Solution to Tutorial 7

Transcription

1 1. (a) We first fit the independence model ST3241 Categorical Data Analysis I Semester II, Solution to Tutorial 7 log µ ij = λ + λ X i + λ Y j, i = 1, 2, j = 1, 2. The parameter estimates are ˆλ = , ˆλ X 1 = , ˆλ X 2 = 0.0,ˆλ Y 1 = and ˆλ Y 2 = 0.0. The likelihood-ratio goodness-of-fit statistic (deviance) for the fitted model is, G 2 = with d.f. = 1 (p-value = ). Thus, the independence model is not a good fit. In this case, ˆλ Y 1 ˆλ Y 2 = , which is estimated log odds of myocardial infarction. In other words, estimated odds of myocardial infarction for either Placebo group or aspirin group is exp( ) = (b) For the saturated model, the parameter estimates are: ˆλ = , ˆλ X 1 = , ˆλ X 2 = 0.0, ˆλ Y 1 = , ˆλ Y 2 = 0.0, ˆλ 11 = , ˆλ 12 = ˆλ 21 = ˆλ 22 = 0.0. The estimated conditional log odds ratio between myocardial infarction and aspirin use is ˆλ 11 +ˆλ 22 ˆλ 12 ˆλ 21 = Therefore, the estimated odds ratio between myocardial infarction and aspirin use is which is the sample odds ratio. exp(0.6054) = = (a) The parameter estimates for the fitted (DV, DP, P V ) model are: ˆλ = , ˆλ D 1 = , ˆλ D 2 = 0, and (b) The fitted values are: ˆλ V 1 = , ˆλ V 2 = 0, ˆλ P 1 = , ˆλ P 2 = = , ˆλ P 11 V DP = , ˆλ 11 = Victims Defendants Death Penalty Race Race Yes No White White Black Black White Black

2 From the fitted values, estimated odds ratio between D and P when V = White is 0.42 and when V = Black, it is Since this is a homogeneous association model, the odds ratios are same at each level of V. Note that, this odds ratio can also be obtained as exp{ˆλ DP DP DP DP 11 + ˆλ 22 ˆλ 12 ˆλ 21 } = exp( ) = The value of this common odds ratio is smaller than 1 indicating that the odds of getting death penalty for a white defendant is smaller than that of a black defendant controlling for the victims race. (c) The fitted marginal table between Death Penalty verdict and defendant s race is Defendants Death Penalty Race Yes No White Black Note that, the marginal table from the fitted values is the same as the one obtained from the sample data. Therefore, the marginal odds ratio is It is greater than 1 indicating that the odds of getting a death penalty for whites are greater than that of blacks if we ignore the victims race. Thus, we observe a completely different type of association in the marginal table than the conditional association controlling for victims race. This illustrates Simpsons paradox. (d) The goodness-of-fit test statistic for this homogeneous association model is G 2 = with d.f. = 1 (p-value = ). Thus, we can conclude that the homogeneous association model is a good fit. (e) The parameter estimates of the fitted (DV, P V ) model are: and ˆλ = , ˆλ V 1 = , ˆλ V 2 = 0.0, ˆλ D 1 = , ˆλ D 2 = 0.0, ˆλ P 1 = , ˆλ P 2 = = , ˆλ P V 11 = This is the conditional independence model between D and P controlling for V. But the estimated conditional log odds ratio between D and V controlling for P is given by 11 = and the estimated conditional odds ratio between P and V controlling for D is given by ˆλ P 11 V = The likelihood ratio goodness-of-fit statistic for this model is, G 2 = with d.f. = 2 (p-value = ). Since the p-value is greater than 0.05, we can conclude that this model is also a good fit. The following table shows the adjusted residuals obtained from this model fit: Victims Defendants Death Penalty Race Race Yes No White White Black Black White Black

3 The adjusted residuals in the partial table for victims race white are little larger compared to those in the partial for victims race black. This suggests that the conditional independence model may not be a very good fit. (f) To test for D P partial association, we need to test the null hypothesis H 0 : λ DP 11 = λ DP 12 = λ DP 21 = λ DP 22 = 0. The likelihood ratio test statistics for this purpose is G 2 [(DV, P V ) (DV, P V, DP )] = G 2 (DV, P V ) G 2 (DV, P V, DP ) = = with d.f. = 2 1 = 1. It has a p-value , which is smaller than 0.05 and thus indicates that the null hypothesis can be rejected. That is, the partial association between D and P is significant. 3. Since {µ ij = nπ ij } satisfies the independence model in a I J table, π ij = π i+ π +j. Therefore, log µ ij = log n + log π i+ + log π +j. The loglinear independence model can be written as (a) From the first model, we can write log µ ij = λ + λ X i + λ Y j. log µ ia log µ ib = (log n + log π i+ + log π +a ) (log n + log π i+ + log π+b) Now from the second model, = log π +a log π +b = log(π +a /π +b ). log µ ia log µ ib = (λ + λ X i + λ Y a ) (λ + λ X i + λ Y b ) = λy a λ Y b. Thus, λ Y a λ Y b = log(π +a/π +b ). (b) All λ Y j = 0 implies that log(π +a /π +b ) = 0 or π +a = π +b for all a and b. Let all π +j = π, then from the relation J π +j = 1 j=1 we have π = 1/J. That is, π +j = 1/J for all j. 4. (a) We first fit the model (AD, DG) of conditional independence between gender and the admission controlling for the department: log µ ijk = λ + λ A i + λ G j + λ D k + λad ik The fitted values are given in the following table: + λ GD jk. 3

4 Whether admitted, male The deviance for the fitted model is with degrees of freedom 6. Thus, it has a p-value of , indicating that the model is a poor fit. Next we fit the homogeneous association model (AD, AG, DG): log µ ijk = λ + λ A i + λ G j + λ D k + λag ij The fitted values are given in the following table: Whether admitted, male + λ AD ik + λ GD jk The deviance for the fitted model is with degrees of freedom 5. Thus, it has a p-value of , indicating that the homogeneous association model is also a poor fit. (b) We compute the adjusted residuals for the model (AD, DG) and report them in the following table: Whether admitted, male We notice that the residuals are large for Department 1, but for other departments the residuals are small. So most likely department 1 is an outlier. (c) We again fit the model (AD, DG) after deleting the data for Department 1. We report the fitted values in the following table: 4

5 Whether admitted, male The deviance for the fitted model is with degrees of freedom 5. Thus, it has a p-value of , indicating that the model is a very good fit and we can conclude that there is no association between gender and admission in the departments. 5