|
内容記述 |
This thesis gives some methods for evaluating statistical data via a<br /> model selection and shows some applications of the methods. Especially,<br /.> plant ecological and worm ethological data with spatial information are<br /> treated. Characteristic problems arising in each data are shown through the<br /> data analysis of explanatory variables in base models.<br /> Firstly, the problem of selecting topographical attribute as the<br />explanatory variable is discussed through contingency table statistics. As an<br /> example, an analysis of observed data about relative altitudes and about<br />distribution of moss is shown. In order to consider the relationship between<br />topography and distribution of plants, data of altitude and mosses in study<br />plots in continental Antarctica were collected by the present author when he<br /> had a chance to visit there. The altitude data was processed as an<br /> explanatory variable of model about the existence of moss. As one of the<br /> important topographical attributes that are calculated from altitude data, <br /> the relief from standard surface was adopted. Under the assumption that the<br /> probability of moss existence is proportional to the value of residuals from<br /> standard surface, the strength of relationship between topography and moss<br /> distribution was obtained by using 2 X 2 contingency table statistics. It<br /> suggested that a simpler standard surface had a better ability, than an<br /> adjusted standard surface, of determining topographical attribute which<br /> strongly related to the moss existence. Then, the standard surface estimated<br /> by robust methods presented in this paper had a little better ability than by<br /> a prevailing least-square method. Totally, the accurate regression methods<br /> were overfitting as a method of determining the standard surface. <br /> Secondly, a modeling with spatial information is treated. The problem<br /> of selecting a model is discussed through linear models and linear mixed<br />models. As examples, the analyses of the distribution of dwarf pine in mixed<br /> forest and of the moving track of nematode are shown. In the analysis of<br /> dwarf pine, data of forest study was collected in central Kamchatka by the<br /> research group in which the present author has participated. There, the size<br /> of stem diameter at the base of the lowest live branch of dwarf pine<br /> individuals and their distance from canopy trees were used for parameters of<br /> growth models. Consequently, as in many previous works of growth model<br /> about canopy trees, an adequateness of using the inverse of distance in<br /> competitive effect terms was shown. In the analysis of nematode, data<br /> collected by an automatic tracking system of the center of gravity about the<br /> individuals at an experiment of learning action to detestable odorant for<br /> nematode was used. The avoidance action of nematode from the odorant was<br /> explained by the factors involving time and coordinate. In the detestable<br /> odorant (2-nonanone) exposed condition, the selected linear model retained<br /> the term of distance with positive coefficient and did not retain the term of<br /> time. On the other hand, the selected linear mixed models retained the terms<br /> of time with positive coefficient and the distance with negative coefficient in<br /> selected models about avoidance action. However, in the solvent (ethyl<br /> alcohol) exposed condition, different variable selection rules lead quite<br /> different results in linear mixed models and stable interpretation cannot be<br /> done. As a result of model selection, it is shown that the linear mixed models<br /> have substantially better reproduction ability of moving track than the other<br /> models. <br /> In summary, by applying statistical methods and models presented in<br /> this thesis, it is shown that many significant results in ecology and ethology<br /> are derived which are useful and suggestive in respective scientific fields. |
要旨・審査要旨 (241.1 kB)
