Optimal Subset Regression Research Articles

An Experiment of a Statistical Downscaling Forecast Model for Summer Precipitation over China

AbstractA combination of the optimal subset regression (OSR) approach, the coupled general circulation model of the National Climate Center (NCC-CGCM) and precipitation observations from 160 stations over China is used to construct a statistical downscaling forecast model for precipitation in summer. Retroactive forecasts are performed to assess the skill of statistical downscaling during the period from 2003 to 2009. The results show a poor simulation for summer precipitation by the NCC-CGCM for China, and the average spatial anomaly correlation coefficient (ACC) is 0.01 in the forecast period. The forecast skill can be improved by OSR statistical downscaling, and the OSR forecast performs better than the NCC-CGCM in most years except 2003. The spatial ACC is more than 0.2 in the years 2008 and 2009, which proves to be relatively skillful. Moreover, the statistical downscaling forecast performs relatively well for the main rain belt of the summer precipitation in some years, including 2005, 2006, 2008, a...

Application of a Bayesian method for optimal subset regression to linkage analysis of Q1 and Q2.

We explore an approach that allows us to consider a trait for which we wish to determine the optimal subset of markers out of a set of p > or = 3 candidate markers being considered in a linkage analysis. The most effective analysis would find the model that only includes the q markers closest to the q major genes which determine the trait. Finding this optimal model using classical "frequentist" multiple regression techniques would require consideration of all 2p possible subsets. We apply the work of George and McCulloch [J Am Stat Assoc 88:881-9, 1993], who have developed a Bayesian approach to optimal subset selection regression, to a modification of the Haseman-Elston linkage statistic [Elston et al., Genet Epidemiol 19:1-17, 2000] in the analysis of the two quantitative traits simulated in Problem 2. The results obtained using this Bayesian method are compared to those obtained using (1) multiple regression and (2) the modified Haseman-Elston method (single variable regression analysis). We note upon doing this that for both Q1 and Q2, (1) we have extremely low power with all methods using the samples as given and have to resort to combining several simulated samples in order to have power of 50%, (2) the multivariate analysis does not have greater power than the univariate analysis for these traits, and (3) the Bayesian approach identifies the correct model more frequently than the frequentist approaches but shows no clear advantage over the multivariate approach.

Effects of crowding on health and social adaptation in the city of Chicago

A double logarithmic linear model was employed to estimate the effects of a variety of ecological variables hypothesized to induce stress and so influence one or more of the 32 endogenous variables which are used to characterize pathology. The pathology indicators were classified into the following five groups: mortality due to disease, mortality associated with social or psychological impairment, rates of morbidity and social disorganization, child and infant birth and death rates, and hospital use statistics. In order to select an appropriate set of explanatory variables, a stress model developed by Cassel (1974) was employed. In this model we distinguished between variables which may function to increase susceptibility to disease or social disorganization, and other variables which may function as buffers to protect individuals from such potential hazards. Among these variables are population density, household crowding, high-rise living, changes in relative population size in the various community areas, occupational diversity, ethnicity, married versus divorced or separated families, education and income. The distinction between supportive and stressor variables is a matter of classificatory convenience, since variables which are continuous, as these are, define the upper and lower boundaries of the conceptual variable they indicate. Thirty-two equations were fit to the twelve exogenous variables finally selected, and using the method of optimal subset regression (Boyce et al., 1974), the best subset was selected. For the four variables of primary concern, the crowding variables, it was found that population density was of some importance for several of the mortality rates, but in an opposite direction, negative rather than positive, from that hypothesized. For the social disorganization variables, however, it was positive, with the more dense areas having higher rates of delinquency, divorce—separation, illegitimacy and venereal disease. Household crowding was positively related to most mortality rates and a number of social disorganization variables, except for the several psychiatric indicators which were negative or zero. The latter suggests that it is isolation which important for psychiatric impairment (as measured by inpatient and outpatient admission rates). High-rise living had some small but significant effects on age-adjusted total death and heart death rates in a positive direction and on delinquency but on few other variables. Change in population density had negative weights on six mortality rates (total deaths, heart deaths, age-specific heart deaths 65+, cancer deaths, stroke deaths and cirrhosis deaths), as well as on the rate of homicide. This suggests that relative loss of population in a particular community area is a significant factor in mortality and is therefore supportive of one aspect of the Cassel theory. In summary, we found that some aspects of the Cassel theory, as respecified in the context of our statistical model, were supported. Population density, whilenot exercising the hypothesized effect on mortality, did so with regards to the social disorganization variables. Household crowding, where significant, tended to have the hypothesized effect. High-rise living had few significant associations, but those associations with total death and heart death rate and delinquency are interesting and merit further investigation. Loss in population appears to be a significant factor in mortality from six diseases as well as in the homicide rate.