2.6 Remove <parameter>

This command is useful if you want to restrict the parameter space and do certain tests. When you remove a parameter the value of the parameter is kept constant during the model fitting process. For example, if you want to test if R is exponentially distributed versus the more general Weibull model (which includes the exponential case when b=1), do the following: First fit the full model H₁, starting with some sensible initial values for the parameters:

   readdata mydata.txt
   model 1
   a = 0.1
   b = 0.5
   fit

This gives you the maximum likelihood under this model. Then, to fit the special case H₀ of b=1:

   remove b
   b = 1
   fit

Then calculate the change in deviance (The deviance of a fitted model is defined as two times the difference between the maximum achiveable likelihood and the maximum likelihood under the fitted model. This means that, you only have to calculate the change in $\ln L$ and multiply that by 2). This test statistic approximately $\chi^2_{n-n_0}$ distributed where n and n₀ is the number of residual degrees of freedom under H and H₀ respectively. For large values of $\chi^2$, H₀ can be rejected.

If there is a large amount of over-dispersion in the data (indicated by values of $\widehat{c}$ larger than 1), then this must be taken into account by doing an F-test instead [1] eq. 9. A useful ms-windows utility for calculating p-values corresponding to different F-values can be found at http://www.icw.com/ncss/shareware.html

Another situation in which the remove command may be useful is in cases where you know a priori the true value of some of the parameters, say $\sigma_{mf}$. By keeping such parameters out of the estimation procedure, you will get more unbiased and efficient estimates of other parameters. If you for example have estimated $\sigma_{mf}$ to, say 5 days, you can incorporate this by entering the following commands:

   fecundity
   sigmamf = 5
   remove sigmamf
   fit

The <parameter>'s are identified by the strings given in section 3.