Objective To investigate the impact of covariates on diagnostic test and assess the correlation between re-peated measurement data,this paper explores innovative modeling techniques of ROC curve.Methods We introduce the new ROC curve method based on generalized linear mixed effects model and apply Bayesian techniques to parameters estimation with Winbugs Software.Further,areas under the ROC curve(AUC)with different values of covariates could be calculated in terms of assessment.Results Cases analysis results indicate the proposed method efficiently explores the repeated measurement data and provides parameters with practical significance,serving as a golden reference.Conclusion The ROC curve based on generalized linear mixed effects models can be effectively used to solve the test accuracy evaluation problem of the repeated diagnostic trials.

Objective　Exploring how to check the proportional hazards assumption of the Cox model,and the solutions to non-proportional hazards between the covariates and the hazard function.Methods　With a example data set of Ⅲc stage ovarian serous cystadenocarcina,illustrating how to use graphical methods to check the proportional hazards assumption of the Cox model.Results　the predictor of post-surgery administering medicine times violated the proportional hazards assumption of the Cox model.Conclusion　when using the Cox model to analyze the predictors of survival time,checking whether the predictors violate the proportional hazards assumption of the Cox model or not should be paid attention to.

Objective　To analyze the partial correlation problems by multilevel bivariate models.Methods　Multilevel bivariate models.Results　The results of correlation analysis with multilevel bivariate models is same with traditional methods,but the former is more quick and simple.Conclusion　Multilevel bivariate models are effective methods in analyzing the partial correlation problems.

Objective　The conditional hierarchical clustering for 1-dimensional(1-d) ordinal data was discussed.Methods　Because the individuals are ordered in 1-d,the conditional matrix was constructed with all elements in the second-diagonal are 1 and the others are 0.Distance matrix of individuals defined by some particular definition.Then the conditional-distance matrix was made for the hierarchical clustering by connecting the conditional matrix and distance matrix.This method was called 1-dimentsional conditional hierarchical clustering.An example was illustrated by this method and a Monte Carol study showed that method was feasible and robust.Results　Compared with the least-squares partition,this method is easy to understand,easy to practice and easy to compute.It also can give us a stable result.Conclusion　Because of the austere theory,the simple thought and the convenient application,it's a good method for the 1-d ordinal data.

Objective　This paper presents formula for determining the sample size required in case-control study to estimate vaccine efficacy VE with adjusting the precision of confidence interval.Methods　Formula to calculate sample size is derived from the principle of relative width of the confidence interval.Results　Prespecified relative width may regulate and control the precision of confidence interval and may reflects magnitude of the sample size.Conclusion　Sample size calculated by formula will assure that the investigator understand accurately the degree of point estimates of VE close to the true value of VE in the population.

Objective Research on variable substitution to non-linear regression forecast model precision's influence, and seek the modelling method that can improve the forecast precision. Methods Based on the data mining,the transform in space and the weighted processing combined method, make full use of information that the primary data provide. Results Given modelling method of combination forecast model based on the data mining. Conclusion Based on data mining's combination forecast model's modelling method can reduce the serious influence that the variable substitution brings and has fully used useful information in the primary data. It obviously improved the accuracy of the prediction model.

Objective To compare the efficiency of χ~2-Fisher's exact test which is one of the competitive null hypothesis approaches with SAM-GS which belongs toself-contained null hypothesis approaches. Methods The two methods were used to analyze a simulation experiment which contained five different scenarios. The results were compared with the simulated initialization,and assessing indexes were calculated to compare the efficiency. Results Under the same conditions,SAM-GS always have higher power than that of χ~2-fisher's exact test. However, the final inference is equivalent, namely if the difference between the two groups are smaller than 0.30,the two methods can not be better to identify differences between them. By contrary, when the differences between the two phenotypes are larger than 0.30, the two ways can both identify differences. Conclusion SAM-GS tends to have slightly higher power thanχ~2-Fisher' s exact test. The two methods can be used for screening enrichment gene sets of gene expression profile.χ~2-Fisher's exact test has the important advantage of being able to analyze multi-class phenotype.

Objective To explore the application of multilevel growth model in community intervention research. Methods The data a-bout blood pressure collected through questionnaire at baseline and 6 months, 12 months after intervention were analysed through multilevel growth model to evaluate the effectiveness of hypertension self-management. Results The blood pressures between the two groups were statistical significance at baseline,and the changing trends also were different. Age could affect diastolic blood pressure (DBP) changing after controlling the other factors. Conclusion Multilevel growth model can analyze the longitudinal data acquired from intervention research flexibly,and the result is more reliable.

Objective To investigate the risk factors of color-ectal cancer and to provide scientific basis for prevention of colorectai canc-er. Methods A 1:1 matched case-control study was carried out in Dalian including 200 cases with colorectal cancer and 200 controls. We analyzedthe data with the univariate analysis and mutivariate conditional logistic re-gressiou,theu calculated the Odds Ratio and the 95% confidence. Results Conditional logistic regression showed that history of constipation, family history of other tumor,intake much bloat food ten yeats ago and easily an-gry were the risk factors of colorectal cancer, and intake much fresh vegeta-bles ten years ago was the protective factors of colorectal cancer. One way analysis showed that following factors were the risk factors of coloroctal cancer:there were pollution factor around home 10 years ago, belly CT test,sleep over twelve o'clock at night,family history of coloroctal cancer, history of pries, appendicitis and the operation history of appendicitis, intake much fry and bake food,seashell and animal liver before ten years ago,of-ten feel strew, don't like to communicate with others,pessimism,don't harmonization with colleague. And iutaking much chicken meat, bean prod-ucts,garlic ten years ago,doing exercise and the frequency of exercise,high income were the protective factors of colorectal cancer. Conclusion Coloroctal cancer was the result of many factors. But the incidence of that in Dalian city was related to history of constipation, family history of other tumor, easily angry, intake less fresh vegetables and intake much bloat food ten years ago.

Objective To investigate the mixed model in bier-archical classification datas and implementing with mixed model in SAS. Methods Hierarchical classification datas exemplify the mixed model u-sing procedure mixed,and compared with traditional general linear model. Results The example shows the same result between the SAS mixed model and the general linear model. Conclusion SAS MIXED can flexi-bly fit and analysis hieraxchical classification datas.