Fishing estimates of species composition and catch per unit effort

We fished in a standardized manner for 4–6 h/d (from about 0730 hours to 1330 hours) for a total of 12 d at each of the four sites sampled in October and November 2003, and 15 d at each of the two study sites sampled in July, August, and September 2005. The commercial fishermen distributed fishing effort throughout the study site each trip in order to sample each portion of the study site each day. Other than being asked to fish in all parts of the study site, the decisions about where and how to fish were left to the fishermen. Each fisherman used techniques (e.g., bait, soak time, type and number of hooks, traps, or sticks used) commonly used in commercial fishing operations. Fishing methods used included trap, handline, and stick gear. Traps (0.6 × 0.6 × 0.5 m) were deployed singly or on a string of two traps, and usually 10 traps were deployed at a time. Handlines consisted of a weight (approximately 1 kg) and two baited hooks on 40 kg-test fishing line, and were fished for set amounts of time, ranging from 5 to 25 min. Sticks were deployed for approximately 1 h at a time on single lines and buoys, and contained five hooks per stick. Sticks are 1-m lengths of steel bar (1-cm diameter) with five baited hooks on 0.3-m-long leaders attached at uniform intervals along the length of the bar. A line is attached to the bar and buoyed at the surface for deployment and retrieval. We usually deployed 10 sets of sticks at a time. Sticks and traps were typically deployed on the bottom for approximately 1 h. Traps were baited with squid Loligo opalescens and anchovies Engraulis mordax, whereas sticks and handlines were baited almost exclusively with squid.

All captured species were measured for total length (TL) and released at location of capture. We collected information on species composition, TL, sex (when possible), and the fishing time and depth at which each unit of gear was fished. Actual depth ranges sampled by the different sampling gears across all sample sites were 5–22 m for handlines, 4–26 m for sticks, and 4–22 m for traps. Additionally, at the two sites sampled in 2005, we placed external dart tags in fishes for use in tag–recapture estimates of population sizes. Dart tags were color-coded based on the type of gear used to catch the fish and the location (low-relief or high-relief site) of release. Mortality of tagged fishes was low because we fished in shallow water and handled captured fish carefully. These same techniques resulted in a handling mortality of 1.4–2.4% in a previous study (Starr and Green 2007).

For stick and trap fishing methods, CPUE was calculated by dividing the number of fish caught on an individual stick or trap by the number of hours the method was deployed. Catch per unit effort of handlines was calculated by dividing the number of fish caught per angler by the time fished. During the study, if the anglers using handlines did not catch fish within 2–3 min, the skipper relocated the boat and fishing continued in a different spot, frequently one that was only a few meters away. These short (<3-min-long) periods were not recorded, or were included as one longer session.

Scuba surveys of species composition, fish density, and size distributions

Scuba surveys were designed and implemented to meet three specific objectives: (1) provide ground-truthing of the habitat types of the study sites used in 2003 and 2005, (2) estimate the density and size structure of the fish assemblage at each of the study areas using visual strip transects, and (3) generate mark–recapture estimates of abundances of fishes in 2005 based on the resighting of tagged fish by divers.

In 2003, surveys at each study site were repeated on 4 d (between 15 October and 24 November). Each of the two study sites surveyed in 2005 was also sampled on 4 d, 2 d before any fishing occurred (13–16 July 2005) and 2 d after fishing occurred (15–19 August 2005). At each site, four target depth zones were identified, ranging from the deepest at the outer edge of the kelp bed to the shallowest at the inner edge. Actual depth ranges sampled by the divers ranged from 4 to 26 m at sites sampled in 2003 and from 8 to 21 m at sites sampled in 2005. Within each depth zone, six replicate transects were surveyed, roughly parallel to shore and following the depth contours of the reef. Transects were positioned end to end with approximately 10 m of space between replicates. This sampling design produced 96 independent replicate estimates of density for each site. Transects surveyed on the first and second day of these pre- and postfishing survey periods were offset by 20–30 m to avoid resampling the same habitat. Transects consisted of two components surveyed simultaneously by a pair of divers—a benthic and a water column survey. The benthic portion of each transect was 30 m long × 2 m wide × 2 m high. The water column component covered an equal volume of water (30 × 2 × 2 m) located approximately 5–7 m above the bottom. On each sampling day, two pairs of divers surveyed an entire site, recording the species and estimated TL of all noncryptic fishes (i.e., excluding fishes such as small sculpins and kelpfishes) observed along each transect. In each diver pair, one diver searched for fishes along the bottom, while the other surveyed the water column. Counts from the bottom and water column surveys were pooled for each transect. Density of fishes from scuba surveys was calculated as the mean number of fish seen per transect, which had a 60-m² footprint over the substrate. Divers visually estimated TL of every fish observed on all transects to the nearest 1 cm. Divers were trained to estimate fish lengths prior to the study by repeatedly estimating the length of models of known length underwater. Examination of the error in these estimates indicates that divers were typically accurate to within 10% of the TL.

In 2005, after the surface tagging of fishes was finished, divers also recorded the numbers of tagged fishes observed. In addition to counting tagged fishes on the visual strip transects, divers conducted roving diver surveys on four other days to estimate tagged-to-untagged ratios of fishes. On those days (22, 23 August; 28 September; and 5 October), pairs of divers divided each study area into shallow and deep halves, and counted tagged and untagged fishes as they swam from one end of a study site to the other so as to avoid recounting individuals they previously encountered. Divers used dive lights to identify and record the color of each tag. Because the tagging effort included midwater species, divers surveyed the bottom and middle of the water column separately for tagged fishes.

Analysis methods: are there relationships among catch-per-unit-effort–density estimates from different sampling methods, and are these relationships strong enough to use one method as a proxy for another?

We performed a robust regression analysis using an MM estimation (Yohai et al. 1991) to examine the strength of regression relationships among CPUE estimates from different fishing methods and the potential for using one method to predict the expected values of another method. Mean CPUE estimates were calculated for each of the sites surveyed by divers and fishermen in both 2003 and 2005. Fishing and scuba samples taken on multiple days at a site were pooled, and the resulting site mean values from the two survey years were combined in the same analysis. These values were tested for normality using the Shapiro–Wilk test and square root transformed to meet assumptions of normal distributions and homoscedasticity. In a regression analysis of this type, which compares estimates of CPUE of one sampling type with another, large measurement errors can be associated with both the dependent and independent variables. As a result, ordinary least-squares regression can be overly sensitive to outliers, particularly in the explanatory variables (leverage points). Robust estimation using the MM estimation method provides estimates of model parameters, which are robust (i.e., not sensitive to small departures from model assumptions) and resistant to outliers in both the explanatory and response variables (SAS, version 9.1.3).

When significant slope values were detected using robust regression, the utility of these relationships (e.g., sticks versus scuba) was further evaluated by calculating confidence intervals (CIs) on estimates of the response variable predicted at various levels of the explanatory variable. Error associated with these predicted values scales with the product of sampling error associated with estimates of the explanatory variable, and statistical error associated with the slope of the regression equation. The upper and lower confidence limits of the CPUE–density estimates were multiplied by the upper and lower confidence limits of the slope to generate a CI on the predicted level of the response variable. Using these extrapolated CIs, it was possible to determine whether a given difference in the explanatory variable would result in a statistically detectable difference in the response variable.

Do differences in catch per unit effort–density from the different sampling methods vary with habitat or depth?

To determine whether relative CPUE–density differed among sampling methods and if any differences varied with habitat and depth, data from the two sites (of low- and high-rock relief) sampled in 2005 were used to test for interactions between sampling method and site, sampling method and depth, and sampling method, site, and depth. A univariate analysis of variance (ANOVA; SAS, version 9.1.3) was used for each of the nine most abundantly sampled species: cabezon Scorpaenichthys marmoratus, kelp greenling Hexagrammos decagrammus, lingcod Ophiodon elongatus, black rockfish Sebastes melanops, black-and-yellow rockfish S. chrysomelas, blue rockfish S. mystinus, gopher rockfish S. carnatus, kelp rockfish S. atrovirens, and olive rockfish S. serranoides. Divers were unable to differentiate olive rockfish from yellowtail rockfish S. flavidus underwater, so those species were treated as a group in the analyses. In the analysis for each of these species, sampling method and site (low versus high relief) were both treated as fixed variables. Third-order and then second-order interaction terms involving the continuous covariate (depth) were sequentially removed from the model when they were nonsignificant, resulting in a reduced model for each species, mean squared error being correctly attributed to the remaining error terms. Prior to analysis, any differences in scale of CPUE–density between habitats were removed from the data by standardizing observations within each habitat across sampling methods (mean = 0; SD = 1).

Do surveys conducted by any particular sampling method more accurately or precisely estimate fish abundance compared with estimates generated by mark–recapture techniques?

We used two methods to estimate population sizes of fishes in each of the 2005 study site areas. First, we multiplied mean density estimates (fish/60-m² transect) from scuba visual transects by the area of each study site to obtain population estimates. This extrapolation approach, which involves taking the mean of multiple transects surveyed on multiple days, provided estimates of abundance that summarize variability occurring across both spatial and temporal scales, and enabled us to generate 95% CIs around mean estimates for all species observed in the scuba surveys. Second, we used the multiple-sample Schnabel mark–recapture method to estimate population size of the more-abundant species inhabiting each study site (Krebs 1989). Total number of tagged individuals of each species was pooled across the different fishing methods. Total number of recaptures was pooled across fishing methods and tagged fishes sighted by divers. The Schnabel method assumes that the population is closed (i.e., minimal emigration), all animals have the same probability of capture (i.e., samples are random), tagging does not affect catchability, and tags are not lost. We assumed that we met the assumptions of the Schnabel method because the majority of the study species have high site fidelity and small home ranges relative to the study areas (Freiwald 2009), the duration of the study was short (reducing the likelihood of emigration from the study area), and the tagging and recaptures pooled across sampling methods reduced sampling bias of any single sampling method (i.e., individuals were sampled randomly). Combining the high proportion (54%) of resightings by divers with recaptures by fishing minimizes the potential effect of tagging on the probability of recapture (i.e., catchability). The Schnabel method has an advantage over alternative mark–recapture (e.g., Peterson) methods in that it enabled us to treat information from multiple days, allowing us to increase the sample size of population estimates. This increased sample size resulted in greater precision of population estimates. However, for the scuba surveys, we had insufficient samples for all fish species resighted to be analyzed using the Schnabel method. Therefore, we invoked the Petersen method for three groups of tagged fishes (kelp rockfish, olive rockfish–yellowtail rockfish complex, and kelp greenling) sighted by divers. All assumptions were the same for both the Schnabel and Petersen methods.

Relative accuracy of the fishing method CPUE estimates and the density estimates from scuba surveys in relation to “total” abundance estimates from the mark–recapture method were evaluated using linear regression. Estimates from each method for the nine commonly sampled species at both the low- and high-relief sites sampled in 2005 were used as replicates. Because abundance estimates extrapolated from scuba data provide a means of scaling the density estimates to correspond to populations within the boundaries of each study site, we again tested for correlation against mark–recapture abundance estimates, in this case testing for a one-to-one relationship (i.e., slope equal to one).

Precision estimates (i.e., ratio of variation to the mean) were also compared among sampling methods and mark–recapture estimates. Confidence intervals (95%) standardized to the mean were used to compare the precision of abundance estimates between the mark–recapture and extrapolation of scuba density estimates.

How do estimates of size structure differ among sampling methods?

Using the data collected from the low- and high-relief sites sampled in 2005, we compared length frequency distributions for the nine commonly sampled species in the study. Two-sample Kolmogorov–Smirnov (K–S) tests were used to compare length frequency distributions for all pairwise combinations of sampling methods, pooling observations from the low- and high-relief sites. Length frequency comparisons were run in three different ways: (1) using all observations from each method regardless of fish length; (2) using only fishes with estimated TLs of 20 cm or greater to account for the fact that fishing methods caught very few individuals below this limit, while scuba divers often saw them in high numbers; and (3) using only length observations of tagged individuals to reduce the possibility of a size-selective sampling bias between methods (i.e., observations came from a known subset of the population).

In these multiple K–S tests, we did not correct (reduce) the critical P-values to avoid type II error (detecting a difference in size distributions when, in fact, there is none) in order to maximize our ability to detect any sampling bias between sampling methods. Here, the consequence of type II error (falsely concluding that there is sampling bias) is considered less egregious than committing type I error (concluding that different sampling methods sample size distributions similarly when, in fact, they do not), leading to the inappropriate comparison of size distributions generated from different sampling methods and possible misinterpretation of sample bias as spatial and temporal differences in size distributions. Moreover, when comparing the relative magnitude of bias among sampling methods, the “significance” (critical P-value) is less important than the relative P-values and, for this reason, we present these relative P-values of the multiple tests. Similarly, pairwise tests were run comparing the low- and high-relief sites with regard to the sensitivity of each sampling method in detecting differences in length frequency distributions between locations.

Do sampling methods differ in their ability to describe assemblage structure?

The four sampling methods were evaluated on their ability to describe nearshore fish assemblages in terms of species richness. In order to determine the relative number of samples required of each sampling method to characterize the entire fish assemblage, species accumulation curves were generated using a resampling procedure to estimate the mean number of species recorded with increasing levels of sampling effort for each method (i.e., numbers of sticks, traps, handlines, or scuba transects; Primer, version 6). Because the relationship between sampling effort and the number of species observed indicates how well an assemblage would be represented given increased or decreased levels of sampling effort, we evaluated how many samples from each method would be required to obtain a greater than 95% probability of seeing at least one individual of a given species. Additionally, we calculated how many of the nine most commonly caught species would be detected in 50 samples of each method.

To determine whether characterization of fish assemblage structure differed among different sampling methods (scuba surveys, traps, sticks, and handlines), we tested for differences in the species composition (i.e., relative CPUE–density) of fishes larger than 20-cm TL of the nine species that were sampled in common by the four sampling methods. Since habitat and depth were suspected to interact with the effectiveness of different sampling methods (see question 2 above), these terms were also included in the analysis. As a result of the limitation on degrees of freedom imposed by testing these effects across a large number of species, multivariate ANOVA could not be employed and a nonparametric method—permutational multivariate ANOVA (PERMANOVA)—was used (Anderson 2001; McArdle and Anderson 2001). In this approach, mean CPUE–density estimates were calculated for each day of sampling for each of the sampling methods at each site. Each of these community samples was standardized (equal total across species) to remove scale differences between sampling methods. A similarity matrix was then calculated, and pairwise distance values between groups of samples were used in an algorithm analogous to the parametric ANOVA approach.

Are There Relationships among Catch-per-Unit-Effort–Density Estimates from Different Sampling Methods, and Are These Relationships Strong Enough to Use One Method as a Proxy for Another?

Robust regression analysis demonstrated significant positive relationships between CPUE–density estimates for seven of the nine species sampled in common by different sampling methods (Table 2). Only CPUE–density for kelp rockfish and lingcod showed no correlation among any of the sampling methods compared. Relationships between CPUE–density estimates for stick versus handline and scuba versus stick were the strongest, there being significant slope values for five and three out of nine species, respectively. When all species were combined into a single overall CPUE–density estimate, only the regression between sticks and handline methods showed a significant relationship. Significant relationships for scuba versus trap were not detected in any species, and only one species (cabezon) showed a significant relationship for scuba versus handline.

Confidence intervals were calculated around predicted CPUE–density values for each of the significant regression equations (Figure 2, graphs on the right). The width of these CIs as a percentage of the predicted value ranged from an average of 12.7% for blue rockfish to an average of 70.0% for cabezon. There was a high degree of overlap in these CIs across the range of observed CPUE–density values for all but one of the species with significant sampling method regressions, such that an observed increase in the value of the predictor variable would not result in the expectation of a significant difference (increase) in the response variable. This was the case whether all species were combined or examined individually (e.g., gopher rockfish; Figure 2C). The one exception to this was blue rockfish, where the relationship between handline and stick CPUE estimates had narrower CIs (Figure 2B). However, using points on the graph as an example, it would require almost a 10-fold increase in handline CPUE from 1.7 to 17.0 to result in a discernibly higher predicted value of stick CPUE.

Do Differences in Catch per Unit Effort–Density from the Different Sampling Methods Vary with Habitat or Depth?

There were strong differences in estimates of CPUE–density of each of the nine common species among the different sampling methods as indicated by highly significant P-values for most species (Figure 3; Table 3). Generally, scuba and handline had greater CPUE than sticks or traps for many species (note the differences in scale of the horizontal axes of Figure 3). Significant three-way interactions between the main effects in the ANOVA model (sampling method, site [i.e., habitat relief and depth]) were detected in four species (kelp rockfish, gopher rockfish, black-and-yellow rockfish, and blue rockfish), suggesting that differences in CPUE among methods are influenced by both depth and relief. As expected, density of gopher rockfish increased and density of black-and-yellow rockfish decreased with increasing depth (Table 3); however, this effect differed among sampling methods, and interpretation of two-way interactions (method × depth, site × depth, or site × method) for these species is not possible given the significant three-way interaction. Two groups (lingcod and the olive rockfish–yellowtail rockfish complex) did show significant two-way interactions between sampling method and depth; in both cases, handline CPUE decreased with depth while scuba density increased with depth. Lingcod also exhibited a significant site × method interaction, suggesting that for this species, relief may influence the effectiveness of sampling methods differentially. The remaining three species (black rockfish, kelp greenling, and cabezon) showed no significant interactions between sampling method and either site or depth.

Do Surveys Conducted by any Particular Method More Accurately or Precisely Estimate Fish Abundance Compared with Estimates Generated by Mark–Recapture?

Fish tagged during the first round of sampling at the two sites surveyed in 2005 were recaptured by fishing methods and observed by divers during a second round of sampling conducted a month later. Of the total number of fishes tagged (1,697), a combined total of 342 were subsequently “recaptured” by one of the sampling methods. The majority of these second observations (54%) were visual observations of tagged fishes by scuba divers either on transect or during random swims through the sample areas. The stick sampling method resulted in the highest rate of fishing recaptures of tagged fishes (13%, or 92 of 687 individuals), followed by handline (8%) and traps (2%). Among species, gopher rockfish had the highest rate of recapture of tagged fishes (15%). Recapture rates were lower for black rockfish (3%) and were lowest among kelp greenling, kelp rockfish, and olive rockfish–yellowtail rockfish (<2%). The number of fishes observed, tagged, and recaptured was similar between the two sites.

We compared CPUE–density estimates from each sampling method for the most commonly sampled species with mark–recapture abundance estimates generated using both fishing and scuba observations (Figure 4). Scuba density estimates had the highest correlation with mark–recapture abundance (Figure 4A; r = 0.811, P = 0.0002), followed by handline CPUE (r = 0.809; P = 0.0003) and stick CPUE (r = 0.684; P = 0.0049). Trap CPUE was not significantly correlated with abundance (r = 0.240; P = 0.3894). Abundance estimates derived by extrapolating scuba densities to the total area of each study site had a higher degree of correlation with mark–recapture abundances than the original density estimates (Figure 4B), and the slope of this relationship was not significantly different from 1 (P = 0.654), indicating a one-to-one correspondence between the two.

We calculated mark–recapture population estimates using recaptures from fishing gear only and also using both fishing recaptures and diver resightings combined (Figure 4C). There was generally close agreement between these two approaches both in terms of the relative abundance of different species and the relative population size of each species at the low-relief site versus the high-relief site.

Based on our sample sizes, precision of CPUE–density estimates was low for all species and sampling methods as indicated by coefficients of variation (CV = 100 × SD/mean) ranging from 100% to 1,700% (Figure 5). Precision was similar among sites and gear for the abundant gopher rockfish and black-and-yellow rockfish, but differed markedly among sampling methods for other species. For all species except blue rockfish, estimates of CPUE–density from scuba had the lowest CV (i.e., highest precision), followed by handlines, sticks, and traps. Precision was lower in the low-relief site than in the high-relief site for the majority of species and across sampling methods, corresponding to the patchy distribution of fishes seen in the discontinuous habitat at the low-relief site.

Lastly, we evaluated the relative precision of the population estimates derived from both extrapolation of scuba densities and from mark–recapture using all sampling methods, in this case by plotting the 95% CIs expressed as a percentage of the mean (CI/mean; Figure 6). Smaller CI–mean ratios reflect higher precision. The precision of abundance estimates generated by the extrapolation of scuba density was generally higher than those from the tag–recapture method. Surprisingly, these precision estimates from extrapolation of scuba density and tag–recapture estimates show an opposite pattern between the low-relief site and the high-relief site: CI–mean ratios of the former tended to be higher at the low-relief site, whereas those of the latter were higher at the high-relief site in all but one case.

How Do Estimates of Size Structure Differ among Sampling Methods?

Comparisons of length frequency distributions generated by the different sampling methods based on all individuals (i.e., all lengths) showed that the strongest differences in lengths (smallest P-values) occurred between scuba and the other sampling methods, especially for the more abundant species (Table 4). These differences reflect the greater number of smaller individuals that is sampled by scuba compared with the other methods. Few differences in length distributions were detected among the fishing methods (sticks, traps, handline) or the less-abundant species. A similar pattern of differences persists when individuals greater than 20-cm TL were compared (Table 4), indicating that the presence of smaller individuals in the length frequency data from scuba surveys was not solely responsible creating dissimilar distributions. However, when length frequency distributions based only on tagged individuals greater than 20-cm TL were compared among sampling methods, differences were less pronounced (larger P-values) for all paired comparisons. Despite these differences in frequency distributions, estimates of mean length of individuals greater than 20-cm TL were very similar among all sampling methods and species in both low- and high-relief habitats (Figure 7).

Comparison of length frequency distributions between the two study sites showed that all the sampling methods tended to show equal sensitivity in detecting potential differences between populations (Table 5). Across the four sampling methods, significant differences in length distributions among the low- and high-relief sites were detected in five of the nine commonly sampled species. Of these five species, only one, kelp rockfish, showed any disagreement between sampling methods. Significant differences were detected by scuba and sticks, which sampled this species in relatively high numbers, but not by traps or handline, where it was recorded less frequently.

Do Sampling Methods Differ in Their Ability to Describe the Structure of Fish Assemblages?

Overall, estimates of the species composition of the sampled fish assemblage (i.e., relative CPUE–density of the nine most abundantly sampled species) differed among the different sampling methods (Table 1; Figure 3). Sticks, traps, and handline methods caught about the same number of species, but the number of individuals of each species caught varied among the sampling methods. Scuba surveys recorded two to three times the number of species and four to fifteen times the number of individuals that were recorded by the fishing methods (Table 1).

Of the 15 species caught with fishing methods, scuba sampled the most species and handline gear sampled the fewest (15 and 11 species, respectively; Table 6; Figure 8). Traps required far more samples than any other sampling method to sample a representative number of species caught (9) and recorded far fewer species for a representative number of samples (50) than any of the other methods, reflecting the selectivity of the sampling method. It also required far more samples to detect 95% of all the species caught (Figure 8). Of the sampling methods examined, scuba recorded, on average, the highest number of species (12) for a given number of samples (50) and required the fewest number of samples (11) to detect a representative number of species, reflecting the higher number of species encountered on a transect sample compared with a stick, trap, or handline. Surprisingly, scuba surveys were remarkably efficient at detecting cryptic species, such as lingcod and cabezon. The number of samples required to obtain a greater than 95% probability of seeing at least one individual lingcod ranged from 26 for scuba (60-m² transects ) to 29 for handlines, 95 for sticks, and 503 for traps (hours fishing). The number of samples required to obtain a greater than 95% probability of seeing at least one individual cabezon ranged from 44 samples for scuba to 161 samples for traps. Cabezon were not caught on handlines.

Differences in patterns of species composition among sampling methods (i.e., relative CPUE–density of the nine commonly sampled species) varied according to depth (PERMANOVA sampling method × depth interaction term: df = 3, P = 0.002). Species composition also differed significantly between the low- and high-relief sites (df = 1; P = 0.005); however, these site differences did not vary according to sampling method (method × site interaction: df = 3, P = 0.435) and the three-way interaction (method × site × depth) was also not significant (df = 3; P = 0.764).

Are There Relationships among Catch-per-Unit-Effort–Density Estimates from Different Sampling Methods, and Are These Relationships Strong Enough to Use One Method as a Proxy for Another?

Relationships among CPUE–density estimates from different sampling methods were found to be statistically significant in the case of many of the common species sampled across sites in Carmel Bay in 2003 and 2005. Generally, regression equations between handline and stick or scuba and stick sampling methods demonstrated the closest fit. The strongest regression (i.e., narrowest CI on the estimate of the slope) occurred in the most-abundant species (blue rockfish), weaker relationships occurring in less-abundant or less-common species such as gopher rockfish and cabezon.

Similar relationships between research fishing CPUE and visual census data have previously been reported in the literature (Richards and Schnute 1986; Haggarty and King 2006); however, the potential for using these regression equations as an aid to stock assessment and management decisions, as yet, has not been fully evaluated. As an example, resource managers may at some point need to incorporate CPUE data from fishing to fill in spatial or temporal gaps in visual census data in order to generate regional stock assessments or to evaluate MPA effectiveness. In this study, we found that the compounding effects of within-sample variance and the error associated with regression equations would result in poor confidence in values translated from one sampling method to another. Even in the case of blue rockfish, which had the narrowest CIs on predicted values, a 100% difference in stick CPUE values between locations could not be used to predict a significant corresponding difference in handline CPUE.

The significant relationships we observed are notable because the data contained a high degree of variability associated with pooling data from different sites, seasons, and years. The predictive capacity of these relationships may be improved by increasing sampling effort in two possible ways. Increasing the number of replicate subsamples (scuba transects, sticks, traps, or handlines) at each site would reduce CIs of the mean values used as predictor variables in the regression relationship. Increasing number of sites (i.e., spatial or temporal replicates) used to characterize the relationship would both increase accuracy and reduce CIs of model parameters and thus allow increased confidence in predicted values of response variables.

Differences in relative CPUE estimates among species obtained from the sampling gear used in this study suggest possible selectivity biases for one or more types of gear (i.e., catching proportionally more or less of a particular species than would be predicted by its actual abundance). However, regression relationships comparing the density estimates for individual species between sampling methods will not be affected by this selectivity bias unless competition for hooks–traps occurs. In this event, CPUE estimates for a particular sampling gear and species may go up or down depending on the local abundance of other species with positively biased catch rates for that sampling gear. It may be possible to model the effects of this type of selectivity bias using mark–recapture data that represent “true” abundance estimates to calibrate the selectivity of various sampling methods for important species. This information, when combined with catch–effort relationships, may allow us to assess whether, at a given level of sampling effort, gear availability is likely to influence catch rates for less-selected species. Whereas increasing sample size will not eliminate these biases, it is possible that higher levels of replication will result in both (1) increased confidence in population estimates and regression relationships among methods, and (2) saturation of sampling effort such that CPUE of less-selected species will not be influenced by density of more-selected species.

The results of this study also suggest the possibility of a cost-benefit analysis that could allow managers to design optimal sampling strategies for characterizing CPUE–density relationships within a region of interest. The analysis would combine a comparison of the effort required to either sample at more sites or sample fewer sites more intensively with a resampling simulation that would provide an estimate of the corresponding reductions in CIs of either the model parameters or the predictor variables. The result would be an estimate of the levels of model sensitivity (i.e., whether a given “effect size” could reliably be translated from one type of data to another) that would result from increased or reallocated sampling effort.

Do the Relative Catch-per-Unit-Effort–Density Estimates Differ According to Habitat or Depth?

Estimates of CPUE–density generated by the different sampling methods were significantly influenced by habitat and depth as evidenced by significant interactions between these terms and sampling methods for many of the species sampled in this study. This result is to be expected given the selectivity of fishing methods for different species and what is known about the specificity of depth ranges and habitat preference of nearshore fishery species (Miller and Lea 1972; Eschmeyer and Herald 1983; Love et al. 2002; Allen et al. 2006). The implications of these interactions are that (1) a single calibration or correction cannot be applied to correct CPUE in the regression model of one sampling method with another if sampling with different methods differs by depth or relief; (2) rather, the depth and relief effects would have to be accounted (i.e., controlled) for by restricting sampling to comparable depths and relief across the geographic range in which the relationship will be applied; or (3) depth- and relief-specific corrections would have to be generated.

Do Surveys Conducted by Any Particular Method More Accurately or Precisely Estimate Fish Abundance Compared with Estimates Generated by Mark–Recapture Techniques?

Significant correlations occurred between mark–recapture abundance estimates and CPUE–density estimates from scuba, sticks, and handlines. Of the various sampling methods, density estimates from scuba showed the highest correlation to mark–recapture estimates, and this pattern was strengthened more when comparing abundance estimates derived from scuba densities by extrapolating them to the area of each study site. This improvement of the correlation by applying a differential scaling factor between the two sites indicates that the area in which tagged and untagged individuals are moving is larger in the low-relief than in the high-relief site and, therefore, that the densities in the low-relief site must be multiplied by a larger number to correctly correlate them to mark–recapture abundance estimates. Although great effort was made to fish uniformly within the same defined area used by divers for scuba transects, it is possible that areas of sand between reef habitat patches, particularly at the northern, low-relief site, would result in differences in the area of available habitat between sites. This suggests that habitat mapping techniques (e.g., multibeam sonar) could be used to improve the correspondence of density or CPUE estimates to mark–recapture abundance estimates and improve translation between these types of data.

The precision of both CPUE–density and abundance estimates from this study was low as evidenced by the poor predictive ability of regression relationships between methods. However, there was greater precision in estimates of some methods (i.e., scuba and handline) than others (i.e., sticks and traps). There was also greater precision in estimates from the contiguous habitat of the high-relief site than from the patchy habitat of the low-relief sites. This suggests again that efforts to integrate relative abundance or CPUE–density estimates among different sampling methods will require information on habitats sampled and further studies using higher replication in suitable and carefully matched habitats to better calibrate relationships between methods.

Tagged fish were resighted at a greater rate using scuba than they were recaptured by any of the fishing methods. This had the result of increasing the sample size of tagged to untagged fish ratios and improving confidence in the resulting abundance estimates generated by the Schnabel method. This presents a considerable complementary benefit of using both diving and fishing to assess relationships between density and abundance: more individuals can be tagged in an area using fishing, and for many species, the proportion of those individuals to the larger population can be better assessed using scuba.

Our interpretations of the relative accuracy of the different sampling methods assume that the abundance estimates generated by the Schnabel tag–recapture method are accurate. We believe we met the key assumptions of this method for the reasons mentioned in the Methods section. However, if populations in the study area were not closed (especially with substantial emigration from the study area) or experienced substantial tag loss, or if tagged fish reduced their likelihood of recapture (catchability), our estimates would likely overestimate the true abundance. If rates of tagging and recaptures were biased by one or more sampling methods such that individuals were not randomly sampled (e.g., only a portion of the size distribution or a particular habitat was sampled), we would have likely underestimated the true abundance of a species and misrepresented the relative abundance of species. Although we designed the study to minimize these sources of error, actual tests of these assumptions will greatly inform this and other assessments of sampling methodology.

How Do Estimates of Size Structure Differ among Sampling Methods?

There was generally good correspondence of mean lengths estimated by each of the fishing methods for fishes larger than 20 cm, and differences in mean length and size frequency between the two sites were detected similarly for most species by all four sample methods. However, size-frequency distributions estimated by scuba for most species were often significantly different from those estimated by the fishing methods. When we used all of the diver data, including fish smaller than 20 cm, it was apparent that length frequencies derived from diver observations spanned a greater range (both larger and smaller fish), and divers often saw more fish at the lower end of the frequency distributions. This bias may be explained in two possible ways: either the visual estimation used by scuba divers introduces greater measurement error into size estimates, or divers are accurately observing a larger population of sizes for most species than that selected by each type of fishing method. When size frequencies are compared using only tagged individuals, which represent a subsample of the actual population with known lengths, size frequencies were more comparable and significant differences were seen less frequently. This suggests the latter of the two possibilities above, that diver observation is less selective than fishing gear with respect to size, and size frequencies estimated by visual census more accurately reflect the natural population.

This has important implications if size- or age-structured models will be used to estimate population status of nearshore species (O'Farrell and Botsford 2005, 2006). On one hand, it confirms the accuracy of length measurements made by divers using visual estimation, and on the other, it suggests a potential hazard of comparing data from different fishing methods due to the size selectivity imposed by different gear types.

Do Sampling Methods Differ in Their Ability to Describe the Structure of Fish Assemblages?

Although fishing surveys can obtain more information about biological parameters of fishes such as weights and ages, scuba divers can visually quantify greater numbers of fish at any given point in time than can be caught and recorded using any of the fishing sampling methods used in this study. In addition, the rates at which the bottom (area) or water column (volume) are sampled are much greater using scuba transects than with any of the fishing methods. Aside from resulting in higher counts of most species, this difference means that the rate at which individual species are encountered is higher on scuba transects (i.e., higher slope of the species accumulation curve; Figure 8) such that fewer samples and less time are required to obtain an estimate of the structure of the fish assemblage in nearshore rocky reef or kelp habitats. For example, the resampling analysis of transect, stick, handline and trap data shows that to encounter the nine most-abundant commercial species, the scuba method would require, on average, 11 transects and 2 h of work for a two-diver team, as opposed to 7, 16, and 18 h of work with sticks, handlines, and traps, respectively.

The total number of species observed using scuba (29) was much higher than was observed using sticks (15), traps (12) or handlines (11), indicating that visual observations are less selective than fishing methods and provide more complete data on presence or absence of individual species and overall community structure. Many of the species that were only observed using scuba were not fished species but are nevertheless important in terms of characterizing the composition of nearshore assemblages, particularly if management goals (e.g., evaluating MPA effects) involve ecosystem-based management. The only fish to be recorded using a fishing method, but not observed by divers, was the spiny dogfish, a highly mobile species seen relatively infrequently in kelp forests.

A common conception among both fisheries scientists and fishers is that certain species that are rare, cryptic, or commonly concealed within the structure of the reef will be undersampled by scuba surveys relative to fishing methods. Surprisingly, lingcod and cabezon were encountered more frequently using scuba than with fishing gear. Resampling simulations showed that for both of these species, far fewer samples would be required to acquire a 95% probability of encounter on a scuba transect than when using any fishing method. One of the reasons for the low encounter rate of some species, however, is that certain fishes occupy relatively specific habitat types. Some species, such cabezon and the brown rockfish S. auriculatus, for example, more commonly inhabit low relief areas that are not abundant in Carmel Bay. Fishing gear is more efficient at sampling some species (e.g., grass rockfish and wolf eel Anarrhichthys ocellatus). All of the species that were caught with fishing gear are commercially important, but some recreationally important species (e.g., sea perches Embiotocidae and California sheephead) were not recorded using fishing gear at the levels of sampling effort employed in this study. For this reason, a sampling program that benefits from the complementary strengths of both fishing gear and scuba sampling will likely result in the most comprehensive description of nearshore assemblages. The combination of fishing to tag and recapture fishes along with scuba tag–recapture surveys may provide an especially good estimate of population abundances of nearshore fishes.