• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Rank regression: an alternative regression approach for data with outliers

    2014-12-08 07:38:34TianCHENWanTANGYingLUXinTU
    上海精神醫(yī)學 2014年5期
    關鍵詞:性健康正態(tài)分布線性

    Tian CHEN, Wan TANG, Ying LU, Xin TU*

    ?Biostatistics in psychiatry (23)?

    Rank regression: an alternative regression approach for data with outliers

    Tian CHEN1, Wan TANG1, Ying LU2, Xin TU1*

    normal distribution, non-normal distribution, linear regression, semi-parametric regression models, rank regression, sexual health

    1. Introduction

    Regression is widely used in mental health research and related services research to model relationships involving health and service utilization outcomes and clinical and socio-demographic factors. Regression models measure changes in the dependent variable in response to changes in a set of independent variables of interest. Linear regression focuses on continuous dependent variables, while other regression models such as logistic and log-linear regression consider noncontinuous dependent variables such as binary and count outcomes. The dependent variable is often called the response, while the independent variables are frequently referred to as the explanatory variables,predictors, or covariates.

    Linear regression is arguably the most popular regression model in practice, because of the ubiquity of continuous outcomes and because it is relatively easy to understand the modeled relationship and interpret the model estimates. Fitting such models is convenient because all major software packages (R,SAS, SPSS and STATA) provide both the model estimates and the diagnostics of the model fit. However, the wide popularity and routine use of the linear regression also creates some problems. Many researchers apply the model without first checking assumptions about the normal distribution of the data underlying the validity of model estimates. The classic normal-based linear regression imposes strong constraints on data, and its estimates are also quite sensitive to departures from assumed mathematical models. Without careful checking of the model assumptions, estimates generated by linear regression models may be difficult to interpret and conclusions drawn from such estimates may be misleading.

    2. Different approaches to deal with non-normal study data in regression analyses

    Classic linear regression assumes a normally distributed response,yi, and models the mean of this response variable as a function of a set of independent variables,xi= (xi1 , xi2 ...., xip)Tas follows∶

    whereβ= (β1, β2, ..., βp)Tis the vector of parameters,nis the sample size,εidenotes the error term,N(μ,σ2)denotes a normal distribution with meanμand varianceσ2, andεi~N(0,σ2) means thatεifollows a normal distribution with mean 0 and varianceσ2. The wellshaped bell curve of the normal distribution is often at odds with the distribution of data arising in real studies,because of its symmetric shape and extremely thin tails(exponential decay). Over the years, various methods have been developed to improve the limitations of the classic linear model. All the different methods can be grouped into 3 major categories.

    One approach is to use mathematical distributions that more closely resemble the data distribution in the study.[1]For example, by positing a t-distribution for the errorεi, the resulting linear model can accommodate data distributions with thicker tails. This is possible because the t-distribution has an additional degree of freedom parameter to control the thickness of the tail.However, like the normal distribution, the t-distribution is also symmetric. To model skewed data distributions,a popular approach is to use the chi-square distribution.Although this parametric alternative broadens the scope of data distributions that can be accommodated, it is still quite limited because mathematical distributions always have more regular shapes than those arising in practice.

    A second popular alternative is to use semiparametric or distribution-free models.[2]Under this approach, no mathematical model is assumed for the data distribution (the non-parametric part) and the relationship betweenyiandxiis represented by the mean ofyiafter adjustment forxi(parametric component). The latter parametric component is implied by the specification of the classic linear regression in (1)and is given by∶

    whereE(yi|xi) denotes mathematical expectation. For those unfamiliar with mathematical expectation, the above expression simply means that the population-level average of the responseyiis a linear function ofxi. This linear relationship is also implicit in the normal-based linear regression in (1). Thus, the semi-parametric linear model in (2) only requires a linear relationship between the response and the set of explanatory variables,thereby offering valid inference for a wide class of data distributions.

    Although significantly improving the utility of linear regression, the semi-parametric model still has limited applications. A major problem is that like the classic model it continues to model the mean of the response.Like the sample mean of a variable, model estimates from this approach can be quite biased when there are extremely large or small values, or outliers, in the response.

    Various approaches have been developed to address this important issue of outliers. A common approach in psychosocial research is to trim outliers using ad-hoc rules. For example, limiting the values of all observations to 3 times the interquartile range when estimating the mean of an outcome (i.e., a ‘trimmed’mean).[3]However, these ad-hoc methods induce artifacts because of their dependence on the specific rules used, and the use of different rules can result in different outcomes.

    Another approach to limiting the influence of outliers is to employ rank tests. The Mann-Whitney-Wilcoxon rank sum test is widely used to compare two groups in such situations. Within the setting of regression analysis, rank regression is a popular approach for dealing with outliers.[4,5]Like the Mann-Whitney-Wilcoxon rank sum test, rank regression does not use the observed responsesyidirectly, but,rather, uses information about the ranking of these observations, thereby yielding estimates that are much less sensitive to outliers.

    3. Simulation studies to compare different approaches

    The data were simulated from a study with one binary variable and one continuous covariate. To show differences across the different methods, we selected a large sample size (n=500) to reduce the effect of sampling variability on model estimates. We performed simulation of data and fitted the different models to the data generated using the R software. All simulations were performed with a Monte Carlo sample size M=1000 and a type I errorα=0.05.

    We simulatedyifrom the following linear model∶

    We then simulated 50 (or 10% of the sample size) values from a uniformU(500, 1000000), ordered them as∶

    and added the valuesu(1)from the uniform to the 50 largest values ofyi, i.e.,

    to form a set of outlying observations, i.e.,

    To assess the robustness of the different methods,we replacedy(451)<y(452)< ...< y(500)in the original sample with the valuesz(451)<z(452)< ...< z(500), and fit the models to the resulting observations∶

    Table 1 shows the estimates ofβ1andβ2, the corresponding standard errors, and type I error rates from fitting the three methods to data simulated from the normal-distributed errorN(0, 1/2) based on 1000 Monte Carlo simulations both with and without included outliers. (The interceptβ0is estimated by the rank regression and so this estimate is missing in the table.) In the table, values in the column titled‘mean’ are the averaged estimates of each parameter over 1000 Monte Caro replications; the ‘asymptotic standard error’ is the model-based standard error; the‘empirical standard error’ is the standard errors of the 1000 estimates of each parameter; and the ‘type I error’is the percent of times the null hypothesis - that the estimated parameter is equal to the true parameter -is rejected. For example, the empirical type I error rates forβ1in the data set without outliers is the percent oftimes of rejecting the nullH0∶β1=1.

    If a model performs well, (a) the averaged value of estimates of each parameter (in the ‘mean’ column)should be close to the true value of the respective parameter; (b) the magnitude of the asymptotic standard error should be close to that of the empirical standard error; and (c) the empirical type I error rate should be close to the nominal value 0.05. As shown in Table 1, in the absence of outliers, all three methods performed well, with the averaged estimates all nearly identical to the true value 1, the asymptotic standard errors all close to their empirical counterparts, and the type I error rate all close to the nominal levelα=0.05.Further, all three methods yielded near identical standard errors, indicating that there is practically no loss of power by using the two robust alternatives instead of the classic linear model for the simulated normal data.

    However, results are very different in the presence of outliers. As shown in the Table 1, both the classic and semi-parametric models yielded extremely large estimates that are un-interpretable, impossibly large standard errors, and type I errors close to 1. In contrast,the rank regression model for bothβ1andβ2generated estimates close to the true value 1, reasonable asymptotic and empirical standard errors that were equal to each other, and type I errors that, though elevated, were close to the nominal 0.05 level.

    Table 2 shows the results of a similar simulation when the data were simulated from t-distributed error, ,instead of from normal-distributed error. In the absence of outliers the mean estimate and type 1 error of the two parameters were acceptable for all three models;however, the empirical standard error was much larger than the asymptotic standard error for the classical and semi-parametric models while these two types of standard error were similar in magnitude in the rank regression model. In the presence of outliers, as was the case in the normal-error simulation, the estimates generated by the classic and semi-parametric models were un-interpretable while those generated by the rank regression model were acceptable. Thus, for data with t-distribution error the rank regression model preforms better than the classic linear and the semiparametric models both in the absence and in the presence of outliers.

    4. A real-life example

    To illustrate the three approaches to dealing with outliers, we use results from a recent randomizedcontrolled study[6]to evaluate the efficacy of a sexual risk-reduction intervention program targeting teenage girls in low-income urban settings who are at elevated risk for HIV, sexually transmitted infections, and unintended pregnancies. The study recruited sexuallyactive urban adolescent girls aged 15 to 19 and randomized them to a sexual risk reduction intervention or to a structurally-equivalent health promotion control group. Assessments and behavioral data were collected at baseline, 3, 6 and 12 months post-baseline.The primary interest of the study was to compare the frequency of unprotected vaginal sex between the two treatment conditions. A difficult problem with the study data was the extremely large values reported by some subjects for their sexual activities. For example, five subjects reported over 100 episodes of unprotected vaginal sex over the past 3 months at the 6 month follow-up. If linear regression is applied directly to this outcome, estimates will be severely biased and become un-interpretable. Alternative models need to be considered when analyzing the data.

    Table 1. Estimates (mean), asymptotic and empirical standard errors, and empirical type I error rates from fitting the classic linear, semi-parametric, and rank regression models to data simulated from normal-distributed errors

    The linear regression for the different methods is specified as follows∶

    whereyiis the number of episodes of unprotected vaginal sex,xi1is the binary indicator for the treatment condition (1 for the intervention and 0 for the control group), andεiis the model error. The model errorεifollows the normal distribution for the classic linear regression, while the distribution is unspecified for the semi-parametric and rank regression methods.

    To highlight the differences in the models we removed zero observations (i.e., individuals who reported no episodes of unprotected sex in the prior three months) and fit all three models (classic linear,semi-parametric, and rank regression) to the remaining data. In addition, we also recomputed the estimates for the classic linear model and the semi-parametric model after trimming the observed responses to decrease the influence of outliers. We trimmed the observed responses of number of episodes of unprotected vaginal sex in the prior three months at 3 times the interquartile range; the 25%, 50% and 75% quartiles were 2, 4, and 10 episodes, respectively, so the interquartile range was 8 (10 - 2) and any observations below -20 (4 - 3*8)or above +28 (4 + 3*8) were considered outliers. There were no observations below -20 so no lower-level trimming was necessary, but all observations above 28 were trimmed to 28.

    Table 3 shows the resulting estimates ofβ1for the treatment condition in the linear model (3) and the corresponding asymptotic standard errors and p-values using the different models. As was the case in the simulation study with outliers, the huge values for the estimates and standard errors using the classic linear and semi-parametric models clearly show that the estimates are profoundly affected by the outliers and,thus, are un-interpretable. In comparison, the classic and semi-parametric methods yielded more reasonable estimates when applied to the trimmed observations.However, results using the trimmed data were still quite different from those generated from the rank regression model; the estimates from the two models that used trimmed data were more than 50% higher than that using the rank regression method and the standard errors were more than double that from the rank regression analysis. Results from the simulation study suggest that rank regression is quite robust against outliers and, unlike models that use trimmed data,are not vulnerable to change when different trimming criteria are employed.

    Table 2. Estimates (mean), asymptotic and empirical standard errors, and empirical type I error rates from fitting the classic linear, semi-parametric, and rank regression models to data simulated from t-distributed errors

    Table 3. Estimates, standard errors, and p-values from fitting the classic linear, semi-parametric,rank regression, classic linear with trimmed outliers, and semi-parametric with trimmed outliers models to the risk-reduction intervention study

    5. Sotfware for alternative linear regression models

    Most major software such as R and SAS has the capability of fitting the semi-parametric linear regression model. In R, there are several packages available for fitting the generalized estimating equations (GEE).Although GEE is an extension of the semi-parametric method for longitudinal data, we may still use these packages for fitting the semi-parametric model to crosssectional data by introducing an ‘ID’ variable that has unique values for each of the observations. For example,if the GEE package is installed, then one may apply the following codes to fit the semi-parametric linear regression model∶

    where y is the outcome and x is the covariate matrix.

    Similarly, SAS also offers ‘Procedures’ for fitting the GEE which can be utilized to provide estimates for semiparametric linear regression models. For example, by adding an ID variable to the SAS data set, we may apply the Procedure GENMOD to fit the semi-parametric model∶

    At the time of writing, SAS does not have the capability to fit the rank regression. For our simulated and real study examples, packages in R were used to fit this robust alternative model. To perform this regression model, first download the R functions from the website∶http∶//www.stat.wmich.edu/mckean/HMC/Rcode/AppendixB/ww.r. Then, we use the following command in R to obtain estimates from fitting the rank regression∶

    where y is the outcome and x is the covariate matrix.

    Note that while SAS is a commercial software package, R is free to download, install, and run. In addition, software for newer statistical methods are generally first available in R. However, unlike SAS, R has no designated technical support so users generally rely on peer-support, web postings, and books for resolving issues concerning applications of specific packages and general data management problems.

    6. Discussion

    Classic linear regression has a number of weaknesses,limiting its applications to real study data. We discussed two robust alternatives, the semi-parametric model and the rank regression model. Although the former yields more valid estimates than the classic linear model, it breaks down when there are extremely large (or small)observations in the response (i.e., the dependent variable). In the presence of such outliers, the rank regression model provides much more robust estimates.Unlike ad-hoc methods such as trimming outliers based on 3 x interquartile range, rank regression generates the same estimates regardless of the actual values of the response as long as the rankings of the observations remain the same. This formal approach not only removes any subjective element in the estimates, but it also makes it easier to compare results of different analyses based on the same study data and to compare results between different studies. Further, the rank regression model is also capable of addressing outliers in the independent variables, although this tutorial only discussed outliers in the response variable.

    Currently, rank regression is only available in some selected software packages such as R - we included sample R codes for fitting this robust regression model in this report to facilitate its use by readers. As this approach becomes more popular, it is likely that other major software giants such as SAS will have similar offerings.

    Unlike the classic and semi-parametric linear regression models, rank regression is only available for fitting cross-sectional data. This is, in part, due to the complexity of computing estimates and asymptotic standard errors. However, as longitudinal studies become the norm rather than the exception in modern clinical research, it will become increasingly important to develop software that can extend this robust model to longitudinal research data and, thus, help investigators more effectively deal with imperfections in real study data.

    Conflict of interest

    The authors report no conflict of interest related to this manuscript.

    Funding

    The preparation of this manuscript was supported in part by DA027521 and GM108337 from the National Institutes of Health.

    1. Kowalski J, Tu XM, Day RS, Mendoza-Blanco JR. On the rate of convergence of the ECME algorithm for multiple regression models with t-distributed errors.Biometrika. 1997; 84∶269-281. doi∶ http∶//dx.doi.org/10.1093/biomet/84.2.269

    2. Tang W, He H,Tu XM.Applied Categorical and Count Data Analysis. Boca Raton, Florida, USA∶ Chapman & Hall/CRC Press. 2012

    3. Schroder EB, Liao DP, Chambless LE, Prineas RJ, Evans GW,Heiss G. Hypertension, blood pressure, and heart rate variability∶ the Atherosclerosis Risk in Communities (ARIC)study.Hypertension.2003; 42(6)∶ 1106-1111. doi∶ http∶//dx.doi.org/10.1161/01.HYP.0000100444.71069.73

    4. Jaeckel LA. Estimating regression coefficients by minimizing the dispersion of the residuals.Ann Math Statist. 1972;43(5)∶ 1449-1458

    5. Jureckova J. Nonparametric estimate of regression coefficients.Ann Math Statist.1971; 42(4)∶ 1328-1338

    6. Morrison-Beedy D, Jones S, Xia Y, Tu XM, Crean H, Carey M. Reducing sexual risk behavior in adolescent girls∶results from a randomized controlled trial.J Adolesc Health.2013; 52∶ 314-321. doi∶ http∶//dx.doi.org/10.1016/j.jadohealth.2012.07.005

    ∶ 2014-10-08; accepted∶ 2014-10-10)

    Ms. Tian Chen is a fifth-year PhD student in the Department of Biostatistics and Computational Biology, School of Medicine and Dentistry, University of Rochester. Her PhD thesis focuses on semiparametric and rank-based statistical models, and variable selection methods for regression models for both cross-sectional and longitudinal data. She has applied these statistical methods in the analysis of mental health and related research.

    等級回歸:離群數(shù)據的另一種回歸方法

    Tian CHEN, Wan TANG, Ying LU, Xin TU

    正態(tài)分布,非正態(tài)分布,線性回歸,半參數(shù)回歸模型,等級回歸,性健康

    Summary:Linear regression models are widely used in mental health and related health services research.However, the classic linear regression analysis assumes that the data are normally distributed, an assumption that is not met by the data obtained in many studies. One method of dealing with this problem is to use semi-parametric models, which do not require that the data be normally distributed. But semi-parametric models are quite sensitive to outlying observations, so the generated estimates are unreliable when study data includes outliers. In this situation, some researchers trim the extreme values prior to conducting the analysis, but the ad-hoc rules used for data trimming are based on subjective criteria so different methods of adjustment can yield different results. Rank regression provides a more objective approach to dealing with non-normal data that includes outliers. This paper uses simulated and real data to illustrate this useful regression approach for dealing with outliers and compares it to the results generated using classical regression models and semi-parametric regression models.

    [Shanghai Arch Psychiatry. 2014; 26(5)∶ 310-316. doi∶ http∶//dx.doi.org/10.11919/j.issn.1002-0829.214148]

    1Department of Biostatistics and Computational Biology, University of Rochester, NY, USA

    2Department of Biostatistics, Stanford University, Stanford, CA, USA

    *correspondence∶ xin_tu@urmc.rochester.edu

    A full-text Chinese translation of this article will be available at www.shanghaiarchivesofpsychiatry.org on November 25, 2014.

    概述: 線性回歸模型被廣泛應用于精神衛(wèi)生和衛(wèi)生服務相關研究。然而,經典線性回歸分析是假設該數(shù)據為正態(tài)分布的,但是很多研究所獲得的數(shù)據并不符合這種假設。解決該問題的方法之一是采用不要求數(shù)據為正態(tài)分布的半參數(shù)模型。但是,半參數(shù)模型對離散數(shù)據相當敏感,因此在處理包含離散值的數(shù)據時產生的估計值是不可靠的。在這種情況下,一些研究者在刪減這些極端值后再進行分析,但是,刪減數(shù)據的事先法則(ad-hoc rules)是基于主觀標準的,所以不同的調整方法就會產生不同的結果。等級回歸為處理包括離散值的非正態(tài)分布數(shù)據提供了更為客觀的方法。本文采用虛擬和實際數(shù)據來闡述這個非常有用的處理離散值的回歸方法,并與采用經典回歸模型和半參數(shù)回歸模型所得出的結果進行比較。

    本文全文中文版從2014年11月25日起在www.shanghaiarchivesofpsychaitry.org可供免費閱覽下載

    猜你喜歡
    性健康正態(tài)分布線性
    我國高職學生性健康知信行量表的構建及應用
    護理研究(2023年20期)2023-10-27 08:16:26
    漸近線性Klein-Gordon-Maxwell系統(tǒng)正解的存在性
    性健康教育,教師怎么做?
    新班主任(2022年4期)2022-04-27 06:20:49
    利用主題活動淺析小班幼兒進行性健康教育的意義和研究策略
    速讀·中旬(2021年12期)2021-10-14 08:05:57
    線性回歸方程的求解與應用
    關注性健康教育 促進兒童青少年健康
    教育家(2018年41期)2018-11-20 11:49:56
    二階線性微分方程的解法
    基于對數(shù)正態(tài)分布的出行時長可靠性計算
    正態(tài)分布及其應用
    正態(tài)分布題型剖析
    日韩,欧美,国产一区二区三区 | 欧美成人免费av一区二区三区| 一本精品99久久精品77| 国产黄片美女视频| 日本五十路高清| 天堂√8在线中文| 欧美成人一区二区免费高清观看| 亚洲欧美精品综合久久99| netflix在线观看网站| 国产高潮美女av| 国产私拍福利视频在线观看| 午夜福利视频1000在线观看| 久久久精品大字幕| 99视频精品全部免费 在线| 村上凉子中文字幕在线| 国产91精品成人一区二区三区| 丰满乱子伦码专区| 成人特级av手机在线观看| 午夜福利高清视频| 人人妻人人澡欧美一区二区| 日韩精品青青久久久久久| 在线观看一区二区三区| 亚洲国产欧美人成| a级毛片a级免费在线| 免费大片18禁| 一区二区三区激情视频| 小说图片视频综合网站| 村上凉子中文字幕在线| www日本黄色视频网| 美女高潮喷水抽搐中文字幕| 俄罗斯特黄特色一大片| 一区二区三区免费毛片| 中文字幕精品亚洲无线码一区| 久久99热6这里只有精品| 99精品在免费线老司机午夜| 波野结衣二区三区在线| 成人特级黄色片久久久久久久| 国产高清视频在线播放一区| 亚洲精品成人久久久久久| 国产私拍福利视频在线观看| 一夜夜www| 男女之事视频高清在线观看| 久久九九热精品免费| 亚洲精品456在线播放app | 91av网一区二区| 亚洲一级一片aⅴ在线观看| av视频在线观看入口| 国产精品一区www在线观看 | 成人三级黄色视频| 国产黄a三级三级三级人| 超碰av人人做人人爽久久| 男插女下体视频免费在线播放| 亚洲美女搞黄在线观看 | 国产视频一区二区在线看| 国产午夜精品久久久久久一区二区三区 | 亚洲最大成人手机在线| 国内久久婷婷六月综合欲色啪| 亚洲va日本ⅴa欧美va伊人久久| 一个人看视频在线观看www免费| 国产精品综合久久久久久久免费| www日本黄色视频网| 国产精品一区二区三区四区久久| 免费看美女性在线毛片视频| 韩国av在线不卡| 观看免费一级毛片| 国产aⅴ精品一区二区三区波| 国产精品乱码一区二三区的特点| 两个人的视频大全免费| 国产精品,欧美在线| 男女视频在线观看网站免费| 亚洲欧美激情综合另类| 国产色婷婷99| 九九在线视频观看精品| 日本与韩国留学比较| 久久久久久久久大av| 长腿黑丝高跟| 欧美成人a在线观看| 亚洲av免费高清在线观看| 色av中文字幕| 免费人成在线观看视频色| 少妇高潮的动态图| 亚洲真实伦在线观看| 亚洲国产欧洲综合997久久,| 亚洲aⅴ乱码一区二区在线播放| 亚洲欧美日韩高清专用| 国产亚洲欧美98| 又爽又黄a免费视频| 国产免费男女视频| 国产爱豆传媒在线观看| 色吧在线观看| 最近视频中文字幕2019在线8| 在线免费十八禁| 成人综合一区亚洲| 亚洲中文字幕日韩| 亚洲成人中文字幕在线播放| 国产精品三级大全| 国产精品嫩草影院av在线观看 | 欧美最黄视频在线播放免费| 一级黄色大片毛片| 最好的美女福利视频网| 亚洲国产欧美人成| АⅤ资源中文在线天堂| 亚洲国产精品成人综合色| 给我免费播放毛片高清在线观看| 亚洲精品456在线播放app | 丰满人妻一区二区三区视频av| 免费大片18禁| 亚洲成人久久爱视频| 最近最新中文字幕大全电影3| 少妇熟女aⅴ在线视频| 大型黄色视频在线免费观看| 亚洲美女视频黄频| 舔av片在线| 最近在线观看免费完整版| 亚洲国产精品久久男人天堂| 成人永久免费在线观看视频| 成人毛片a级毛片在线播放| 国产高清激情床上av| 午夜亚洲福利在线播放| 很黄的视频免费| 99热精品在线国产| 亚洲色图av天堂| 日本色播在线视频| 在线观看美女被高潮喷水网站| 欧美色视频一区免费| 99久久成人亚洲精品观看| 日本-黄色视频高清免费观看| 国产精品三级大全| 精品久久久久久成人av| 动漫黄色视频在线观看| 成年女人永久免费观看视频| 亚洲avbb在线观看| 欧美区成人在线视频| 国产国拍精品亚洲av在线观看| 国产成人aa在线观看| 国产成人一区二区在线| 香蕉av资源在线| 亚洲最大成人手机在线| 午夜免费男女啪啪视频观看 | av视频在线观看入口| 国产在视频线在精品| 免费在线观看影片大全网站| 日本a在线网址| 午夜久久久久精精品| 一区二区三区高清视频在线| 深爱激情五月婷婷| 国产综合懂色| 日韩一本色道免费dvd| 美女xxoo啪啪120秒动态图| 99久国产av精品| 欧美日韩综合久久久久久 | 老女人水多毛片| 色哟哟哟哟哟哟| 乱人视频在线观看| 男女那种视频在线观看| 亚洲自拍偷在线| 午夜福利在线观看吧| 国产成人a区在线观看| 神马国产精品三级电影在线观看| 日韩欧美精品免费久久| 国产伦在线观看视频一区| 日本 av在线| 在现免费观看毛片| 国产精品一区二区三区四区免费观看 | 久久人人爽人人爽人人片va| 最新在线观看一区二区三区| 欧美区成人在线视频| 久久久久久国产a免费观看| 久久久国产成人精品二区| .国产精品久久| 亚洲内射少妇av| 久久草成人影院| 亚洲人成网站在线播放欧美日韩| 制服丝袜大香蕉在线| 亚洲美女搞黄在线观看 | 国产一区二区亚洲精品在线观看| 国产久久久一区二区三区| 免费在线观看影片大全网站| 性插视频无遮挡在线免费观看| 天堂网av新在线| 色在线成人网| 性色avwww在线观看| 国产av麻豆久久久久久久| 一级a爱片免费观看的视频| 成人无遮挡网站| 国产激情偷乱视频一区二区| 国产免费av片在线观看野外av| 乱系列少妇在线播放| 一个人看视频在线观看www免费| 亚洲av.av天堂| 18禁黄网站禁片午夜丰满| 少妇高潮的动态图| 日本一二三区视频观看| 亚洲va日本ⅴa欧美va伊人久久| 免费不卡的大黄色大毛片视频在线观看 | 成年免费大片在线观看| 欧美三级亚洲精品| 久久久久久久精品吃奶| 午夜免费男女啪啪视频观看 | www.www免费av| 成人精品一区二区免费| 99久久中文字幕三级久久日本| 别揉我奶头~嗯~啊~动态视频| 国产探花在线观看一区二区| 日韩欧美 国产精品| 久久久久久久午夜电影| 国内精品宾馆在线| 国产久久久一区二区三区| 伊人久久精品亚洲午夜| 一级a爱片免费观看的视频| 听说在线观看完整版免费高清| 美女 人体艺术 gogo| 在线国产一区二区在线| 男人舔奶头视频| 亚洲av一区综合| 成年免费大片在线观看| 欧美绝顶高潮抽搐喷水| 精品久久久久久久久亚洲 | 久久精品国产亚洲av涩爱 | 免费黄网站久久成人精品| 在线国产一区二区在线| 国产熟女欧美一区二区| 亚洲 国产 在线| av在线天堂中文字幕| 最新在线观看一区二区三区| 在线国产一区二区在线| 午夜福利在线在线| 综合色av麻豆| 精品午夜福利视频在线观看一区| 国产精品1区2区在线观看.| 高清日韩中文字幕在线| 午夜爱爱视频在线播放| 大又大粗又爽又黄少妇毛片口| 久久久午夜欧美精品| 岛国在线免费视频观看| 欧美成人免费av一区二区三区| 亚洲国产高清在线一区二区三| 国产91精品成人一区二区三区| 麻豆国产av国片精品| 亚洲午夜理论影院| 男女下面进入的视频免费午夜| 好男人在线观看高清免费视频| 波多野结衣巨乳人妻| 欧美+亚洲+日韩+国产| 日韩av在线大香蕉| 国产欧美日韩精品一区二区| 精华霜和精华液先用哪个| 色播亚洲综合网| 乱码一卡2卡4卡精品| 九九在线视频观看精品| 日本精品一区二区三区蜜桃| 色视频www国产| 一个人看的www免费观看视频| 国产精品久久久久久精品电影| 色精品久久人妻99蜜桃| 精品久久久久久成人av| 一区二区三区激情视频| 亚洲内射少妇av| 两人在一起打扑克的视频| 国产精品99久久久久久久久| 国产高清激情床上av| 一本一本综合久久| 99热精品在线国产| 久久人人精品亚洲av| 亚洲七黄色美女视频| 一卡2卡三卡四卡精品乱码亚洲| 亚洲人成网站在线播放欧美日韩| 国产男人的电影天堂91| 午夜激情欧美在线| 国产午夜精品论理片| 男女下面进入的视频免费午夜| 男人的好看免费观看在线视频| av专区在线播放| 亚洲乱码一区二区免费版| 亚洲黑人精品在线| 国产大屁股一区二区在线视频| 国内少妇人妻偷人精品xxx网站| 舔av片在线| 99热这里只有精品一区| 成人特级黄色片久久久久久久| 一级毛片久久久久久久久女| 尤物成人国产欧美一区二区三区| 大型黄色视频在线免费观看| 69人妻影院| 啦啦啦观看免费观看视频高清| 日本 av在线| 51国产日韩欧美| 天天一区二区日本电影三级| 色吧在线观看| 成人性生交大片免费视频hd| 国产精品不卡视频一区二区| 一a级毛片在线观看| www.www免费av| 免费一级毛片在线播放高清视频| 最好的美女福利视频网| 欧美成人一区二区免费高清观看| 小蜜桃在线观看免费完整版高清| 日韩欧美精品v在线| 中亚洲国语对白在线视频| 亚洲三级黄色毛片| 国产精品福利在线免费观看| 啪啪无遮挡十八禁网站| 真实男女啪啪啪动态图| 22中文网久久字幕| 一个人免费在线观看电影| 欧美另类亚洲清纯唯美| 久久香蕉精品热| 女生性感内裤真人,穿戴方法视频| 中出人妻视频一区二区| 亚洲电影在线观看av| 我要看日韩黄色一级片| 亚洲欧美精品综合久久99| 亚洲电影在线观看av| 综合色av麻豆| 久久精品国产亚洲av天美| 精品一区二区三区人妻视频| 中文字幕熟女人妻在线| 日韩欧美在线二视频| 亚洲成人久久性| 少妇裸体淫交视频免费看高清| 色尼玛亚洲综合影院| 成人毛片a级毛片在线播放| 国产伦人伦偷精品视频| 一本精品99久久精品77| 日韩精品青青久久久久久| 亚洲国产精品成人综合色| 中文字幕久久专区| 精品久久久久久,| 久久久久九九精品影院| 欧美高清成人免费视频www| 久久这里只有精品中国| av女优亚洲男人天堂| 丰满的人妻完整版| 别揉我奶头~嗯~啊~动态视频| 亚洲,欧美,日韩| 丰满的人妻完整版| 国产成人av教育| 又黄又爽又免费观看的视频| 亚洲精品色激情综合| 成人鲁丝片一二三区免费| 欧美区成人在线视频| 九九热线精品视视频播放| 直男gayav资源| 在线观看av片永久免费下载| 舔av片在线| 色尼玛亚洲综合影院| 日韩人妻高清精品专区| 亚洲黑人精品在线| 日韩一区二区视频免费看| 精品久久久久久久久亚洲 | 伦理电影大哥的女人| 国产黄a三级三级三级人| 亚洲人成网站在线播| 久久热精品热| 免费搜索国产男女视频| 99精品在免费线老司机午夜| 久久精品夜夜夜夜夜久久蜜豆| 午夜爱爱视频在线播放| h日本视频在线播放| 亚洲无线在线观看| 国内久久婷婷六月综合欲色啪| 久久欧美精品欧美久久欧美| videossex国产| 日韩在线高清观看一区二区三区 | 人妻少妇偷人精品九色| 又黄又爽又免费观看的视频| 精品久久国产蜜桃| 麻豆av噜噜一区二区三区| 国产精品亚洲美女久久久| 日韩欧美一区二区三区在线观看| x7x7x7水蜜桃| 人妻制服诱惑在线中文字幕| 国产男人的电影天堂91| 欧美极品一区二区三区四区| 亚洲成人久久性| 亚洲aⅴ乱码一区二区在线播放| 国产激情偷乱视频一区二区| 色综合亚洲欧美另类图片| 久99久视频精品免费| 成年版毛片免费区| 国产精品爽爽va在线观看网站| 黄色丝袜av网址大全| 婷婷精品国产亚洲av在线| 97热精品久久久久久| 最好的美女福利视频网| 国产精华一区二区三区| 熟妇人妻久久中文字幕3abv| 99在线人妻在线中文字幕| 色尼玛亚洲综合影院| 久久人人精品亚洲av| 国产成人影院久久av| 全区人妻精品视频| 国产精品嫩草影院av在线观看 | 性插视频无遮挡在线免费观看| .国产精品久久| 午夜老司机福利剧场| 在线天堂最新版资源| 毛片女人毛片| 亚洲一区二区三区色噜噜| av女优亚洲男人天堂| 日本免费一区二区三区高清不卡| 久久99热6这里只有精品| 精品人妻视频免费看| 免费看日本二区| 男人舔女人下体高潮全视频| 乱码一卡2卡4卡精品| 亚洲18禁久久av| 日日夜夜操网爽| 久久久久久大精品| 两性午夜刺激爽爽歪歪视频在线观看| 丝袜美腿在线中文| 在线观看免费视频日本深夜| 久9热在线精品视频| 一区二区三区高清视频在线| 久久99热6这里只有精品| 日本色播在线视频| 精品久久久久久久久久久久久| 99热只有精品国产| av专区在线播放| 日韩强制内射视频| 老司机深夜福利视频在线观看| 最近中文字幕高清免费大全6 | 国产精品女同一区二区软件 | 久久久色成人| 日韩一本色道免费dvd| 日本撒尿小便嘘嘘汇集6| 少妇人妻一区二区三区视频| 欧美高清性xxxxhd video| 久久久成人免费电影| 国产又黄又爽又无遮挡在线| 色哟哟·www| 久久久久久久久中文| 制服丝袜大香蕉在线| 老熟妇乱子伦视频在线观看| 男插女下体视频免费在线播放| 12—13女人毛片做爰片一| 一级黄片播放器| 国产午夜精品久久久久久一区二区三区 | 午夜精品久久久久久毛片777| 男人舔奶头视频| 亚洲精品456在线播放app | а√天堂www在线а√下载| 亚洲在线自拍视频| 久久久久久国产a免费观看| 久久久久免费精品人妻一区二区| av视频在线观看入口| 又紧又爽又黄一区二区| 真人一进一出gif抽搐免费| 97碰自拍视频| 欧美日韩亚洲国产一区二区在线观看| 亚洲av免费在线观看| 免费一级毛片在线播放高清视频| 国产伦人伦偷精品视频| 日本成人三级电影网站| 精品一区二区三区视频在线| 村上凉子中文字幕在线| 在线播放国产精品三级| 国产精品三级大全| 伊人久久精品亚洲午夜| 亚洲欧美清纯卡通| 97碰自拍视频| 亚洲久久久久久中文字幕| 日韩欧美三级三区| 国产亚洲精品综合一区在线观看| 亚洲黑人精品在线| 人妻夜夜爽99麻豆av| 在线看三级毛片| 男人舔女人下体高潮全视频| 亚洲欧美日韩高清在线视频| 我要搜黄色片| 亚洲中文字幕日韩| 很黄的视频免费| 最新中文字幕久久久久| 成人美女网站在线观看视频| 偷拍熟女少妇极品色| 黄色日韩在线| 国产美女午夜福利| 欧美bdsm另类| 成人美女网站在线观看视频| 国产精品女同一区二区软件 | 国产美女午夜福利| 精品久久久久久久久亚洲 | 极品教师在线免费播放| 免费无遮挡裸体视频| 1024手机看黄色片| 亚洲乱码一区二区免费版| 欧美国产日韩亚洲一区| 国国产精品蜜臀av免费| 少妇高潮的动态图| 国产淫片久久久久久久久| 日韩欧美一区二区三区在线观看| 成人欧美大片| 欧美zozozo另类| 不卡一级毛片| 欧美丝袜亚洲另类 | 观看免费一级毛片| 国产精品自产拍在线观看55亚洲| av天堂中文字幕网| 欧美xxxx性猛交bbbb| 婷婷色综合大香蕉| 日韩大尺度精品在线看网址| 精品免费久久久久久久清纯| 999久久久精品免费观看国产| 免费在线观看成人毛片| 国产高清激情床上av| 欧美潮喷喷水| 91久久精品国产一区二区成人| 日本-黄色视频高清免费观看| 国产视频内射| 色吧在线观看| 亚洲无线观看免费| 欧美xxxx黑人xx丫x性爽| 国产精品久久久久久久电影| 非洲黑人性xxxx精品又粗又长| 精品国产三级普通话版| 99久久精品一区二区三区| 欧美性感艳星| 天天一区二区日本电影三级| 成人午夜高清在线视频| 国内精品一区二区在线观看| 欧美一区二区亚洲| 中国美白少妇内射xxxbb| 亚洲最大成人av| 亚洲无线观看免费| 深爱激情五月婷婷| 少妇熟女aⅴ在线视频| 丝袜美腿在线中文| 亚洲经典国产精华液单| 亚洲精华国产精华精| 欧美区成人在线视频| 国产亚洲av嫩草精品影院| 两个人的视频大全免费| 一进一出抽搐动态| 婷婷精品国产亚洲av在线| 成人鲁丝片一二三区免费| 一级a爱片免费观看的视频| 欧美极品一区二区三区四区| 色综合亚洲欧美另类图片| 国产 一区 欧美 日韩| 免费不卡的大黄色大毛片视频在线观看 | 日韩欧美三级三区| 联通29元200g的流量卡| 91久久精品国产一区二区三区| 日韩欧美在线乱码| 美女xxoo啪啪120秒动态图| 看免费成人av毛片| 国产精品久久视频播放| 22中文网久久字幕| 直男gayav资源| 舔av片在线| 国产熟女欧美一区二区| 性欧美人与动物交配| 波多野结衣高清作品| 免费看光身美女| av女优亚洲男人天堂| 国内揄拍国产精品人妻在线| 真人做人爱边吃奶动态| 天天躁日日操中文字幕| 嫩草影视91久久| 天堂影院成人在线观看| 欧美一区二区亚洲| 麻豆久久精品国产亚洲av| 亚洲av电影不卡..在线观看| 成人特级av手机在线观看| 日本成人三级电影网站| 亚洲一区高清亚洲精品| 久久人人爽人人爽人人片va| 99久久九九国产精品国产免费| 狂野欧美白嫩少妇大欣赏| 亚洲综合色惰| 亚洲无线观看免费| 欧美色视频一区免费| 日韩精品有码人妻一区| 18+在线观看网站| 男插女下体视频免费在线播放| 搡老岳熟女国产| 日日摸夜夜添夜夜添小说| 欧美xxxx黑人xx丫x性爽| 欧美日本亚洲视频在线播放| 国产色婷婷99| 欧美黑人巨大hd| 人人妻人人澡欧美一区二区| 午夜精品一区二区三区免费看| 欧美在线一区亚洲| 日韩精品青青久久久久久| 真实男女啪啪啪动态图| 69人妻影院| 亚洲 国产 在线| 国产 一区 欧美 日韩| 日日干狠狠操夜夜爽| 欧美+亚洲+日韩+国产| 久久人妻av系列| 一夜夜www| 91午夜精品亚洲一区二区三区 | 综合色av麻豆| 久久精品91蜜桃| 精品99又大又爽又粗少妇毛片 | 亚洲av五月六月丁香网| 一级黄片播放器| 亚洲18禁久久av| h日本视频在线播放| 色视频www国产| 精品久久久久久久久av| h日本视频在线播放| 欧美丝袜亚洲另类 | 国产精品乱码一区二三区的特点| 国产成人aa在线观看| 国产精品免费一区二区三区在线| 久久久久国产精品人妻aⅴ院| 在线免费观看不下载黄p国产 | 韩国av一区二区三区四区| 男人和女人高潮做爰伦理| 草草在线视频免费看| 欧美绝顶高潮抽搐喷水| 三级毛片av免费|