DOI: 10.13128/BAE14715
Effects of switching between production systems in dairy farming
Antonio Alvarez1, Carlos Arias2
1 University of Oviedo and Oviedo Efficiency Group
2 University of León and Oviedo Efficiency Group
* Corresponding author: carlos.arias@unileon.es.
Date of submission: July 17th, 2014
Abstract. The increasing intensification of dairy farming in Europe has sparked an interest in studying the economic consequences of this process. However, empirically classifying farms as extensive or intensive is not a straightforward task. In recent papers, Latent Class Models (LCM) have been used to avoid an adhoc split of the sample into intensive and extensive dairy farms. A limitation of current specifications of LCM is that they do not allow farms to switch between different productive systems over time. This feature of the model is at odds with the process of intensification of the European dairy industry in recent decades. We allow for changes of production system over time by estimating a single LCM model but splitting the original panel into two periods, and find that the probability of using the intensive technology increases over time. Our estimation proposal opens up the possibility of studying the effects of intensification not only across farms but also over time.
Keywords. Dairy farms, intensiﬁcation, latent class model, panel data.
JEL Codes. C23, D24, Q12
5. Empirical analysis of intensification and efficiency
5.1. Evolution of intensification over time
5.2. The effect of intensification on dairy farm efficiency
The number of dairy farms in the European Union has fallen dramatically in recent decades and continues to decline. Given that farm output remains roughly at quota level, over the same period the average size of dairy herds has increased steadily. At the same time, genetic and management improvements in dairy cattle have permitted large increases in milk production per cow. These structural changes have provided the basis for the propagation of intensive systems of production in the dairy sector. In fact, intensification is often mentioned as a feature of structural change, reinforcing the interest in the analysis of the impact of intensification on productivity and efficiency. The relationship between structural change and intensification has an interesting aspect in our empirical example since institutional and geographical reasons make it difficult to grow by accruing more land. Some of the farms are small, with fragmented land and located in hilly areas (del Corral et al., 2011). Obviously, the cessation of these farms does not increase the land available for the farms that stay in production and that are mainly located in flatter coastal areas.
Extensive dairy farming, on the other hand, consists of producing milk using mainly onfarm produced forage with low stocking rates. Fostering production using extensive systems has often been an explicit goal of agricultural policy, justified by factors such as environmental soundness, improved animal welfare, or use of abundant land in some areas. In contrast to this, many dairy farms in Europe have gone in the opposite direction, adopting more intensive production systems.
Despite the importance of the intensification process and the intense policy debates it has generated, few papers have studied the intensification of dairy farming using economic analysis. Most articles have adopted a technical perspective, describing the physical changes of the process (e.g., Simpson and Conrad, 1993), while very few have analyzed the economic consequences of this process (some exceptions are Alvarez et al., 2008, 2010, and Nehring et al., 2011).
From an empirical point of view, the coexistence of both extensive and intensive farms implies that there are two different technologies in the sector. This runs contrary to the assumption of a common technology for all farms which is the most frequent in the production literature. However, there is also awareness among researchers of the estimation bias that arises if such an assumption is unrealistic. For this reason, several approaches have been followed in dairy sector studies in order to account for the likely existence of different technologies. The most basic one is to drop a number of farms from the sample on the grounds that they may operate under a different technology (Tauer and Belbase, 1987). A second approach is to split the sample into several groups based on some observable farm characteristics. For example, Hoch (1962) divided the sample into two groups based on the location of farms, while Newman and Matthews (2006) considered two different technologies depending on the number of outputs each farm produces.
Classifying farms as intensive or extensive is not as straightforward as it might appear. For example, Nehring et al. (2011) used the number of cows per hectare in order to split the sample. However, the stocking rate partition does not fully describe the production system as other aspects such as the productivity of cows or the share of concentrates in the feed ration could also be taken into account. In this paper we avoid an exante classification of farms as extensive or intensive by estimating a Latent Class Model (LCM). This model assumes that several unknown technologies (classes) have generated the sample, and allows for the estimation of the parameters of the different technologies plus the probability that each observation has been generated by a specific technology.
To the best of our knowledge, the starting point of previous models in the literature is a setup and estimation proposal that assumes that the probability of each observation belonging to a class (i.e. using an intensive or extensive technology) is constant over time. This implies that changes during the period of analysis, no matter how dramatic, do not lead to a farm being labeled as belonging to a different class (use of a different technology). Such an assumption becomes increasingly untenable as the number of observed periods gets larger.
The objective of the present paper is twofold. First, we wish to determine whether the intensification process that has been taking place in recent decades has come to an end or whether dairy farms are still switching from extensive to more intensive production systems. For this purpose, we make a simple methodological proposal to circumvent the assumption of class probability being constant over time. Second, we are interested in analyzing the effects of intensification on farms’ efficiency. In particular, for given inputs we would like to know whether intensive farms have the potential to produce more output than extensive farms, and if so, the degree to which intensive farms achieve such potential.
In order to fulfill these two objectives, in the empirical section of the paper we use a panel of dairy farms in Northern Spain to illustrate the feasibility of estimating a LCM with timevarying probabilities. Our objectives are firstly to measure the changes over time in the probability of belonging to a class (technology) and secondly to analyze the effects on production potential and efficiency of dairy farm intensification1. Unlike previous studies, our methodological proposal allows us to look specifically at the farms that might have tilted towards a more intensive production system in the period of analysis.
The organization of the paper is the following. In the next section we describe the model. In section 3, we present the data and the empirical model. In section 4 we show the econometric estimation of the latent class models. In section 5 there is a discussion of the empirical results. The paper ends with some conclusions.
We use a LCM to analyze the extent of intensification in our dataset. The initial step is to check whether the LCM identifies different technologies and if those technologies represent different degrees of intensification. The starting point is a loglinear stochastic production frontier (Orea and Kumbhakar, 2004) such as:
(1)
where x is a vector of inputs, y a single output, v a symmetric random disturbance, and u a onesided random disturbance that measures technical inefficiency. Subscript i (i=1,…, N) denotes firms, t (t=1,…,T) denotes time, and subscript j (j=1,…,J) indicates a technology in a finite set. The vertical bar means that there is a different production function (different parameters) for each class j. We assume that, conditional on each class, the random disturbances v and u follow a normal and a truncated normal random distribution respectively.
In a latent class model we need to consider three likelihood functions. The first is the likelihood function of a firm i at time t belonging to class j:
(2)
where Θj represents the set of parameters of technology (class) j, and g denotes the likelihood function of a production frontier (Kumbhakar and Lovell, 2000). The second is the likelihood function of a firm i conditional on class j, obtained as the product of the likelihood functions in each period.
(3)
Finally, the unconditional likelihood function of firm i is calculated averaging the likelihood conditional on each class using the prior probabilities of class membership Pij as weights:
(4)
Prior probabilities can be interpreted as the probabilities attached to class j membership (Greene, 2005). These prior probabilities can be parameterized using a multinomial logit model:
(5)
where zi is a vector of “separating variables” and δj a vector of parameters to be estimated. The “separating variables” are related to the adoption of a technology and, as a result, can be used as explanatory variables of the prior probability of using that technology.
After estimation of the model in (1) by maximum likelihood a “posterior probability” can be computed as:
(6)
As equation (6) shows, the prior probability of class membership for each farm is weighted by the empirical likelihood that the farm belongs to that class. This implies that the ability of each technology to explain the observed production of a farm is incorporated in the calculation of the posterior probability. In a sense, the estimates obtained with the parsimonious parametric model of the prior probability are complemented with information on individual fit to provide a more accurate evaluation of the probability of class membership. In fact, the posterior probability is considered the best estimate of class membership (Greene, 2005), and as such, the value of this probability is the criterion that we will use for determining whether a farm is using a particular technology.
As mentioned above, a subtle feature of the LCM for panel data is that prior probabilities are modeled as time invariant. In practical terms, timeinvariant probabilities amount to assuming that changes in farms over time do not affect the classification of a farm as intensive or extensive. However, we expect some farms to change the use of inputs during the period of analysis in ways that suggest tilting towards intensive farming.
Our aim, therefore, is to circumvent the assumption of timeinvariant probabilities in the empirical analysis. For that purpose, we estimate a single LCM model for the whole period of analysis but split the observations for each farm into two periods, where the first period roughly corresponds to (t=1,…,T/2) and the second period to (t=T/2+1,...,T). In this approach, farm i is considered to be a different farm in each of the two periods. As a result, the probability of class membership is constant for a given farm during each of the two periods but can change from the first period to the second. Alvarez and del Corral (2010) report timevarying probabilities of class membership in the conventional Panel LCM through a slight modification of equation (6). They use as weighting factor of prior probabilities the likelihood function of each observation in each year (LFijt in equation 2) instead of the likelihood function of each farm (LFij in equation 3). This computational choice allows them to report time variation of a feature of the model, latent class probability, assumed to be constant over time for estimation purposes. Therefore, the time variation of probability reported is akin to variation around a constant mean. Instead, our estimation proposal avoids such problems by allowing class probability to change from the first sub period to the second while being constant in each sub period.
In principle, it is possible to divide the original sample in three or more sub periods. However, the time length of the panel shortens accordingly, causing econometric and numerical problems. In fact, the limit of this increasingly finer division in sub periods would be to estimate a pooled model. This amounts to treating the dataset as a cross section of farms instead of taking into account the existence of a panel. In this case, in each observation farm i is treated as a different farm, thereby allowing probabilities of class membership to vary freely over time. The downside is that we disregard the information provided by observing the same farm in several periods and that the computed probabilities of class membership are not necessarily parsimonious. Indeed, it would be possible to observe some farms moving repeatedly from one class to the other. All in all, this approach seems prone to numerical problems and to difficulties in interpreting the results. In fact, in the empirical application that we propose the LCM with pooled data failed to converge.
The data used in the empirical analysis consist of a balanced panel of 128 Spanish dairy farms observed over the 12 year period from 1999 to 2010. In general, the units in the sample are small to mediumsized family farms. In Table 1, we show the evolution of the average value of a set of variables related to intensification such as milk, number of cows, land, stocking rate, milk yield per cow, milk per hectare and purchased concentrate per cow.
Table 1 shows that, on average, farm output has grown during the sample period (64%) through an increase in the number of cows (44%) but much less so in land (16%). The stocking rate (23%), milk yield per cow (13%), milk per hectare (41%) and purchased concentrate per cow (6%) all grow in the period analyzed. In summary, these descriptive statistics depict a process of farm growth with intensification.
Table 1. Evolution of intensification features (sample means).
Year 
Milk (l) 
Cows 
Land (ha) 
Cows per hectare 
Milk per cow (l) 
Milk per hectare (l) 
Purchased concentrate per cow (kg) 
1999 
233250 
33 
18 
1,97 
6823 
13657 
3291 
2000 
253326 
35 
18 
2,02 
7005 
14365 
3266 
2001 
279077 
37 
18 
2,17 
7200 
15896 
3370 
2002 
312127 
40 
18 
2,35 
7419 
17783 
3434 
2003 
313470 
41 
18 
2,38 
7292 
17675 
3451 
2004 
332461 
43 
18 
2,41 
7501 
18420 
3465 
2005 
348243 
43 
19 
2,42 
7729 
19027 
3483 
2006 
361358 
44 
19 
2,41 
7877 
19323 
3470 
2007 
360131 
45 
19 
2,41 
7841 
19089 
3554 
2008 
362113 
45 
20 
2,39 
7661 
18765 
3557 
2009 
363012 
46 
20 
2,40 
7591 
18912 
3537 
2010 
383172 
47 
21 
2,43 
7699 
19304 
3489 
% increase 19992010 
64% 
44% 
16% 
23% 
13% 
41% 
6% 
The empirical specification of the production function is translog. We have chosen a flexible functional form in order to avoid imposing unnecessary a priori restrictions on the technologies to be estimated. The empirical counterpart of equation (1) is the following translog production frontier:
(7)
The dependent variable (y) is the production of milk (liters). We have considered only one output since these farms are highly specialized (more than 90% of farm income comes from dairy sales). Five inputs are included: (x1) number of cows, (x2) purchased feed (kilograms), (x3) ‘farm expenses’ (includes expenditure on inputs used to produce forage crops, namely seeds, sprays, fertilizers, fuel, and machinery depreciation), (x4) ‘animal expenses’, such as veterinary, medicines, milking and other expenses, and (x5) land. All monetary variables are expressed in constant euros of 2004. Additionally, 11 time dummy variables (TDm=1 if t=m, TDm=0 otherwise ) were introduced to control for factors that affect all farms in the same way in a particular year but which vary over time, such as weather (the excluded period is 1999).
Prior to estimation each input was divided by its geometric mean. In this way, the first order coefficients of the Translog production function can be interpreted as output elasticities evaluated at the geometric mean of the inputs.
The prior probabilities of class membership are assumed to be a function of two “separating variables”: the natural logarithm of the stocking rate (cows per hectare) and the natural logarithm of purchased concentrate feed per cow. Since prior probabilities are modeled as timeinvariant in each sub period, the separating variables are averaged over time within each period.
In this section, we report the main results of the estimation of two latent class models by maximum likelihood:
a) Panel model, i.e. using panel data and timeinvariant class membership probabilities.
b) Splitpanel model, i.e., estimating a single LCM model but allowing the probabilities of class membership to differ over time. This is achieved by splitting the sample into two periods where the probability of class membership is constant within each period but can change for the ‘same’ farm from the first period to the second. In other words, the estimation proceeds by treating each farm in the second period as a different farm.
In both models two latent classes were found.2 As mentioned above, we make the prior probability of belonging to a latent class a function of two variables: ‘cows per hectare of land’ and ‘purchased concentrate per cow’. Since these variables measure the degree of intensification of a dairy operation, we have labeled as “intensive” the latent class which shows positive estimates of the coefficients of both ‘separating’ variables in the prior probability equation. The estimates of the prior probability function of the intensive class for both models are shown in Table 2.
Table 2. Prior probability equation for the intensive class.
Panel model 
Splitpanel model 

Constant 
11.906* 
13.112** 
ln(cows/land) 
1.7919** 
.8632* 
ln(concentrate/cows) 
1.3510* 
1.5513** 
*, **, Significantly different from zero at 0.05 and 0.01 significance levels respectively
Standard errors reported in Tables A1 and A2 in the Appendix
Next, the farms were classified as Intensive using the highest (greater than 0.5) estimated posterior probabilities (equation 6) since these provide the best estimates of class membership for an individual (Greene, 2005). In Table 3 we show descriptive statistics of the two groups (intensive and extensive) for the two models estimated.
The descriptive statistics of each group roughly agree with the labels we gave to the classes based on the effects of intensification variables on the probability of being in each class. As expected, intensive farms have larger values of key variables such as milk per cow, milk per hectare and feed per cow. Intensive farms are also larger in terms of milk production but are rather similar to extensive farms in terms of land. In our view, the explanation for this result is that marginal increases of land are unlikely to be an option for farmers due to the fact that most abandonments take place in less favoured (mountainous) areas while remaining farms are mainly located in the coastal plain, so that the land available after some farms shut down cannot be used by the remaining farms. For this reason, farmers who wish to increase production need to use more feed per cow and in some cases buy more productive cows, thereby becoming more intensive.
Table 3. Characteristics of the estimated production classes (sample means).
Panel model 
Splitpanel model 

Intensive 
Extensive 
Intensive 
Extensive 

Observations 
804 
732 
798 
738 
Milk (l) 
365442 
280884 
371525 
274994 
Cows 
42.7 
40.3 
43.2 
39.9 
Land (ha) 
18.0 
19.8 
18.3 
19.6 
Cows per hectare 
2.45 
2.16 
2.4 
2.1 
Milk per cow (l) 
8129 
6745 
8202 
6677 
Milk per hectare (l) 
20329 
14779 
20428 
14718 
Purchased concentrate per cow (Kg) 
3533 
3352 
3579 
3304 
In Table 4 we report the output elasticities for the two groups evaluated at the geometric mean of the sample. The differences in the elasticities across groups can be seen as evidence of different technological characteristics. The complete set of estimated parameters of the production functions are reported in Tables A1 and A2 in the Appendix.
Table 4. Output elasticities evaluated at the geometric mean of the sample.
Panel model 
Splitpanel model 

Intensive 
Extensive 
Intensive 
Extensive 

Cows 
.7176** 
.4626** 
.7491** 
.4553** 
Feed 
.2969** 
.3623** 
.2737** 
.3685** 
Farm expenses 
.0405** 
.0718** 
.0508 ** 
.0767** 
Animal expenses 
.0296** 
.0964** 
.0206 * 
.0789** 
Land 
.0368** 
.0339** 
.0214 
.0295* 
*, **, Significantly different from zero at 0.05 and 0.01significance level, respectively
All elasticities, with one exception, are significantly different from zero at conventional levels of significance. Despite the different assumptions behind the two models, the output elasticities evaluated at the sample geometric mean are similar between them. The similarity between the estimates and farm classifications obtained suggests that the differences between the two models (time varying versus constant latent class probability) are a significant but moderate feature of the sample.
On the other hand, it is interesting to note that there are wide differences between the parameters of the two latent classes (within each model). For example, the output elasticity with respect to cows is almost twice as large in the intensive group as it is in the extensive group. On the other hand, the output elasticity of feed is always larger in the extensive group. These different elasticities imply large differences in marginal productivity of inputs across technologies (extensive or intensive), especially for cows and feed.
5. Empirical analysis of intensification and efficiency
In this section we use the results of the estimation of the LCM to analyze a set of issues with important policy implications. First, we are interested in studying the evolution of the intensification process over time. Second, we want to analyze the differences in technical efficiency between intensive and extensive farms.
5.1. Evolution of intensification over time
We want to check if the probability of adopting the intensive technology increases over time. We should note that this analysis is only possible in the splitpanel model proposed in the present paper and not in the conventional panel LCM.
For this purpose, we regress the posterior probabilities of being in the intensive class against individual dummies (fixed effects) and a binary variable (dt) that takes the value zero for the first half of the period analyzed and one for the second half. This is an unconditional analysis over time and is clearly different from the conditional analysis of prior probabilities that could be achieved by including a time trend in equation (5). In the conditional analysis, prior probabilities can change over time, keeping input use and separating variables constant. In the unconditional analysis performed here, the posterior probability varies over time due to changes in the use of key inputs such as feed, cows or land.
The equation to be estimated is the following:
(8)
where j denotes the latent class and w is a random disturbance assumed to follow a normal distribution with zero mean and constant variance. In each class, we have estimated the posterior probability for each individual (i=1,…,128) and for each period (t=1,..,12). Expression (8) represents a general proposal with J different classes. In this case, there would only be J1 free equations since the dependent variable, the posterior probability, adds up to one. As we are considering only two latent classes (Intensive and Extensive) in our setting, we only have one relevant equation. We choose to estimate the equation corresponding to the posterior probability of the Intensive Class.
Table 5 shows the estimated coefficient of the time binary variable (bj) for the Posterior probability of Intensive Class in the Splitpanel model. This coefficient is positive and significantly different from zero, indicating that the probability of being classified as an intensive farm is larger in the second period. We interpret this result as evidence of “intensification” of dairy production in our sample over the period analyzed.
Table 5. Analysis of the evolution over time of posterior probabilities.
Coefficient (bintensive) 
Standard Error 
tstatistic 
.0363 
.0099 
3.28 
Fixed effects not shown
In our view, the average change over time of the probability of belonging to the intensive class provides evidence of farms in the sample tilting towards an intensive production system. Additionally, the change in class probability between the two periods allows for some descriptive analysis of the subset of farms that have switched production system in the conventional sense of crossing the threshold defined by a class probability of 0.5. In particular, 13 extensive farms became intensive in the second period, while 8 farms switched from intensive to extensive. In Table 6 we show the characteristics of the farms that changed to a different production system.
Table 6. Characteristics of the farms that switch production system over time.
From extensive to intensive 
From intensive to extensive 

Period 1 
Period 2 
Period 1 
Period 2 

Farms 
13 
13 
8 
8 
Milk (l) 
279777 
383108 
269831 
301237 
Cows 
37.2 
42.9 
35.3 
42.5 
Land (ha) 
16.5 
16.6 
14.8 
17.7 
Cows per hectare 
2.1 
2.4 
2.4 
2.4 
Milk per cow (l) 
7061 
8381 
7456 
7004 
Milk per hectare (l) 
15796 
21242 
18246 
17191 
Feed per cow (Kg) 
3483 
3824 
3307 
3204 
The farms that become intensive in the second period reflect the typical transformation pattern: increase in the stocking rate and feed per cow, resulting in higher milk per cow and per hectare. On the other hand, the farms that switch from the intensive to the extensive class keep the stocking rate unchanged but reduce the amount of feed per cow, lowering milk per cow and per hectare. We would like to stress two features of this exercise: first, its descriptive and exploratory nature since the 0.5 probability cutoff point that defines intensive or extensive farms is arbitrary; second, the intensification measured by the increase in the probability of being in a latent class is an average effect compatible with farms moving in both directions.
5.2. The effect of intensification on dairy farm efficiency
In this section, we want to explore the effect of intensification on production efficiency. Two questions are addressed. First, are intensive farms more efficient than extensive farms? Second, which technology is more productive (i.e., does one of the two frontiers lie above the other one)?
The level of technical efficiency can be calculated as the ratio between current output and potential output, as defined by the technological frontier. An outputoriented index of technical efficiency can be computed as:
(9)
Subscript j in equation (9) indicates that the technical efficiency index can be calculated with respect to each of the Latent Class frontiers (Orea and Kumbhakar, 2004). In our case, we can thus consider two different frontiers. Table 7 shows the average technical efficiency for the two technologies.
Table 7. Average efficiency of the estimated classes (intensive/extensive).
Panel Model 
SplitPanel Model 

Intensive Frontier 
Extensive Frontier 
Intensive Frontier 
Extensive Frontier 

Full sample 
.92 
.92 
.91 
.93 
Intensive farms 
.94 
.96 
.94 
.96 
Extensive farms 
.88 
.89 
.87 
.89 
As can be seen, for the full sample the average level of technical efficiency is quite similar both between models (Panel vs. SplitPanel) and between latent technologies (Intensive frontier vs. Extensive frontier). However, if we consider the two groups of farms separately, a very interesting result is found: intensive farms have a higher level of technical efficiency in all the four frontiers. Additionally, the average technical efficiency is higher in the extensive frontier than in the intensive one.
This last result seems to indicate that the latent frontier of the Intensive Group dominates the other, that is, for any given set of inputs it is possible to produce more output with the intensive technology. We try to shed some light on this issue by calculating the difference in frontier output between the two frontiers using the actual inputs of the farms:
(10)
where is the (log of) frontier output of farm i at time t evaluated at the intensive (extensive) technology. Table 8 shows the average value of the differences between the frontiers (Dit) calculated using the actual inputs of the farms for the full sample as well as for the two classes.
As the differences between the frontiers are calculated using natural logs, Dit can be interpreted approximately as the percentage difference of potential output between both frontiers. The intensive frontier is, on average, 9.6% above the extensive frontier in the panel model, while this difference is 10.8% in the splitpanel model.
The relationship between technical efficiency and intensification can be explored further by analyzing the correlation between the estimated indexes of efficiency and the probability of being in the intensive class. As technical efficiency is timevarying and so is the probability in the splitpanel model, we have included a time dummy as a control variable in the regression. By doing so, we avoid confounding the effect of the change of probability over time with the effect of time.
Table 8. Average difference between the frontier output of the two technologies.
Panel model 
Splitpanel model 

Full sample 
.0962 
.1080 
Intensive farms 
.0883 
.0988 
Extensive farms 
.1049 
.1180 
Table 9. Relationship between technical efficiency and intensification.
Panel model 
Splitpanel model 

OLS 
OLS 
OLS with farm dummies 

Intensive frontier 
Extensive frontier 
Intensive frontier 
Extensive frontier 
Intensive frontier 
Extensive frontier 

Probability of intensive class 
0.0607** 
0.0658** 
0.0786** 
0.0752** 
0.0545** 
0.0439** 
Time dummy 
0.0010** 
0.0009** 
0.0014** 
0.0007* 
0.0013** 
0.0008** 
*, **, Significantly different from zero at 0.05 and 0.01significance level, respectively
In Table 9, we show the results of regressing the level of technical efficiency against the probability of being in the intensive class and a time dummy that takes the value 1 for the second period. We show the results for both models (Panel and Splitpanel) and for both frontiers (Intensive and Extensive). In the Splitpanel model we perform two different estimations: standard OLS and OLS with farm dummy variables. The inclusion of individual effects is not possible in the conventional Panel model because the probability of class membership is constant over time.
We find common patterns of results for both models and estimators. The probability of being in the intensive class increases the level of technical efficiency with respect to both frontiers in the two models and for both estimation methods. The coefficient of the time dummy indicates that the index of technical efficiency estimated using the extensive frontier increases over time while the index of technical efficiency estimated using the intensive frontier decreases over time. This result is probably due to different patterns across frontiers of the yearly shifts of the production frontier measured by the coefficients of the time dummy variables. The time dummy coefficients of the intensive frontier (Tables A1 and A2 in the appendix) show a clear upward trend at the beginning of the period followed by a fall in the final years. The upward shift of the production frontier over time is compatible with decreasing technical efficiency if the movements of the frontier are due to productive improvements of a subset of leading farms while other farms do not move immediately towards the shifted frontier. On the other hand, the extensive frontier features smaller and erratic shifts over time. In our view, it is not surprising to observe farms approaching, on average, a production frontier with no sudden upward shifts.
Additionally, the Splitpanel model provides a subtle result: the probability of being in the intensive class changes across farms and increases over time (on average). In other words, there are two sources of variation. The coefficient of the probability using plain OLS is estimated using both sources of variation. However, the coefficient of the probability using OLS with farm dummy variables is estimated using only the over time variability. The results show that the estimates of the coefficient of probability are 0.0545 (intensive) and 0.0439 (extensive) when using only changes over time, while the same estimates increase to 0.0786 (intensive) and 0.0752 (extensive) when using both crosssection and time variation of the posterior probability. In summary, it seems that the bulk of the change in efficiency caused by changes in the probability of being in the intensive class can be attributed to changes over time of this probability.
The assumption of timeinvariant prior probability of a latent class can be circumvented by estimating a single LCM but splitting the original panel into two or more periods. This proposal allows for the estimation of the technology of different production systems without an assumption that becomes increasingly untenable as the period of years analyzed increases. On the positive side, our proposal is simple and straightforward. On the negative side, it is not part of a formal model. As a result, there is no clear rationale for choosing the number of periods in which the original panel should be split.
By splitting the original panel into two periods, we find in our empirical application that the probability of being in the “intensive dairy” class increases over time. This result can be interpreted as evidence of dairy farming intensification over the sample period.
We find differences in technical efficiency if we split the sample in terms of the production system using the posterior probability of each latent class. More precisely, the average technical efficiency is higher for farms that belong to the intensive latent class. Additionally, the intensive frontier dominates the extensive frontier, indicating that the intensive technology is more productive that the extensive one. This result can be seen as an economic rationale for the observed trend towards the intensification of dairy farms.
What are the policy implications of these findings? Given that the intensive technology is more productive and that intensive farms are more efficient, i.e. produce closer to their frontier than extensive farms, it seems that the trend towards intensification will continue in the near future. This trend can be seen as a problem if the environmental consequences of intensification are taken into account.
The new reform of the Common Agricultural Policy may also affect farmers’ technology choices. On the one hand, the phasing out of milk quotas may result in higher production. In this sense, intensive farms will find it easier to boost production since they do not depend heavily on forage. At the same time, productivity and efficiency gains associated with intensification can be a plus on a more competitive environment. On the other hand, the new direct payment scheme which will move towards a uniform payment per hectare may lower the incentives to adopt intensive systems.
Alvarez, A., del Corral, J., Solis, D. and Pérez, J.A. (2008). Does Intensification Help to Improve the Economic Efficiency of Dairy Farms? Journal of Dairy Science 91(9): 36933698.
Alvarez, A. and del Corral, J. (2010). Identifying Different Technologies Using a Latent Class Model. Extensive versus Intensive Dairy Farms. European Review of Agricultural Economics 37(2): 231250.
del Corral, J., Perez, J.A. and Roibas, D. (2011). The Impact of Land Fragmentation on Milk Production. Journal of Dairy Science 94(1): 517525.
Greene, W. (2005). Reconsidering Heterogeneity in Panel Data Estimators of the Stochastic Frontier Model. Journal of Econometrics 126(2): 269303.
Hoch, I. (1962). Estimation of Production Function Parameters Combining TimeSeries and CrossSection Data. Econometrica 30(1): 3453.
Kompas, T. and Nhu Che, T. (2006). Technology Choice and Efficiency on Australian Dairy Farms. Australian Journal of Agricultural and Resource Economics 50(1): 6583.
Kumbhakar, S.C. and Lovell, C.A.K. (2000). Stochastic Frontier Analysis. Cambridge: Cambridge University Press.
Newman, C. and Matthews, A. (2006). The Productivity Performance of Irish Dairy Farms 19842000: A Multiple Output Distance Function Approach. Journal of Productivity Analysis 26(2): 191205.
Nehring, R., Sauer, J. Gillespie, J. and Hallahan, C. (2011). Intensive versus Extensive Dairy Production Systems: Dairy States in the Eastern and Midwestern U.S. and Key Pasture Countries the E.U.: Determining the Competitive Edge. Selected Paper prepared for presentation at the Southern Agricultural Economics Association Annual Meeting, Corpus Christi, TX.
Orea, L. and Kumbhakar, S.C. (2004). Efﬁciency Measurement Using a Latent Class Stochastic Frontier Model. Empirical Economics 29(1): 169183.
Simpson, J. R. and Conrad, J.H. (1993). Intensification of Cattle Production Systems in Central America: Why and When. Journal of Dairy Science 76(6): 17441752.
Tauer, L. and Belbase, K. P. (1987). Technical Efficiency of New York Dairy Farms. Northeastern Journal of Agricultural and Resource Economics 16(1): 1016.
Table A.1. Estimation of the Panel Model.
Variable 
Class 1 
Class 2 

Coefficient 
Std. Error 
Coefficient 
Std. Error 

Constant 
12.5724 
.0158 
12.542 
.0152 
Cows (lnx1) 
.7176 
.0213 
.4626 
.0241 
Feed (lnx2) 
.2969 
.0137 
.3623 
.0152 
Crop (lnx3) 
.0405 
.0067 
.0718 
.0080 
Animal (lnx4) 
.0296 
.0091 
.0964 
.0110 
Land (lnx5) 
.0368 
.0104 
.0339 
.0125 
(lnx1)2 
.2030 
.1445 
.0854 
.1096 
(lnx2)2 
.0257 
.0486 
.1737 
.0612 
(lnx3)2 
.0575 
.0152 
.0523 
.0149 
(lnx4)2 
.0731 
.0247 
.0069 
.0320 
(lnx5)2 
.2857 
.0398 
.0624 
.0542 
lnx1lnx2 
.3277 
.0641 
.0663 
.0732 
lnx1lnx3 
.0999 
.0395 
.0103 
.0301 
lnx1lnx4 
.3738 
.0532 
.0394 
.0561 
lnx1lnx5 
.2055 
.0584 
.0263 
.0657 
lnx2lnx3 
.1055 
.0238 
.0608 
.0219 
lnx2lnx4 
.0960 
.0321 
.0827 
.0369 
lnx2lnx5 
.1288 
.0405 
.0160 
.0413 
lnx3lnx4 
.0307 
.0149 
.0025 
.0164 
lnx3lnx5 
.0426 
.0196 
.0196 
.0231 
lnx4lnx5 
.0619 
.0275 
.0138 
.0282 
TD00 
.0206 
.0129 
.0072 
.0172 
TD01 
.0366 
.0131 
.0046 
.0173 
TD02 
.0358 
.0130 
.0243 
.0175 
TD03 
.0313 
.0131 
.0181 
.0173 
TD04 
.0519 
.0131 
.0038 
.0171 
TD05 
.0707 
.0132 
.0359 
.0172 
TD06 
.0952 
.0132 
.0366 
.0173 
TD07 
.0839 
.0132 
.0378 
.0175 
TD08 
.0529 
.0134 
.0156 
.0175 
TD09 
.0559 
.0134 
.0275 
.0177 
TD10 
.0661 
.0139 
.0032 
.0172 
Sigma 
.0909 
.0078 
.1537 
.0064 
Lambda 
1.0506 
.3484 
2.7329 
.3866 
Prior probability equation 

Constant 
11.906 
5.2881 

Ln(cows/land) 
1.7919 
.6078 

Ln(concentrate/cows) 
1.3510 
.6600 
Table A.2. Estimation of the SplitPanel Model.
Variable 
Class 1 
Class 2 

Coefficient 
Std. Error 
Coefficient 
Std. Error 

Constant 
12.5754 
.0135 
12.536 
.0146 
Cows (lnx1) 
.7491 
.0203 
.4553 
.0211 
Feed (lnx2) 
.2737 
.0148 
.3685 
.0149 
Crop (lnx3) 
.0508 
.0068 
.0767 
.0078 
Animal (lnx4) 
.0206 
.0093 
.0789 
.0118 
Land (lnx5) 
.0214 
.0115 
.0295 
.0136 
(lnx1)2 
.2943 
.1448 
.0405 
.1092 
(lnx2)2 
.0121 
.0460 
.0964 
.0645 
(lnx3)2 
.0732 
.0147 
.0410 
.0134 
(lnx4)2 
.0530 
.0263 
.0447 
.0331 
(lnx5)2 
.2786 
.0400 
.0407 
.0609 
lnx1lnx2 
.3067 
.0613 
.0549 
.0693 
lnx1lnx3 
.1158 
.0375 
.0217 
.0316 
lnx1lnx4 
.2953 
.0530 
.0704 
.0591 
lnx1lnx5 
.2457 
.0563 
.0135 
.0596 
lnx2lnx3 
.0825 
.0226 
.0131 
.0217 
lnx2lnx4 
.0526 
.0331 
.0418 
.0396 
lnx2lnx5 
.1130 
.0410 
.0562 
.0452 
lnx3lnx4 
.0275 
.0151 
.0003 
.0160 
lnx3lnx5 
.0263 
.0199 
.0053 
.0224 
lnx4lnx5 
.0350 
.0274 
.0306 
.0289 
TD00 
.0201 
.0126 
.0013 
.0162 
TD01 
.0376 
.0129 
.0108 
.0174 
TD02 
.0351 
.0133 
.0238 
.0188 
TD03 
.0283 
.0130 
.0151 
.0166 
TD04 
.0471 
.0132 
.0099 
.0165 
TD05 
.0660 
.0128 
.0408 
.0167 
TD06 
.0917 
.0130 
.0457 
.0167 
TD07 
.0790 
.0129 
.0463 
.0172 
TD08 
.0531 
.0131 
.0143 
.0171 
TD09 
.0563 
.0131 
.0292 
.0172 
TD10 
.0697 
.0134 
.0007 
.0170 
Sigma 
.0883 
.0067 
.1497 
.0060 
Lambda 
1.1750 
.3155 
3.0344 
.4630 
Prior probability equation 

Constant 
13.112 
4.2444 

Ln(cows/land) 
.8632 
.3943 

Ln(concentrate/cows) 
1.5513 
.5247 
1 The present paper is related with a strand of literature linking technological choices with efficiency. As an example, Kompas and Nhu Che (2006) studied the effect of different technologies on the efficiency of dairy farms by including in the inefficiency model a set of variables reflecting technological choices.
2 We also tried to fit a model with three classes but it did not converge.
DOI: http://dx.doi.org/10.13128/BAE14715
This work is licensed under a Creative Commons Attribution 4.0 International License (CCBY 4.0)
Tel. (0039) 055 2757700 Fax (0039) 055 2757712
Email: info@fupress.com