Skip to contents

When fish are scarce, a catch or survey sample can record zero for an age even if fish of that age are present. The default lognormal family uses positive observations and treats zeros as missing values. Poisson-link delta (PLD) families use recorded zeros to inform the predicted catch or index. PLD can be used for catch, survey indices, or both.

The observation model

Let \(Y_i\) be a recorded catch or survey index for an age and year, and let \(\widetilde{Y}_i\) be its prediction from the population model. The catch prediction comes from the Baranov catch equation. The survey prediction is abundance at sampling time scaled by catchability. Poisson-link delta (PLD) uses this prediction to explain both the chance of recording a positive value and the size of positive observations (Thorson 2018, see also the sdmTMB Poisson-link delta guide).

tinyAM adapts Thorson’s parameterization by using the assessment prediction as the overall mean. Each fleet or survey \(j\) has a soft detection scale \(\widetilde{Y}_{50,j}\) (the predicted catch or index at which half of observations are expected to be zero). The prediction and the scale use the observation’s units.

Encounter. Conditional on the population states and parameters, recording a positive value is a binomial outcome with one trial and probability \(p_{+,i}\). The probability of recording zero is the chance of zero events under a Poisson distribution with mean \(\lambda_i\).

\[ \lambda_i = \log(2)\frac{\widetilde{Y}_i}{\widetilde{Y}_{50,j}}, \qquad p_{0,i} = \Pr(Y_i = 0) = e^{-\lambda_i}, \qquad p_{+,i} = \Pr(Y_i > 0) = 1-p_{0,i}. \]

Doubling the prediction from the soft detection scale to twice that level lowers the chance of recording zero from 50% to 25%. The encounter component uses a complementary log-log link for the probability of a positive.

\[ \operatorname{cloglog}(p_{+,i}) = \log\!\big[-\log(1-p_{+,i})\big] = \log\lambda_i. \]

Positive observations. The positive component uses a log link for its arithmetic mean \(m_{+,i}\), with \(\log m_{+,i} = \log\widetilde{Y}_i - \log p_{+,i}\). Dividing by the probability of a positive keeps the average across zeros and positives equal to the assessment prediction.

\[ m_{+,i} = \frac{\widetilde{Y}_i}{p_{+,i}}, \qquad p_{+,i}\,m_{+,i} = \widetilde{Y}_i. \]

The positive distribution can be lognormal, Gamma, or generalized gamma. For example, with a lognormal positive component,

\[ \log Y_i \mid Y_i > 0 \sim N\!\left(\log m_{+,i}-\frac{\sigma_i^2}{2},\ \sigma_i^2\right). \]

Here \(\sigma_i\) controls the spread on the log scale. Subtracting \(\sigma_i^2/2\) gives the positive observations an arithmetic mean of \(m_{+,i}\). The default, positive-only lognormal() family uses the assessment prediction as its conditional median. Observation family reference

Likelihood. A recorded zero contributes the log probability of recording zero. A positive contributes the log probability of recording a positive plus the log density of its size. Let \(g_{+,i}\) be the chosen positive density on the log-observation scale.

\[ \ell_i = \begin{cases} \log p_{0,i}, & Y_i = 0,\\[4pt] \log p_{+,i} + \log g_{+,i}(\log Y_i), & Y_i > 0. \end{cases} \]

Apply PLD to Northern cod

Northern cod has 230 recorded catch zeros and 268 Fall 2J3KL survey zeros at ages 2–20+ in 1983–2025. This example compares fits using the default lognormal and the Poisson-link delta lognormal.

ncam <- fit_tam(cod_obs, years = 1983:2025, ages = 2:20,
  N_settings = list(process = "off", init = "exp"),
  rec_settings = list(process = "rw"),
  F_settings = list(process = "ar1", mu_form = ~ 1, mean_ages = 5:14),
  M_settings = list(process = "ar1", mu_supplied = ~ I(0.3),
    age_breaks = c(2, 5, 8, 21), mean_ages = 5:14),
  catch_settings = list(sd_form = ~ 1),
  index_settings = list(sd_form = ~ 1, q_form = ~ mono(q_block)),
  silent = TRUE
)

ncam_pld <- update(ncam,
  catch_settings = list(family = pld_lognormal()),
  index_settings = list(family = pld_lognormal()),
  silent = TRUE
)

The lognormal fit treats recorded zeros as missing in the likelihood. Under PLD, a recorded zero contributes the fitted probability of observing zero. Truly missing observations contribute to neither likelihood. The lognormal model predicts a median for the positive observations. PLD predicts an arithmetic mean across zeros and positives.

Soft detection scales

Series Soft detection scale Lower 95% Upper 95% Units
Commercial catch 0.70 0.61 0.81 thousand fish
Fall 2J3KL survey 26.86 23.54 30.64 thousand fish

For each series, a fitted mean near its soft detection scale gives a zero probability near 50%. The table uses the same units as the matrices below.

Which observations are likely to be zero?

The numbers in the cells are the observations, in thousands of fish. The colour shows the probability of zero calculated from the fitted mean. Darker cells have higher probabilities. Grey cells mark missing observations. Hover over a cell for the full recorded value, fitted mean, and probability. In the cell labels, k means one thousand of the displayed units.

The catch recorded 230 zeros, with 221 expected under PLD. The survey recorded 268 zeros, with 263 expected.

How many zeros occur each year?

Dots count the ages with a recorded zero. The line sums the fitted probabilities of zero over the same sampled ages. Years without observations leave a gap.

Compare the assessments

Only six catches and two survey observations are positive at age 20+. The 20+ estimates therefore draw heavily on the fitted population dynamics. The abundance plot shows point estimates on a log scale so differences are visible when the stock was low. Shading marks the low-stock years from 1992 to 2010.

For 2025 spawning biomass, the change from lognormal to PLD is +35%.

Northern cod dashboard

Forecast held-out observations

For each target year from 2006 to 2025, the hindcast fits through the previous year and forecasts one year with terminal F held constant. Log-RMSE and MASE compare predicted medians against the same positive survey observations for both models. A MASE below one means the model had a lower mean absolute error on the log scale than a forecast that carried forward the last available positive observation, evaluated on the same records.

hind <- fit_hindcast(ncam, ncam_pld, folds = 20, proj_settings = list(F_mult = 1))
#> Models differ in which Poisson-link delta observations can be scored. Compare
#> Brier scores only for matching cases.

The result includes held-out predictions and fold diagnostics. The table uses only positive survey observations with usable predictions from both models.

Log-RMSE MASE
Lognormal 0.73 0.92
PLD 0.67 0.87

All folds passed numerical checks. PLD reduced both errors on the shared positive survey observations. Both models had MASE below one.

The hindcast also scores PLD’s forecasts of whether each held-out record was zero. The Brier score averages the squared difference between the forecast probability of zero and the outcome (1 for a zero and 0 for a positive). Lower scores are better, with zero indicating perfect probability forecasts on these records. PLD scored 0.05 for catch and 0.06 for the survey.

AIC comparisons require a common likelihood for the same observations. Here the families use different sets of observations because of the zeros. Conditional AIC is currently available for the all-lognormal fit only.

References

Thorson, J. T. 2018. Three problems with the conventional delta-model for biomass sampling data, and a computationally efficient alternative. Canadian Journal of Fisheries and Aquatic Sciences, 75(9), 1369–1382. doi:10.1139/cjfas-2017-0266.