milk production (other products: yogurt, custard and cream).

The representative feed block firm is an actual compound feed manufacturer producing about 200,000 tons of feed per year for different animal species. For the transport of feed ingredients and finished compound feed, only road transport is considered. Over the period 1999–2003, total compound feed production in the Netherlands fell by 17 per cent (to 12.2 million tons), though production of compound feed for cattle fell by only 6 per cent (to 3.4 million tons). By 2003, out of 115 small manufacturers, 92 manufactures were producing less than 25,000 tons annually and accounted for only 5 per cent of Dutch compound feed production ( Product Board Animal Feed, 2005 ). These tiny manufacturers cannot deal on their own with some important control measures considered in this study (such as measures relating to design of dedicated production lines, Table  1 ). Therefore, we calculated the average firm size (output of 240,000 tons a year) from the 23 largest factories only. For the same reasons given in the previous paragraph, we chose a real factory with annual production of 200,000 tons, the closest to this average, as being representative.

Data regarding the farm block were mainly obtained from published studies, Dutch handbooks and reports (e.g. Research Institute for Animal Husbandry, 2002 , 2003 ). Cost estimates refer to a Dutch dairy farm that is the average of two size classes: small (up to 35 cows) and medium (from 36 to 90 cows), which together comprise 85 per cent of dairy farms in the Netherlands (CBS, 2005). The typical farm is characterised as having a fixed milk quota, 50 milking cows with a 30 per cent replacement rate and a 305-day milk production of 8,500 kilograms per cow; 25 ha of grassland/pasture (10 ha of which are located on clay soil) and 5 ha of fodder crops (maize). By changing these farm characteristics, cost estimates can be adjusted according to farm size.

In evaluating the extra costs of the control measures shown in Table  1 , it was important to distinguish between two types of measures: specific interventions and new methods of managing process control. Specific interventions include the purchase of new equipment, the physical modification of production facilities and other physical changes in production procedures, many of which require capital investments (e.g. ‘dedicated production lines for feed for particular animal species’ in the feed block, Table  1 ). Costs of the specific interventions are defined as the sum of depreciation, capital utilisation and maintenance costs and costs of labour and other inputs if applicable. New methods for managing process control involve the development, introduction and updating of different practices of process evaluation and further improvement. Examples of such practices are ‘written procedures and instructions’ for all stages of feed production and ‘perfect level of best farm practice, all best practices are performed’ (this includes procedures for feed and animal traceability, milking of cows, storage and use of antibiotics and farmer training by certified persons). Costs of new methods for managing process control are calculated by adding annualised initial costs and recurrent costs. Initial costs are associated with planning, developing documentation, new software, training personnel and hiring additional labour; these costs are expressed as annual payments over a 15-year period, including interest. Recurrent costs are associated with improving hygiene, better quality inputs, recording and monitoring, training, additional checks and testing, internal and third-party audits and certification by a legal authority.

2.4. Model specification

• Z , total extra costs for the chain, Euros per ton of milk;

• I₁ , number of action points that include measures related to measures corresponding to other action points via additional links;

• I₂ , number of action points that include measures not related to measures corresponding to other action points;

• L , number of additional links that relate measures corresponding to different action points; each link has two options: link option 1 and link option 2 (see example in what follows);

• M(N) , number of measures within action points that are related, via link option 1(2), to measures corresponding to other action points;

• J , number of measures within action points that are not related to measures corresponding to other action points via additional links;

• x_im(n) , measure m(n) within action point i , ∀ i ∈ I₁ ;

• x_ij , measure j within action point i , ∀ i ∈ I₂ ;

• y_lm(n) , link option 1(2) within additional link l to relate measures m(n) within action points to measures corresponding to other action points;

• c_im(n) , extra costs of measure m(n) within action point i , ∀ i ∈ I₁ ;

• c_ij , extra costs of measure j within action point i , ∀ i ∈ I₂ ;

• δ _m(n)(j) , linking coefficients that relate measures m(n)(j) corresponding to different action points with each other;

• γ _m(n)(j) , linking coefficients that relate measures m(n)(j) corresponding to different action points with each other when one of these measures is a ‘package’ measure (see example in what follows);

• W_k , required food safety level k ( k = 1 refers to microbiological, k = 2 refers to chemical food safety);

• w_im(n)k , increase in food safety level k achieved by implementing measure m(n) within action point i , ∀ i ∈ I₁ ;

• w_ijk , increase in food safety level k achieved by implementing measure j within action point i , ∀ i ∈ I₂ .

This ILP model chooses the least-cost combination of control measures (from amongst action points in I₁ and I₂ ) to meet the requirements ( W_k ), for k food safety levels, given the effectiveness of measures ( w_imk , w_ink , w_ijk ) and their extra costs ( c_im , c_in , c_ij ). With the exception of the ‘package’ measures (see what follows), the order of a measure's entrance into the optimal (least-cost) plan does not matter (as it might in practice).

The model comprised 245 variables and 150 constraints. It was specified in a Microsoft Excel spreadsheet and solved using the add-in optimisation software ‘What'sBest’ (Lindo Systems Inc., 2000), which includes integer optimisation, making use of a branch-and-bound algorithm. ⁶

The main groups of activities in the model are: The main groups of constraints are as follows.

• Group A: Control measures corresponding to action points in the dairy chain.

• Group B: Additional links, each of which has two options. Some measures relating to one action point have logically exclusive or necessary relationships with measures relating to another action point. An example of an exclusive relationship of measures in the farm block is that the measure ‘no grazing at all, feeding in the pen during the whole year’ (one of the measures relating to the action point ‘grazing of pasture after manure application’) cannot logically be chosen in the optimal (least-cost) plan together with the measure ‘no contact with neighbour's cows, double fence’ (one of the measures relating to the action point ‘contact with neighbour's cows while grazing’). These two measures refer to different ways of keeping cattle on the farm during the year, namely, stalled, and stalled and pasturable. To specify this relationship, the link ‘keeping of cattle’ comprises two options, i.e. ‘stalled keeping of cattle’ (link option 1) and ‘stalled and pasturable keeping of cattle’ (link option 2) in the model.

• Group 1: Constraints linking control measures within each action point [equations ( 2.1 ) and ( 2.2 )] ensure that only one of the control measures available within an action point can be selected in an least-cost plan.

• Group 2: Constraints formulating logical relationships between activities included in the model are subdivided into four subgroups:

  1. Relating link options 1 and 2 with each other (e.g. ‘stalled keeping of cattle’ and ‘stalled and pasturable keeping of cattle’) within an additional link to ensure that only one of the link options is chosen in the least-cost plan [equation ( 3.1 )].

  2. Relating control measures with a logically matching link option within an additional link [equations ( 3.2a ) and ( 3.2b )], for example, linkage between the control measure ‘no contact with neighbour's cows, double fence’ and the link option 2 ‘stalled and pasturable keeping of cattle’.

  3. Relating control measures corresponding to different action points with each other [equation ( 3.3 )]. For instance, the control measure ‘no acquisition of cattle, i.e. having a ‘closed herd’ and the control measure ‘periodic veterinary checks of cattle, including checks with regard to purchased cattle’ (which correspond to ‘acquisition of cattle’ and ‘veterinary checks of cattle’, respectively) are related to avoid both of them being selected together in a least-cost plan.

  4. Relating control measures corresponding to different action points with so-called ‘package’ control measures [equation ( 3.4 )]. The ‘package’ measures imply compliance with a given QA system, which requires the implementation of a given set of control measures. Thus, a ‘package’ measure can logically be chosen only after all measures required by this ‘package’ measure are included in the optimal plan. ‘Complying with GMP ⁺ or another certified QA system equivalent to GMP ⁺ ’ in the farm block is an example of a ‘package’ control measure that can enter the least-cost plan when other control measures in the feed block such as ‘ingredients are traceable to their transporter/manufacturer or farm of origin via documentation/recording/coding’ and ‘ingredient manufacturers comply with GMP ⁺ or another certified QA system equivalent to GMP ⁺ ’ have already been included. In other words, the farm block can get compound feed produced according to a certified QA system only after it is available in the feed block.

• Group 3: Food safety constraints ensure the fulfilment of food safety requirements [equation ( 4 )].

3. Results and discussion

For different required levels of improving food safety, the model was run for the set of 186 control measures corresponding to 75 action points along the chain. In each run, the model selects the least-cost control measure at each action point, given the fixed food safety target level for the chain. The measures selected at each action point along the chain comprise the optimal (least-cost) combination of measures.

The model chose 15 different combinations of measures, as chemical and microbiological levels were simultaneously tightened in steps of 5 points until a maximum food safety level was reached (67 microbiological points and 69 chemical points). Simultaneous tightening of both levels is assumed to result in an overall increase in the food safety level. The combination when 0-point chemical and microbiological food safety levels were required was defined as the default situation. Results are shown relative to this default situation, represented by the origin in Figures  1–5 .

3.1. Improving food safety per block

Figures  1 and 2 illustrate the contribution of each block to improving chemical and microbiological food safety, respectively, at each food safety level for the chain. The required levels (as reflected in the constraints of the model) of both chemical and microbiological food safety on the horizontal axes are plotted against the achieved level (as reflected in the output of the model) of chemical (Figure  1 ) and microbiological (Figure  2 ) food safety after optimisation on the vertical axis.

Figure  1 shows that the feed block contributes most to improving chemical food safety in least-cost plans at all food safety levels, especially at relatively low required levels (5–35 points). For the 5-point and 35-point required levels, the feed block's contribution is 100 and 78.5 per cent, respectively, of the total points achieved by reducing chemical hazards. By contrast, the contributions of the farm and dairy-processing blocks increase slowly to reach 7.1 and 14.4 per cent of the total contribution, respectively, at the 35-point level. However, when higher levels of food safety improvement are required, the role of the farm block increases. At the maximum food safety level, the feed, farm and processing blocks' contributions are 58.8, 27.1 and 14.1 per cent of the total, respectively. These results indicate that relatively small improvements in chemical food safety can be attained in a cost-efficient way with the active involvement of the feed block only. However, to achieve higher levels, more effort from both the feed and farm blocks is needed.

Figure  2 shows that for relatively low required levels of food safety (5–25 points), the contributions of the feed and dairy-processing blocks to microbiological improvements in food safety are almost equal in the optimal strategies, whereas the farm block's contribution is small, below the 35-point required level. However, from the 40-point level onwards, the farm block's contribution to improving microbiological safety gradually increases. At the maximum level of food safety improvement, the farm block contributes to reducing microbiological hazards almost as much as the dairy-processing block (34.5 versus 41.7 per cent).

Overall, the results of Figures  1 and 2 suggest that using a chain approach for improving the food safety of the consumed pasteurised milk, the farm block's participation in cost-effective solutions is significant only for reaching rather high target levels of food safety improvement.

3.2. Cost-effectiveness of different strategies for improving food safety in the chain

Figure  3 reports the numbers of control measures implemented by each block (vertical axis) for different food safety improvement levels (horizontal axis), with the corresponding extra costs. The number of control measures represents the number of measures that are replaced (within an action point) in the optimal plan by more effective ones (with a higher food safety coefficient) compared with the default situation. ⁷ However, they do not include substitution between more effective measures within the action points, in cases where more than two control measures can be chosen (see Table  1 for examples).

Extra costs rise steadily as required levels of food safety improvement increase, reaching €44.37 per ton of milk at the maximum improvement level. At this level, 70 more effective control measures are implemented: feed = 24, farm = 24, dairy processing = 22. We discern three stages in this process of gradual replacement of control measures. Each stage is characterised by a particular block (or blocks) being responsible for implementing more of the changes and by certified systems of improving food safety entering the least-cost plan.

At the end of the first stage (0–35 points), extra costs are €1.43 per ton of milk. In this stage, cost-effective combinations of control measures mainly include measures from the feed and dairy-processing blocks. In particular, to reach a required level of 30 points, the feed and dairy-processing blocks have to implement 16 and 12 measures, respectively, whereas the farm block implements only five measures. Additional analysis of optimal plans (not shown in the figure) within this stage shows that, within the feed block, measures relating to compound feed ingredients are quite cost-effective. They enter the optimal (least-cost) plan for achieving levels that are lower than 20 points. Examples of such measures are ‘ingredients are traceable to their transporter/manufacturer or farm of origin’, ‘ingredient manufacturers comply with GMP ⁺ or another certified QA system equivalent to GMP ⁺ ’, ‘separate storage of different ingredients (premixes, additives and veterinary medicines), clear identification’ and ‘chain control programme’ (i.e. audit programme of feed ingredient suppliers). Furthermore, within both the feed and dairy-processing blocks, the majority of measures leading to better production practices and hygiene conditions, such as ‘control of adequate cleaning and disinfection of production equipment and premises’ and ‘written procedures and instructions’ at all steps of production and transport (i.e. new methods of process control), enter the least-cost plan in the first stage. All measures aimed at improving raw milk transport enter in this stage. This allows the complete package of measures required for certified raw milk transport to enter the least-cost set. Thus, the certification of the whole package becomes cost-effective at the 30-point improvement level.

The second stage (35–45 points) involves a further cost increase of €2.84 per ton of milk. In this stage, the feed block needs to implement most of the extra control measures. At the 45-point required level, nearly half the measures are implemented in the feed block. In addition to managerial measures of improving food safety, measures relating to (rather expensive) changes in feed production and transport procedures (that require some capital investments) enter the optimal plan in this stage. Examples of these measures include ‘dedicated production lines for feed for particular animal species’, ‘filters for air intakes for cooling’ and ‘separate vehicles for feed with critical additives and medicines’. Furthermore, at the 45-point level, it becomes cost-effective for compound feed producers and transporters to ‘comply with GMP ⁺ or another certified system equivalent to GMP ⁺ ’. This means that the complete package of measures relating to certified compound feed supply enters the optimal strategy.

Within stage 2, the number of measures improved by the farm block increases from 5 to 11, indicating again the increasing role of the farm block when higher levels of improving food safety are required. ‘Acquisition of cattle whose herd health status is known’, ‘participation in existing disease monitoring programmes’, ‘control of cleaning and disinfection of farm equipment and premises’ are examples of such farm block measures.

In the third stage (50 points to the maximum), maximum food safety improvement raises extra costs per ton of milk further by €40.1. The number of measures in the farm block more than doubles (from 11 to 24 measures) when the level of food safety improvement rises from 45 points to the maximum. At the same time, the dairy-processing and feed blocks have to implement only 7 measures and 1 measure, respectively. In addition, in this stage, the whole package of measures relating to certified raw milk supply [from farms producing according to KKM ⁸ or a QA system for farm production based on Hazard Analysis Critical Control Point (HACCP) principles] enters the optimal plan. Although most of the measures relating to the package of certified raw milk supply are already in the optimal combination at the 60-point level, the complete package enters into just the last two combinations for the 65-point level and the maximum level. ‘Periodic veterinary checks of cattle, including checks with regard to purchased cattle’ and ‘perfect level of best farm practices performance, all best practices are performed (100%)’ are the measures that complete the package by entering the optimal plan for a 65-point level. Among others, measures that also enter the optimal plan at this moment are ‘drug residue screening test in the case of doubt about withdrawal period’ and ‘purchase of cattle followed by quarantine’. However, at this step (60–65 points), when the complete package comes into the least-cost plan, a slight increase in food safety level raises the extra costs considerably (from €21.14 to €35.75 per ton of milk). Furthermore, it is interesting to note that it is only when reaching the maximum food safety improvement level that it becomes cost-effective for the dairy-processing block to implement extra measures regarding raw milk, such as ‘chain control programme’ (i.e. audit programme of raw milk suppliers).

Figure  3 suggests that in the first two stages of improving food safety, the feed and dairy-processing blocks have to change the bulk of measures at relatively low extra cost; by contrast, the farm block has to implement many measures only in the final stage, leading to very high extra costs. This indicates that, also in terms of the number of implemented control measures, the farm block makes a large contribution only when rather high levels of improving food safety are required. In this context, the results are justified by the current situation in the Dutch dairy chain, which matches the third stage described in this study. Compared with the feed and dairy-processing blocks, the farm block lagged behind somewhat in adopting measures; only during the last few years have food safety improvements focused more on the farm block (e.g. developing and introducing farm-level QA schemes). However, achieving these high levels comes at a high cost. Thus, introducing all extra measures on the farm may or may not be worthwhile, depending upon actual requirements for food safety. This aspect indicates the importance of finding a good balance between achieving food safety improvements and the costs incurred by chain participants.

The lines plotted in Figure  4 trace out extra cost functions for each block and for the entire chain, connecting the least-cost points of achieving particular levels of improving food safety. The results show that the cost distribution per block changes as levels of improving food safety are tightened. For quite low levels of food safety improvement (corresponding to the first stage), the feed and dairy-processing blocks incur higher extra costs than the farm block. At the 45-point level (second stage), the cost distribution over blocks is more balanced: extra costs per ton of milk for the feed, farm and dairy-processing blocks are €1.09, €1.81 and €1.37, respectively. However, after the 55-point level, more than 60 per cent of the total extra chain costs of improving food safety are incurred by the farm block. The highest share of the farm block in total costs is 77.1 per cent (€27.54 per ton of milk), incurred at the 65-point level. Thus, the high levels of food safety improvement reached in the third stage represent a considerable cost increase for the farm block. However, the rise in the share of the farm block in the total extra costs is somewhat higher than the rise in its contribution to improvements in chemical and microbiological food safety (Figures  1 and 2 ).

Figures  3 and 4 show that about two-thirds of the maximum food safety improvement can be reached at a relatively low extra cost that is rather evenly distributed over the blocks, but to come close to the maximum levels involves a steep rise in cost, which falls mainly on the farm block.

Figure  5 shows the three-dimensional cost surface, combining the step cost functions for non-synchronised increases in chemical and microbiological food safety levels. Whereas Figures  1–4 , in which reductions in these two hazards are synchronised, involved just 15 model runs, this figure required 225 model runs. Each point on the surface represents the cost of a least-cost set of control measures that have to be implemented to reach particular levels of improving chemical and microbiological food safety. The results demonstrate that costs rise more steeply near the maximum level for microbiological food safety improvement than for chemical food safety improvement. Furthermore, these findings suggest that from a whole-chain cost perspective, it is more cost effective to focus on improving chemical and microbiological food safety simultaneously. This is because many measures affect both chemical and microbiological food safety.

3.3. Sensitivity analysis

Sensitivity analysis was conducted to determine how robust the results are with respect to the chain food safety coefficients. The basic scenario, presented so far, uses simple averages of the experts' block weights to translate average values of food safety coefficients per block into chain food safety coefficients. Three sensitivity analyses were performed. In the first, only the average partworths within each block were changed. In the other two, we kept the average partworths from the basic scenario and changed only the relative weights of the blocks.

In the first scenario, for chemical and microbiological food safety separately and within each block, we identified the expert whose opinions diverged the most from the average. He was the one for whom the sum of absolute deviations from the average values of the food safety coefficients was greatest. In the second scenario, equal weights were used for each block. In the third scenario, extreme weights for each block were used. They were derived by identifying the most dissenting expert, as in the first scenario. In this scenario, the weights used were 0.05, 0.84, 0.11 for chemical food safety and 0.21, 0.74, 0.05 for microbiological food safety, for the feed, farm and dairy processing blocks respectively (whereas in the basic scenario, these weights were 0.47, 0.38, 0.15 for chemical food safety and 0.18, 0.49, 0.33 for microbiological food safety, for the respective blocks). ⁹

Table  2 shows the effect on each block's contribution to improving chemical and microbiological food safety and to the corresponding extra costs of the different chain food safety coefficients under the three scenarios. Results are presented for the required food safety levels in the chain at 30, 65 and 100 per cent of the maximum food safety level achievable in each scenario (corresponding to approximately the 20-point, 45-point and the maximum-point levels in the basic scenario). It should be noted that the extra cost of attaining the given levels according to the least-cost plans is rather similar across the four scenarios.

Under the first scenario, each block's contribution to improving chemical and microbiological food safety, and the distribution of corresponding extra costs among the blocks, are rather similar to those in the basic scenario at almost all required levels of improving food safety in the chain. The only noticeable difference is a higher contribution of the farm block and a smaller contribution of the processing block to improving chemical food safety.

Assigning equal relative weights of each block in the second scenario implies that, compared with the basic scenario, (i) for chemical food safety, the relative weight of the feed block decreases by about 15 per cent and the relative weight of the dairy-processing block increases by about 15 per cent; (ii) for microbiological food safety, the relative weight of the feed block increases by 15 per cent and the relative weight of the farm block decreases by about 15 per cent. These changes in relative weights result in similar changes in the contribution of the blocks to chemical and microbiological food safety for each required level of improving food safety. For example, compared with the basic scenario, to attain 30 per cent of the maximum level in the second scenario, the contribution of the feed block to improving chemical food safety decreases by about 15 per cent, and the contribution of the dairy-processing block increases by about 15 per cent. These changes also affected the distribution of the extra chain costs among the blocks at 30 and 65 per cent, but the extra chain cost distribution for the maximum level of improving food safety is very similar to the basic scenario.

In the third scenario, the farm block becomes heavily involved in improving microbiological and especially chemical food safety. This is explained by an extremely high relative weight of the farm block in this scenario, which results in greater values of chain food safety coefficients of the measures in the farm block. At the same time, these changes cause a significant increase in the farm block share in the total extra chain cost, in particular at low levels of food safety improvement. These outcomes are quite different from the results of the basic scenario.

The sensitivity analysis shows that results of this study are quite robust with respect to changes in the partworths attached to control measures, but are sensitive to changes in relative block weights used to derive chain food safety coefficients. Not surprisingly, scenarios 2 and 3 indicate that a higher (lower) weight for a particular block may lead to a higher (lower) contribution of this block to improving food safety, with an increase (decrease) in this block's share in total extra chain costs. Although there was near consensus among the experts on the ranking of the blocks ( Valeeva et al ., 2005b : 1608), their individual weight diverged. In taking the average, as we have done in the basic scenario, we aim for the ‘most likely’ values.

3.4. Economies of scale in food safety provision along the chain

The preceding sections show that reaching high levels of food safety improvement involves high extra costs for the chain as a whole, with an obvious cost disadvantage faced by the farm block. These results are obtained for a chain consisting of representative participants. Previous studies ( Antle, 2000 ; Nganje and Mazzocco, 2000 ) found important economies of scale in compliance costs for food safety regulations for a particular chain participant in the meat industry. Here, we can only explore the effects of economies of scale in improving food safety in the chain for the farm sector, for which relevant data are available. Cost data for all the types of food safety improvement considered in this study are not currently available for different sizes of compound feed and dairy-processing plants. This would require active collaboration from the industry.

The 50-cow farm was taken as the representative farm size in the previous sections. A scenario in which the farm block is represented by a 250-cow farm can demonstrate the largest possible impact of farm scale on costs of improving food safety in the chain. The chain was analysed with a 250-cow ‘representative’ farm, using the ILP model in the same way as for the 50-cow scenario. Complete details are given in Valeeva (2005) .

The results demonstrate that, for both farm sizes, relatively low required levels of improving food safety (up to 35 points or 50 per cent of the maximum) can be achieved at low extra costs for the farm block as well as for the chain as a whole (€0.10 vs. €0.34 and €1.42 vs. €1.09 per ton of milk, for the farm block and the whole chain in the 50- and 250-cow scenario, respectively). The remarkable outcome is that, whereas the large-scale farm assumption decreases the total extra chain costs by 23 per cent, it increases the share of extra costs incurred by the farm block nearly 2.5 times, resulting in a considerable rise (from 7.1 to 30.7 per cent). By contrast, the farm block's contributions to improving chemical and microbiological food safety increase from 7.1 to 14.2 per cent and from 7.6 to 20.0 per cent, respectively. These results indicate that the large-scale farm assumption results in a more equal distribution of extra costs and food safety contribution among chain participants at low levels of food safety improvement in the chain. The observed distributional changes in the 250-cow scenario can be explained by increased cost-effectiveness of certain control measures in the farm block, compared with measures in the feed and dairy-processing block. These farm-level measures refer to herd management (i.e. ‘participation in monitoring programmes’) and other conditions (i.e. ‘general/personal hygiene conditions on the farm’ and ‘adequate cleaning and disinfection’).

To attain food safety improvement levels near the maximum in the 250-cow scenario, the extra costs per ton of milk incurred by the farm block are €4.77, or about 15 per cent, lower than in the 50-cow scenario. The breakdown of this difference by type of farm-level control measures reveals an extra cost decrease of €0.80 (4.0 per cent), €1.04 (29.5 per cent) and €2.93 (36.9 per cent) per ton of milk associated with improvement of herd management, water management and other conditions, respectively. Although the large-scale farm leads to a 10 per cent decrease in total extra chain costs, this does not affect the overall extra cost distribution among chain participants relative to the 50-cow scenario.

Overall, these findings suggest that large-scale farms have the advantage of economies of scale for the entire chain at all levels of food safety improvement. The decrease in total extra chain costs of food safety improvement varies from 25.5 per cent at the 50-point level to 10.0 per cent at the maximum level. However, irrespective of farm scale, extra costs borne by the chain remain rather high when high levels of food safety are required; compared with other blocks, the farm block continues to face a big cost disadvantage. It does not seem feasible that the farm block is capable of bearing this unless it is compensated in some way by the other blocks or by an increase in the end product price.

Evidently, economies of scale would also be important up- and downstream within the dairy chain. It may be expected that total extra chain costs would increase when small-scale participants form the chain, and vice versa. However, additional research is needed to investigate how different structures of food chains affect costs of food safety provision in the chain and their distribution among chain participants.

4. Conclusions and discussion

This paper provides insight into the optimal (least-cost) strategies for improving food safety along the production chain for consumed fluid pasteurised milk. The strategies involve control measures that can be implemented at certain action points. The measures are different in nature, containing both specific interventions and new methods of managing process control. For these measures, the optimisation model combines economic and food safety data to explore the trade-offs in achieving a series of required levels of chemical and microbiological food safety improvements in the chain. Although data on the effectiveness of control measures represent the opinions of experts, and not objective observations on the efficacy of the measures, our approach turns out to be a helpful tool for exploring the cost-effectiveness of different strategies for improving food safety. Given the rather small number of experts taking part, sensitivity analysis was performed with respect to the subjective scale for measuring food safety improvements. This analysis indicates that model results are rather robust at higher levels of food safety improvement, but less so at lower levels. The relative weights per block used to derive the scale have a large impact on the results obtained in this study, as could be expected.

Our findings show that for simultaneous improvement of chemical and microbiological food safety in the chain, the extra cost function is upward sloping. Extra costs per ton of milk, relative to the default situation, can be up to €44.37. This can be a substantial component of the milk price at the end of the chain (compared with the dairy-processing factory gate price of €570 per ton of milk in the Netherlands in 2004 ¹⁰ ). Attaining two-thirds of the maximum food safety improvement can be achieved at relatively low extra cost (€4.27 per ton of milk), which is rather evenly distributed among the feed, farm and dairy-processing blocks. To reach this level, the feed block contributes mainly to improving chemical food safety, whereas the dairy-processing and farm blocks contribute mainly to improving microbiological food safety. However, greater levels of food safety improvement can only be achieved at considerably higher extra costs, with a more than 60 per cent share from the farm block. Most of the measures in the farm block need to be implemented only if higher levels of improving food safety are required. For reaching these high levels, a five-fold increase in farm scale would reduce extra costs incurred by the farm block by nearly 15 per cent, resulting in a 10 per cent decrease in the total extra chain costs. However, this would not result in significant changes in the distribution of extra costs in the chain.

This study only relates different levels of food safety to extra cost. It does not examine the net benefits resulting from the implementation of a particular set of measures. In fact, the minimum cost of achieving a certain level of food safety may still exceed the gross benefits obtained. Therefore, further research is needed to determine the total consumer benefit of each level of food safety improvement, in order to determine at which level the net social benefit (consumer benefit minus least-cost strategy for the chain) is greatest. Additionally, high costs incurred by the farm—which are usually not reflected in a higher farm-gate milk price—suggest that future research should investigate how these costs should be compensated (price premium or subsidies) by the other chain participants when retail milk prices increase.

When interpreting the results of this research, the assumption that there is a regulator leading the chain in achieving a higher level of food safety must be borne in mind. Because of this assumption, it is rather difficult to link any of the optimal (least-cost) plans discussed in the previous section to the current level of food safety improvement in the dairy chain, which involves the decisions of independent chain participants. Also, the current level might differ among firms within the same stage of the chain. Nevertheless, the findings provide insights into contexts in which a regulator may be present. In the context of regulations, in which government performs as regulator, the results can be helpful for evaluating whether food safety policy should be designed for the chain as a whole and not in separate stages. In this sense, the outcomes can be of particular interest for countries where the level of food safety needs to be improved considerably. Also, this study can be relevant for retailers and other downstream stakeholders in the food system. As suggested by Hennessy et al . (2001) , they may have strong incentives to engage in caretaking actions in earlier stages because of asymmetric liability hazard in the food chain. Furthermore, the outcomes of this study can be useful in a special case of more aligned chains when downstream operators may want to own other upstream suppliers, partly in order to coordinate food safety improvement along the chain.

In addition, it should be noted that as well as results relating to a chain with leadership, this approach provides essential results for each chain participant. Though our approach focuses on optimal strategies for improving food safety for the entire chain, it also allows each block to be modelled separately and, therefore, yields cost-effective strategies per block.

The LP model used assumes that the food safety improvement level achieved by implementing a certain combination of control measures is the sum of the measures' effects on improving food safety. In fact, such combinations along the chain often result in a non-additive increase in the level of improving food safety. So, evaluation of the increase in the level of improving food safety would ideally need to consider the combined effect of measures. To our knowledge, the literature on the combined effect of different measures is rather limited. Furthermore, it refers mainly to evaluating the combined effect of combinations of possible interventions (and not new methods of managing process control) implemented in a certain stage of the chain. However, within the model, it was possible to some extent to account for the multi-effect of measures that refer to the control of a few chemical or microbiological hazards or chemical and microbiological hazards combined.

In this study, scarce information and much uncertainty about the costs of many control measures, mainly in the feed and dairy-processing blocks, may have led to a rather approximate estimation of economic input for the model. Moreover, a few aspects were not completely taken into consideration while estimating costs. First, for some control measures, food safety improvement costs also included the costs related to other goals of chain participants, which are difficult to separate out. For example, control measures involving developing, implementing and maintaining traceability systems by chain participants facilitate traceback for safety and quality, improve supply management and increase the efficiency of production processes throughout the chain. Secondly, as can be seen from the model, the total extra costs of achieving a certain level of improving food safety are obtained by adding the extra costs of the measures needed to be implemented. Costs resulting from synergy effects while implementing some combinations of measures are considered only for those measures for which the information on possible changes in costs is available.

In conclusion, our framework is intended to contribute to analysing the cost-effectiveness of strategies for improving food safety, which are of interest to producer–processor supply chain, and not to address consumer or social benefits that accrue from applying these strategies. Still, the study does not cover the entire chain. A possible development of the model would be to include the retailer/catering stage and the final consumption stage. The role of individual consumers in ensuring food safety would be of particular interest. Proper food-handling practices by consumers can help prevent problems caused by some food safety hazards. In this regard, cost-effective food safety improvements could be made by consumers themselves. To examine this also requires techniques that deal with consumer decision making. Another possible extension of the approach is to consider measures to control other existing hazards or ‘new’ hazards that may emerge. Moreover, it can be adjusted to investigate the least-cost strategies for the dairy chain with another end product or for other chains like pork or beef. Further research could then compare cost effectiveness of each chain participant and the cost distribution among participants across chains. However, it would be possible only where, in terms of main participants, the chain structure is similar to that of the dairy chain.

Acknowledgements

The authors would like to thank the Dutch insurance company Achmea, the Dutch Dairy Organization (NZO) and the Mesdagfonds foundation for the funding of this research. The authors express their gratitude to the editor and to three anonymous referees for their valuable comments, which substantially improved the quality of the paper.

References