Simultaneous estimation of biomass models for 13 tree species: effects of compatible additivity requirements

Abstract

A desirable feature of biomass models distinguishing different tree components is compatible additivity of the component functions. Due to forcing of parameter estimates, such additivity is achieved at an expense of precision of the component functions. This study aimed to analyse the loss of precision incurred by forcing of parameters in tree biomass models due to (i) additivity constraints, (ii) combining global and species-specific parameters, and (iii) estimating component functions simultaneously as a system instead of as individual equations. Based on biomass data from 697 trees including 13 different species, we estimated a set of compatibly additive, nonlinear biomass models using simultaneous estimation and compared these with less restricted model systems. In line with other similar studies, the overall model system explained 88%–99% of the variation in individual biomass components. Compared with the unrestricted model, restricting parameters to obtain compatible additivity resulted in a change in RMSE of –0.6% to 5.4%. When restricting parameter estimates using both species-specific and global parameters, RMSE increased by 1.7%–13.1%. Estimating model parameters using simultaneous estimation (nonlinear iterated seemingly unrelated regression, NSUR) increased model bias compared with ordinary least squares estimation (OLS) for most biomass components. Contrary to expectations, NSUR estimation did not lead to a reduction in the standard error of estimates.