Comparison of Fish Communities at Random and Nonrandom Locations in a Sand-Bed River

Obtaining accurate and reliable information on distribution and population status of many imperiled species requires implementation of large-scale, long-term monitoring plans. Proper framing of monitoring objectives is crucial to ensuring adequate data are collected (Yoccoz et al. 2001). Often, large-scale monitoring programs rely on “convenience sampling” (sensu Anderson 2001) because accessing probabilistically chosen sampling locations can be difficult or impossible. Monitoring programs with nonrandom selection of sampling locations have been criticized because the areas sampled may not be representative of the monitoring area, and therefore inferences to the overall population may be biased and misleading (Anderson 2001; Ellingson and Lukacs 2003). When possible, convenience sampling should be avoided; however, critics often lack empirical data on the reliability of long-term monitoring programs based on convenience sampling (see Hutto and Young 2003). Thus, there is a need to assess the reliability of data collected by monitoring programs that rely on nonrandom site selections.

A program in place since 1993 for monitoring Rio Grande Silvery Minnow Hybognathus amaraus (RGSM), a federally endangered species, consists of 20 sites monitored monthly by performing 18–20 seine hauls at each site (USFWS 2010). The 20 sites were chosen from more than 100 possible locations based on access, land ownership, and spatial distribution, and on how efficiently surveys can be conducted (USFWS 2010). These sites, though carefully chosen to be representative of the river as a whole, were not randomly selected, and any data collected on the status of RGSM may not be representative of the population. The goal of the RGSM monitoring program is to provide overall status and trend updates on the RGSM population in the middle Rio Grande (MRG) River, New Mexico. Thus, a single riverwide estimate (and associated error) of catch per unit effort (CPUE) is generated for each survey to compare from month to month and year to year. A secondary goal of the program is to collect information on the fish community in general, as several other native fish species still occur in the MRG and may be at risk.

We hypothesized that standardized (but nonrandom) RGSM monitoring sites in the Rio Grande would produce a fish community similar to that in randomly selected sites and thereby increase the credibility of the long-term RGSM monitoring program. We made this hypothesis because in the Pecos River fish community data collected at nonrandom sites are similar to those from randomly selected sites (Archdeacon and Davenport 2013), and both the Pecos River and Rio Grande monitoring programs rely on similar methods of seining at the mesohabitat level (Dudley et al. 2014). These systems are both turbid, sand-bottomed streams with relatively few obstructions, making them ideal for seining (Rabeni et al. 2009). Mesohabitats in sand-bed streams are distributed fairly homogeneously, and thus the fish associated with those mesohabitats occur homogeneously as well (Kehmeier et al. 2007). Indeed, anecdotal evidence from many other research projects conducted in the MRG suggested that fish communities were generally similar, regardless of location or sample gear. Because the targeted species, small-bodied minnows, are widespread within restricted geographic ranges (Bestgen and Platania 1991; Hoagstrom et al. 2008) and the occurrence of mesohabitats is relatively homogeneous among sites, we believe that mesohabitat-based seining will produce similar riverwide estimates of CPUE and fish community, regardless of sampling sites.

Our objective was to examine the representativeness of fish community data, and specifically RGSM data, collected from nonrandom sites. We compared the fish community data from the standard RGSM monitoring sites with those from 20 randomly chosen sites within the MRG. We hypothesized that the species richness, community similarity measures, and RGSM CPUE would be similar between the two data sets. Thus, we compared the overall species richness between the surveys, compared fish community similarity, and examined the CPUE of RGSM between the two data sets. From our findings, we make recommendations on the credibility of the data collected for RGSM monitoring.

STUDY SITE

The MRG extends from downstream of Cochiti Dam to Elephant Butte Dam and is divided into four reaches by irrigation diversion dams: Cochiti Reach, Angostura Reach, Isleta Reach, and San Acacia Reach (Figure 1). The MRG in the Angostura, Isleta, and San Acacia reaches is a large, sand-bottomed stream that currently contains the only remaining natural population of RGSM. From Elephant Butte Dam to the Angostura Diversion Dam is approximately 301 km, 270 km of which is upstream of Elephant Butte Reservoir. Areas below the San Acacia and Isleta diversion dams frequently dewater or have substantially reduced flows between June and November each year, affecting as much as nearly 82 km in 2012. Since 2007, 32–82 km have dewatered each year, except in 2008, when no dewatering occurred (Archdeacon et al. 2015). Dewatering occurred prior to 2007, but no reliable records of extent and duration of dewatering exist. Periodic dewatering might have contributed to the decline of several native species (Bestgen and Platania 1991; Hoagstrom et al. 2010). The long-term RGSM monitoring program has 20 nonrandom sites, with five sites in the Angostura Reach, six sites in the Isleta Reach, and nine sites in the San Acacia Reach. No sites were located in the Cochiti Reach because of land ownership.

METHODS

Site selection.

We used a generalized random-tessellation stratified (GRTS) sampling scheme to select 20 base sampling locations and 20 “overdraw” sites from a line segment running down the center of the channel in package spsurvey (Kincaid and Olsen 2013) in program R (R Development Core Team 2014). Generalized random-tessellation stratified sampling is a spatially balanced probability sampling design that allows inferences of the entire sampling area (Stevens and Olsen 2004). Additionally, the random sample is still spatially balanced even when base sites are replaced with overdraw sites. We limited the area of inference to the 270 km downstream of Angostura Diversion Dam, ending at the mouth of the Rio Grande at Elephant Butte Reservoir. Tribal land comprises approximately 40 km of those 270 km; we rejected six sites that fell on tribal land, as no access would be granted, and hence no sites for the long-term monitoring program are located on tribal land. Each rejected site was replaced with the next site in the overdraw list. No further sites were rejected for any reason. Thus, we limit our inferences to nontribal land in the MRG, between Angostura Diversion Dam and upstream of the head of Elephant Butte Reservoir.

Fish sampling.

We used the same fish sampling protocol in the random surveys as in the nonrandom surveys. The protocol is intended to minimize the influence of other variables on detection probability, which can bias estimates of both CPUE and occupancy, by using standard gears, standardized amounts of sampling, and standardized habitats and by avoiding collecting samples during extreme flow events. Nonrandom surveys were conducted by an independent crew not associated with the random survey crews. Fish community data were collected during two time periods, once from each GRTS site between June 17 and July 17, 2014 (hereafter the “summer random” survey), and again between September 30 and October 15, 2014 (hereafter the “autumn random” survey). Fish data from the nonrandom surveys were collected on July 6–8, 2014 (hereafter “summer nonrandom” survey) and October 6–8, 2014 (hereafter “autumn nonrandom” survey). The autumn nonrandom survey is an important component of the recovery plan: it occurs after any summer dewatering and provides the primary annual estimate of RGSM population trends and site-occupancy rates, both of which are used as part of the criteria for down-listing or de-listing (USFWS 2010). We collected fish at each site with a 3.0 × 1.0 m seine (3.2 mm mesh size). Two people rapidly drew the seine downstream through a single mesohabitat (riffle, run, shoreline run, pool, shoreline pool, backwater, and shoreline backwater) of uniform depth and velocity. Site lengths were limited to <400 m and sampled upstream to downstream. We measured the length of each seine haul (m), multiplied by seine width (typically 2.5 m), and summed the area seined for each site. Summer random surveys averaged 448 m² (standard error [SE] = 30.7), whereas summer nonrandom surveys averaged 511 m² (SE = 5.1). Because two sites were dewatered during the summer nonrandom surveys, the RGSM CPUE was zero at those two sites, and no area was sampled. Autumn random surveys covered an average 436 m² (SE = 13.2), whereas autumn nonrandom sampling covered 503 m² (SE = 11.1). We attempted to replicate the sampling allocation of mesohabitats to be the same as in the long-term monitoring plan (Dudley et al. 2014), which includes four runs and shoreline pools each, two backwaters, two pools, and two riffles; any remaining samples were taken in shoreline runs to reach 18–20 seine hauls. Not all sites contained all habitats (particularly riffles), and we were not able to exactly replicate the mesohabitat allocation; when necessary, we sampled shoreline run habitat to reach 20 seine hauls. We counted and released all fish in each seine haul. We calculated CPUE for each site individually as the number of RGSM per 100 m².

Data analysis.

Currently, RGSM are rare and CPUE data, even at the site-level, are positively skewed and contain many zeros. To compare CPUE values between data sets, we used the two-step combination model method of Fletcher et al. (2005), combining ordinary and logistic regression to create expected values of CPUE and associated bootstrapped confidence intervals. First, we transformed the data into site-level presence–absence data and used an intercept-only generalized linear model with a binomial error distribution to model the mean and SE of the probability of RGSM presence. Next, we used an intercept-only linear model to model the mean and SE of log-transformed CPUE data (only from occupied sites). We combined the models to estimate the expected CPUE value as follows:

where is the expected occupancy and calculated as

and is the expected CPUE and calculated as

Therefore, is the intercept from the presence–absence generalized linear model, is the intercept from the log-transformed CPUE linear model, and is the mean square error from the log-transformed CPUE model (Fletcher et al. 2005). We used parametric bootstrapping to create 80% and 95% confidence intervals, where random values for were drawn from a normal distribution with mean = and standard deviation = SE of from the generalized linear model. Similarly, random values for were drawn from a normal distribution with mean = and standard deviation = SE of from the log-transformed linear model. We created random values for by resampling the log-transformed CPUE (with replacement) n times, where n is the sample size of the log-transformed linear model. We performed 10,000 bootstrapped samples and used the 2.5th, 10th, 90th, and 97.5th percentiles to create the 80% and 95% confidence intervals.

The intercept-only models do not include covariates and provide only the mean and estimate of variance. This assumes the predicted CPUE and occupancy are constant across all sites. While this is almost certainly untrue and covariates could be used to build better descriptive models, we chose not to include any covariates. Standardization of the monitoring methods attempts to minimize changes in detection probability over time and across sites, making possible the comparisons and detections of trends. Further, the main goal of the monitoring program is to provide a monthly estimate of CPUE for managers to see trends and overall status of RGSM. While inclusion of covariates is important for analyses examining how and why CPUE changes over time, creating reach- and mesohabitat-specific estimates of CPUE would be confusing and unnecessary to meet the overall RGSM monitoring goals.

We used a Morisita Index of Similarity (hereafter C) to compare composition similarity (Kwak and Peterson 2007). The C index compares the relative abundance of species in two different communities. A value of 1.0 indicates identical communities. We used the package vegan 2.0–10 (Oksanen et al. 2013) in program R 3.0.3 (R Development Core Team 2014) to calculate C.

RESULTS

Rio Grande Silvery Minnow Comparisons

Summer random and nonrandom surveys showed similar RGSM CPUE and occupancy rates. Summer random surveys collected RGSM at 9 (45%) of the 20 sites (95% confidence interval, 25.8–65.8%). Raw RGSM CPUE scores were 1.94, 1.87, 0.61, 0.60, 0.45, 0.23, 0.21, 0.20, and 0.18 for summer random surveys. Summer nonrandom surveys collected RGSM at 11 (55%) of the 20 sites (95% confidence interval, 34.2–74.2%). Raw RGSM CPUE scores were 1.92, 0.92, 0.82, 0.74, 0.63, 0.56, 0.55, 0.40, 0.20, 0.20, and 0.19 for summer nonrandom surveys.

Autumn random and nonrandom surveys showed similar RGSM CPUE values but different occupancy rates. Autumn random surveys collected RGSM at 25% of the 20 sites (95% confidence interval, 11.2–46.9%). Raw RGSM CPUE scores were 0.49, 0.42, 0.28, 0.21 and 0.21, with a total of only seven RGSM collected at those five sites in the autumn nonrandom surveys. Autumn nonrandom surveys failed to collect any RGSM at the 20 sites (95% confidence, 0.0–16.1%).

There was also little difference in expected RGSM CPUE from the random and nonrandom surveys. The combination model–derived RGSM CPUEs were similar for the summer random and nonrandom surveys (Figure 2), showing overlapping confidence intervals and no meaningful biological differences in means (0.32 compared with 0.37). Both surveys showed a marked decline in RGSM CPUE from summer to autumn. Expected CPUEs of RGSM were also different between the autumn random and nonrandom surveys because the autumn nonrandom survey found zero RGSM (Figure 2). However, the autumn survey mean expected density was 0.08 and was composed of seven RGSM (Table 1).

DISCUSSION

The two surveys established that the fish communities were essentially identical, especially in the autumn surveys. Both the overall richness and composition were similar, and both surveys recorded a reduction in species richness from summer to autumn. Both surveys also suggest a very uneven fish community dominated by a single species (Red Shiner). Major differences in composition between the summer surveys were the large numbers of Channel Catfish Ictalurus punctatus in the summer random surveys, probably resulting from the extended sampling period as the majority were age-0 fish recruiting to the gear, and the many more Western Mosquitofish in the summer random survey. In both sets of surveys, many more Longnose Dace Rhinichthys cataractae were collected in the nonrandom locations than in the random surveys. Both surveys also showed a substantial decrease in RGSM CPUE from the summer to autumn surveys. Long periods of low flow occurred in 2014, and about 42 km of the MRG dried (Archdeacon et al. 2015). These conditions probably promoted a community dominated by tolerant species such as Red Shiners and Western Mosquitofish, thus increasing similarity in fish communities.

We found little evidence that the long-term nonrandom locations for monitoring the Rio Grande fish community and trends in RGSM CPUE were biased compared with random survey locations in the MRG. The long-term nonrandom locations were representative of RGSM occupancy and CPUE in the MRG in summer and autumn of 2014. We found no biologically meaningful differences in RGSM CPUE between the two surveys. Rather, the CPUE and the occupancy rates for RGSM between the two surveys were very similar considering the small sample size, the use of independent crews, and the fact that the random survey crews had never sampled at the locations previously. We also failed to find any convincing differences in the composition structures—in fact, the autumn surveys were nearly identical in composition. We conclude that data collected at the long-term nonrandom sites for fish community, RGSM CPUE, and RGSM occupancy are representative of the MRG despite not being sampled from random sites.

However, we did find some differences in the fish composition and RGSM CPUE between the two surveys. The lack of any RGSM in the autumn nonrandom survey clearly differs from the autumn random survey results, but the difference between 0 and 0.08 fish/100 m² is so small as to be irrelevant from a recovery standpoint. Rio Grande Silvery Minnow are a short-lived species such that the majority of fish found belong to a single cohort (Horowitz et al. 2011). Expected CPUEs vary greatly within a year due to recruitment and mortality but also vary more greatly among years (Dudley et al. 2014) than seen in longer-lived species. For perspective, the U.S. Fish and Wildlife Service augments the RGSM population with hatchery-reared fish each autumn following the autumn nonrandom surveys to bring the reach-specific density to approximately 1.0 fish/100 m² (Archdeacon 2014). For 2014, >270,000 RGSM were released, with >700,000 released since autumn 2012. In contrast, RGSM were present at all 20 nonrandom sites in 2008, with a CPUE of 7.96 fish/100 m² (868 collected) during autumn nonrandom surveys (Dudley and Platania 2009). Clearly, during years of high spring runoff and little summer dewatering (e.g., 2008), RGSM can be abundant (Dudley et al. 2014). Therefore, the difference between 0 and 0.08 fish/100 m² is not biologically meaningful.

The most striking dissimilarity for RGSM between the two surveys is the occupancy rates of the autumn surveys. The 95% confidence intervals only slightly overlap, indicating moderate evidence of a statistical difference. Biologically, 0% compared with 25% of sites occupied by RGSM is meaningful. However, given that the autumn random survey captured only seven RGSM (three singles and two doubles), the difference is insubstantial and does not indicate that nonrandom sites are underestimating occupancy (or CPUE) by a meaningful amount. For perspective, one recovery criterion for down-listing RGSM is that 75% of the nonrandom sites must be occupied in the autumn surveys (USFWS 2010). Coupled with the summer results, we fail to show any evidence that the random sites produced CPUEs or occupancy rates for RGSM meaningfully different from those found in nonrandom sites; both sets of surveys showed a statistically significant decline in RGSM CPUE from summer to autumn.

The survey design probably accounts for the similarity in results between the random and nonrandom sites, while the site selections are the reason for the comparably few differences. Mesohabitats for RGSM are not rare in the MRG and are continually changing as the sand bed shifts. Both the nonrandom and random survey sites contain appropriate mesohabitats for RGSM. This is a major point to emphasize—during wetter years, RGSM are one of the more abundant fishes and are much easier to capture than in drier years (Dudley et al. 2014). The lack of any RGSM in the autumn nonrandom survey is not because the sites are in the wrong locations, but because the RGSM in the river are so few they are very hard to detect. From a recovery standpoint, RGSM should not be rare or hard to find. Conversely, Longnose Dace is a cobble-riffle obligate species (Mullen and Burton 1995), a rare mesohabitat in the MRG that is generally found in the upper sections of river between Isleta and Angostura diversion dams. Indeed, no random sites were selected there, compared with three nonrandom sites. Unsurprisingly the random surveys collected far fewer Longnose Dace than the nonrandom surveys. Thus, the similarity between the surveys is that the mesohabitats for RGSM are uniformly distributed throughout the MRG, though the fish itself may not be present at every site because of rarity.

Our model assumed constant probability of occupancy across sites and constant CPUE across sites; we also had to assume that detection probability was constant across all surveys to make meaningful comparisons. Because detection probability varies by species, season, gear, site, and environmental variables, we had to assume the standardized sampling methods reduced the variation to a level that did not influence our results. We believe it more probable that the true fish communities and the RGSM CPUE were similar between random and nonrandom surveys and that detection probability was also similar between surveys, and we disagree that variable detection probability artificially produced similar communities and RGSM CPUE even though they differed.

Because species are detected imperfectly, the autumn nonrandom survey captured no RGSM although RGSM were present in the system. This could be because no RGSM were at the nonrandom sites, suggesting that the nonrandom sites could be biased (although the summer nonrandom sites showed more sites occupied by RGSM than did the random sites). We believe another possibility is more likely: RGSM were so rare and the detection probability so low that nonrandom surveys failed to collect any by chance, while a few were collected at the random sites by chance. Our data support the second possibility, as it seems unlikely the species composition from the two surveys would be so similar if the nonrandom sites were biased.

Two additional problems may have influenced our results and conclusions. First, the nonrandom sites were surveyed within days, but we were not able to complete the random surveys in such a short temporal span. Evidence that this influenced our results is the large numbers of age-0 fish recruiting to the gear in the summer survey, primarily Channel Catfish and River Carpsuckers Carpiodes carpio. Had we completed the random surveys over a shorter time period, the results probably would have shown even more similar fish communities in the summer surveys. The extended sampling did not appear to affect the autumn surveys, probably because young-of-year fish fully recruited to the gear by that time.

Secondly, we had little statistical power to detect changes because of the sample size. Quadrupling the sample size to 80 sites for each set of random and nonrandom surveys (to cut confidence intervals in half) would be expensive and difficult to perform (and still would probably not result in any statistical differences between the summer random and summer nonrandom surveys). Increasing effort in the autumn surveys would probably have resulted in more similar RGSM CPUE, presumably by increasing the probability that the autumn nonrandom survey would collect at least a few individuals. Further, we emphasize that 20 sites is about the feasible limits for a monthly monitoring program in the MRG. Forty sites might be possible, but the increase in statistical power and precision is minimal, whereas 80 sites would not be possible without either increasing staff or spreading the surveys over a longer temporal period; both require unrealistic levels of funding.

Critics of convenience sampling often lack empirical evidence that nonrandom samples are not representative of the population or community. Here, we have shown that nonrandom samples are representative of the fish community in general and representative of RGSM CPUE but might underestimate occupancy rates of RGSM when the population is extremely small. However, we argue that the underestimation of RGSM occupancy is biologically negligible when viewed in context of RGSM CPUE, as none of the random survey evidence indicates the RGSM population is actually common or widespread in the MRG, as it had been historically (Bestgen and Platania 1991) and is now required for recovery (USFWS 2010). Our results corroborate another study on small-bodied fishes in the Pecos River (Archdeacon and Davenport 2013), and we have shown that nonrandom locations may not be have a biased community structure in sand-bed rivers. We believe this is because surveys are based on seining in discrete mesohabitats and the mesohabitats occur with regularity throughout the MRG, as sand-bed rivers are relatively homogeneous. We are not suggesting that newly implemented monitoring programs abandon probabilistically chosen sites for easy-access sites. New monitoring programs should implement some sort of random process for site selection; however, many long-term fish monitoring programs that rely on nonrandom sites probably still provide useful fish community data. We recommend that managers of long-term monitoring programs examine the differences between current sampling sites and randomly chosen sites to ensure that the data collected are representative of the target population or community.