SSM Estimates Using Synthetic Data.

Before leaving last week, I gave Ken some synthetic data to try to fit. He got the q values pretty well, but really missed the intitial distribution and migration during the first few time steps. My "true" state of nature had the stock spread out over the entire range, no movement until day 100, then constant movement after day 100 (i.e., step size fixed at 7 units per day toward the natal stream). I set up the efforts to match the efforts for 1990 (i.e., ocean troll fishery in early timesteps, very concentrated net fisheries in Puget Sound later, and very low impact sport fisheries throughout the year). This set-up gave model catches very similar to CWT recoveries observed for 1990 (for Voights Creek stock).

The SSM estimated that the entire stock was located in the north ocean at the start of the simulation, and then distributed itself over the range during the first few time steps. It got the catches pretty close.

This is a little discouraging because the problem seems to be related to the fact that the recovery information comes first from the ocean fisheries and then from the inside net fisheries. In other words, the effort is not distributed evenly over all the time steps--the fisheries operate only where the fish are. Thus, at the start of the year there is no significant fishery in the inside waters to provide info that there are some fish there. Thus, the SSM just assumes they are all located in the north ocean. If a new fishing pattern is established for a future year, then the SSM model of the fish distribution will still be way off and will result in very incorrect catches.

We set up 12 different types of conditions (2 migration patterns x 2 initial distributions x 3 fishing patterns) for me to test all the models, not just the SSM. I promised to try to get this done by the next meeting.

----------- end Norris email to Frever -------------------


3. Update on code development (Troy Frever).

Troy reported that he has ported the C++ code from his Unix system to his new PC. This took more time than expected due to slight differences in the C++ compilers. He still hopes to have a prototype model in PSC Chinook Model configuration by the end of October.


4. Update on effort data collection (Jim Scott).

Jim reported that he provided Ken with the ocean troll effort data. Ken checked it over and, with some minor exceptions, was fairly close to the data he used for his dissertation work on the Humptulips stock. Jim's next step is to finish getting the "inside" sport and net effort data together.


5. Update on State Space Model (SSM) work (Ken Newman).

Ken could not attend the meeting, but provided the following email report on estimating parameters for the Humptulips stock using the additional data supplied by Jim Scott.

-------------------- Ken Newman email report ---------------------------

Thanks to Jim Scott getting the effort data for 86 to 91 I've now got estimates of the SSM parameters for 6 years for the Humptulips recoveries.

The model structure:

   Initial dist'n = Beta(Init_alpha, 2.0) on Brookings to North. BC (12 regions)

   Mortality = Survival from time of release to beginning of fishing
               q_US
               q_Canada

   Movement = Beta(Move_alpha, 3.0) "advection-diffusion"

Notes:
 1. I've added northern BC back in since Allan Hicks found that
    there were a fair number of recoveries in that region for some
    of the years (especially 1988).
 2. still using different q's for US and Canada even though Jim
    Scott got effort "standardized" to boat days; I've still got
    a "scaling" factor for Canada effort twice that of US
 3. natural mort during season = 0

-----------------------------------------------------------------------
Resulting estimates:
       Initial Surv q_US q_Can Move
        alpha to start alpha -Log likelihood

1986: 2.96 5.95 6.51 4.93 8.58 5675.4
1987: 3.14 1.51 3.51 4.44 7.46 573.9
1988: 281.54 6.58 18.16 2.98 10.06 3593.6
1989: 2.94 4.40 11.97 11.82 7.36 2658.1
1990: 2.36 2.83 9.78 4.71 9.44 2462.3
1991: 2.55 6.26 8.79 4.35 9.51 3175.6

Except for 1988 the Initial_alpha parameter is varying between 2.4 and 3.1- not too wild.

Initial survivals ranging from 1.5% to 6.6%.

Given the different scaling, one could double q_Canada to put things on the same scale:

      q_US q_Canada
1986 6.51 9.86
1987 3.51 8.88
1988 18.16 5.96
1989 11.97 23.64
1990 9.78 9.42
1991 8.79 8.70

-----------------------------------------------------------------------

1988 stands out as a "strange" year in many ways:

 1. the initial distribution is shoved way north
 2. the only year with "less" effective gear, assuming boat days
    are truly equivalent.
 3. highest initial survival

Just looking at the recovery data for that year, there are a relatively high number of late season recoveries in the far north regions. This is what is most affecting the results I believe. To account for these catches a sizeable portion of the population must be present that far away that late- one way of "forcing" this is to shove a lot of fish way north at first. It could be that fixing the 2nd parameter = 2.0 in the initial dist'n is causing this. Allowing this parameter to be estimated may smooth things out.

I think it's important to try out different initial dist'n and movement modules though before drawing too many conclusions.

---------------- end Ken Newman email report ----------------

There was a lengthy discussion of Ken's results which focused on the following four general items.

Item 1. Some members were wondering how correlated the q-values were with the initial alpha values. Jim Scott recalled that Ken previously reported that the covariance matrix didn't indicate much correlation. Some members wondered how much trouble it would be to generate the response surface for these two parameters given the others fixed.

Item 2. The variability in the q-values seemed pretty high. Jim Norris suggested it might be possible to assume constant q's across years and then fit a model to all six years.

Item 3. Two thoughts on evaluating the predictive capabilities of the SSM. First, select one of the six years as a base year, estimate parameters, then predict each of the five other years. Second, select five of the six years as base years, estimate "average" parameters for those five years, then predict the remaining year. This is similar to what Rich Comstock did with the PM Model. Also, Jim Norris suggested that we need to get Ken the inside (i.e., non-ocean) effort data so he can try the SSM on the Voights Creek stock. This will provide an opportunity to work with a stock that has a more unidirectional movement pattern and a more complicated initial distribution.

Item 4. Most members would like to have some way of relating the "move alpha" parameter to an intuitive description of the migration process. Perhaps something like a plot of "average" daily step size for each day (given the estimated initial distribution). This would help identify any unrealistic predictions of the migration model.


6. Update on model comparisons program (Jim Norris).

Overview. Jim reported that he generated 16 synthetic datasets and supplied them to Ken for SSM parameter estimation. When Ken attempted to use the data, he discovered an error in Jim's data aggregation routine. Jim fixed the error and generated new datasets, but the SSM estimates could not be completed prior to this meeting. All datasets were generated in deterministic mode (i.e., no variability for any model processes). The 16 datasets included all combinations of four "treatments" and two "levels" per treatment:

Migration Pattern
  -- constant daily rate (13 units/day starting on Aug 12);
  -- increasing daily rate (0 units/day until Jun 28, then increasing linearly from 0 to 10.5 units/day by Sep 11).

Harvest Plan
  -- Base (similar to 1986);
  -- Low Ocean (Ocean Troll effort = 15% of Base).

Data Aggregation By Time
  -- 30 days per aggregation period;
  -- 7 days per aggregation period.

Data Aggregation By Space
  -- 15 statistical regions;
  -- 6 region groups (RG):
        RG1 = WCVI 25-27 (North WCVI)
        RG2 = WCVI 21-24 (South WCVI)
        RG3 = WA 4B - WA 6 (St J de F)
        RG4 = WA 6b - 9 (WA 6B9)
        RG5 = WA 10 - 11 (South Sound)
        RG6 = Puyallup River

Fishing Effort For Base Harvest Plan. For ocean troll effort, he used the data presented by Jim Scott at the July 30 meeting. For effort patterns in the net and sport fisheries, he used educated guesses based on his own fish tickets from 1986 and published regulations. The gillnet effort was concentrated in the Strait of Juan de Fuca during weeks 33-36 (August sockeye salmon fishery) and in South Puget Sound during the September coho fishery). Some sport effort occurred throughout the year, with elevated effort in the Strait of Juan de Fuca during August-September and in south Sound in September-October.

Initial Distribution. All model runs used the same initial distribution of the cohorts: normal distribution with mean = 550 (southern part of WCVI) and sigma = 150.

Effect of School Size. The model comparison program configures each cohort into individual schools that migrate independently. To test for the effects of the number of schools on the simulated catch distribution, Jim compared simulated catch distributions using 10, 25, 50, and 75 schools per cohort for several model configurations. There were some minor differences between 10 and 25 schools, but virtually no differences between 25, 50, ad 75 schools. Thus, he used 25 schools per cohort to generate synthetic data.

Base Harvest Plan Catch Distribution. For the "base" harvest plan and the "constant" migration pattern the simulated catch distribution (25 schools/cohort) was similar to that of the 1986 CWT recoveries for the Voights Creek stock. For the "increasing" migration pattern, the J de F Net fishery catch increased and the SS Net fishery catch decreased.

                                       Model Simulation %
  Fishery Est CWTs ( %) Constant Mig Increasing Mig

  Ocean Troll 2,953 (37%) 35% 35%
  J de F Net 1,465 (19%) 19% 26%
  J de F sport 548 ( 7%) 6% 7%
  6B9 Net 82 ( 1%) 0% 0%
  6B9 Sport 117 ( 2%) 0% 1%
  SS Net 2,481 (31%) 36% 27%
  SS Sport 272 ( 3%) 3% 3%

FRAM Analysis. Jim demonstrated the latest version of the model comparisons program. The program makes a "BaseModel" run using one of the 16 model configurations, estimates FRAM parameters, and then simultaneously runs the BaseModel (under ANY configuration) and FRAM (using the estimated parameters) to compare FRAM predicted catches with "true" (i.e., BaseModel) catches in a future year. A "selective" fishery plan can be used in the "future" year. The "selective" harvest plan reduces harvest rates or catchability coefficients for the Unmarked Cohort in the Ocean Troll and all Sport fisheries to 15% of the "base" harvest plan. When the BaseModel was run in deterministic mode for future years, FRAM catch predictions were very close to the "true" values for all harvest plan configurations (base, low ocean, and selective). When the BaseModel was run with variability in the Initial Distribution and Migration processes, FRAM predictions were further from the "truth." Too few runs were completed prior to the meeting to draw further conclusions.

Next Step. The next step will be to add PM model parameter estimation and to quantify goodness-of-fit to the "true" model for FRAM, PM, and SSM models during forward simulations.


7. Next meeting.

The next meeting is scheduled for Tuesday October 6 at 9:00 am at NMFS Montlake.