A set of olive mills were identified across Extremadura and subjected to the two-phase cluster sampling technique, where a target population is grouped into clusters simultaneously selected at random to represent the regional oil-producing zones (Llerena et al., 2009). From these clusters, a random sample of mills was selected. This representative selection comprised 60 (50.4%) of the total 119 olive mills located in Extremadura during the 2005/2006 campaign. Surveys conducted in mills collected organizational and oil processing information as well as samples from already bottled oils to be analyzed in our laboratory. Surveys were performed following a random itinerary system that divided the region into homogeneous olive oil transformation areas.

The data matrix comprised 60 representative virgin olive oil mills of Extremadura. The surveyed 30 variables were related to processing method, olive harvesting method, olive variety, product quality control, differential payment based on oil quality, yield, or other parameters related to olive production and quality. Most of these variables were dichotomous, and coded in the data matrix for subsequent statistical analysis as ⁻¹ if negative, and + 1 if positive. Others were continuous variables, such as milling capacity, percentage of packaging, duration of the season or olive oil quantity produced. In these cases, four value ranges (categories) were established for each of these variables, and real values were classified into these categories, so that their corresponding entries in the data matrix were of an ordinal type.

At each mill, a single oil sample was taken from the tank being bottled at that moment for physicochemical and organoleptic analysis as a representative sample of the olive oil commercialized by this industry. Each sample was stored at –30 °C until analysis. The procedure followed for the analysis corresponds to the EC Regulation N° 2568/91, and measures free acidity (%), peroxide index (meqO₂·kg⁻¹), the ultraviolet extinction coefficient at 270 and 232 nm (K₂₇₀ and K₂₃₂), sitosterol (%), erythrodiol and uvaol (%), waxes (mg·kg⁻¹), as well as an organoleptic assessment.

The compiled data was used to calculate a modified “Overall Quality Index” (OQI) as proposed by Gutiérrez (1989). The index modification was imposed by current legislation which mandates that a sensory evaluation cannot be assigned a numerical value (as with the original OQI), but instead the median of the oils’ fruitiness and negative attributes. Since it is currently impossible to include the sensorial analysis results in OQI calculations, the present work's results are solely based on physicochemical parameters. Thus, we calculated a Physicochemical Overall Quality Index (PCOQI) adapted from Gutierrez (1989) and defined by the following formula:

Where %A corresponds to acidity, K₂₇₀ to absorbance at 270 nm, and PI to peroxide index (meqO₂·kg⁻¹).

The data were analyzed using two specific chemometrics and experimental design software packages: Parvus to perform a principal components analysis, and Nemrod-W (Nemrod-W Software, 2008) to design the experiments and statistically analyze results. Significant differences (p < 0.05) between data sets were evaluated by one-way ANOVAs followed by Duncan's multiple range test when necessary, carried out using software package SPSS version 17.0 for Windows.

Principal component analysis (PCA) is a statistical method of reducing the dimensionality of data by means of a covariance analysis among variables. It is especially well suited for data sets which involve many variables. When these variables have been measured for a set of samples and are plotted spatially, the result is a cloud of values in a multidimensional space. PCA can be used to automatically identify the trends exhibited by these samples and the measured variables, extracting directions in the space along which the cloud is densest.

A statistical analysis was performed using the experimental design software package, Nemrod-W. Each response can be described by the following first-order model: 1 η = β 0 + β 1 . X 1 + β 2 . X 2 + β 3 . X 3 + β 4 . X 4 + β 5 . X 5 + β 6 . X 6 + β 7 . X 7 + β 8 . X 8 + β 9 . X 9 + β 10 . X 10 + β 11 . X 11

where η is the theoretical response function, X_j are the coded variables of the system, and β₀ and β_j are the true model coefficients.

The observed response y_i for the i-th experiment is: 2 y i = η i + e i

where e _i is the error.

The model coefficients β₀ and β_j are estimated by least squares fitting the model to the experimental results obtained at the design points. The notation b₀ and b_j were used for the estimates of these coefficients. The computed values of the responses were then given by ŷ: 3 ŷ y = b 0 + b 1 . x 1 + b 2 . x 2 + b 3 . x 3 + b 4 . x 4 + b 5 . x 5 + b 6 . x 6 + b 7 . x 7 + b 8 . x 8 + b 9 . x 9 + b 10 . x 10 + b 11 . x 11

The model's goodness of fit was tested by comparing the variance due to the lack of fit with the pure error variance using the F-statistic. The model is considered adequate if the variance, due to the lack of fit, is not significantly different from the pure error variance.

Table 1 lists the physicochemical results from each olive oil sample collected at every surveyed mill. Mean olive oil acidity was 0.6%. Values greatly varied, with only 50% of samples registering values between 0.3% and 0.8%. Over 76% of samples showed acidities below the maximum allowed for EVOO. One of the samples (1.7% total) presented an above 2% acidity value.

With respect to peroxide index values, all samples presented values below the maximum allowed by European Regulations for qualities fit for bottling. It was noteworthy that 75% of the samples showed values below 10 meqO₂·kg⁻¹.

The average for the K₂₃₂ absorbance values was 1.95, with almost 97% of samples showing below the legislated maximum for EVOO: only two samples (3.4% of total) presented a value above 2.50. With respect to K₂₇₀ values, over 86% of the samples registered values below the legislated maximum for EVOO. Only two samples presented values above 0.25. With regards to Delta K values, all samples were within the legislated limits.

All sitosterol values satisfied the legislated minimum of 93.0%, with the average being 94.7%, and only one sample showing the minimum value allowed by law.

Triterpene alcohols, erythrodiol and uvaol registered a mean value of 2.6%, but showed above 4% values in 10% of the oils. Only one sample recorded a value above the legislated maximum. This same sample had a wax content of 125 mg·kg⁻¹. It is likely that these values corresponded to the character of the predominant oil varieties in its area of origin and/or poor processing practices rather than potential fraudulent admixtures.

With regards to the sensorial analysis, only 35% of samples registered fruitiness intensities above zero and lack of defects, thus corresponding to the EVOO category. While 58.3% of the samples were designated as VOO according to the sensorial analysis, only 56.6% also satisfied the required physicochemical criteria for this category. In addition, 10% of the oils corresponded to the LOO category which is unfit for bottling, although only 6.7% of them fell into this category solely due to their sensorial characteristics. Nevertheless, it should be noted that, at the time the samples were analyzed, Regulation 2568/94 classified oils as LOO when the median defect was 2.5 of intensity, and this value was later increased to 3.5 when the EC Regulation N° 640/2008 of 4 July 2008 came into effect.

Table 2 shows the average values of the analyzed parameters with oils grouped under the correctly assigned commercial category. The mean free acidity of the EVOO category (35% of samples) was 0.3%, with 50% of values between 0.2 and 0.4, while the VOO category showed a mean value of 0.7. The corresponding values of the LOO category were widely spread, indicating that the parameter primarily responsible for their lower rating corresponded to the intensity of their negative attributes in the sensorial analysis.

The oil peroxide index values were significantly different (p < 0.01) among the three categories, with the lowest values corresponding to the top category. There were no significant differences, however, among the values of K₂₃₂ or erythrodiol and uvaol. For K₂₇₀, sitosterol, and waxes, the LOO rated oils showed significantly higher values than the other two categories.

The calculated values of PCOQI were significantly different (p < 0.05) among groups, although the formula used to define this index would not correctly classify oils since two LOO category samples registered values above 8.2, emphasizing the relevance of the sensorial analysis in rating the quality of olive oils.

The above results indicate that the main reason for olive oils not qualifying for the top category is their performance in the sensorial analysis. In particular, although their physicochemical parameters might satisfy the thresholds set in the legislation, they may present negative attributes in the tasting panels at intensities which do not allow them to be classified as highest quality. The causes of this sensorial impairment may be varied, and require an analysis of the processing at each mill to identify the main cause responsible for the loss in virgin olive oil quality.

In the present study, PCA was used as an exploratory analysis to search for clustering within the set of variables and samples, hence reducing the number of variables needed to explain their relationships. Data sets were subject to different pre-treatments, and the procedure applied implemented the NIPALS (Nonlinear Iterative Partial Least Squares) algorithm. Figure 1 shows the plotting of loadings and scores on the first two principal components’ space. In this figure, column auto scaling has been applied to the data (to equalize the variances), and oil samples and variables are represented by producer identification number and name code, respectively.

The results revealed no clustering trend in either variables or samples. The oil samples were fairly evenly distributed over the experimental space, and only producers 89, 94, and 103 were notably distant from the origin of the PCA plot.

The variables with greatest leverage values (distance from the origin of the first two principal component axes) were selected as the most descriptive of the sample set. The greater the leverage value and its corresponding projection onto each principal component axis, the greater the amount of information it carries about the system for each of the two principal components represented. Figure 1 shows that variables such as differential payment for clean fruit and for variety, and type of weighing carry no relevant information for the purposes of the study since they are positioned close to the origin. In contrast, 11 variables were pointed out by the screening study as explaining most of the relevant physicochemical information in the systom. These variables, the scores corresponding to the dichotomous responses, and the number of positive responses for each variable are listed in Table 3.

Surveys were carried out at 60 olive oil mills, collecting information on 30 parameters related to the processing method. Virgin olive oil samples were analyzed by measuring the parameters as described in the Data Analysis. The resulting models from each response were only found to be significant for PCOQI and acidity, but unexpectedly not for the organoleptic assessment. These two indicators were influenced by ground/top separation of fruits, differential payment based on acidity, and type of process (continuous/press).

Other variables did not affect the final product quality, one of them being ownership type. It could be expected that links between cooperative olive mills and their suppliers would be more internalized than in privately owned mills since suppliers themselves are partners and have the guarantee that each annual crop has a buyer (Moyano and Núñez-Nickel, 2007). The fact that this does not seem to be the case may reflect a lack of involvement from growers in the marketing of the end product, and in particular, a certain lack of concern about its final quality, since they see the milling industry as a mere transformer of their olives (Montegut-Salla et al., 2007).

Differentiated payment based on fruit characteristics (such as fruit cleanliness, oil yield, variety, acidity, etc.) had no statistically significant influence on the final quality of the oil; Neither did specific steps during the transformation process such as the use of decanting wells prior to shipment to the warehouse, the establishment of quality control procedures, or even the sensorial analysis which, although vital for marketing, were not found to influence the final quality of the oil. This may be because the initial separation of ground- from tree-harvested fruit outweighs other variables in oil production. In particular, since this separation would be the first step in the production chain, it may already establish whether or not the final olive oil will be of high quality, so that the following steps in the process have no further statistical significance (Jimenez et al., 1998).

Other variables studied showed no statistically significant influence on the analytical parameters: content in erythrodiol and uvaol, sitosterol, and waxes, the spectrophotometric measurements (K₂₃₂ and K₂₇₀), and the peroxide index.

Influence of variables on PCOQI. Only one variable (ground/top separation of fruits) significantly influenced the PCOQI indicator. Table 4 lists the estimated values of coefficients (weights associated with each variable considered in the model) and results of the t-test significance analysis. These results are graphically presented by two histograms on Figure 2. Plotting on the left shows the weight of each of the 11 variables taking the highest level as reference, and on the right, displays the estimates of the coefficients corresponding to these variables in the model explaining the quality indicator (PCOQI). As shown in Fig. 2, only coefficient b2 reaches significance.

These results confirm other findings in the literature (Jiménez et al., 2003) which state that the separation of ground- from tree-harvested fruits is crucial in order to obtain the highest quality olive oils, since the fall from the tree and contact with the soil promotes a marked increase in free fatty acid content, the variable which has the greatest weight in the PCOQI formula.

Of all surveyed mills, 20% continue to harvest fruit without classification (Table 3), which leads to final lower quality oils which have to be marketed as VOO. Several of these mills attributed the lack of separation to logistic reasons (absence of facilities to carry it out) and to growers refusal to perform this separation since they are usually not paid extra for this quality improvement. This latter aspect was clearly reflected on mills which paid differentially for ground-harvested and tree-harvested olives producing virgin olive oils with lower free acidity content.

It can then be concluded that olive oil producers who separate ground from top fruit during harvesting score higher on PCOQI. This seems to be related to the general oil quality increase trend registered in Extremadura and other producing regions during the last 15 years. For instance, Extremadura has gone from applying ground-top separation at only 19% of its mills (Uceda et al., 2001) to 63% in 2006 (Llerena et al., 2009). A similar increase has been described for Andalusia, from only 11% of its mills in the period from 1993-1999 (Uceda et al., 2001) to 58% in the period from 2001-2002 (De Toro-Jordano, 2002).

This quality increase derives from a greater awareness of its importance in marketing the final bottled EVOO product. Progress is still slow because many mills are not oriented towards the final consumer market. Their focus is on competing to capture the producer end of the chain since they think the VOO they produce will sell on the bulk market at a price only slightly cheaper than a good EVOO (Moyano and Núñez-Nickel, 2006). It is also possible that, rather than a lack of awareness from growers and mill owners, the quality differentiated production of oils may be economically unviable due to limited production of raw material in the local area served by the mill and the consequent limited processing capacity of the mill itself.

Influence of variables on acidity. Acidity is related to the proportion of free fatty acids in the oil, and related to the physical integrity of the fruit before being processed (Panzanaro et al., 2010). The data screening revealed three variables significantly affecting acidity: ground/top fruit separation, differential payment based on acidity, and production method. Table 4 lists the estimated values of coefficients (weights associated with each variable considered in the model), and Figure 3 graphically presents the results in two histograms. Plotting on the left shows the weight of each of the 11 variables taking the highest level as reference, and on the right displays the estimated values of coefficients corresponding to the weight of these variables in the model explaining the quality indicator (acidity). It can be observed that oils with the highest acidity values correspond to those where the process used is batch production by pressing and which lack olive separation and differential payment based on acidity.

Acidity is related to what has been examined in the previous subsection, since this parameter has the greatest influence on the value of PCOQI. Not only do other analytical parameters, such as peroxide index or K272, have less influence in the formula, but their values are similar in VOO and EVOO oils.

García et al. (2006) estimated that a 20% proportion of ground-harvested fruit is enough to cause the free acidity of the resulting oil to exceed the ceiling of the EVOO category (established at 1.0%), although with the current lowering of the maximum acidity to 0.8%, this proportion is reduced to 2.5% to 10.0%, depending on the quality of ground-harvested olives (Montaño, 2010).

A second variable that affected the quality of the final product was differential payment for ground/top harvested fruit. This clearly represents an incentive for growers to separate the two qualities of olives. With this payment difference, the mill shows that it is willing to obtain different oil qualities (extra virgin and virgin), and is therefore willing to pay somewhat more to growers who bring fruit separated by harvesting method.

The third relevant variable in the production of low acidity virgin olive oils was the production system, which was second in terms of weight after ground/top separation of fruit. This finding is consistent with those previously reported in the literature (Di Giovacchino et al., 1994; Uceda et al., 2001) stating that the continuous systems facilitate the production of higher quality oils than batch press systems. Although theoretically any system which protects fruit quality is potentially capable of producing extra virgin olive oils, in practice this is not so. The reason is that the intrinsic characteristics of olive harvesting campaigns, where variables such as cleanliness, work load, olive pulp and oil areas of contact (batch filters versus stainless steel containers), etc., lead to the batch press system significantly undermining the theoretical quality of the final olive oil versus the better maintenance of quality in continuous extraction systems.