| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
ANIMAL GENETICS |
Department of Animal Science, Iowa State University, Ames 50011
 |
Abstract |
|---|
 Top Abstract IntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
Key Words: Automatic Feed Dispensers • Errors • Pigs • Quality Standards
|
Introduction |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
; De Haer et al., 1992) are used to automatically measure individual feed intake on group-housed growing pigs. These feeders are frequently used in genetic nucleus herds and on research farms; however, data from electronic feeders have been found to contain errors (De Haer et al., 1992; Eissen et al., 1998). Errors are largely the result of malfunctions that are inherent to the feeders because of the environment in which they operate, which is moist, dusty, and contains corrosive gases. In addition, the playful and aggressive nature of pigs can cause equipment malfunctions. Although many steps are taken to minimize these malfunctions, they still occur and cause errors in recorded data from these feeders. Eissen et al. (1998) developed nine criteria to identify errors in data from IVOG (Insentec, Marknesse, The Netherlands) electronic feeders. In concept, these criteria also can be used to identify errors in data from Feed Intake Recording Equipment (FIRE; Osborne Industries Inc., Osborne, KS) electronic feeders, which are predominantly used in the United States, but criteria must be adapted to accommodate feeder differences. Current software provided with FIRE feeders uses four criteria by default to filter errors. More comprehensive criteria are needed. Eissen et al. (1998) proposed that identification of errors could be used to diagnose feeder malfunctions. To accomplish this goal, an analysis of factors that are associated with errors is needed to better understand the types of errors that occur and what causes them. Therefore, the objectives of this study were to 1) develop more comprehensive criteria to identify errors in data from FIRE feeders; 2) quantify the frequency of errors across three consecutive experiments; and 3) identify factors associated with errors. |
Materials and Methods |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
Data used to develop criteria were provided by the National Pork Board (Clive, IA) and came from the Maternal Line Evaluation Program (MLEP) experiment (Moeller et al., 2004). These data (Table 1) were from 591 crossbred pigs that represented six breeds (one sire line by six maternal lines), two sexes (barrows and gilts), and two replicates (Replicates 2 and 3). Data from Replicate 1 were unavailable. Growth and feed intake were measured on each pig, starting and ending at an average BW of 50.9 and 114.3 kg.
|
), and the feeder recognized individual pigs via an electronic transponder in an ear tag. Each feeder contained a feed trough that was suspended and continually weighed by a load cell. When a pig entered the feeder, the weight of the feed trough, pig transponder number, time, and feeder number were recorded (Table 2). Weight of the feed trough and time were recorded when a pig exited the feeder. Data were recorded and stored electronically at the feeder until they were downloaded to a computer using software provided with the feeders. The software used four criteria by default to filter errors from the raw data to create summary reports. The raw, unfiltered data were used in this study.
|
. However, this approach is time consuming and could result in errors. To avoid these potential errors, data from three pens with excessive feed disappearance in unidentified visits (>90 kg) were discarded. Feed disappearance over the test period was obtained by summing recorded intake of unidentified visits, excluding visits for which recorded intake was less than –20 g or greater than 2,000 g, which were assumed to be invalid. The threshold of 90 kg for excessive feed disappearance was subjectively determined based on its frequency distribution across pens.
Sixteen criteria were developed to identify errors in individual visits (Table 3). Development of these criteria started by computing seven variables in which errors could occur for each visit in the complete data set (see Table 2 for an example): feed intake per visit (FIV); occupation time per visit (OTV); feeding rate per visit (FRV); feed trough weight differences between subsequent visits in time; and time differences between subsequent visits. Because weight and time differences involve two visits, each variable was separated into two variables to distinguish between the leading (leading weight difference [LWD] and leading time difference, [LTD]) and following (following weight difference [FWD] and following time difference [FTD]) visits in time. Only variables FIV, FRV, and FWD were used by Eissen et al. (1998) to identify errors in data from the IVOG feeder. Following the calculation of these variables, data from pigs that were removed early from the test were discarded.
|
). A total of 16 criteria were developed, which will be referred to by the descriptive acronyms listed in Table 3. Criteria were established based on frequency histograms of the variables, knowledge of the feeder, or based on criteria developed by Eissen et al. (1998). Frequency of Errors
To quantify the frequency and variability of errors, the criteria of Table 3 were used to code each visit in the MLEP data and in two other experiments conducted by the National Pork Board, the Quality Lean Growth Modeling Program (QLGMP; Robison et al., 2000) and the Genetics of Lean Efficiency Program (GLEP; Newcom et al., 2002), for the presence (1) or absence (0) of each error type. An additional 1/0 variable was created to indicate presence of at least one error of any type (overall error, OE). Frequencies of each error type and of two error types occurring in the same visit were computed for unidentified and identified visits from the MLEP data. The co-occurrence of errors was used to understand relationships among error types.
All three experiments (MLEP, QLGMP, and GLEP) were conducted at the Minnesota Pork Producers Association test station in New Ulm using the same FIRE feeders, starting with the QLGMP and ending with the GLEP. Data from the first replicate were not available for all three studies, and data from 11 feeders (15.7%) in the QLGMP, three feeders (6.5%) in the MLEP, and one feeder (4.3%) in the GLEP were discarded because of excessive (>90 kg, see explanation above) feed disappearance from unidentified visits. Remaining data from the QLGMP (Table 1) were from 893 pigs that represented six genetic lines, two sexes (barrows and gilts), four diets (varied by lysine level), and two replicates. Remaining data from the GLEP (Table 1) were from 237 purebred pigs that represented two breeds (Yorkshire and Duroc), two sexes (barrows and gilts), and one replicate.
Factors Associated with Errors
To identify factors associated with each error type, a logistic regression model (Dobson, 2002) was fitted to each error type and to OE in identified visits from the MLEP data. The 17 response variables (presence, y = 1, or absence, y = 0, of an error type or overall error) were assumed to have a binomial distribution: y ~ Bin (1, p) where p is the probability that y = 1. A linear predictor (
= Xß + Zu; –
), along with the logit inverse link function (exp() (1 + exp())–1) were fitted to the data so that:
 |
The following model was fitted for the linear predictor ():
 |
where µ = intercept, Ri = fixed effect of the ith replicate, Sj = fixed effect of the jth sex, b1xk = linear effect of the kth day in the test period, b2x2k = quadratic effect of the kth day in the test period, Fli = effect of the lth feeder in the ith replicate, assumed random with Fli ~ N(0,
2F), Pmli = effect of the mth pig within the lth feeder in the ith replicate, assumed random with Pmli ~ N(0,2P), Dki = effect of the kth day in the ith replicate, assumed random with Dki ~ N(0,2D), and Wnli = effect of the nth week within the lth feeder in the ith replicate, assumed random with Wnli ~N(0,2W).
Breed also was originally fitted but was dropped because it was not significant (P > 0.90). The GLIMMIX macro of SAS (SAS Inst., Inc., Cary, NC) was used to fit the model. Restricted/residual pseudo likelihood was used to estimate the variance components (Wolfinger and O’Connell, 1993). Convergence was not reached for five error types (FIV-0, OTV-lo, OTV-hi, FRV-hi-FIV-lo, and FRV-0), likely because of low frequencies. The CONTAIN option was used to calculate denominator degrees of freedom for F-tests of the fixed effects. After solutions were obtained, the linear predictor was converted into the probability of an error occurring using the logit inverse link function:
 |
|
Results and Discussion |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
To identify errors in individual visits, thresholds were assigned to variables related to feed intake, occupation time, and feeding rate per visit, and to consistency of weight and time data between subsequent visits in time.
Feed Intake Per Visit.
Three criteria were used to identify errors in the variable FIV, with FIV-lo identifying visits with negative FIV, FIV-hi identifying visits with too high FIV, and FIV-0 identifying visits of 0 s duration with greater than expected FIV (Table 3). Visits were classified as FIV-lo if FIV was less than –20 g, which was based on measurement error. Brown and Van der Steen (1990) reported that the average difference between recorded and actual FIV from FIRE feeders was 0.6 g (SD = 8.5 g). Measurement error can result in FIV = –20.7 g (= –2.5 SD) if no feed was consumed, so visits with FIV <–20 g were identified as errors for FIV-lo. Eissen et al. (1998) also used –20 g for this criterion. An age-dependent criterion, which varied from +20 to +30 g, was used by Nienaber et al. (1990). They assumed that recorded consumption less than these values were false visits or caused by insignificant "nibbling." The FIRE software used –300 g by default for this criterion.
Error type FIV-hi was designed to identify visits where recorded feed intake was larger than that physically possible. The histogram for FIV (Figure 1) was used to identify the point in the tail of the distribution where the frequency showed a substantial decrease, following the removal of very extreme values. This point (FIV = 2,000 g) was used as the threshold to identify errors of this type. The same threshold was used by Eissen et al. (1998), but the FIRE software used 1,500 g by default. Histograms were used similarly for other error types.
|
) was used. Eissen et al. (1998) used thresholds of ± 10 g.
Occupation Time Per Visit.
Two criteria were developed to identify errors in the variable OTV (Table 3). These error types were not used by Eissen et al. (1998), although they can occur in data from IVOG feeders. Error type OTV-lo was designed to detect visits where the recorded exit time was less than the recorded entrance time, which also was used as the default value by the FIRE software. Such errors can occur as a result of electronic malfunctions caused by electric storms or power fluctuations. Error type OTV-hi was designed to detect visits that lasted longer than reasonable. Based on the histogram of OTV, 1 h was chosen as the threshold. Software for the FIRE feeder used 2 h by default.
Feeding Rate Per Visit.
Five criteria were developed to identify errors in the variable FRV (Table 3). Error types FRV-hi-FIV-lo, FRV-hi-strict, and FRV-hi were designed to detect visits where recorded feed consumption was faster than is reasonable. Error type FRV-hi-FIV-lo was applied to visits with low intake (0 < FIV < 50 g). Such visits could result in high FRV if a pig took a mouthful of feed and left the feeder quickly. To avoid coding such visits as errors, FRV-hi-FIV-lo was given a larger threshold (500 g/min) than FRV-hi-strict and FRV-hi. The threshold for FRV-hi-FIV-lo was based on the histogram of FRV for these visits and was similar to the 600 g/min used by Eissen et al. (1998).
Eissen et al. (1998) detected a pattern in the data where visits with a negative value for feed intake were often directly preceded or followed by a visit with an elevated positive feed intake. They recommended using a more stringent criterion on FRV for the visit associated with the positive feed intake. Error type FRV-hi-strict was used to identify such visits, with a threshold of 110 g/min (Table 3), which was taken from Eissen et al. (1998). For all other visits, FRV-hi was used to detect unreasonably large FRV, with a threshold of 170 g/min, which was based on the histogram of FRV. This threshold was similar to the 150 g/min used by Eissen et al. (1998) but larger than the 100 g/min used by McDonald and Nienaber (1994).
Error type FRV-0 was used to detect visits where no feed was consumed and occupation time was long (Table 3). Based on the histogram of OTV for visits where FRV = 0 g/min, a threshold of 500 s was chosen, which was twice that used by Eissen et al. (1998). Error type FRV-lo was used to detect visits where recorded feed consumption was slower than reasonable. The threshold of ± 2 g/min was based on the histogram of FRV of all visits, except those where FRV = 0 g/min. This threshold was the same as that used by Eissen et al. (1998), but less than the 10 g/min used by McDonald et al. (1991). Nienaber et al. (1990) also identified visits with this type of error but did not specify the threshold used.
Consistency Between Subsequent Visits.
Error types LWD-lo, LWD-hi, FWD-lo, FWD-hi, LTD-lo, and FTD-lo (Table 3) were used to detect unreasonable differences between feed trough weights and time from subsequent visits in time. Because these differences involved two visits, two criteria (leading and following) were used to code both visits. Error types LWD-lo and FWD-lo were used to detect negative weight differences using a threshold of –20 g, which was based on measurement error.
Error types LWD-hi and FWD-hi were used to detect weight differences that were larger than reasonable (Table 3). A threshold of 1,800 g was chosen based on the operation of FIRE feeders. The FIRE feeder weighs only the feed trough, which contains a maximum of 1,000 g of feed. When the amount of feed in the feed trough is less than a level set by the user (usually 500 g) a volume of feed is dispensed from a bin above the trough. The weight of the feed dispensed is usually about 450 g, but it depends on the density of the feed. Feed can be dispensed between or during visits. If the feeder detects a transponder when it dispenses feed, the amount of feed dispensed is added to the entrance weight of the trough. If the feeder cannot weigh the amount dispensed because the pig moves the feed trough, then it uses the dynamic portion calibration (DPC) value as an estimate of feed dispensed. The DPC value is based on a rolling average weight of previously dispensed feed that could be measured. The threshold of 1,800 g for weight differences between subsequent visits was based on the feeder dispensing a maximum of four times (450 g x 4). For example, if a pig finished a visit and there was 200 g of feed left in the trough, the feeder would dispense feed once between visits, bringing the quantity of feed in the trough to 650 g. If the next pig that entered the feeder ate 2,000 g, the feeder would dispense three times during the visit, which would increase the entrance trough weight of that visit by 1,350 g (3 x 450 g). Thus, the weight difference between visits would be 1,800 g.
Eissen et al. (1998) used a threshold of ± 20 g, which was based on measurement error, to identify errors based on the difference in trough weight between subsequent visits. They did not distinguish between the leading and the following visit but gave both visits the same error type. Additional criteria were not used by Eissen et al. (1998) because of the different design and operation of the IVOG feeder. The IVOG feeder weighs a feed hopper that functions as both a trough and a storage bin. As a result, feed is dispensed to the feeder fewer times than with the FIRE feeder. In addition, when feed is dispensed, a separate visit is recorded, unlike for FIRE feeders. As a result, comparisons between feed trough weights between subsequent visits in time should only differ by measurement error for the IVOG feeder.
A negative time difference between subsequent visits in time was used as the criterion for LTD-lo and FTD-lo (Table 3). Eissen et al. (1998) did not use these two criteria.
Frequency and Nature of Errors in Unidentified Visits
The frequency of unidentified visits varied considerably between the three experiments and ranged from 6.2 to 67.2% (Table 1). This was much greater than the frequency (0.95%) found by Eissen et al. (1998). The percentage of unidentified visits increased in the order in which the projects were completed (QLGMP, MLEP, and GLEP), but the percentage of pens discarded because of excessive feed disappearance in unidentified visits decreased. The reason for the increase in the percentage of unidentified visits is not clear, but the decrease in discarded pens is likely the result of the increasing ability of the managers to operate the feeders. Pens were discarded because pigs were eating but not being identified by the feeder. For example, power outages can cause the identification system to become out of tune on some feeders. As a result, all visits to that feeder will be unidentified until it is tuned. If this problem is not fixed for several days, then feed disappearance would surpass the 90-kg threshold, and data from that pen would be discarded. As the manager better understands feeder malfunctions, the number of pens discarded should decrease. If this is true, then the number of unidentified visits may not be as important as feed disappearance during these visits.
The MLEP data set was used to further investigate the nature of unidentified visits. For this data set, mean feed intake per unidentified visit was 3.4 g, and 85.1% had feed intakes that were less than 20 g, which can be attributed to measurement error. The frequency and co-occurrence of error types for unidentified visits are given in Table 4 (above diagonal). Error types FIV-lo, FIV-0, and FRV-hi-FIV-lo were common in unidentified visits (12.1, 14.4, and 3.4%), which can be attributed to rodent activity. For example, if a 25-g mouse entered the feed trough, an unidentified visit would be recorded, where FIV = –25 g and OTV would be small, resulting in FIV-lo. If the visit lasted 0 s, then the visit also would be coded for FIV-0. This seems to occur frequently because approximately half the unidentified visits coded for FIV-0 also were coded for FIV-lo. When the mouse exits the feeder, FRV-hi-FIV-lo could occur because the visit would last a short time with FIV = 25 g. Eissen et al. (1998) also found that FRV-hi-FIV-lo was common in unidentified visits (16.2%), but FIV-lo and FIV-0 were not common (0 and 0.3%), likely due to differences in feeder operation.
|
Percentages of the 16 error types in identified visits for the three experiments (QLGMP, MLEP, and GLEP) are given in Table 5. The overall error rate varied from 4.33 to 18.74%. Eissen et al. (1998) reported an overall error rate of 5.7% in identified visits and also found considerable variation between groups of pigs on the same feeders. For each error type, the percent varied considerably between projects (CV of up to 144%), except for OTV-lo. Eissen et al. (1998) also found considerable variation in the proportion of each error type between subsequent groups of pigs in the same pen. The proportion of errors between projects tended to decrease by their order of completion (QLGMP, MLEP, and GLEP). With the QLGMP representing the first experience of National Pork Board personnel with electronic feeders, the decreasing error rates could reflect the increasing ability of the managers to operate the feeders. Changes are, however, confounded with a decreasing number of pigs per pen (Table 1), which also could decrease error rates.
|
). Literature values range from 29 (Eissen et al., 1999) to 34.6% (Eissen et al., 1998). The MLEP had a smaller proportion of daily feed intake records with errors than the GLEP, despite the fact that the MLEP had a larger proportion of overall errors. This indicates that errors were spread over more days in the GLEP than in the MLEP.
To develop a better understanding of the nature and relationships between different error types, frequencies of each error type and of two error types occurring in the same visit in identified visits from the MLEP data are listed in Table 4 (below diagonal). Examples will be used to illustrate how a given feeder malfunction can result in multiple error types in the same visit.
Feed Intake Per Visit.
Error type FIV-lo occurred at a relatively high frequency (0.8%, Table 4). Of the 2,319 visits that contained FIV-lo, 8.6 and 22.1% also contained LWD-lo and FWD-lo, respectively. This was similar to the value of 26.6% for FIV-lo combined with LWD or FWD found by Eissen et al. (1998). Error type FIV-lo can occur if the feed trough is weighed too heavy at the end of a visit, too light at the beginning of a visit, or if the pig dropped an object in the feed trough. Incorrect trough weights could be caused by a load cell that is not calibrated properly or that is malfunctioning or by a trough that is not hanging freely. This happens when there is debris under the trough or when the trough touches the frame of the feeder. If the trough was weighed too light at the start of a visit, it could result in FWD-lo, as demonstrated by Example 1 of Table 6, using data modified from Table 2. If the trough was weighed too heavy at the end of a visit, it could result in LWD-lo, as demonstrated by Example 2 of Table 6. The fact that FWD-lo and FIV-lo occurred more often together than LWD-lo and FIV-lo (Table 4) suggests that entrance weight was incorrect more often than exit weight, although both occurred.
|
). Error type FIV-hi could result if the trough was weighed too heavy at the start of the visit, too light at the end of the visit, by excessive feed wastage, or if the DPC value was used and it overestimated the amount dispensed. Error type FWD-hi could result concurrently if the trough was weighed too heavy at the beginning or the DPC value was used and was too large (see Example 3, Table 6). This happened 97% of the time. If the trough was weighed too light at the end, then LWD-hi could result concurrently with FIV-hi (see Example 4 in Table 6); this happened 30% of the time. These results suggest that entrance weight was incorrect more often than exit weight, consistent with results based on FIV-lo.
Visits with high intake (FIV-hi) also tended to have a high feeding rate (FRV-hi-strict and FRV-hi), as demonstrated by the data (Table 4) and in Examples 3 and 4 of Table 6. Eissen et al. (1998) also found this relationship. Error type FIV-0 rarely occurred simultaneously with other error types.
Occupation Time Per Visit.
Error types OTV-lo and OTV-hi occurred at low frequencies in identified visits and occurred mostly together with LTD-lo (80 and 52%) and FTD-lo (20 and 19%; Table 4). Error type OTV-lo can occur when the time at the feeder does not match the time at the computer, and the user resets the time while a visit is in progress. Error type OTV-hi would be identified if recorded exit time was too large, recorded entrance time was too small, or if a pig were to lie with its head in the feeder. If the recorded entrance time were too small, then it would tend to also cause FTD-lo, as demonstrated by Example 5 of Table 6. If recorded exit time were too large, then it would tend to cause LTD-lo, as demonstrated by Example 6 of Table 6. Error type LTD-lo occurred more frequently with OTV-hi than FTD-lo (Table 4), which indicates that exit time was recorded incorrectly more often than entrance time. In addition, OTV-hi occurred frequently with FRV-0 and FRV-lo (24 and 53%, respectively), which could occur if time were recorded incorrectly by a large degree (6,000 s), but a normal amount of feed was consumed (FIV = 199 g; Table 2), as shown in Example 5 of Table 6. More likely, however, co-occurrence of these error types is caused by pigs that lie with their head in the feeder, while eating very little.
Feeding Rate Per Visit.
Percentages of error types relating to feeding rate in identified visits ranged from 0.03 to 1% (Table 4). Error types FRV-hi-FIV-lo and FRV-0 occurred infrequently and rarely with other error types. Error types FRV-hi-strict and FRV-hi reflected visits that had a large feeding rate and could be caused by overestimating feed intake or by underestimating occupation time. These error types occurred frequently with FWD-hi. Error type FRV-hi also occurred frequently with LWD-hi. Relationships of FRV-hi with LWD-hi and FWD-hi are demonstrated by Examples 3 and 4 of Table 6, and are similar to the relationships of FIV-hi with LWD-hi and FWD-hi, which suggests that these errors were caused by an overestimation of feed intake. Error type FRV-lo reflected visits that had a small feeding rate and could be caused by underestimating feed intake, overestimating occupation time, or by pigs lying in the protective crate and eating little. This error type was the most frequent of the feeding rate errors but infrequently occurred with other error types.
Consistency Between Subsequent Visits.
Percentages of error types relating to weight and time differences between subsequent visits ranged from 0.39 to 1% (Table 4). Error types LWD-hi and FWD-hi frequently occurred together, which could happen if the DPC were too large and used in both visits. Error types LTD-lo and FTD-lo also frequently occurred in the same visit.
Factors Associated with Errors
A logistic regression analysis was used to identify factors associated with each error type, including random and fixed effects. Results from five error types (FIV-0, OTV-lo, OTV-hi, FRV-hi-FIV-lo, and FRV-0) will not be presented because convergence was not reached.
Random Effects.
Estimates of variances for the four random effects that were fitted are presented in Figure 2. The largest percentage of variability (32 to 57%), after correcting for fixed effects, was explained by week within feeder and replicate for all error types, except LTD-lo and FTD-lo. This effect explains week-to-week variation in the amount of errors within a feeder and could result from temporary feeder malfunctions. Eissen et al. (1998) also found that feeders with the most errors had a period of at least 2 wk with a relatively large proportion of errors. These results indicate the importance of timely detection and correction of feeder malfunctions.
|
) in error types. This variable accounts for day-to-day variation in the amount of errors within a replicate across feeders. Error types FRV-lo, LTD-lo, and FTD-lo were most affected by this factor, which could be explained by electrical storms. High temperatures could increase FRV-lo as a result of pigs lying in the protective crate to stay cool.
Variance explained by feeder within replicate ranged from 0 to 38% (Figure 2). This factor explains feeder-to-feeder variation in error rates within a replicate and can be caused by recurrent malfunction of individual feeders. Error types LWD-lo and FWD-lo were affected most by this factor, which can result from feeders that frequently weigh incorrectly.
Variance explained by pig within feeder and replicate ranged from 4 to 25% (Figure 2) and may reflect behavior differences among pigs. For example, aggressive pigs could cause more errors. Error types FIV-lo, FIV-hi, FRV-hi-strict, and FWD-hi were most affected by this factor. Error type FIV-lo could result if a pig enters the feeder, does not eat, and plays with the feed trough. If a pig wastes a significant quantity of feed, this could result in FIV-hi and FRV-hi-strict.
Fixed Effects.
Probabilities of an error occurring for OE, FIV-lo, FIV-hi, FRV-hi, FRV-lo, LWD-lo, and LWD-hi as a function of test day, replicate, and sex are plotted in Figures 3 to 6. Error type FRV-hi-strict was not plotted because the probabilities were small (p < 0.001). Graphs for FWD-lo and FWD-hi were similar to those for LWD-lo and LWD-hi (Figure 6) and are not shown. The effect of sex was significant (P < 0.05) for FRV-hi-strict, FRV-hi, FRV-lo, and LWD-hi. Males caused more errors than females for all error types, except for LTD-lo and FTD-lo. The effect of replicate was only significant for FTD-lo, and there were more errors of this type in Replicate 2. The effect of test day tended to be significant for overall error (P = 0.057), but was not significant for LTD-lo, and FTD-lo. The quadratic effect of test day was left in the model if P < 0.10, which occurred for FIV-lo, FRV-hi-strict, FRV-lo, LWD-lo, and FWD-lo. In general, the number of errors increased with test day, except for FRV-lo, LWD-lo, and FWD-lo. Eissen et al. (1998) also observed this increasing trend and suggested that using a feeder for a longer period without maintenance increased the number of errors. It could, however, also be explained by the increasing weight of the pigs as an experiment progressed. The number of errors decreased with test day for FRV-lo, LWD-lo, and FWD-lo. These errors could be caused by pigs that play with the feed trough and eat little, especially in feeders where the trough does not hang freely, which could explain the large pen effect in Figure 3 for LWD-lo and FWD-lo. These errors may occur more frequently at the beginning of the test period because younger animals are more active and playful and the amount of time the feeder is not occupied for feed consumption is greater.
|
|
|
|
Electronic feeders are used mainly in genetic nucleus herds to improve feed conversion over and above what can be achieved by selecting on lean growth rate. The added benefit of measuring feed intake could be negated by the inaccuracies caused by errors. In order to achieve the added benefit of measuring feed intake, the number of errors must be decreased and accurately edited. This could also decrease the environmental variation of feed intake, which would increase heritability and genetic improvement of feed conversion. The 16 criteria developed in this study can be used to edit feed intake data after an experiment is completed, as proposed by Eissen et al. (1999). A subsequent paper develops a method of editing and correction of feed intake data for errors based on the 16 criteria. The ability of editing methods to accurately estimate feed intake traits is to be investigated in a subsequent article (Casey, 2003).
Although the current paper shows that errors can be identified in data from electronic feeders, it is likely that the 16 criteria developed in this study do not identify all errors and will incorrectly classify some visits as errors. In general, thresholds for the criteria described in Table 3 were designed to be conservative in order to maximize the probability of detecting errors. This resulted in more incorrect identification of errors, but we decided that this was preferable. Future work could include relaxing the criteria and comparing results with this study. The criteria also could be improved by developing age-dependent thresholds, as Nienaber et al. (1990) used for the lower threshold of FIV. An age-dependent threshold also could be used for the maximum amount of feed consumption per visit (FIV-hi).
Results from this study indicate that errors occur in data from electronic feeders, and the number of such errors is substantial, even in the experiment that had the fewest errors (GLEP). Eissen et al. (1999), however, showed that an accurate estimate (r = 0.96) of ADFI across the growing period can be obtained even if only 30% of daily records are useable, as defined by having no visits with errors. Errors will, however, have a greater effect on the accuracy of feed intake trait records when shorter test periods are considered.
Although error rates were affected by feeder, pig, weather, day within the test period, and sex, management of the feeder seemed to be the main factor. Because feeder management is so vital to data quality, a list of problems and recommendations, which can be used to ensure proper functioning of electronic (FIRE) feeders, are given in Table 7. In addition, the 16 criteria described in Table 3 could be used during an experiment to diagnose feeder malfunctions, as proposed by Eissen et al. (1998). A computer could easily generate daily reports that contain the frequency of each error type occurring in data from each feeder. These reports could then be used to diagnose feeder malfunctions based on the knowledge of the error type and the factors that are associated with the error type. Fixing feeder malfunctions quickly will decrease the number of errors and will increase the accuracy of measuring feed intake.
|
|
Implications |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
|
Footnotes |
|---|
2 Current address: PIC, 3033 Nashville Rd., Franklin, KY 42134.
3 Current address: 4900 Berkeley Place, Univ. of California, Irvine 92697.
4 Correspondence: 225C Kildee Hall (phone: 515-294-7509; fax: 515-294-9150; e-mail: jdekkers{at}iastate.edu).
Received for publication February 15, 2004. Accepted for publication January 24, 2005.
|
Literature Cited |
|---|
|
TopAbstractIntroductionMaterials and MethodsResults and DiscussionImplicationsLiterature Cited |
|---|
This article has been cited by other articles:
|
|
 |
|
  W. Cai, D. S. Casey, and J. C. M. Dekkers Selection response and genetic parameters for residual feed intake in Yorkshire swine J Anim Sci, February 1, 2008; 86(2): 287 - 298. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
