Compares several selection curves fitted with
select_Millar to the same data by the (quasi-) Akaike
information criterion corrected for small sample sizes (AICc and
QAICc).
Arguments
- ...
results of
select_Millar(class"select_Millar") fitted to the same data, or one list of such results. The names of the arguments (or of the list) are used as model names, otherwise the curve types (rtype).- c_hat
optional; overdispersion factor. By default (
NULL), it is estimated from the model with the most parameters (see Details).- c_hat_method
method to estimate
c_hat:"deviance"(default; deviance / degrees of freedom) or"pearson"(Pearson chi-square statistic / degrees of freedom).
Value
A data frame with one row per model, sorted by the QAICc, with the
columns model, rtype, number of parameters (npar),
deviance, degrees of freedom (dof), c_hat,
AICc, QAICc, and the difference to the lowest QAICc
(delta).
Details
Catches of gillnet and hook experiments are often overdispersed,
i.e. more variable than expected from the multinomial model of
select_Millar. The QAICc accounts for this by dividing
the log-likelihood by the overdispersion factor c_hat
(Burnham and Anderson 2002):
QAICc = -2 logL / c_hat + 2 K + 2 K (K + 1) / (n - K - 1),
where logL is the log-likelihood of the model, K the number of
parameters (plus 1 for c_hat if c_hat > 1), and n the
number of observations, here the number of non-zero cells of the
catch matrix (length classes x mesh sizes). The models can only be
compared with the same c_hat, which is therefore estimated from
the most general model (with the most parameters; if several models
have the most parameters, the one with the lowest deviance) and set to 1
if it is smaller than 1. The AICc is the same criterion without overdispersion
(c_hat = 1, K without the additional parameter).
The log-likelihood is the multinomial log-likelihood of the catches
conditional on the total catch per length class, without constant
terms. Therefore, only differences between models fitted to the same
data are meaningful. The model with the lowest QAICc (delta = 0)
is the best supported model; models with delta < 2 have similar
support (Burnham and Anderson 2002).
References
Burnham, K. P., Anderson, D. R., 2002. Model selection and multimodel inference: a practical information-theoretic approach. 2nd edition. Springer, New York.
Millar, R. B., Holst, R., 1997. Estimation of gillnet and hook selectivity using log-linear models. ICES Journal of Marine Science: Journal du Conseil, 54(3), 471-477.
Examples
data(gillnet)
fits <- list(
norm.loc = select_Millar(gillnet, x0 = NULL, rtype = "norm.loc", plot = FALSE),
norm.sca = select_Millar(gillnet, x0 = NULL, rtype = "norm.sca", plot = FALSE),
lognorm = select_Millar(gillnet, x0 = NULL, rtype = "lognorm", plot = FALSE),
binorm.sca = select_Millar(gillnet, x0 = c(55, 4, 65, 4, 3),
rtype = "binorm.sca", plot = FALSE),
bilognorm = select_Millar(gillnet, x0 = c(4, 0.2, 4.2, 0.1, 2),
rtype = "bilognorm", plot = FALSE))
select_Millar_compare(fits)
#> model rtype npar deviance dof c_hat AICc QAICc delta
#> 1 bilognorm bilognorm 5 545.5352 72 7.576878 22124.56 2931.706 0.0000000
#> 2 binorm.sca binorm.sca 5 550.4941 72 7.576878 22129.52 2932.361 0.6544772
#> 3 lognorm lognorm 2 704.2809 75 7.576878 22276.67 2945.845 14.1390214
#> 4 norm.sca norm.sca 2 772.7613 75 7.576878 22345.15 2954.883 23.1770946
#> 5 norm.loc norm.loc 2 862.9241 75 7.576878 22435.31 2966.783 35.0768303
## AIC of a single fit (without overdispersion)
AIC(fits$lognorm)
#> [1] 22276.52