Even-aged stand mortality estimation through a two-step regression process

It is argued that a conflict between data and models is inherent in many mortality functions. Equations are utilised which have the property that however short a chosen growth period, then some loss of stems/ha must occur. But permanent sample plot records frequently contains data which exhibit no mortality over several years. If these data are discarded then mortality must be overpredicted. If they are retained, difficulties can occur in residual plot analyses and fitting of models, through the creation of bimodal residual distributions and failure to produce average residual values near zero. To resolve this, a two-step modelling strategy is suggested. A logistic regression predicting the probability of stems/ha death is first constructed. A mortality equation is then built, but only utilising data where death has occurred over a period. Estimates from this second model are then reduced by a factor, equivalent to the probability of death occurring, acquired from the logistic equation. An example involving Pinus radiata data in New Zealand is given. It is argued that better behaved prediction emerges through adoption of the system and clearer information on stand death dynamics is obtained. The system can be utilised stochastically, if desired. (C) 1998 Elsevier Science B.V.