Survey Methodology and Data Collection

The 2015 VIMS fishery-independent survey was specifically modified to test for a vessel effect and, thus, differed slightly from normal survey operations. Under normal conditions, two different vessels survey a portion of the survey domain on a survey leg in a given year. The survey generally occurs in May to minimize temporal variability among years and the time among survey legs is minimal (i.e., 2–4 days). A survey leg is defined as the time it takes a vessel to complete the predetermined number of stations and return to port. All vessels participating in the surveys are required to be similar in length, have a minimum horsepower of 800, and have sufficient space to complete the required sampling. In 2015, three vessels were selected to participate in the survey to test for a vessel effect. Vessel characteristics are provided in Table 1. Three consecutive survey legs were completed by the vessels, with each vessel sampling the entire survey domain. For a balanced design, each vessel surveyed a similar number of stations per strata and for the entire domain (Table 2, Fig. 1).

TABLE 2.

Number of stations and percentage of stations with at least one sea scallop caught included in all analyses by strata and vessel for the 2015 VIMS survey.

The 2015 survey was conducted from the Virginia/North Carolina border to Block Island, RI, in May and June of 2015 (Fig. 1, Table 1). The survey domain consisted of a combination of NEFSC core scallop survey strata within the MAB resource area and additional strata that have been identified as important scallop habitat (NEFSC 2018). Sea scallop strata are delineated by depth and latitude (NEFSC 2018). The sampling framework was a stratified-random design. Stations were randomly allocated to strata using a hybrid approach consisting of both proportional (station allocation based on stratum area) and Neyman allocation techniques using the prior year's scallop survey catch data in weight and number of animals (Cochran 1977).

One of the contemporary approaches used to examine for the effect of vessel is a paired vessel field study, where vessels fish in close proximity using identical fishing gear to reduce spatial heterogeneity in catch rates (Cadigan & Dowden 2010, Cadigan et al. 2006, Fowler & Showell 2009, Miller 2013, Rudders & DuPaul 2010). Several constraints, such as limited staff and equipment, as well as the financial burden associated with deploying multiple vessels during a survey leg precluded the use of this approach for the survey. The assumption with a stratified-random sampling design is that catch within a stratum is more homogenous, reducing within strata variability (Cochran 1977, Helser et al. 2004). With the theory of homogeneity within strata in mind, having each vessel sample a similar number of randomly selected stations within each stratum was deemed as a valid approach to test for vessel effect. Scallops are also generally considered to be sessile animals, which minimized concerns regarding changes in scallop distribution and abundance associated with migration among strata during the survey (Smith & Rago 2004). Ehrich (1991) diverged from the paired vessel design to determine whether a vessel effect existed for haddock (Melanogrammus aeglefinus) catch between two research vessels. Vessels in Ehrich's (1991) study completed randomly selected stations within predetermined areas with known concentrations of haddock to deal with the high variability in haddock catch rates that had been observed in the survey.

At each station, vessels towed a NEFSC standard survey dredge with a 2.4 m dredge frame equipped with 5 cm rings, 10.2 cm diamond twine top, and 3.8 cm diamond mesh liner. Survey tows followed a standard protocol with a 15 min tow time at a speed-over-ground of approximately 1.9–2 m/sec (3.9–4 kts). A standard tow distance following speed and time protocols is approximately 1.8 km, although variability in tow distance can occur due to environmental conditions (i.e., wave height, wind speed, wind direction, and tide), as well as vessel speed, which can also be affected by the same environmental conditions. Starting and end coordinates of a tow were recorded by the chief scientist in the bridge using a GPS. Vessel speed and location were also automatically recorded every 3 sec during a tow from the GPS data string. A Star Oddi™ inclinometer was placed on the dredge to record dredge angle and was used to determine realized bottom contact time. Tow distance was calculated after completion of a survey using bottom contact time, vessel speed, and coordinate data. Captains were instructed by the chief scientist to set the survey dredge at approximately 0.9 km ahead of the station location and tow through the station on the same course. The direction of a tow was generally made in the direction of the next station to be sampled. Deviations from where and what direction a tow was completed were due to a number of factors including navigational hazards (i.e., ship wrecks or hangs that would damage the survey gear), weather conditions that made sampling on deck difficult, and the presence of fixed fishing gear. All modifications to how a station was sampled were done in consultation with the chief scientist and the vessel remained in the strata, where the station was originally allocated for the entirety of a tow.

Standardized catch sampling was conducted in the same manner as described by DuPaul and Kirkley (1995). All scallop catch was placed into bushel baskets and counted to quantify total catch at a given station. This included all sizes of live scallops, including damaged scallops (i.e., the shell was punctured, crushed, or broken). To obtain an estimate of the number and size of scallops caught at each station, either the total scallop catch or a subsample, depending on catch volume, was measured from the umbo to the shell margin to the nearest millimeter (mm). If a subsample was selected for length measurements, one to three baskets were randomly selected from the total catch after being placed in bushel baskets. The number of baskets selected for subsampling was scaled to the total scallop catch at a station: all scallops were measured if scallop catch was 1 basket or less, 1–5 baskets of total scallop catch equaled one basket selected for subsampling, 5–20 baskets equaled 2 baskets selected for subsampling, and more than 20 baskets equaled 3 baskets selected for subsampling. The size frequency of the entire catch at a station was determined by applying an expansion factor (number of baskets caught/number of baskets measured in a subsample) to the sampled number caught-at-length. The number of scallops caught-at-length measured in a subsample without applying an expansion factor is referred to as the unexpanded catch-at-length.

Data

Station-level catch data for all successful survey tows were queried for sea scallop catch (number of animals) and length data (number-at-length) for 28 strata. A successful tow is defined as a tow, where the survey dredge fished correctly, the dredge did not flip over, there were no hangs with the gear as it was hauled in, and no fixed fishing gear was caught on the dredge. Successful tows included stations, where zero scallops were caught in the survey dredge. Strata with less than three valid survey tows were excluded from analysis (n = 6), leaving station-level data from 22 strata for analysis. The unexpanded number-at-length data were summed to obtain the total number of scallops at each station by vessel for the pooled catch analysis. This resulted in 588 station-level observations for analysis (Table 2). The unexpanded number-at-length data in 1 mm length bins were used for the catch-at-length analysis (Fig. 2). Stations with zero catch were removed for length analysis (n = 525 stations). The number of length bins for each vessel was relatively similar, with 156 length bins for Vessel 1, 149 for Vessel 2 and 146 for Vessel 3. Using the unexpanded scallop data with an offset to account for differences in subsampling rates is the recommended approach for analyzing count data, where variability exists in sampling at the observation level (Cadigan & Dowden 2010; Cadigan et al. 2006, Holst & Revill 2009, Maunder & Punt 2004, Zhang & Chen 2015, Zuur et al. 2009).

Survey stations were assigned to rotational access areas to account for commercial effort and removals by the fishery, as well as growth. Fishery removals may impact the catch and catch-at-length of vessels completing later survey legs as scallops were removed from the population that would have been available to the survey if the rotational access area was not open to the fishery. Scallops also exhibit faster growth in rotational access areas compared with scallops outside of these areas (Hart & Chute 2009, NEFSC 2018). Rotational area designation is illustrated in Figure 1. Stations within a rotational area were coded as a 1 and stations outside rotational access areas were coded as 0 for analysis.

Figure 2.

Sea scallop relative length frequency distributions by strata and vessel for the 2015 VIMS survey.

Although standard survey protocols control for tow distance, variability in tow distance can occur for several reasons, as stated previously. Significant differences in the mean tow distance among vessels was detected with a one way ANOVA (P < 0.01). An additional offset term was included in analyses to account for this, as these differences may affect scallop catch and variability existed at the station-level (Cadigan and Dowden 2009, Cadigan et al. 2006, Helser et al. 2004, Moriarty et al. 2020, Zuur et al. 2009).

Pooled Catch Analysis

Generalized linear models were developed to determine whether there was a significant difference in the total catch of scallops among vessels. Modeling approaches to estimate a vessel effect for cooperative research surveys have been conducted via generalized linear mixed models, where models were developed with vessel as a fixed and random effect (Cooper et al. 2004, Helser et al. 2004, Thorson & Ward 2014). Results from these studies indicated that modeling vessel, as a random effect, was preferred and vessel was a source random variation when estimating biomass. The data precluded considering vessel as a random effect because three vessels participated in the survey and the threshold for including a variable as a random effect is to have a minimum of five levels of a variable (Bolker 2008, Zuur et al. 2009). For analysis, models were developed treating vessel as a fixed effect to test for the effect of vessel. Generalized linear models allow for flexibility in specifying the error structure of the response variable, which, in this case, was count data, and allow for the incorporation of additional variables beyond vessel that may also impact scallop catch like rotational area and strata (Zuur et al. 2009).

The response variable was the unexpanded number of scallops caught. Predictor variables considered were: vessel (a categorical variable with 3 factor levels), strata (a categorical variable with 22 factor levels), and rotational area (a categorical variable with 2 factor levels). An offset term to account for subsampling of catch and tow distance at the station-level was also included the following methods from Millar (1992), Cadigan et al. (1996), and Cadigan and Dowden (2010). The general GLM was:

where, g is the log link function, u_v,s,a is the expected mean number of scallops caught by vessel ν in strata s and rotational area a. α is the intercept, and x_i is an offset term In(towdistance_i • subsampling fraction_i) at station i. The negative binomial error distribution was chosen after preliminary analysis with a Poisson error distribution indicated the data were overdispersed (Zuur et al. 2009). The negative binomial distribution is also shown in Eq. 1, where y_v,s,a is the number of scallops caught by vessel ν in strata s and rotational area a, and k is the overdispersion parameter.

All combinations of covariate candidate models were developed using the MuMIn R package (Barton 2009) and the glm. nb function in the MASS R package with maximum likelihood for parameter estimation (Venables & Ripley 2002). A manual forward model selection process was also completed to confirm modeling results and identify any other potentially useful models. Models were compared with the Bayesian information criterion (BIC) and confirmed with the Akaike information criterion (AIC) due to the tendency of AIC based model selection to favor more complicated models (Burnham & Anderson 2002). The model with the lowest BIC was selected as the best fitting model (Burnham & Anderson 2002). Models with a BIC within three units of the BIC_min and an AIC within three units of the AIC_min were considered equally plausible as preferred models (Bolker 2008, Raftery 1995). Overall significance of model terms was determined with the ANOVA function in the car R package by applying a Wald's test (Fox & Weisberg 2019). Model goodness-of-fit was assessed with residual diagnostics including a Q–Q plot, residuals against fitted model values, and residuals against all explanatory variables.

Tukey's honest significance test (HSD) was used to conduct post hoc pairwise comparisons to test for significant differences among vessel factor levels for the preferred model(s), if the vessel term was included in the preferred model(s) (Major et al. 2017, Miller 1981, Nanda et al. 2021, Roop et al. 2018, Zuur et al. 2009). The null hypothesis for this test is that there are no differences between parameter mean estimates and the test uses confidence intervals to determine significance among specific pairs of factor levels. This approach can be preferred over a significance value from an ANOVA for categorical variables (Nanda et al. 2021). The glht function in the multcomp R package was used to carry out the tests (Hothorn et al. 2008). All analyses were completed with the statistical program R (R Core Team 2020). Statistical significance (α) was equal to 0.05 for all analyses.

Catch-At-Length Analysis

Generalized additive mixed models were developed to test for differences in the number of scallops caught-at-length among vessels. This family of models allow for flexibility in modeling nonlinear relationships between a response variable and predictor variables through the use of smoothing functions (Hastie & Tibshirani 1990, Pedersen et al. 2019, Wood 2017). Generalized additive mixed models can also incorporate categorical and continuous predictor variables and various error distributions. The response variable for these models was the unexpanded number of scallops-at-length caught in 1 mm length bins. Predictor variables considered were: length (mm), stratum, rotational area, vessel, and an interaction of length and vessel. Thin plate regression smoothers were used for both the length and interaction terms. The interaction term was modeled as a factor smooth interaction following examples from Wood (2017) and Pedersen et al. (2019), where the smoother of length was modeled by the vessel term (length, by = vessel) to allow for smoothers to be estimated for each vessel level. The vessel term was also included as a categorical predictor to allow for parametric parameter estimates for Vessels 2 and 3, and to test for significance differences between the reference vessel level (Vessel 1) and the other two vessel levels. All other terms entered into the models as categorical variables with the same number of factor levels as described for the GLM. A random effect smoother was used for the random effect term of station, as this approach allows for the significance of the random effect to be determined (Pedersen et al. 2019). Station was included as a random intercept to account for the spatial autocorrelation of scallops caught at a given station (Pinheiro & Bates 2000, Zuur et al. 2009). The same offset term applied in the GLM was also used for the GAMMs to account for differences in subsampling of catch and tow distance at the station level. Generalized additive mixed models were developed with a negative binomial error structure. The general GAMM was:

where, g, i, n, s, a, x_i , and α have the same notation as in Eq. 1. µ_i,n_,s,a, is the expected mean number of scallops-at-length as a function of the smooth term s for length l and the smooth terms for the interaction of length by vessel (s(l, by = n)), strata s, and rotational area a. g_i is the random effect of station i that is assumed to be independent and identically distributed with a mean of 0 and variance of . The negative binomial distribution is also shown in Eq. 2, where y_l,n,s,a is the number of scallops caught by length l and vessel n, in strata s and rotational area a, and k is the overdispersion parameter.

Candidate models were developed with manual forward and backward selection procedures for all combinations of variables. Model selection procedures were also similar to those used for the GLM, where preferred models were determined based BIC values and confirmed by AIC. Model fit was also examined with residual diagnostics, similar to the GLM. Models were fit with restricted maximum likelihood using the mgcv R package and the gam function (Wood 2011). Overall significance of both smooth and parametric model terms was determined with the ANOVA function in the mgcv R package by applying a Wald's test (Wood 2011). A Tukey's HSD test was also completed for the preferred model(s) to test for significant differences among vessel factors parametric estimates following the same approach outline for the GLM. The predicted number of scallops at-length was estimated with 95% confidence intervals for the preferred model(s) and plotted to assess the length frequency distributions by vessel.

Survey Characteristics

The entire survey was completed over the course of 43 days from May 16, 2015 to June 28, 2015 (Table 1). There was an 11-day delay between legs 1 and 2, and leg 3 departed after a 7-day hiatus (Table 1). Trip duration was similar among vessels (Table 1). The average number of stations with positive scallop catch rates ranged from 92% to 96% and the percentage of positive catches was generally consistent across strata (Table 2). Catch per unit effort was variable among strata, whereas relatively consistent across vessels within strata, and 95% confidence intervals overlapped for all strata (Fig. 3). Four strata showed variability in catch per unit effort among vessels. Strata 9 and 34 tend to have more variable catch rates due to the patchy distribution of scallops and being located at northern and southern extents of the MAB resource. Strata 18 and 19 are located within rotational access areas that were open to fishing activity in 2015 during the course of the survey.

Figure 3.

Catch per unit effort (standard survey tow) in number of sea scallops by strata and vessel with 95% confidence intervals.

TABLE 3.

Candidate generalized linear models for the pooled scallop catch analysis ranked by BIC.

Pooled Catch Analysis

Seven candidate GLM were developed for the pooled catch analysis (Table 3). Model glm6 was the preferred model based on BIC, whereas model glm7 had a similar AIC value (ΔAIC = 1.6) compared with the preferred model. For the preferred model, predictors of scallop catch were rotational area and stratum (Table 4); both terms were significant (P < 0.01 for both terms). Although model glm7 was not considered a preferred model based on selection criteria, summary information for this model is included because AIC values were comparable to model glm6. In model glm7, both stratum and rotational area were again significant predictors of scallop catch (P < 0.01 for both terms), whereas vessel was not a significant predictor (P = 0.30). There were also no significant differences among the reference vessel factor level, Vessel 1, and either of the other two vessel factor levels (Vessel 2 P = 0.11 and Vessel 3 P = 0.57). Tukey's HSD test was completed for model glm7, where vessel was included as a predictor. There was no significant difference in mean scallop catch among all combinations of vessels (all P > 0.24). The manual forward selection procedure produced similar model results. Model diagnostics and partial effect plots for predictors in model glm6 indicated an acceptable goodness-of-fit (Fig. 4).

Catch-At-Length Analysis

In total, 10 GAMM were developed for the catch-at-length analysis (Table 5). Models gam7 and gam9 were identified as preferred models based on BIC model selection criteria, although model gam9 had a significantly higher AIC compared with model gam7 (ΔAIC = 204.8). Predictors in model gam7 included stratum, rotational area, and the length:vessel interaction term. All smooth interaction terms were significant (all P < 0.01), indicating the smoothers were significantly different from zero (Table 6, Fig. 5). Both stratum and rotational area were significant predictors of scallop catch-at-length (P < 0.01 for both terms), whereas the vessel length interaction term was not significant (P = 0.58). Parametric vessel term estimates indicated the effect size for Vessels 2 and 3 were small and there was no significant difference between the reference vessel (Vessel 1) and either Vessel 2 or Vessel 3 (Table 6). The overall shape and effective degrees of freedom for each vessel specific smoother were similar (Fig. 5). The summed effects plots with 95% confidence intervals by vessel indicated no differences in catch-at-length across the length range as confidence intervals for all three vessels overlapped across the entire length range (Fig. 6). The predicted catch-at-length plots by vessel and strata also confirmed no differences in catch-at-length, with 95% confidence intervals overlapping across the length range sampled for all vessels (Fig. 7). The random effect of station was significant (P < 0.01) and the estimated variance was 1.43 (SD 0.004). A Tukey's HSD test was completed for model gam7. There was no significant difference between the mean scallop catch-at-length and all combinations of vessel factor levels (all P > 0.90). The other preferred model, gam9, did not have the length:vessel interaction term or the vessel term as predictors and indicted stratum, length, and rotational area effected scallop catch-at-length. Model diagnostics for both preferred model were acceptable.

TABLE 4.

Parameter estimates and P values for predictors of stratum and rotational area from the preferred GLM (glm6) for the pooled scallop catch analysis.