Interpreting Multiple Regression: R-squared, Estimates, and Intervals, Lecture notes of Mathematical Statistics

Download Interpreting Multiple Regression: R-squared, Estimates, and Intervals and more Lecture notes Mathematical Statistics in PDF only on Docsity! Interpretation in Multiple Regression Topics: 1. R-squared and Adjusted R-squared 2. Interpretation of parameter estimates 3. Linear combinations of parameter estimates variance-covariance matrix standard errors of combinations standard error for the mean We will use the final model from last time to illustrate these concepts. Summaries of the model - least squares estimates with standard errors given below in parentheses:
logit  proportion    2.71  0.89  log  duration 0.57  I 0.38  .14  0.24 = 0.65 with 44 degrees of freedom R-squared = 0.6068029 R-squared and Adjusted R-squared: The R-squared value means that 61% of the variation in the logit of proportion of pollen removed can be explained by the regression on log duration and the group indicator variable. As R-squared values increase as we ass more variables to the model, the adjusted R-squared is often used to summarize the fit as it takes into account the the number of variables in the model. Adjusted R-squared = 1 - Mean Square Error /Total Mean Square where Mean Square Error is 2 from the regression model and the Total mean square is the sample variance of the response ( sY 2  2 is a good estimate if all the regression coefficients are 0). For this example, Adjusted R-squared = 1 - 0.65^2/ 1.034 = 0.59. Intercept: the intercept in a multiple regression model is the mean for the response when all of the explanatory variables take on the value 0. In this problem, this means that the dummy variable I = 0 (code = 1, which was the queen bumblebees) and log(duration) = 0, or duration is 1 second. For queenbumblebees, with visits of 1 second, we are 95% confident that the mean logit(proportion of pollen removed) is between  2.71  2.02  0.38 or between - 3.49 to - 1.93. The Student t quantile 2.02 is based on 44 degrees of freedom; qt(.975, 44). To convert back to the original units, we can take the inverse of the logit transformation. I.e. if logit(p) = log(p/(1-p)), then p = exp(x)/(1 + exp(x)). To get the confidence interval for the proportion just apply the inverse transformation. So for queen bumblebees with visits lasting 1 second, we are 95% confident that the mean proportion of pollen removed is between 0.03 and 0.13. [exp(-3.49)/(1 + exp(-3.49)) to exp(-1.93)/(1 + exp(- 1.93))] Note: while this is the interpretation of the intercept, we are extrapolating. Regression Coefficients: Typically the coefficient of a variable is interpreted as the change in the response based on a 1-unit change in the corresponding explanatory variable keeping all other variables held constant. In some problems, keeping all other variables held fixed is impossible (i.e. A quadratic model, or the model with different slopes for queen and worker bees). For this example, we have the estimated coefficient of log(duration) is 0.89. Because we have taken the log transformation of duration, the interpretation of the coefficient is easier to understand by looking at a doubling of duration (review page 208 chapter 8). A doubling of the duration of visit corresponds to a β1 log(2) change in the mean logit(proportion of pollen removed) or 0.89*log(2) = 0.62. The 95% confidence interval for β1 is 0.89  2.02  0.14 or 0.61 to 1.17. The interval under the doubling of duration is obtained by multiplying this interval by log(2). So a 95% confidence interval for the change in the mean logit(proportion pollen removed) is 0.42 to 0.81. Further simplification is not possible. Dummy variable coefficients: A 1 unit change for a dummy variable implies going from level 0 to level 1, so the the interpretation of the dummy variable coefficient is the amount by which the mean logit(proportion) for worker bees exceeds the mean logit(proportion) for queen bumble bees. i.e the logit of the proportion pollen removed for worker bees is 0.56 higher than the logit for queen bumble bees. A 95% confidence interval for the amount is 0.09 to 1.05. (this is the case for parallel regression lines; if we still had the interaction variable we could not make this statement, since the interaction of the dummy*log(duration) cannot be held constant). In the model derivation, we said that the intercept plus the dummy variable coefficient corresponded to the intercept for the worker bees, which is estimated as - 2.71 + .57 or - 2.14. This can be translated ac to the original scale as we did the intercept for the queen bumble bees. As this has a more interesting meaning, let's find a confidence interval for β0 + β2. To do this we need to find the standard error for a linear combination. Linear Combination of Parameters To find the variance (and then standard deviation) of the estimator of β0 + β2 we need to take into account the individual variances plus how the estimates will vary together from sample to sample (their covariance). The variance of the sum is the sum of the variances plus 2 times the covariance. We can get the covariance from the correlation of the estimates (recall the correlation is the covariance divided by the product of the standard deviations, so the covariance is the correlation times the product of the standard deviations. Since the standard deviations are unknown, we use the estimated covariance matrix calculated using the standard errors. In the Results options for Regression, check