Analysis of Data from the At-Sea Data Collection Project

David B. Sampson
Oregon State University
Last modified on 8 Aug. 2002

<< pdf version (2.6 mb) >>

The Enhanced Data Collection Project (EDCP), administered by the Oregon Department of Fish and Wildlife, collected data on discard rates for groundfish species and bycatch rates of prohibited species (salmon and Pacific halibut). From late 1995 to early 1999 EDCP observers collected discard data from 235 fishing trips by 25 trawl vessels that voluntarily participated in the program. Besides these observer data, vessels in the program kept enhanced logbooks that recorded retained and discarded catch. These logbooks had data from 866 trips by 44 vessels, including most of the trips with observers. Provided logbook data are acceptably accurate, collecting discard data using logbooks could be a cost-effective supplement to an observer program for measuring discard rates and total discards. Comparisons of tow-by-tow logbook discards with the corresponding observer discards indicated substantial inaccuracies in the logbook information. However, when averaged across tows, trips, and vessels, the discard rates (discard/catch) from the logbooks were lower than the observer discard rates, but predictably so. Generalized linear models were used to determine the major factors contributing to variability in the discard rates. The models indicated tremendous vessel-to-vessel variability in discard rates. Principal components analysis (PCA) was applied to trip-level landings and species compositions from the entire groundfish trawl fleet to summarize the fleet-wide characteristics of fishing trips. Analyses of the PCA scores from the trips that were in the EDCP compared to the scores from all other trips indicated that the EDCP trips probably were not representative of the fleet at large.

In 1995 the Oregon Trawl Commission (OTC) contracted with the Oregon Department of Fish and Wildlife (ODFW) to begin the Enhanced Data Collection Project (EDCP). The major goals of the project were to collect data on trip-limit induced discard rates for primary groundfish species, discard rates for other groundfish species, and bycatch rates of prohibited species such as Pacific halibut and salmon. From late 1995 to early 1999 EDCP at-sea observers collected discard and bycatch data from over 200 fishing trips by trawl vessels that voluntarily participated in the program. In addition to these observer data, boats participating in the program kept enhanced logbooks in which the skippers recorded their retained and discarded catch. These enhanced logbooks have data from over 800 trips, including most of the trips with observers.

For the project described in this report the Commission contracted with Oregon State University to analyze the EDCP databases, which were provided by ODFW on CD-ROM. The analyses compare the discard rates measured by the EDCP at-sea observers with the discard rates reported by the corresponding enhanced logbooks. The overall goal was to examine the feasibility of using logbooks, in conjunction with some observer coverage, to measure discard rates and total discards for use in stock assessments and by management. The main objectives of the analyses were: (a) to evaluate the accuracy of the skippers' estimates of discarded catch; (b) to estimate average discard rates for the major species and determine the factors that contribute to variability in the discard rates; and (c) to evaluate whether the boats that volunteered to participate in the EDCP were representative of the fleet at large.

The ODFW staff organized the EDCP data into three primary files. The "obs_ttow" file has tow-by-tow data collected by the observers on discards of the main groundfish species and the bycatch of halibut and salmon. It contains data from a total of 2,172 tows from 235 fishing trips by 25 different fishing boats. The "obs_tckt" file has trip-by-trip summary data from the observed trips on the total amounts of discards of the main groundfish species and the bycatch of salmon and halibut. This file also contains data from the fish tickets (the official landing receipts) on the retained catches of the main groundfish species. The "logbook" file contains the skippers' tow-by-tow estimates of retained and discarded catch of the main groundfish species and the bycatch of salmon and halibut. These enhanced logbook data are available for 7,400 tows from 866 trips by 44 different boats. Enhanced logbook data are available in the database for all but 9 of the observed trips, but 10 boats failed to report discards in their logbooks when there was an observer on board. Evidently the skippers of these boats mistakenly thought that their logbook data were unnecessary given that an observer was monitoring the discards.

During preliminary screening of the database files we uncovered a number of discrepancies in the data including some missing observer discard records (due to incomplete species composition sample information) and some key-punch errors (e.g., halibut lengths entered in inches instead of cm). We notified ODFW staff of the discrepancies, which they investigated using the original EDCP paper records and subsequently rectified.

The preliminary screening and data exploration also identified certain aspects of the data that, if ignored, could lead to incorrect interpretations. For example, notes associated with the observer data records indicated instances where the observers' recorded discards were "visually estimated"; in many cases the amounts record by the observer were identical to the amounts listed in the corresponding logbook. These data grossly overstate the accuracy of the logbook's discard information. Also, the observer notes sometimes indicated that the discard information for specific tow sequences on a trip represented combinations of tows. The discards reported for the last tow of the sequence included the discards from the earlier tow(s), which reported no discards. In preparing the data for analysis we removed these false zeroes from the observer data and combined the logbook data from these tow sequences so that we could correctly match the observer data with the logbook data. A final issue regarding the observer data was the presence of large numbers of data records where the observer reported that "unidentified fish" were discarded. In preparing the data for analysis we included a flag variable to identify observed tows that had discards of unidentified fish.

The at-sea observers on the EDCP project collected discard data for a wide range of groundfish species, many more than we could reasonably evaluate in our analyses. For this report we limited our detailed statistical analyses to the following species (or groups of species): "All Fish", Dover sole, sablefish, longspine and shortspine thornyheads, English sole, rex sole, petrale sole, sanddab, all Sebastes, canary rockfish, Pacific ocean perch, widow rockfish, yellowtail rockfish , lingcod, and Pacific hake. To facilitate data handling we analyzed the species in groups: deepwater complex (Dover sole, sablefish, and longspine and shortspine thornyheads); flatfish (English sole, rex sole, petrale sole, and sanddab), rockfish (all Sebastes, canary rockfish, Pacific ocean perch, widow rockfish, and yellowtail rockfish), and other fish (lingcod and Pacific hake). We configured the statistical models to include a Species ID factor that allowed us to maintain separate parameter estimates for individual species while analyzing the species collectively.

Evaluating the Accuracy of the Logbook Discard Data

For measuring the accuracy of the discard data reported in the logbooks we created a data file that matched the tow-by-tow logbook discard data with the corresponding observer discard data. We then examined two aspects of these combined data: the proportion of tows for which an observed discard of a given species (or species group) was also reported in the corresponding logbook,

and the ratio of the amount of discard reported in the logbook over the amount observed,

This type of evaluation assumes that the observer data are correct, but in many cases (e.g., tows with discards of unidentified fish or discards that were "visually estimated") this assumption is not entirely valid.

The %Agreement proportions were calculated for each species on a trip-by-trip basis for all trips having matched logbook and observer data. The Log/Obs Ratios were calculated for each species on a tow-by-tow basis for all tows with matched logbook and observer data for which both the logbook and the observer reported some discards.

The %Agreement proportions were formally analyzed using logistic regression methods as implemented in the GLIM Release 4 statistical software program (Francis et al. 1993). The logistic regression model can be expressed as

Logistic regression is an accepted method for the analysis of proportions and other forms of binomial data (Cox and Snell 1989). The Log/Obs Ratios were analyzed using the GLIM software based on the following type of regression model,

The observed Logbook Discards were treated as normally distributed random variables, as in standard regression or analysis of variance.

To assess the statistical significance of different variables and factors that might influence the dependent variables (the %Agreement proportions and the Log/Obs Ratios) we conducted forward-stepwise analyses (Draper and Smith 1966) that examined the relative improvement in fit that resulted from adding explanatory variables to the statistical model. In GLIM the so-called "deviance" statistic, measures how well a statistical model fits the observed data. If the model is correctly specified (i.e., it includes the correct variables in the correct formulation), then the deviance is approximately distributed as a chi-square random variable and changes in the deviance can be used in an F ratio to evaluate the change in goodness of fit (McCullagh and Nelder 1983),

where the deviance and degrees of freedom in the denominator are from the more complete model. The F ratio can be compared to the corresponding theoretical F distribution to evaluate the statistical significance of the change. If a change in deviance is large compared to the background variability (as measured by the deviance divided by the degrees of freedom of the more complete model), then the change is unlikely to be due to random chance alone and the P value for the F ratio statistic will be near zero.

For the %Agreement observations we fit a series of three logistic models to each species. The first model had a single parameter representing the overall proportion agreement about discarding, the second had a separate parameter (proportion) for each boat, and the third had a parameter that allowed %Agreement to vary with the size of the observed discard. For the Log/Obs Ratios we fit a model with a single overall ratio, another with a separate ratio for each boat, and a third with a separate ratio for each trip. We examined two additional models: one to evaluate the influence of unidentified fish discards (Ufish) and the other to evaluate the influence of questionable data (Qtow). The Ufish variable for a tow was set to one if the unidentified fish discards were 10% or more of all the fish discards for that tow; otherwise it was set to zero. The Qtow variable for a tow was set to one if the logbook discard was exactly the same as the observer discard or if the observer discard amount was evenly divisible by 100 (e.g., exactly 400 rather than 402, or exactly 10,000 rather than 10,013).

Estimating the Discard Rates

Because the observers did not measure the amount of catch retained from each tow, the observer data did not provide any direct measurement of discard rates at the level of individual tows. For our analysis of the observer data we derived discard rates for each groundfish species on a trip-by-trip basis using the ratio

This discard rate measures the fraction of the catch that was discarded (as opposed to the amount of fish discarded per hour of fishing or per trip). We calculated similar discard rates on a trip-by-trip basis from the logbook data. We excluded trips with incomplete discard data, ones with unobserved tows or where the logbook had tows with blank (as opposed to zero) discards. We did not attempt to explore discard rates at the level of individual tows. However, the logbook data could be used for calculating tow-by-tow discard rates because they include tow-by-tow estimates of retained catch.

We analyzed transformed discard rates using the following type of regression model,

Prior to running the regression analysis we applied to the discard rates the so-called "angular transformation", as shown in the left-hand portion of the equation, which is a standard technique for converting proportions so that the resulting values have a distribution that is more similar to the normal distribution, on which standard regression methods are based (Zar 1974). We did not consider it appropriate to use logistic regression with the discard rate data because the proportions were based on estimated weights rather than counts.

To evaluate the importance of various factors that might influence the discard rates we conducted forward-stepwise analyses that examined the relative improvement in fit that resulted from adding explanatory variables to the statistical model. In our analysis of the observer discard rates we examined factors for Boat, Port, Gear type (shrimp trawl, sole trawl, bottom trawl, bottom trawl with rollers), Area (Pacific Marine Fishery Commission statistical area), Year, and season (quarter of the year, Qtr). Also, we tested the importance of factors that identified large (>10%) discards of unidentified fish (Ufish) and tows with questionable data (Qtow), as in our analysis of logbook discard accuracy, and we fit models that included continuous variables (as opposed to discrete factors) for the size of the catch (Lbs) and the tow time (Hrs). In our analysis of the logbook discard rates we considered factors for boat, year, and quarter, plus a factor to identify trips that had an observer on board (Obs) and a continuous variable for the size of the catch (Lbs).

Halibut Bycatch Rates

The observers on the EDCP project almost always recorded the number of halibut caught and discarded. The observers were instructed to give this task, along with the recording of bycatch of salmon, the highest priority. During most tows with halibut bycatch the observers measured and recorded the individual lengths, and sometimes the weights, of a sample of the halibut. If small numbers of halibut were caught, all were measured. In the EDCP "halibut" file the ODFW staff included a field with the estimated weights of those individual fish for which length data were available, and another field with the estimated total weights of the entire halibut catch, including fish that were not measured. In the enhanced logbooks used with the EDCP program the skippers were allowed to record their halibut bycatch either in terms of the numbers or pounds of fish. Some reported both measures, but they reported halibut bycatch in terms of weight more frequently than in terms of numbers of fish.

During preliminary examinations of the data we determined that the estimated halibut weights tended to overstate the weights of those halibut that had been weighed. For example, the estimated weight of a 110-cm fish was on average about 2.4 kg heavier than the weight of the 110-cm fish observed during the EDCP project. We used the individual halibut length and weight data from the halibut file to derive coefficients for a length-weight relationship, which we then used to estimate individual halibut weights for those fish with length measurements. From the individual halibut weights for each tow we estimated an average halibut weight for each tow. We estimated the total halibut weight for each tow by multiplying the number of halibut caught on the tow times the estimated average halibut weight for that tow. For the 37 observed tows that caught halibut but none were measured, we used an overall average halibut weight.

Because trawl-caught halibut cannot be retained legally, it is not meaningful to define a halibut bycatch rate in terms of the catch of halibut. All of the catch is always discarded. For our analysis of halibut bycatch we defined two measures for the trip-level bycatch rate, one based on the hours of trawling,

and the other based on the weight of all the landed fish,

Histograms of the halibut bycatch data indicated appreciable numbers of tows and trips that caught no halibut. As a consequence, in our formal analysis of the halibut bycatch we modeled the catch rates using the so-called "delta-distribution" (Pennington, 1983, 1996). The proportion of trips that caught halibut was treated separately from the magnitude of the halibut catch (in numbers or pounds of halibut), and the overall bycatch rate was simply the product of the two components. We used logistic regression methods to formally analyze tabulations of the proportions of trips that caught halibut. To analyze the ratios of halibut catch over hours of trawling (or pounds of fish landings) we first logarithmically transformed the ratios and then applied standard regression methods.

To evaluate the importance of various factors that might influence halibut bycatch, either the proportion of trips catching halibut or the magnitude of the non-zero halibut catches, we conducted forward-stepwise analyses that examined the relative improvement in fit that resulted from adding explanatory variables to the statistical models. In our analysis of the observer data we examined factors for boat, port, gear type, area, year, and a flag to identify tows with questionable data (Qtow), and continuous variables for the size of the fish landings (Lbs) and the tow time (Hrs). In our analysis of the logbook data we considered factors for boat, year, and quarter, and a factor to identify trips that had an observer on board and a continuous variable for the size of the fish landings (Lbs). We excluded from the analyses observer data from trips that had unobserved tows and logbook data from trips for which there were missing logbook discards, unrecorded trawl hours, or no landings of fish.

Evaluating Whether the EDCP Data are Representative

The fishing boats that participated in the EDCP program did so on a voluntary basis. As a consequence, they cannot be considered a random sample of the trawl fleet at large; it may be inappropriate to extrapolate the results from the EDCP program to the entire fleet, despite the pressing need for reliable estimates of discards. To evaluate whether the boats and fishing trips covered by the EDCP program were representative of the trawl fishery in general, we compared traits of the trips covered by the EDCP program with the traits of the rest of the groundfish trawl trips that landed their catch in Oregon. We obtained fish ticket files from ODFW for the period covered by the EDCP program (16 Nov. 1995 through 31 Dec. 1999) and tabulated on a trip-by-trip basis the total fish landings and the proportions landed by species. The tabulations included the following 21 species or species groups: Pacific hake, English sole, petrale sole, Dover sole, rex sole, sanddab, arrowtooth flounder, small rockfish, large rockfish, Pacific ocean perch, widow rockfish, yellowtail rockfish, canary rockfish, shortspine thornyheads, longspine thornyheads, grenadier, sablefish, lingcod, Pacific mackerel, jack mackerel, and skate). We then applied Principal Components Analysis (PCA) to these landings and species composition data to derive PCA scores, which we treated as summary statistics to characterize each trip. Because the species composition numbers are proportions (restricted to the range 0 to 1) we transformed them using the angular transformation prior to application of PCA. In an auxiliary file we included flag variables that identified trips covered by EDCP observers or EDCP logbooks. We compared the PCA scores of the EDCP trips with the PCA scores of the rest of the trips.

Because we did not have ready access to fish ticket information for trips that landed their catch in California or Washington, our analysis was restricted to trips that landed in Oregon. Also, the fish ticket files that we obtained from ODFW did not include shrimp trawl landings of groundfish, but relatively few of the EDCP trips were aboard boats using shrimp trawls.

The fishing trips that had enhanced logbooks and reported discards landed a total of 11.6 million pounds of fish, caught during 5,134 non-zero tows (Table 1). The logbooks for these trips reported total fish discards of 4.7 million pounds (exclusive of halibut and salmon), most of which was Pacific hake (24.8%), shark (17.8%), or sablefish (10.6%). The rockfish species as a group were 12.8% of the reported discards. The observers reported fish discards (exclusive of halibut and salmon) totaling 2.8 million pounds, from 2,102 tows. The largest category of observed discards was unidentified fish (29.8%). Most of the identified discards were Pacific hake (15.4% of all fish discards), shark (11.3%), and sablefish (7.6%). Rockfish species were 11.9% of the observed discards.

Evaluating the Accuracy of the Logbook Discard Data

More than half of the tows for which observer data were available could not be matched to corresponding logbooks, primarily because nine boats never reported discards when they had an observer on board. However, matched logbook and observer discard data were available from 15 boats for 919 tows during 118 fishing trips (Table 2). Discards of fish were reported in the logbooks or observed on 916 of the tows, and the logbooks reported fish discards for 889 of the 913 tows (97.4% agreement) for which the observer reported fish discards. With respect to identifying that individual species had been discarded on a tow, there was in general a fairly poor level of agreement between the logbooks and the observers, except for "high-profile" species such as sablefish (61.4% agreement), canary rockfish (52.2% agreement), and Pacific hake (53.9% agreement). Some species were rarely identified as discards in the logbooks. For example, only 2.7% of the tows with Pacific ocean perch discards and only 2.8% of the tows with English sole discards were correctly identified in the logbooks.

Oddly, the logbooks reported fish discards for three tows for which the observer reported none (Table 2, the column labeled "Lg no Ob"). These may represent reporting errors on the part of the skipper. For a substantial number of tows the logbooks reported discards of unidentified rockfish or unidentified fish that the observer was apparently unable to apportion to individual species. However, the moderate numbers of logbook reports of sablefish, Pacific hake, and shark discards that were not seen by the observers are difficult to explain. For our analysis we assumed that the observer records were correct and that the logbooks contain fair numbers of false-positive records of discards.

The ratios of the discards reported in the logbooks over the discards reported by the observers (Table 2) provides another measure of the accuracy of the logbooks (assuming the observer data are completely accurate). When summed over all tows, including those where the logbooks and observers disagreed that there were discards, the total logbook discards for all fish was only 0.79 of the observer discards. For most individual species the discards reported in the logbooks were considerably smaller than the discards reported by the observers. When the discard amounts were limited to those tows where the logbooks and observers agreed that there were discards, the ratios of logbook discards over observer discards were generally much closer to 1.0, as it would be if the logbook discards (and observer discards) were perfectly accurate.

Our formal analysis of the %Agreement between logbooks and observers that discards occurred (Table 3) in general indicated significant variability amongst the boats. Even with the "All Fish" discards, for which the overall %Agreement was almost 100%, there was one boat that had 0% agreement and another with only 50% agreement (Figure 1). At the level of individual species there was even more variation in the %Agreement values for the different boats. Except for the "All fish" category, there were significant improvements in the fits of the logistic regression models when they included a term for the size of the observed discards (the rows labeled "Boat + ObsDisc" in Table 3). Furthermore, the coefficients on the ObsDisc term were positive for all species except sanddabs (where the coefficient was not significantly different from zero), indicating that the %Agreement tended to increase as the observed discards increased. Large discards were more likely to be reported in the logbooks than small ones.

The formal analysis of the Log/Obs ratios, for tows that had non-zero discards reported by both the logbooks and observers, also indicated highly significant variation amongst the boats for the "All fish" category as well as for the individual species, except for those in the flatfish group (Table 4). Furthermore, there was very significant variation in the Log/Obs ratios at the trip-to-trip level for individual boats (Table 4, the rows labeled "Trip | Boat"). When factors were added to identify tows with more than 10% of the observed discards reported as unidentified fish (ufsh), there was a significant improvement in fit for "All Fish", the Deepwater Complex group and the Rockfish group (Table 4, the rows labeled "Ufsh.ObsDisc"). Also, the Log/Obs ratios for the tows with unidentified fish were in general larger than the ratios for the other tows. This result is consistent with the idea that the unidentified fish discards included quantities of the focal species, so that the reported discards of the focal species were too small. When factors were added to identify tows with questionable observer data (exactly the same as the logbook data or discards that were simple multiples of 100), there was a significant improvement in fit for the "All Fish" category but not for the other groups (Table 4, the rows labeled "Qtow.ObsDisc"). For the "All Fish" model the Log/Obs ratio for the tows without questionable data was 0.67, whereas the ratio for the tows with questionable data was 0.93. Most of the questionable tows were ones having logbook discards that were exactly the same as the observer discards, and many of these also were simple multiples of 100.

Estimating the Discard Rates

Over the course of the entire EDCP program there were 416 trips with complete logbook data (no tows with missing discard information) and there were 205 trips that were fully observed (no tows with missing discard information). We used the discard data from these trips to derive trip-level estimates of discard rate averages and standard deviations (Table 5). In general, the discard rates based on the logbooks were smaller than the rates based on the observers (Figure 3) by a factor of 0.92. The ratio of the logbook discard rate over the observer discard rate was variable among the different species, ranging from a low of 0.16 for petrale sole to 1.7 for widow rockfish. The discard rates from the logbooks were generally more variable than the rates based on the observer data, with the coefficients of variation from the logbook data being about 1.2 times larger than the coefficients of variation from the observer data.

Our formal analysis of the transformed discard rates derived from the observer data (Table 6) indicated highly significant variation among the boats for the "All Fish" category and for all of the individual species groups. The factors Port and Area were also important and significant explanatory variables. The factors for trips having tows with unidentified fish (Ufsh) and those having tows with questionable data (Qtow) were often significant, but had much less explanatory power than other variables, as measured by the rank of the adjusted R² statistic. The variables for the size of the total catch (Lbs) and the tow time (Hrs) were also significant but low ranking explanatory variables. We examined interactions between Boat and the other factors to evaluate in greater detail the importance of the other factors (Table 6, right-hand columns). This analysis indicated a strong Gear effect on the boat-specific discard rates for the "All Fish" category, a strong seasonal effect (Qtr) on the rates for the deepwater complex and flatfish species, a strong Area effect on the rates for the rockfish species, and a strong Year effect on the rates for the other species (lingcod and Pacific hake).

The analysis of the transformed discard rates derived from the logbook data (Table 7) also indicated highly significant variation among the boats for the "All Fish" category and for all of the individual species groups. Our analysis of interactions between Boat and the other factors indicated a strong Year effect on the boat-specific discard rates for the "All Fish" category and all the individual species groups except rockfish, for which the seasonal effect (Qtr) ranked higher than the Year effect. The factor identifying trips with observers on board (Obs) was generally significant, either as a main effect or as an interaction with Boat, which suggests that skippers and crews may have altered either their discarding practices or discard reporting practices when an observer was present.

Halibut Bycatch Rates

Weight and length data were available from 711 individual halibut measured by the EDCP observers. From these data we derived the following weight - length relationship, which we used for estimating weights for halibut with observed lengths but no weights,

For halibut with no observer length or weight data we used the overall average halibut weight, 7.65 kg (16.86 lb).

During the entire EDCP program halibut were caught on 733 of the 1,859 observed tows (39%) and were reported in logbooks for 690 of the 2,956 tows (23%) from the 395 trips that had complete logbook information on discards and hours of trawling (Table 8). Interpretation of the halibut bycatch data in the logbooks is complicated because some fishers reported their halibut bycatch as numbers of fish whereas others reported it as total weight of halibut. From those trips where both observers and logbooks reported halibut bycatch it was possible to make direct tow-by-tow comparisons of the numbers of halibut caught (Fig. 5a) and the weight of the halibut bycatch (Fig. 5b). When the logbooks reported halibut bycatch, the reported numbers of fish caught were almost identical to the numbers counted by the observers (78.4% agreement to within one fish), but on 152 of the 319 tows (48%) where the observers reported halibut bycatch, the corresponding logbooks listed zero halibut bycatch. The non-zero bycatch weights reported in the logbooks were much more variable than the corresponding bycatch numbers and in general were underestimates (57.2% too small) of the weights based on the observer data.

Our formal analysis of the halibut bycatch rates indicated that boat-to-boat variations were the dominant source of variability in the proportions of tows that caught halibut, both for the observer data (Table 9) and the logbook data (Table 10). The factor Boat was also the main source of variability in the logbook data for the four variables that we used to measure the positive halibut bycatch rates (numbers of halibut per trawl hour, weight of halibut per trawl hour, numbers of halibut per weight of gross fish landings, and weight of halibut per weight of gross fish landings). With the observer data the factor Boat was identified as the primary explanatory variable for the bycatch rates based on trawl hours, whereas Area was the primary explanatory variable for the bycatch rates based on the weight of the fish landings and Boat was the 2nd most influential variable. The factor Port was also an important source of variability in both data sets and for all five of the dependent (Y) variables examined. Oddly, the variable for trawl hours (Hrs) was not a significant explanatory variable for the proportions of tows that caught halibut, although one would expect that long tows would more frequently encounter halibut. As measured by the mean squared error (deviance/degrees of freedom), the most precise measure of the overall positive halibut bycatch rate was the one based on the numbers of halibut per pound of fish landings for the observer data and the one based on pounds of halibut per pound of fish landings for the logbook data.

Evaluating Whether the EDCP Data are Representative

To facilitate our analysis of fishing trip characteristics and whether the trips covered by the EDCP were representative, we divided the trip-level groundfish trawl landings in Oregon during the study period into three categories. "Hake" trips landed more than 10,000 pounds of fish that were at least 50% Pacific hake; "big" trips landed more than 10,000 pounds of fish that were less than 50% hake; and "small" trips landed less than 10,000 pounds of fish (Table 11). EDCP logbooks or observers covered none of the hake trips, but ODFW samplers regularly monitor the shore-based fishery for hake in Oregon for discards and bycatch. The hake trips accounted for 79% of the overall landings of fish in Oregon and 31% of the groundfish trips. In terms of their overall landings of fish and groundfish the trips covered by EDCP logbooks or observers tended to be big trips more than small trips, especially for the trips covered by logbooks. There was equal coverage by observers of big- and small-trip landings of the deepwater complex (Dover sole, thornyheads, and sablefish); observers covered the small-trip landings of rockfish more than the big-trip landings.

Principal Components Analysis (PCA) was applied to the trip-level data on overall landings and species composition (percent by weight) to derive a small number of summary statistics for each trip. The PCA for the data from the big trips (Fig. 6a) was moderately successful at simplifying the data, with the first three PCA axes accounting for 15.4%, 12.0%, and 8.0% of the variability in the data. Ideally, the first two or three principal components would account for 50% or more of the variability in the data being analyzed, but for these data the first three components only "explained" 35% of the variability. Similarly, the PCA applied to the data from the small trips (Fig. 6b) was only somewhat successful. The first principal component accounted for 13.7% of the variance, the second for 10.4%, and the third for 9.0%. Although the two data sets (big trips versus small trips) were analyzed separately, both analyses produced similar groupings of trips (lower panels of Fig. 6a and 6b). For example, trips that were dominated by the deep-water species had large positive scores for PCA1 and PCA2.

We calculated simple t-statistics to formally compare the PCA scores from the trips that were part of the EDCP with those that were not. For the big trips we found statistically significant differences (P<0.01) between the PCA1 scores from the trips that had enhanced logbooks versus those that did not, and between the PCA3 scores from the trips covered by observers versus those that were not. These results indicate that the among the trips making big landings of groundfish in Oregon, the trips covered by the EDCP made landings that had different species compositions than the landings by trips that were not covered under the EDCP. For the small trips we found statistically significant differences (P<0.01) between the PCA1 scores and PCA3 from the trips that had enhanced logbooks versus those that did not, as well as for the trips covered by observers versus those that were not. Among the trips making small landings of groundfish in Oregon, the trips covered by the EDCP made landings that had different species compositions than the landings by trips that were not covered under the EDCP. These results provide no support for the presumption that the trips covered by the EDCP were representative of the fleet at large.

Collecting information on discards and bycatch is a time-consuming and costly process. However, unless reasonably accurate information is available on fishery removals (landings plus discards), it is very likely that stock assessment will produce estimates of exploitable stock size that are biased, or at least highly uncertain. Logbooks provide a relatively inexpensive mechanism for the collection of large quantities of data on discards. Our direct comparisons of the logbook discard data with the observer discard data, however, generally indicated that the enhanced logbooks were not very accurate either in terms of their identification of discards or their estimates of the amount discarded. Non-reporting of discards seemed especially to be a problem with unimportant species such flatfish, which are routinely discarded when they are of unmarketable size. Non-reporting of discards of Pacific ocean perch was also a particular problem, presumably because fishers have been discarding small amounts of this species for so long that they are no longer even aware of the practice. Given the intensity of debate in recent years over the catch quotas and trip limit regulations for thornyheads, the non-reporting of discards of the two thornyhead species was surprisingly large (62% non-agreement for longspine thornyheads, 70% non-agreement for shortspine thornyheads).

The estimates of discarding derived from the logbook data were not accurate at the level of individual tows (e.g., Fig. 1 and 2). However, the discard rates derived from the logbook data by averaging across tows and trips were fairly comparable to those based on the observer data (e.g., Fig. 3), although they were generally biased low. It certainly should be feasible to develop bias adjustment factors that could be used to inflate the logbook discard rates so that they more closely resembled the discard rates recorded by the at-sea observers. Extensive collections of discard data from logbooks would greatly increase the sample size and thus could provide much more precise estimates of discard rates. Because the discard rate estimates from the logbook data were generally more variable than the discard rates derived from the observer data (Table 5), it will take two or three trips with logbook data to achieve the same degree of precision as a single observed trip. Except for Pacific hake (discard rate ~= 100%), the discard rates from both the logbook and observer data were extremely variable, with coefficients of variation ranging from 98% (sanddab) to 791% (petrale sole) for the logbook data, and from 76% (sablefish) to 611% (widow rockfish) for the observer data. Such high levels of variability imply that to obtain reliable estimates of discard rates it will be necessary to sample discards from very large numbers of trips.

Large boat-to-boat variations in the data were a dominant feature in the results from all of our analyses. This was the case in the accuracy of the logbooks discards and the estimates of discard rates derived from the logbook data and the observer data. Data collected from a one set of boats could be markedly different from the data collected from a different set of boats, which implies that large numbers of boats will need to be sampled in any program that monitors discards.

If future programs use fishers' logbooks to record discards the programs should provide the fishers with more training and practice to identify species and estimate discard amounts. In our tow-by-tow comparisons of the logbook data with the observer data it appeared that discrepancies between the two sometimes were the result of differences in species identification, for example, with the skipper reporting discards of miscellaneous rockfish while the observer reported several individual rockfish species. Also, a training program could encourage the fishers to be more aware of discarding so that small discards do not go unnoticed, as apparently occurred with the EDCP logbooks. Finally, the training program should help the fishers learn how to estimate discard amounts, especially with those species that are not landed. With species that are routinely discarded (e.g., dogfish) the fishers never receive any feedback to help them learn whether they correctly estimated the weight of the fish they discarded. With retained marketable species they have an opportunity to learn from their bad estimates, for example, when the processor pays them for 1000 pounds of fish but they thought they caught 2000 pounds.

Three individuals provided crucial support with this project. Yong-Woo Lee, Ph.D. candidate in Fisheries at OSU, helped with data screening and statistical analyses, and Jodene Summers and Clay Creech, support staff at the ODFW office in Newport, were very responsive at tracking down information to answer questions about the data. Also, discussions with Keith Matteson and Mark Saelens (ODFW, Newport) helped me understand the data collection process. The Oregon Trawl Commission provided funding for this data analysis project. The Commission, in collaboration with the Oregon Department of Fish and Wildlife, the Pacific States Marine Fisheries Commission, and the Northwest Fisheries Science Center of the National Marine Fisheries Service, supported the Enhanced Groundfish Data Collection Project, which collected and compiled the data that we analyzed. Numerous individuals were involved in producing these data: the at-sea observers and shore-side support staff, and the skippers and crew of the participating vessels. We are grateful to them all for their hard work.

• Cox, D.R. and E.J. Snell. 1989. Analysis of binary data, Second Edition. Chapman and Hall, London. 236 p.
• Draper, N.R. and H. Smith. 1966. Applied regression analysis. John Wiley, New York. 407 p.
• Francis, B., M. Green, and C. Payne. 1993. The GLIM sytem, release 4 manual. Clarendon Press, Oxford. 821 p.
• McCullagh, P. and J.A. Nelder. 1983. Generalized linear models. Chapman and Hall, London. 261 p.
• Pennington, M. 1983. Efficient estimators of abundance, for fish and plankton surveys. Biometrics 39:281-286.
• Pennington, M. 1996. Estimating the mean and variance from highly skewed marine data. Fish. Bull. 94:498-505.
• Zar, J.H. 1974. Biostatistical analysis. Prentice-Hall, Englewood Cliffs, NJ. 620 p.