IDEAS THAT WE WILL NEED FROM CHAPTER 5 Chapter 5 is about the simultaneous behavior of two or more random variables. For example, we might be concerned with the probability that a randomly selected person will have a body mass index at most 35 AND a systolic blood pressure at most 140. Body mass index (Y1) is one random variable, and systolic blood pressure (Y2) is another. The values 35 and 140 are arbitrary. As we slide these values around, we can create a function called the joint (cumulative) distribution function: đšđ1, đ2 (y1,y2) = P(Y1 īŖ y1 AND Y2 īŖ y2) Much of chapter 5 is devoted to methods for calculating and manipulating these functions. Fortunately for us, we will only need to be able to calculate the means and standard deviations for linear combinations of multiple variables. That is, we will frequently be concerned with functions of the form W īŊ a1Y1 īĢ a2Y2 īĢ ... īĢ akYk īĢ b Suppose that each Yi has expected value denoted by īi . What can we say about E (W ) ? RESULT #1: E (W ) īŊ a1ī1 īĢ a2 ī2 īĢ ... īĢ ak īk īĢ b īŊ a1E (Y1 ) īĢ a2 E (Y2 ) īĢ ... īĢ ak E (Yk ) īĢ b While I am deliberately avoiding developing the machinery for a proof of this important result, it is not surprising that expectations distribute over linear operations on multiple variables in the same way that they distribute over linear functions of single variables. Example 1. You are scoring college applicants on the basis of their High School GPA (Y1) and their total SAT score (Y2). The scoring function you use is S īŊ 10Y1 īĢ Y2 / 40 . If applicants have a mean High School GPA of 3.1 and a mean total SAT of 1100, what is the mean for the values of S? The mean is 10*3.1 + 1100/40 = 58.5 Example 2. The mean time to service a rental car when it is returned is 25 minutes. 10 cars are returned during the day. What is the expected total time to service all of them? Intuitively, you know that this must be 25*10 = 250 minutes. Formally, let Y1 be the time to service the first car, Y2 the time to service the second, etc. total īŊ T īŊ Y1 īĢ Y2 īĢ ... īĢ Y10 Then E(T) = 25 + 25 + ….+ 25 = 10*25 = 250 In addition to getting the mean, we also need to know the standard deviation. This works out reasonably easily IF we can assume the variables are independent. Definition: Let the random variable đ1 have c.d.f. đšđ1 , and let the random variable đ2 have c.d.f. đšđ2 . Let the two random variables together have the joint c.d.f. đšđ1 ,đ2 . The two random variables are said to be independent if, for all đĻ1 đđđ đĻ2 , we have đšđ1 ,đ2 (đĻ1 , đĻ2 ) = đšđ1 (đĻ1 )đšđ2 (đĻ2 ). If đ1 and đ2 are not independent, they are said to be dependent. Note: If the two random variables are continuous, then independence also implies that the joint p.d.f. is factorable into the product of the individual (marginal) p.d.f.’s. If the two random variables are discrete, then independence implies that the joint p.m.f. is factorable into the product of the individual (marginal) p.m.f.’s. 1 Example: Consider the joint p.d.f. for two continuous r.v.’s đ1 and đ2 given by 1 1 đđ1 ,đ2 (đĻ1 , đĻ2 ) = 2đđ2 đđĨđ {− 2đ2 [(đĻ1 − đ1 )2 + (đĻ2 − đ2 )2 ]} , for −∞ < đĻ1 < +∞, and −∞ < đĻ2 < +∞. This joint p.d.f. may be factored into a product of two normal p.d.f.’s with different means but the same standard deviation. Thus the two r.v.’s đ1 and đ2 are independent. Example: p. 253, 5.52 Example: Consider the joint p.m.f. for two discrete r.v.’s đ1 and đ2 given by 1 đđ1 ,đ2 (đĻ1 , đĻ2 ) = 36, for (đĻ1 , đĻ2 ) ∈ {1, 2, 3, 4, 5, 6}, or đđ1 ,đ2 (đĻ1 , đĻ2 ) = 0, otherwise. This joint p.m.f. may be factored into the product of 1 đđ1 (đĻ1 ) = 6, for (đĻ1 ) ∈ {1, 2, 3, 4, 5, 6}, or đđ1 , (đĻ1 ) = 0, otherwise; and 1 đđ2 (đĻ2 ) = 6, for (đĻ2 ) ∈ {1, 2, 3, 4, 5, 6}, or đđ2 , (đĻ2 ) = 0, otherwise. Thus the two r.v.’s đ1 and đ2 are independent. Example: p. 251, 5.45 The most important consequence of independence is that the variances add in a sensible way. If the variables were not independent, then the variance of the linear combination would also include terms involving covariances of pairs of the Y’s. RESULT #2: If the variables are independent, V (W ) īŊ a12V (Y1 ) īĢ a22V (Y2 ) īĢ .... īĢ ak2V (Yk ) . Notice that the additive constant, the ‘b’ term, disappears, just as it did for the single variable case in chapters 3 and 4. Important Note: Random sampling implies that the measurements taken on separate members of the random sample will be independent random variables. Whether or not variables are independent can be an interesting question. For example, we are fairly sure that body mass index is NOT independent of blood pressure – those people with extremely high BMI are more likely to have high blood pressures. Similarly, High School GPA is not independent of total SAT. Notice that this does not mean that those with high HS GPA must have high total SAT, but only that they are more likely to have high total SAT. On the other hand, the HS GPA’s of two randomly selected applicants are independent because the results of the first applicant will not give any information about the results of the second applicant. Example 2, continued. Suppose that the standard deviation for a single car’s service time is 6 minutes. Also, assume it is reasonable that the times for different cars are independent. a. What is the standard deviation for T = total time to service the 10 cars? Solution: Each car has a time with variance īŗ 2 īŊ 62 īŊ 36 . Hence, the variance for T is V (T ) īŊ V (Y1 ) īĢ V (Y2 ) īĢ ... īĢ V (Y10 ) īŊ 10 ī´ 36 īŊ 360 and the standard deviation is 360 īŊ 18.97 2 b. Suppose that T īŊ T /10 is the mean time in the sample of 10 cars. What is the expected value and standard deviation for T ? Using the basic results from chapters 3 and 4, we know E (T ) īŊ E (T ) /10 īŊ 25 And V (T ) īŊ V (T ) /102 īŊ 3.6 The standard deviation for T is 3.6 īŊ 1.897 It is extremely important to notice that the standard deviation for T was much smaller than the standard deviation for any single car. This fundamental property of averages is the principle that makes statistical inference possible! Problem A. Suppose that the time between service calls on a copy machine is exponentially distributed with a mean īĸ . Let T be the sample mean for records of 36 independent times. a. Express E (T ) and V (T ) as expressions using īĸ . b. Suppose the sample mean is 20.8 days. Provide a common sense estimate of īĸ . Assuming that the distribution of T is nearly bell-shaped, is it likely that the true value of īĸ differs from your estimate by more than 10 days? The behavior of the the sample mean ( T ) is of great concern to people doing practical statistics. Does it provide a stable estimate of the population mean ī? Results #1 and #2 yield the following: RESULT #3. Suppose that Y1 , Y2 ,..., Yn are independent random variables all having the same mean ī and variance īŗ 2 . Then Y has mean ī and variance īŗ 2 / n . proof: Using result #1, E (Y ) īŊ E (Y1 / n īĢ Y2 / n īĢ ... īĢ Yn / n) īŊ ī / n īĢ ī / n īĢ ... īĢ ī / n īŊ ī Using result #2 and the assumption of independence, V (Y ) īŊ V (Y1 / n īĢ Y2 / n īĢ ... īĢ Yn / n) īŊ īŗ 2 / n2 īĢ īŗ 2 / n2 īĢ .... īĢ īŗ 2 / n2 īŊ īŗ 2 / n Notice that we only needed independence for the variance calculation. MOMENT GENERATING FUNCTIONS (This is actually covered in Chapter 6.) Another useful property of independent random variables is that the expectation distributes over PRODUCTS. Notice that expectations always distribute over +/-, but it takes the special case of independent to allow them to distribute over multiplication! RESULT #4: If the variables are independent, then E ( g1 (Y1 ) ī´ g2 (Y2 ) ī´ ... ī´ gk (Yk )) īŊ E( g1 (Y1 )) ī´ E( g2 (Y2 )) ī´ ... ī´ E( g k (Yk )) There are many useful consequences of this, and we will mostly be concerned with the implications for the moment generating function of the linear combination W. E(etW ) īŊ E(exp īģīĨ a jYj īĢ bīŊ) īŊ E(eta1Y1 ī´ eta2Y2 ī´ ... ī´ etakYk ī´ etB ) īŊ etb ī E (e ta jY j ) īŊ etb ī m j (a jt ) 3 That is, the moment generating functions will multiply! Example 3. Suppose that Y1 , Y2 ,..., Yn are independent random variables having the exponential distribution with mean īĸ . a. What is the distribution of T īŊ Y1 īĢ Y2 īĢ ... īĢ Yn ? b. What is the distribution of T īŊ T / n ? Example 4. Suppose that Y1 is a binomial variable with parameters n1 , p and Y2 is an independent binomial variable with parameters n2 , p . Note that the values of p are the same, but not necessarily the values of n . What is the distribution of T īŊ Y1 īĢ Y2 ? Would the same result hold if the values of p were allowed to differ? Problem B. Suppose that Y1 is a normally distributed variable with parameters ī1 , īŗ12 and Y2 is an independent normally distributed variable with parameters ī2 , īŗ 22 . What is the distribution of T īŊ Y1 īĢ Y2 ? 4