258 
Fishery Bulletin 11 6(3-4) 
and years. The 60% reduction typifies data for several 
RSTs that have low numbers of marked and recaptured 
fish because of site limitations or low abundances. The 
fourth scenario had information from 8 strata, strata 
12-19, removed from the first year of the data set and 
had a 60% reduction across all marked and recaptured 
individuals across all strata and years. This modifica¬ 
tion emulates several RSTs that did not operate dur¬ 
ing summer months initially because it was thought 
that salmonid migration had ceased during this time 
period (see history of the Marsh Creek data set below). 
It was later realized that small, but significant, num¬ 
bers of individuals did migrate throughout the summer 
and RSTs now operate during this time period. Simu¬ 
lated scenarios and known parameters used to create 
the data are presented in Supplementary Table 2 and 
Supplementary Figures 5-8 (online only). Each model was 
tested with these 4 scenarios. 
We illustrate the relative performance of the 4 com¬ 
peting models, using 2 RST data sets. Marsh Creek 
and Big Creek are tributaries to the Middle Fork of 
the Salmon River located in central Idaho, have com¬ 
parable salmon populations, but have “good” and “poor” 
quality data sets, respectively. 
Marsh Creek is a third-order tributary with a RST 
located at a river transect that has geographic and hy¬ 
drographic features conducive to continuous operation 
throughout the majority of the migratory season with 
high capture efficiencies (Venditti et al. 1 ). In addition, 
the RST has been operating for 21 years. Median day 
of installation is March 20 and removal is November 3. 
Within this period, the RST operates a median of 97% 
of the days. Gaps in operations are typically short for 
reasons such as icing and passing thunderstorms. Dur¬ 
ing the first few years of operations, the RST was re¬ 
moved during the summer for 2.5 months until it was 
discovered that fish emigrated in that time period, too. 
This gap in the record motivated the fourth scenario in 
the simulations described above. Annual total catch of 
Chinook salmon has fluctuated more than 2 orders of 
magnitude from 846 to 91,719 fish. The high capture 
efficiencies, abundant amount of juvenile salmonids 
captured, and longevity of RST operation present a 
nearly ideal RST mark-recapture data set (Suppl. Fig. 
9) (online only). 
Big Creek, in comparison, is a fourth-order tributary 
located in the Frank Church River of No Return Wil¬ 
derness. The RST was first installed in 2007 in a reach 
with widths from 30 to 40 m, resulting in low capture 
efficiencies (Copeland et al. 2 ). Trap efficiencies are 
much lower than those seen in Marsh Creek and can 
be quite variable (Suppl. Fig. 10) (online only). Annual 
total catch of Chinook salmon has fluctuated from 5167 
to 33,308 fish. Median day of installation is March 11 
and removal is November 10. Within this period, the 
RST operates a median of 80% of the days. There are 
frequent short gaps in operations for reasons such as 
icing and thunderstorms that produce concomitant 
turbidity and debris. In addition, there are substan¬ 
tial gaps in operations during snowmelt. For example, 
adding up gaps in service >7 days, the RST missed a 
median of 55 days in mid to late spring, ranging from 
20 to 75 days. The exception was 2007, the first year 
of operation, but the RST was not installed until May 
21. The Big Creek RST data exemplifies nearly all po¬ 
tential pitfalls possible at RSTs in terms of sparse and 
missing data. Scenarios like those at the Big Creek 
RST are what motivated this study. 
We focused on the results for the 2014 emigration 
year. During 2014, the Marsh Creek RST was deployed 
on March 22, 1 week later than the earliest date the 
trap had been deployed in its 21 years of operation. 
In 2014, the Marsh Creek RST ran continuously with 
high capture and recapture efficiencies throughout the 
trapping season and was removed on October 31 when 
snow and ice prevented operation. In total during 2014, 
the Marsh Creek RST missed 3 weeks of the trapping 
season at the tail ends when few fish were suspected 
to be migrating. In 2014, the Big Creek RST was de¬ 
ployed on March 13 and operated for roughly 5 weeks 
before high water prevented the RST from operating 
from April 11 to June 13. On June 14, the trap was re¬ 
deployed and ran continuously until snow and ice pre¬ 
vented operation on November 9. During the 8-week 
outage, it was known that age-1 Chinook salmon over¬ 
wintering in upper portions of Big Creek migrated out 
of the system during this time because this behavior 
had been observed in past years and at other RSTs in 
Idaho. In addition, when the Big Creek RST was op¬ 
erating, RST capture and recapture efficiencies where 
low owing to the site limitations described above. The 
Marsh Creek and Big Creek RST operations in 2014 ex¬ 
emplify the 1) complete and abundant mark-recapture 
data and 2) sparse and missing mark-recapture data. 
Model implementation 
All models were implemented with the statistical pro¬ 
gram JAGS (vers. 4.0.0; Plummer, 2003) run through 
the program R interface, vers. 3.2.2 (R Core Team, 
2015) with the R2jags package vers. 0.5-7 (Su and Ya- 
jima, 2015). The complexity of the models inhibited 
calculating an exact posterior distribution. As such, 
Markov chain Monte Carlo (MCMC) simulations were 
implemented in JAGS to sample from the joint poste¬ 
rior distributions of all parameters to approximate a 
posterior distribution. Three parallel chains initiated 
at random values were run for each model. Chains 
were run for a total of 500,000 iterations and the first 
100,000 iterations were discarded and the remaining 
iterations were thinned by a factor of 100. The final 
sample size for each chain comprised 4000 values. 
MCMC posterior distributions were visually inspected 
for multiple peaks and Gelman-Rubin test statistics 
were calculated to ensure chain convergence. Multiple 
peaks in the posterior distribution or Gelman-Rubin 
test statistics >1.1 were subject to nonconvergence 
and chains were run for additional iterations to try to 
achieve convergence. 
Model performance was evaluated by comparing the 
