Interpretation Comparing Response Probabilities
An alternative way of summarizing effects in cumulative logit models directly uses the cumulative probabilities for Y. For example, to describe the effect of a quantitative variable x, we could compare for a particular j at different values of x, such as the maximum and minimum values. To describe effects of a categorical predictor, we compare for different categories of that predictor. We can control for other quantitative variables in the model by setting them at their mean. We can control for other quantitative variables in the model by making the comparison at each combination of their values. When there are several qualitative variables, We could, instead, merely set them all at the means of their indicator variables, mimicking the treatment of quantitative control variables. Similarly, we can describe effects on the individual category probabilities .Using the lowest and highest categories of Y, we could report the maximum and minimum values of and over the set of predictor values ,reporting the values of the predictors that provide these extremes.
For example, consider the uniform association model ( 3.12 ) applied to Table 3.1.The estimated probability that astrology is judged to be “very scientific” decreases from 0.090 for those with less than a high school education to 0.015 for those with a graduate degree. The estimated probability that astrology is judged to be “not at all scientific” increases from 0.507 for those with less than a high school education to 0.867 for those with a graduate degree. Figure 3.5 portrays the estimated category probabilities on opinion about astrology at the five education levels. The height of the lowest portion ( darkly shaded ) of each bar is the estimated probability of response “ not at all scientific” ; the height of the medium bar is the cumulative probability for categories “ not at all scientific” and “sort of scientific,” so the middle portion of each bar portrays the estimated probability of response “sort of scientific.” The more lightly shaded top portion of each bar portrays the estimated probability of response “very scientific.” With multiple predictors, such a comparison can be made at the maximum and minimum values of each predictor, with the other predictors set at their means.
For a continuous predictor a comparison of probabilities at maximum and minimum values of is not resistant to outliers on . When severe outliers exist, it is often preferable to use the lower and upper quartiles of instead. A comparison of estimated probabilities at the quartiles summarizes an effect over the middle half of the data (on ) and is not affected by outliers. Alternatively, a standard approximation for the rate of change of a probability in the logistic regression model also applies with ordinal logit models. The instantaneous rate of change in as a function of explanatory variable , at fixed values for the other explanatory variables, is
For example, suppose that for the effect of = number of education in a particular application. Then, at predictor values such that , an increase of 1 in while keeping fixed the other predictors corresponds to approximately a 0.150(0.60)(0.40)=0.036 estimated increase in .
We have seen that interpretations can also focus on standardized effects for the conditional distribution of an underlying latent response variable Y*. Alternatively, standardized effects can refer to the marginal distribution of Y*, as is often done in ordinary regression. This is discussed in Section 5.1.3. Yet another type of interpretation focuses on standardizing other measures. For example, Joffe and Greenland ( 1995 ) showed how to convert estimated regression coefficients into estimates of standardized fitted probabilities, probability differences, and probability ratios.
น่าจะตอบสนองการเปรียบตีทางเลือกของการสรุปผลในแบบจำลอง logit สะสมโดยตรงใช้น่าจะสะสมสำหรับ Y เช่น เพื่ออธิบายผลของตัวแปรเชิงปริมาณ x เราอาจเปรียบเทียบสำหรับเจเฉพาะที่แตกต่างกันค่าของ x เช่นค่าสูงสุด และต่ำสุด อธิบายผลกระทบของ predictor แตก เราเปรียบเทียบสำหรับประเภทต่าง ๆ ของ predictor ที่ เราสามารถควบคุมตัวแปรอื่น ๆ เชิงปริมาณในรูปแบบ โดยการกำหนดที่หมายถึงการ เราสามารถควบคุมตัวแปรอื่น ๆ เชิงปริมาณในรูปแบบ โดยทำการเปรียบเทียบในแต่ละชุดของค่า เมื่อมีตัวแปรเชิงคุณภาพหลาย เราสามารถ แทน เพียงตั้งที่หมายถึงตัวแปรแสดงสถานะของพวกเขา นั่งรักษาตัวแปรควบคุมปริมาณได้ ในทำนองเดียวกัน เราสามารถอธิบายผลกระทบน่าจะแต่ละประเภท ใช้ประเภทต่ำสุด และสูงสุดของ Y เราสามารถรายงานค่าสูงสุด และต่ำสุดของ และ ผ่านชุดของค่า predictor รายงานค่าทำนายให้สุดขั้วเหล่านี้For example, consider the uniform association model ( 3.12 ) applied to Table 3.1.The estimated probability that astrology is judged to be “very scientific” decreases from 0.090 for those with less than a high school education to 0.015 for those with a graduate degree. The estimated probability that astrology is judged to be “not at all scientific” increases from 0.507 for those with less than a high school education to 0.867 for those with a graduate degree. Figure 3.5 portrays the estimated category probabilities on opinion about astrology at the five education levels. The height of the lowest portion ( darkly shaded ) of each bar is the estimated probability of response “ not at all scientific” ; the height of the medium bar is the cumulative probability for categories “ not at all scientific” and “sort of scientific,” so the middle portion of each bar portrays the estimated probability of response “sort of scientific.” The more lightly shaded top portion of each bar portrays the estimated probability of response “very scientific.” With multiple predictors, such a comparison can be made at the maximum and minimum values of each predictor, with the other predictors set at their means.For a continuous predictor a comparison of probabilities at maximum and minimum values of is not resistant to outliers on . When severe outliers exist, it is often preferable to use the lower and upper quartiles of instead. A comparison of estimated probabilities at the quartiles summarizes an effect over the middle half of the data (on ) and is not affected by outliers. Alternatively, a standard approximation for the rate of change of a probability in the logistic regression model also applies with ordinal logit models. The instantaneous rate of change in as a function of explanatory variable , at fixed values for the other explanatory variables, is For example, suppose that for the effect of = number of education in a particular application. Then, at predictor values such that , an increase of 1 in while keeping fixed the other predictors corresponds to approximately a 0.150(0.60)(0.40)=0.036 estimated increase in .We have seen that interpretations can also focus on standardized effects for the conditional distribution of an underlying latent response variable Y*. Alternatively, standardized effects can refer to the marginal distribution of Y*, as is often done in ordinary regression. This is discussed in Section 5.1.3. Yet another type of interpretation focuses on standardizing other measures. For example, Joffe and Greenland ( 1995 ) showed how to convert estimated regression coefficients into estimates of standardized fitted probabilities, probability differences, and probability ratios.
การแปล กรุณารอสักครู่..
