Retrospective analysis with short time blocks

[jysullivan]

There is an issue accommodating short time blocks at the end of the model time series in a retrospective. Specifically it isn't possible to estimate parameters when time blocks are peeled back to < 3 years.

During a meeting 5/21/19, @scdressel and I discussed 4 alternatives to address this issue. To illustrate the problem and alternatives, I will use the example of natural mortality (M) time blocks from 1980-1998, 1999-2014, and 2015-2017.

Alternative 1: Remove the last time block and run the retrospective with only 2 time blocks. E.g. Assume time blocks for M are 1980-1998 and 1999-2017.

Not preferred because this is a different model. It is the easiest to code.

Alternative 2: For retrospective peels that results in time blocks <= 2 yrs, fix the parameter to the estimate obtained with 3 yrs.

Not preferred because it could mask poor retrospective behavior. Also challenging to program.

Alternative 3: For retrospective peels that results in time blocks <= 2 yrs, remove time block. So for peel 0, the time blocks are 1980-1998, 1999-2014, and 2015-2017, for peel 1 they are 1980-1998 and 1999-2016, and for peel 2 they are 1980-1998 and 1999-2015.

Preferred alternative. It is a compromise between Alternatives 1 and 4 that closely matches how time blocks are added to the model (they have to be at least 3 yrs in length before being considered).

Alternative 4: If the peels don't run, do not report them in the figures or include them in the calculation of Mohn's rho.

Second preferred alternative. Concern that we lose this information completely.