Building on the logic of hypothesis testing T-tests

Building on the logic of hypothesis testing: T-tests ______________________________________________ 1) Introduce the t-test and explain when it should be used 2) Define Directional Hypotheses (one-tailed t-tests) and contrast them with ‘Non-Directional Hypotheses’ (two-tailed t-tests) that were described in Chapter 8. 3) Learn how to find tcrit for directional hypotheses 4) Highlight a second method for making decisions regarding the null hypothesis: the p-value method 5) Learn to calculate a p-value and review important points about using this method 6) Demonstrate one measure for estimating whether an experimental effect is large or small (Cohen’s D) 7) Advanced Topic: Calculating  and the Power associated with a hypothesis test 8) Outline the steps for conducting Hypothesis Tests using SPSS The Animal Cracker Packer ______________________________________________ Abby is the manager of an animal cracker factory. She is concerned that her aging cracker packer might need to be replaced. Each bag of crackers is supposed to weigh 454 grams. Abby would like to conduct a hypothesis test but she faces one big obstacle: she does not know the population variability (σ), which means she cannot use the formula for zobs that we learned in Chapter 8. Q: Have we ever faced a similar problem? A: Yes; when we wanted to calculate confidence intervals and we did not know σ. Q: What did we do then? A: Substituted s for σ. Q: Is that all? A: Yes, no wait. We also substituted t for z. Q: Is that what Abby should do now? A: You betcha. Why do you think this chapter is called t-tests? Limitations on hypothesis testing with z-scores ______________________________________________ 1. Our sample must be large (> 30) in order for the Central Limit Theorem to ‘kick in’. 2. Even if our sample is not large, we can still be OK as long as our population is distributed normally. • These limitations can be overcome, simply by collecting more data, if necessary. However… 3. σ must be known. a. This is a more serious limitation because it is almost impossible to know σ in the real world. Because most hypothesis tests use t as the test statistic, statisticians generally refer to hypothesis tests as t-tests. What do we do if σ is unknown? ______________________________________________ Just like we did with confidence intervals: 1. Use s as an estimate of σ 2. Use t as our test statistic instead of z The basic steps for conducting t-tests 1. Determine the value for tcrit 2. Calculate tobs 3. Compare tobs with tcrit Step 1: Determining tcrit ______________________________________________ tcrit depends on the degrees of freedom; df = n-1 If α = .05, and n = 25, tcrit = t(.05, 24) = Step 2: Calculating tobs ______________________________________________ zobs tobs The only difference is that s replaces σ (of course if you know σ, just go ahead and use σ). ______________________________________________ Step 3: Comparing tobs with tcrit ______________________________________________ Exactly the same as before. Steps for completing a t-test ______________________________________________ 1) 2) 3) 4) Specify the NULL hypothesis (HO) Specify the ALTERNATIVE hypothesis (HA) Designate the rejection region by selecting . Determine the critical value of your test statistic o Use appropriate degrees of freedom 5) Use sample statistics to calculate test statistic. o tobs = 6) Compare observed value with critical value: o If test statistic falls in RR, we reject the null. o Otherwise, we fail to reject the null. 7) Interpret your decision regarding the null o What do your data imply regarding the question that motivated your experiment? Abby Lyons’ Animal Cracker Packer ______________________________________________ Abby samples the next 25 cracker bags packed by the machine to determine whether it is putting 454 g of crackers in each bag. The sample statistics are as follows: M = 462.4 g; s = 16 g. Is the machine properly filling cracker packages? Set  = .05. Step 1: Ho:  = 454 g Step 2: Ha:   454 g Step 3: α = .05 Step 4: tcrit( = .05, df = 24) = ±2.064. Step 5: tobs = = = 8.4 / 3.2 = 2.625 Step 6: Because the tobs falls in the rejection region, we would reject the null. t (24) = 2.625, SEM = 3.2. Step 7: Comparing tobs and tcrit: The Cracker Packer ______________________________________________ Oooh! She's angry! ______________________________________ My statistical prowess has landed me a coveted internship working for 'The Crocodile Hunter'! Steve (he let's me call him Steve) has found a new group of crocs living at a golf course in Perth. He wants to know how the length of an average adult in this group compares to other adults in the area. Owing to his many years of experience with crocs, Steve knows that  = 21 feet. Because catching and measuring crocs is dangerous – even for the Crocodile Hunter – Steve is only able to capture and measure five crocs (M = 24, s = 1.6); I was VERY far away at the time. Does this sample provide enough evidence to reject the null hypothesis? Assume  = .05. Non-Directional vs. Directional Hypothesis Tests ______________________________________________ Non-directional test – a hypothesis test in which observing a sample mean in either the upper or lower tail of the sampling distribution would be theoretically meaningful (a/k/a two-tailed tests) • Null and alternative hypotheses Ho:  = some value Ha:   some value • Rejection region split between two tails: α/2 in upper tail, α/2 in lower tail. Directional test – a hypothesis test in which observing a sample mean in only one of the two tails of the sampling distribution would be theoretically meaningful (a/k/a onetailed tests) • Null and alternative hypotheses Two possibilities… • Rejection region located entirely in one tail; EITHER α in upper tail OR α in lower tail. Directional Hypothesis Tests: What happens at the drive-thru? ______________________________________________ Let’s pretend that Abby leaves the cracker packer factory for the fast-paced, high-pay, take-no-prisoners life of a fast food restaurant manager. Abby is considering whether or not to buy a new intercom system for the drive-thru, so she rents one to test it out. What if Abby rented the new intercom to reduce errors? If the new intercom… …Abby would … Increased errors keep the old intercom Did not change the number of errors Decreased errors keep the old intercom buy the new intercom What if Abby rented the new intercom to increase sales? If the new intercom… …Abby would … Increased sales buy the new intercom Did not change sales keep the old intercom Decreased sales keep the old intercom Statistical Hypotheses and the Rejection Region for Directional tests: Decreasing Errors at the Drive-Thru ______________________________________________ Decreasing errors: the only meaningful result would be if the sample mean fell at the extreme low end of the sampling distribution. In this case: Ho:   some value Ha:  < some value And the rejection region would look like this: Statistical Hypotheses and the Rejection Region for Directional tests: Increasing Sales at the Drive-Thru ______________________________________________ Increasing sales: the only meaningful result would be if the sample mean falls at the extreme high end of the sampling distribution. In this case: Ho:  ≤ some value Ha:  > some value And the rejection region would look like this: Critical values for One- and Two-Tailed Tests with the same  ______________________________________  Level  = .10  = .05  = .01 Alternative Hypotheses One-Tailed Two-Tailed (n=25) (n=25) Lower-Tail Upper-Tail t < -1.318 t > 1.318 t < -1.711 or t > 1.711 t < -1.711 t > 1.711 t < -2.064 or t > 2.064 t < -2.492 t > 2.492 t < -2.797 or t > 2.797 Are you more likely to reject the null for a one-tailed test or a two-tailed test? How do you decide which one to do? Choosing between one- and two-tailed tests ______________________________________________ By default, you should run a two-tailed test. You should only run a one-tailed test if your research question demands it; that is, if there is no practical or theoretical significance to a result in one tail of the sampling distribution. Cracker Packer Example: One-Tailed Test (RR in upper tail) Ho:   454 Ha:  > 454 Problem? OR One-Tailed Test (RR in lower tail) Ho:   454 Ha:  < 454 Problem? OR Two-Tailed Test Ho:  = 454 Ha:   454 Problem? Finding tcrit for Directional Hypotheses ______________________________________________ Steps for conducting t-tests: including directional tests ______________________________________________ 1) Decide whether you are conducting a one- or a twotailed test. 2) Specify the NULL hypothesis (HO) o 2-tailed: µ = some value; o 1-tailed: µ ≤ or ≥ some value 3) Specify the ALTERNATIVE hypothesis (HA) o 2-tailed: µ ≠ some value o 1-tailed: u > or < some value 4) Designate the rejection region by selecting . 5) Determine the critical value of your test statistic (remember to use appropriate df) o 2-tailed: /2 in the tail o 1-tailed: α in the tail 6) Use sample statistics to calculate test statistic. o tobs = 7) Compare observed value with critical value: o If test statistic falls in RR, we reject the null. o Otherwise, we fail to reject the null. 8) Interpret your decision regarding the null o What do your data imply regarding the question that motivated your experiment? Comparing the results of one- and two-tailed t-tests ______________________________________________ Big Bad Lou as a One-Tailed Test You lost a lot of money at the track and were forced to become the personal statistician of notorious underworld crime boss “Big Lou”. Big Lou wants to know if his son “Moderately-Sized Lou” is stealing from his gambling operation. Before Lou, Jr. took over the operation, it used to gross $3500 per night (µ). “Big Lou tells you, “I don’t care if he is grossing more than $3500, I only care if he’s grossing less. Got it?!” At this point, you could give Big Lou a big lecture regarding the theoretical considerations that guide the choice between one- and two-tailed tests, but I would not be so bold; in these matters, you should let your conscience be your guide. You sample the gross earnings of the casino over the next 25 nights. The average of the sample is $3338; s = 450. Will “Moderately-Sized Lou” be sleeping with the fishes if we set  = .05? Step 1: Big Lou has asked us to conduct a one-tailed test with the entire rejection region in the lower tail. Thus, our null and alternative hypotheses will be as follows: Step 2: Ho:   3500 Step 3: Ha:  < 3500 Big Bad Lou as a One-Tailed Test: continued ______________________________________________ Step 4:  = .05 Step 5: tcrit (α=.05, df = 24; 1-tailed) = -1.711. Step 6: tobs = = M − 0 3338 − 3500 450 / 25 s/ n = -162 / 90 = - 1.8 Step 7: Our observed t falls in the rejection region. Therefore, we would REJECT the null: t (24) = -1.8, SEM = 90. Step 8: Big Bad Lou as a Two-Tailed Test ______________________________________________ Although it would be unwise for you to challenge Big Lou’s decision to run a one-tailed test, the same is not true for Mrs. Lou (who really wears the pants in this underworld family). She loves her baby boy and wisely asks you to conduct a two-tailed test, just to see what would happen. After all, wouldn’t Lou Jr. deserve a big raise if receipts from the gambling operation increased rather than decreased? Bear in mind that, just like with selecting a value for α, the time to make a decision regarding whether to run a one- or two-sampled test is BEFORE you have seen the data. Step 1: Because we have decided to conduct a two-tailed test, our statistical hypotheses would be as follows: Step 2: Ho:  = 3500 Step 3: Ha:  ≠ 3500 Step 4:  = .05 Step 5: tcrit(α=.05, df = 24, 2-tailed) = -2.064. Step 6: The observed value of our test statistic does not change: tobs = -1.8 Big Lou as a 2-tailed test: continued _____________________________________________ Step 7: Our observed t DOES NOT fall in the rejection region. Therefore, we would FAIL TO REJECT the null: t (24) = -1.8, SEM = 90. Step 8: ______________________________________________ Ethics: Be judicious when making a choice. Choose before you see your data! Problems with the critical value method ______________________________________________ Critical value method: Decisions about the null hypothesis are rendered by comparing the observed value of a test statistic with a critical value. Problems with the critical value method: • All-or-none decision Do drivers in Cambridge, MA run more red lights than average? Sample tcrit tobs Mean Professor 4.4 2.093 2.0930000000000000001 Click Professor 4.3 2.093 2.0929999999999999999 Clack Introducing the p-value method ______________________________________________ Observed Significance level (p-value) – the probability (assuming Ho is true) of observing a sample mean that is at least as EXTREME as the observed sample mean. P-value method. Decisions about the null are made by calculating the probability of observing a sample mean at least as far from µ0 as our sample mean (i.e., the p-value), and comparing that probability with α. Using the p-value method: Big Lou Question ______________________________________________ 1. For a one-tailed test: a. Find the area in the tail beyond zobs. b. If the p-value is i. Less than α, reject the null. ii. Greater than α, fail to reject the null Big Lou question: zobs = -1.80, a. Area in the tail = .0359; that is the p-value. b. We would: Reject the null if α = .10 Reject the null if α = .05 Fail to reject the null if α = .01 2. For a two-tailed test: a. Double the area in the tail beyond zobs. Why? b. Use the same decision rule Big Lou question: zobs = -1.80, a. .0359 x 2 = .0718 b. We would: Reject the null if α = .10 Fail to reject the null if α = .05 Fail to reject the null if α = .01 Important points about the p-value method ______________________________________________ 1) The p-value method and the critical value method will always produce the same decision regarding the null. So, you only need to do one. 2) The p-value method gives us more information than the critical value method, in that it tells us where our sample mean fell in the sampling distribution. • marginally significant results • resolving Click and Clack’s debate 3) The p-value does not imply something about how likely your experimental result is to replicate. 4) A lower p-value does not imply a ‘bigger effect’ even though – all things being equal – a lower p-value implies a larger difference between M and µ0.BO • What is the easiest way to increase the p-value of an experimental result? ______________________________________________ BO = beatable offense Effect sizes - Cohen’s D ______________________________________________ Effect size – a statistical procedure for determining the magnitude of an experimental manipulation. Cohen’s D = d  d = the difference between the sample mean and the expected value of µ. M M Cohen’s d 0 < d < .2 .2 < d < .8 d > .8 Evaluation Small Medium Large Effect size for the ‘Big Lou’ question: Cohen’s d = d s M = 3338 − 3500 = 162 450 450 = .36 This represents a medium-sized effect. What that means for Moderately-sized Lou depends on how Big Lou interprets moderate. Steps for Conducting a Hypothesis Test: Final ______________________________________________ 1) Decide whether you are conducting a one- or a twotailed test. 2) Specify the NULL hypothesis (HO) o 2-tailed: µ = some value o 1-tailed: µ ≤ or ≥ some value 3) Specify the ALTERNATIVE hypothesis (HA) o 2-tailed: µ ≠ some value o 1-tailed: u > or < some value 4) Designate the rejection region by selecting . 5) Determine the critical value of your test statistic o 2-tailed: /2 in the tail; 1-tailed: a in the tail 6) Use sample statistics to calculate test statistic. o tobs = M −  s / n 0 7) Make a decision regarding the null and report the result using proper notation: o Critical Value method: • Reject the null if |zobs| > zcrit • Other wise, fail to reject the null o P-value method: • Reject the null if p <  • Fail to reject the null if p >  o Calculate and interpret Cohen’s d (optional) 8) Interpret your decision regarding the null in terms of your original research question. Little Rascals Example ______________________________________ The yield of alfalfa from six test plots is as follows: 2.1 2.5 1.8 2.6 2.9 1.9 tons per acre. The test plots were selected randomly sampled and that the population distribution of alfalfa yields is normal. Conduct a hypothesis test to determine how the yield of these test sites compared with the average yield of 1.5 tons per acre? (Hint: s = .434). Set α = .01. Advanced Topic: Calculating  and Power ______________________________________ 1) We assume that the null is false. 2) Calculate the minimum (maximum) value needed to reject the null. 3) Because H0 is false, we must (somewhat arbitrarily) choose a value for . 4) Calculate the area that falls below (above) the value derived in Step 2. Calculating : Under the Bridge Example ______________________________________ The person in charge of traffic management is concerned that too many cars are using the Coolidge Bridge (bridge to Northampton). The bridge was only designed to carry 900 cars per day. A sample of the number of cars that use the bridge in a 90-day period yielded an average of 980 cars with a standard deviation of 350. Do these data suggest that traffic on the bridge exceeds capacity? Calculate  and determine the power of the test. Ho:  = 900 Ha:   900 Critical z =  1.96 Observed z = = = REJECT the null. 80 / 36.89 = 2.17 Calculating : Under the Bridge Example ______________________________________ 1) What is the minimum value needed to reject H0 if it is true? Min Value = 900 + 1.96 (350 / 90) 900 + 1.96*36.89 900 + 72.31 = 972.31 2) Calculating : What is the probability of obtaining a score of 972.31 or lower if H0 is false? Must assume a value for . What would you recommend? z = = -7.69 / 36.89 Area(Tail -.21) = Power = 1 -  = -.21 .4168 = 1 - .4168 Not too bad; .60s is considered good. = .5832