MODELING SPECIES‐HABITAT RELATIONSHIP IN THE MARINE ENVIRONMENT A RESPONSE TO GREGR (2004)

I thank Gregr (2004) for providing valuable comments and criticisms of my article (Hamazaki 2002). This provides me an opportunity to clarify some of the methodologies I missed. I agree with Gregr that theoretical and methodological issues regarding a prediction model should be discussed plainly and openly so that we may learn from each other, and avoid common pitfalls. I also agree with all of his methodological concerns; however, I think Gregr missed the most important issues about building a prediction model: (1) the objective of building a prediction model, and (2) understanding the data source and sampling design. A prediction model is built to estimate/predict distribution of marine mammals from a set of environmental variables. The model is based upon statistical regression (e.g., multiple regression, logistic regression, general additive models) between the distribution of marine mammals and predictor/environmental variables (e.g., SST). Additionally, regression models can be used to (1) estimate/compare the effects of predictor variables on distribution of marine mammals, and (2) to test hypotheses, which would enhance understandings of ecological relationships and processes that influence distributionlhabitats of marine mammals. However, using regression models for these objectives requires different sets of requirements that are not necessarily required for predictions: (1) every sampling location must have an equal chance of being surveyed ( i e . , random sampling), and (2) {though this is not a statistical requirement] ecological/biologicaI relationships between marine mammal distribution and predictor variables should be theorized beforehand (Williams 1997, Guchery et aZ. 2001). Biased sampling would lead to biased estimation of the effects and significance of predictors. Statistical null hypothesis testing is an empirical testing of supporting or not supporting a particular theory, and thus null hypothesis testing without a solid theory is meaningless (Cherry 1998, Johnson 1999). Significant statistical relationship does not prove a causal relationship. It is most important that research data sampling should be designed to answer specific research objectives/questions (Cherry 1998, Johnson 1999). Most prediction modeling studies utilize data that are not specifically designed for these objectives (e.g. , Gregr and Trites 2001, Hamazaki 2002). This is dangerous data dredging (Anderson et dl . 2001, Johnson et al. 2001). It is meaningless and dangerous to conduct statistical analyses and interpret the results if the data are not designed to answer the research questions. This would also mean that available data determine research questions/ objectives. In my study I used sighting survey data that were collected by a systematic sampling method ( i e . , non-random line transect survey) to estimate abundance of species in areas they are known tofrequent ( i e . , violation of requirement ( 1 ) above). [Very few sighting surveys are conducted in areas where the species sought are known t o be yare or absent.] Predictor variables were selected for convenience of availability ( i e . , violation of requirement (2) above). [A prediction model is practically useless if the predictor variable data are not easily obtainable (Hamazaki 2002)]. The sighting survey also does not provide certainty whether the species are observed ar their preferred environmental conditions (Hamazaki 2002). I also have not found a theory describing functional relationship between the distribution of marine mammals and the predictor variables. Thus, it is obvious that the data are not designed to investigate ecological relationships or processes of distribution/habitats of marine mammals. Even, the data are severely limited for prediction purposes (Hamazaki 2002). In the light of the inappropriateness of using the data to investigate ecological relationships and hypothesis testing, most of Gregr’s concerns related to these objectives are