1 Models
The underlying model takes the form:
\[ W_{i} = \alpha L_{i}^{\beta}e^{z_{i}}\]
where \(\alpha\) and \(\beta\) are intercept and slope parameters (on log scale), \(L_{i}\), \(W_{i}\) are the length and weight, repectively, of individual i, and \(z_{i} \sim N(\mu,\sigma^2)\)
This model assumes log normal observational error (normal/gaussian on the log scale). This is important when the fitted model is used to predict weights from length
This model can be extended to :
\[ W_{ji} = \alpha L_{ji} ^ {\beta_j} e^{z_i} \] for season j, j= 1, …, 4 (spring, summer, fall, winter) or sex j, j = 1, …, 3 (0, 1, 2) or extended further to include season:sex combinations. These models are all nested and can therefore be tested using standard statistical methods.
1.1 Relationship between Normal and Log Normal distribution
If
\[Z \sim ~ LN(\mu, \sigma^2)\] then
\[ln(Z) \sim N(\mu,\sigma^2)\]
Note that each distribution uses the same parameters but their interpretations are different. For example the interpretation of \(\mu\) and \(\sigma^2\), under the normal distribution, are mean and variance:
\[E(ln(Z)) = \mu\] \[Var(ln(Z)) = \sigma^2\]
but under the log normal distibution the mean and variance are:
\[E(Z) = e^{\mu + \frac{\sigma^2}{2}}\] \[Var(Z) = (e^{\sigma^2} - 1)(e ^{2\mu + \sigma^2})\]
This is important when predicting weight from length.
1.2 Fitting
For this type of model it is natural to take logs and fit using ordinary least squares, since \(z_{i} \sim N(\mu,\sigma^2)\)
\[ ln(W_i) = ln(\alpha) + \beta ln(L_i) + z_i\] where \(ln(W_i)\) is regressed on \(ln(L_i)\) to estimate \(\ln(a)\), \(\beta\), and \(\sigma^2\)
1.3 Prediction
Now to estimate Weight using this fitted model
\[ \begin{aligned} E(W_{i}) &= \hat{W_i} \\ &= E({aL_i^\beta}e^{Z_{i}}) \\ &= E({aL_i^\beta}) E(e^{Z_{i}}) \\ &= {aL_i^\beta}e^{\sigma^2/2} \\ &= e^{ln(a) + \beta ln(L_i)}e^{\sigma^2/2} \\ &= e^{a + \beta ln(L_i) + \sigma^2/2} \end{aligned} \]
It is common, although incorrect, to see
\[\hat{W_i} = e^{a + \beta ln(L_i)}\] omitting the term \(\sigma^2/2\).
By omitting this term the resulting fitted value of weight will be biased toward smaller (in weight) individuals.