This action might not be possible to undo. Are you sure you want to continue?
Statistical inference involves: (a) Estimation (b) Hypothesis Testing Both involve using sample statistics, say, population parameter (µ ).
, to make inferences about the
There are two types of estimates of population parameters: point estimate interval estimate A point estimate is a single number used as an estimator of a population parameter. The problem with using a single point (or value) is that it might be right or wrong. In fact, with a continuous random variable, the probability that X is equal to a particular value is zero. [P(X=#) = 0.] We will use an interval estimator. We say that the population parameter lies between two values. Problem – how wide should the interval be? That depends upon how much confidence you want in the estimate. For instance, say you wished to give a confidence interval for the mean income of a college graduate: You might have: 100% confidence 95% confidence 90% confidence 80% confidence 0% confidence that the mean income of a college grad is between $0 and $∞ $35,000 and $41,000 $36,000 and $40,000 $37,500 and $38,500 $38,000 (a point estimate)
The wider the interval, the greater the confidence you will have in it as containing the population parameter.
n (sample size) is a very small fraction of N (population size). we will use Zα rather than Za/2 . 3 . in most situations. ----------------------------------------------------------------------------- 40934634. [when n≥30. ----------------------------------------------------------------------------We will not have to worry about the finite population correction factor since. for the sake of simplicity.doc p. However.Confidence Interval Estimators To estimate μ. the α error should be split in half since we are constructing a two-sided confidence interval. we use: X ± Zα σ X X ± Zα σ n (1-α) confidence where we get Zα from the Z table. we use s as an estimator of σ] To be more precise.
the population mean would lie in intervals constructed by the same procedure (same n and same α).000 X at 95% confidence: 14.000 ± 1.doc p.804 ----------------------------- $14.196 contains the true population parameter. in 95 samples out of 100.EXAMPLE: Take-home salary of a NYC supermarket clerk n = 100 = $14. Thus. Or. it is not a random variable. it is incorrect to say that there is a 95% chance that the population mean will “fall” in this interval.804 to $14.96 14.000 σ = $1. μ. Remember – the population parameter (μ ) is fixed.196 Interpretation: We are 95% certain that the interval from $13. 1000 100 How many statisticians does it take to change a light bulb? [one plus or minus three] 40934634. 4 .000 ± 196 $13.
P = 0. Source: http://www-micro.Confidence Intervals (CI) Statistics means never having to say you're certain.96 SD) 99.doc p. 5 .7% of samples fall between ±3 SD There is less than a 1 in 20 chance of any sample falling outside ±2 SD (95% CI.html 40934634. In a normal distribution: • • • 68% of samples fall between ±1 SD 95% of samples fall between ±2 SD (actually + 1.msb. P = 0.01).uk/1010/1011-18.le.05) and less than a 1 in 100 chance of any sample falling outside ±3 SD (99% CI.ac.
920 $11.000 What is the true average take home pay of a part-time college student? (a) use a 95% confidence interval estimator.080 ↔ $18.000 ± 3.EXAMPLE: A researcher wishes to determine the average take-home pay of a part-time college student. (b) use a 99% confidence interval estimator.doc p.000 100 15. In this problem we use s as an unbiased estimator of σ: E(s) = σ σ= ∑ (X i =1 n i − µ)2 N s= ∑(X i =1 n i − X )2 n −1 95% Confidence Interval Estimator: X ± Zα s X X ± Zα s n 15.000 S = $20.000 ± 1. 6 .96 20 .920 40934634. He takes a sample of 100 students and finds: X = $15.
000 ± 2. there are two ways of reducing the size of the confidence interval: 1..doc p.150 $9. a more efficient sampling procedure (e. However. this depends on the variability of the population.(b) 99% confidence interval estimator: 15.000 ± 5. The researcher can control n.000 100 Note: for a given (1-α ) confidence interval. 2. Use a larger n.575 15. 7 . 40934634.150 20 .g. stratification) may help. Use a smaller s? Of course.850 ↔ $20.
Zα = 2.78 years 17.66 years 17.03 years 17.32. 8 .97 years ↔ 19.EXAMPLE: Average life of a GE refrigerator.22 years ↔ 18.645 18 ± 0.66 (e) 68% Confidence α = .78 (d) 90% Confidence α = .doc p.4 4 100 4 100 4 100 4 100 −∞ ↔ +∞ 16. Zα = ∞] (b) 99% Confidence α = .05.96 18 ± 1. Zα =1.01.575 18 ± 2.60 years ↔ 18.0 18 ± 1.0 18 ± 0.40 years 40934634. Zα = 1. n = 100 = 18 years s = 4 years X X ± Zα s n [n≥30] (a) 100% Confidence [α = 0.645 18 ± 1.575 18 ± 1. Zα = 1.03 (c) 95% Confidence α = .96 18 ± 0.10.34 years ↔ 18.
A sample size of 121 is sufficiently large.50 ± 1.5% -1.86 years The .doc p.96 1.96.96 to -1.96 14.EXAMPLE A company is interested in determining the average life of its watches (μ). 14. The Z value that is associated with 95% probability is +1. 9 . The 2 121 2 121 is the standard error of the mean. 40934634.5% The central limit theorem assures us that the sample means follow a normal distribution around the population mean when n is large.50 years s = 2. An employee randomly samples 121 watches and finds that X = 14.36 14.50 ± .14 years ←———→ 14.00 years We shall construct a 95% confidence interval about the mean.36 is the sampling error (this is also sometimes referred to as the margin of error).96 2. Thus. 2.
10 . 40934634.By the way.doc p. given the above confidence interval. would you suggest that the company run ads stating that their watches have an average life of 15+ years? This is the kind of question that we ask when we do hypothesis testing: We test a claim about the population parameter.
96 30.000 miles s = 2.000 100 50.000 miles Construct a 95% C.712.I.E of μ 50.50 $48.000 100 EXAMPLE: Average Income of a College Graduate n = 1.600 X = $50.E of μ 30.000 s = $20.I.000 ± 2. 392 miles 2.000 ± 392 29.287.608 miles ↔ 30.50 ↔ $51.EXAMPLE: Tire Manufacturer n = 100 X =30.000 ± 1.50 40934634.doc p.000 Construct a 99% C.575 20 .000 ± 1287. 11 .
There is a tradeoff. but the only facts available to estimate the true parameter are those provided by the sample. H1 : The alternative hypothesis. the hypothesis is rejected. Two types of error can occur: STATE OF NATURE H0 Is True DECISION H0 Is False β Error (Type II Error) GOOD Do Not Reject H0 Reject H0 GOOD α Error (Type I Error) An α error occurs if we reject H0 when it is true.HYPOTHESIS TESTING A hypothesis is made about the value of some parameter. If the statistic differs from the hypothesis made about the parameter. 40934634. H0 : The null hypothesis. This contains the hypothesized parameter value which will be compared with the sample value. the possibility of making a β error increases.doc p. This will be “accepted” only if H0 is rejected. a decision must be made as to whether or not this difference is significant. H0: μ = # As we try to reduce α . it cannot be rejected. If its. If not. 12 .
If we make it extremely difficult to convict criminals because we do not want to incarcerate any innocent people we will probably have a legal system in which no one gets convicted. On the other hand.doc p. 5%) but NOT zero. The company is willing to live with a few defects per 1. Unfortunately. complete certainty) but “beyond reasonable doubt.. i. with a lot of confidence. we can never have both: there is a tradeoff. this is a woman who is terrified of making the error of “acceptance”.e.” they are very likely to make the error of acceptance. Ideally. But.” Social: A woman wants to be sure that she marries the right person. Business: a company purchases chips for its computers.” She has a friend who is exactly the opposite. she will probably end up rejecting a large number of suitors who would make great husbands. if they are too liberal in what they accept and assume everything is “sampling error.e. if we make it very easy to convict. they may end up rejecting too many shipments.000 chips. In statistical terms. the error of accepting when she should reject.000. This is why our legal system does not require a guilty verdict to be “beyond a shadow of a doubt” (i. she will be making the error of “rejection. All of this is very similar (in fact exactly the same) as the problem we had earlier with confidence intervals. we would love a very narrow interval. practically. 40934634. She has thousands of criteria and if a suitor is missing even one trait on her list she will reject him. 13 . It purchases them in batches of 1. and therefore has virtually no criteria.To understand the tradeoff between the α and ß errors think of the following examples: Legal: our legal system understands this tradeoff. then we will have a legal system in which many innocent people end up behind bars.. She is likely to make the error of “acceptance” and very unlikely to make the error of rejection. Of course. Statisticians solve this problem by trying to limit the alpha error to some small value (say. How many defects? If it randomly samples 100 chips from each batch and rejects the entire shipment if there are ANY defects. She is terrified of not finding a husband.
985 ± 1.05 level.1 and 990.We can actually test a hypothesis using the confidence interval estimators we already learned. How likely is this? This is equivalent to testing at the .9 hours You are 95% sure that the true μ is somewhere between [“is covered by the interval”] 979. You find: X = 985 hours s = 30 hours Construct a 95% CIE and decide whether the company’s claim should be rejected or not.88 hours 979. Suppose. of 1. you take a sample of 100 batteries and test them.000 hours.000 hours REJECT THE CLAIM [H0] 40934634. EXAMPLE: The Duracell Battery Company claims that its new Bunnywabbit batteries have a life.doc p. Notice that if we were to standardize the sample average of 985 … Z= 985 −1000 30 100 = -15/3 = -5 … we would find that it is 5 standard deviations away from the mean.9 hours. 14 .96 30 100 hours 985 ± 5.000 hours H1: μ ≠ 1. on the average.1 hours ↔ 990. H0: μ = 1.
2. 15 .96 1.5% [EXPLAIN – region of “acceptance” and region of rejection] 40934634.doc p.5% -1.96 2.
If you recall. DRAW A PICTURE! Determine the actual value (computed value) of the test statistic. etc. i. Establish the critical value or values of the test statistic needed to reject H0. 4. alternative hypothesis. This suggests that there is a 5% chance that the interval does not contain the population mean. F. The α refers to the alpha error above. H0 is the null hypothesis and H1 is the 2. We are. we have made an error.e.. 3. (Normally. 6.05 or 0.doc p. in fact. Make a decision: Reject H0 or Do Not Reject H0.01 are used) Select the test statistic: e. significance level of 0. The sample evidence is converted into a Z-score – in other words. Z= X − µH σ/ n If n is large.g. it is standardized – using the hypothesized µ value. H0 is a hypothesis about the value of some parameter. t.Steps in Hypothesis Testing You test a hypothesis—an assertion or claim about a parameter by using sample evidence. Z. 1. 95% certain that the interval we created contains the population mean. we used Zα when we constructed confidence intervals. true. Specify the level of significance (α) to be used. 40934634. 16 . say.. s can be used in lieu of σ . Formulate H0 and H1. 5. χ 2. This level of significance tells you the probability of rejecting H0 when it is.
This is why the region of rejection is divided into two tails. 17 . 40934634.doc p.H0: μ = # With a “two-tail” hypothesis test. H1 contains two possibilities: μ > # OR μ < #. α is split into two and put in both tails. Note that the region of rejection always corresponds to H1.
EXAMPLE: A pharmaceutical company claims that each of its pills contains exactly 20. You sample 64 pills and find that X =20.00 mgs H1: µ ≠ 20.00 mgs 2.00 .80 64 = . and s = .00 milligrams of Cumidin (a blood thinner). H0: µ = 20.5% -1. ] The Z value of 5 is deep in the rejection region. 18 .96 2.80 mgs.50 − 20 .05 If we took the above data and constructed a 95% confidence interval: 95%.50 ± 1.1 0 = 5 [ .80 64 = .10 This is the standard error of the mean.5 0 . Should the company’s claim be rejected? Test at α = 0.05.5% Z= 20 .doc p.96 1. Confidence Interval = 20. reject H0 at p < . Therefore.96(.50 mgs.10) 40934634.
19 .304 mg ←———→ 20.05 (. When you do a two-tail test at an alpha of . you will be using the same Z-values.20.96. ± 1. Indeed.doc p. The hypothesis test in this example is called a Two-tail Test. because the region of rejection is split (equally) into the two tails of the distribution.05 significance level) you will come to exactly the same conclusions as when you construct a twosided 95% confidence interval.696 mg Note that 20. 40934634.00 mg is not in this interval.
The machine is supposed to deliver exactly 12.96 2. you will want to do a two-tail test. If it is not thick enough. The customer inserts two dollars and the coffee machine delivers exactly 12 ounces of premium coffee.00 centimeters it will not fit into the hole in which it is supposed to be inserted. Suppose you manufacture coffee machines. Here are some examples: Suppose a company makes a claim about the thickness of a bolt.00 ounces. The bolt must have a diameter of exactly of 10. 20 .2. Suppose a single pill is supposed to contain 200 mgs. If it is thicker than 10. If it delivers more than 12. Whether it is fast or slow. We will reject whether the bolt is too thick or too thin. the α error is split into two with α /2 going into each tail.00 ounces. if you buy a watch. Too much medication is a problem since 40934634.doc p.5% More About Two-Tail Tests: With a two tail test.00 ounces of coffee. the owner of the machine will be upset since it will affect his profit margins.00 centimeters. you want it to be accurate. you have a problem with it.96 1. customers using the machine will be cheated. If it delivers less than 12. For instance.5% -1. then it will be loose and cause problems. of a very powerful heart medication. you want it to be exact. When there are problems with either too much or too little.
21 . The pill must have exactly 200 mgs. of the medication to work.it will kill the patient. too little and it will not work and the patient dies. Too much and too little are both problems.doc p. 40934634.
If the sample evidence indicates an average life of less than 10 years. 40934634. This is done when you are only concerned with one side. For example. 22 . we have to test to make sure that we are not looking at sampling error. The entire α error is on the left (the “less than”) side. The problem is only in one direction. If the sample evidence indicates impurities of more than 100 PPM. The α error is on the right side. too much is a problem but too little is NOT a problem. say you marry someone who tells you that she is worth one million dollars. we have to test to make sure that we are not looking at sampling error.5% -1. It is doubtful that you will be upset if you find out she is actually worth ten million dollars. For example. or vice versa.645 One-Tail Tests: When we do a one-tail test.doc p. No customer will be upset if the peanut butter has 25 PPM of impurities. Here are some examples of one-tail tests: Suppose a company claims that its parts have a life of at least 10 years. The problem is only in one direction. Suppose a peanut butter company claims that its peanut butter has no more than 100 parts per million (PPM) of impurities (I do not think you want to know what has been found in peanut butter). the α error is put all into one tail of the probability distribution. No customer will be upset if the part lasts for 15 years.
05.50 = –6 [ 9 324 = .645 Z= 97 −100 9 324 = −3 .50 This is the standard error of the mean. Should the company’s claim be rejected? Test at α = 0.EXAMPLE A company claims that a box of its raisin bran cereal contains at least 100 raisins. An inspector working for the Federal Trade Commission (FTC) takes a random sample of 324 boxes of cereals and finds that: X = 97 raisins. ] The value of -6 is deep in the rejection region. Therefore.doc p. reject H0 at P < . and s = 9 raisins.05 40934634. 23 . H0: µ ≥ 100 H1: µ < 100 [Notice that H1 “points to” the region of rejection in the picture.] 5% -1.
If we took the above data and constructed a (two-sided) 95% confidence interval: 95%. In this course. Confidence Interval = 97 ± 1.02 raisins ←———→ 97. however. The hypothesis test is a one-tail test so we use a critical value of -1.98 raisins In this case. there is such a thing as one-sided confidence intervals but we will not deal with them in this course.50) 96.doc p. 24 . the confidence interval uses the Z-values of ± 1.96(.] 40934634.645. [It should be noted. we will always construct two-sided confidence intervals.96 since it is two-sided and the alpha must be divided up into the two tails.
05 and a one-tail test: 5% -1.5% α =. 25 .05 1.645 OR .96 2.α =.5% -1.96 1.645 40934634.05 and a two-tail test: 2.doc p.
26 .01 and a two-tail test: .33 40934634.005 2.doc p.01 and a one-tail test: .01 -2.005 -2.575 α =.33 OR .575 .α =.01 +2.
This action might not be possible to undo. Are you sure you want to continue?