Managerial Psychology Laboratory T Test and ANOVA Notes

Managerial Psychology Recitation 8: T test and ANOVA Statistics does all kinds of stuff to describe data Talk about baseball, other useful stuff We can calculate the probability. Good for gambling Build bridge, likely to collapse? Calculate probabilities Another tool is statistical inference – cover today Designing models – not discuss Imagine observe set of things that co-occur People work for you – tall, short, M, F, MIT, Harvard, ME, CS, vary all kinds of things Also Observe how good they are, See how well perform Predict new person – how will they perform If prediction mechE vs CS makes 12% better employees? Can find out all these things with data Very hard to intuit who is better/worse Statistics is very useful because it describes very accurately this structural aspect of data Can try to understand what is the structure of the data Statistical inference Question we ask: we have sample, some info, we ask, are these things different from each other for real? Hypothesis testing T-test, F-test Say Females in this class 8 average on niceness scale, males 7 Are they different? Yes, everything in life is different, even 8, 8.1 different What can we conclude? -not a lot – don’t know scale Why can’t infer? -error might be big enough to encompass both No doubt 8 bigger than 7 Question - is 8 more than 7 in meaningful way? Or just tiny, meaningless way Is f nicer than m in this class, or can we infer to rest of world? Other example that don’t have variance due to error Trying to estimate, adjust for variance, ask can we make inferences Could be biased sample. What does that mean? Could be this class, one person comes has all these nice friends, gets them to come. Could be all kinds of stories. If self selection, inferential statistics can’t help us All things talked about earlier in semester, outside range of inferential statistics Assumes sample is correct Test if valid like or not Stochastic world, lot of random error, come from all kinds of things -bad survey technique -bad random sampling -coincidence Lot of stuff happening all the time All things happen we can’t record Weather could have diff effect Bad grade yes, good grade yes Because we can’t make everything, error All kinds of reasons we expect variance All inferential statistics about variance The size of the mean difference is the same? (slide) Blue – different Red – much more cautious to say these are really different Find by taking mean difference and variance Imagine story about women, men Almost no overlap, then can say this class, women nicer than men, then think about applying it to the rest of the world In science, we can never prove anything Cannot go everywhere and find out, prove always true Why? Or can we? -can’t say 100% probability, don’t know all the possibilities, evidence We are limited, we can never say something is true Enough to have 1 counterexample to prove something false Honest or dishonest Can find 1 instance you were dishonest, and decide you are dishonest Might quibble If women nicer than men 90% of places, really say not true? Standard way of testing theories is to assume they are right and try to find a case where they are wrong Hypothesis testing does – very backwards way of looking at life Come up with opposite theory from mine and prove it wrong Therefore promote own theory That’s what inference is H0, H1 H0 hyp we are not interested in H1 is hypothesis we are interested in. H0 allows us to show it is wrong How? H0 includes an = sign. (slide) My real hypothesis, H1, women nicer than men. In that case, H0 is what? Women as nice as men, or less nice Always important H0 includes =, = we can reject Important cover whole universe of possible outcomes with H0 and H1 combined Otherwise disproving H0 means nothing about H1 H0 could be women as nice as men. H1 would be? -women nicer or meaner I say w as nice or nicer. H1? Women not as nice Always take opposite Remember H0 one we don’t believe in, but the one we test -why easier to prove false than true? One thing, counterexample can prove false. One supporting example means nothing, impossible to check all cases, every instance Could never check entire universe Think about it as fully true, if one false, false Flukes are a separate topic, issue Other point, can show things are not the same. Only thing can disprove is a = sign. If I say larger, can’t actually prove that Can’t prove something larger Can prove not same, That’s the exercise we do Examples H1, H0 Say you have coin, Suspect it is fake What is H0, H1? -test if fair -H0 coin not fake What probability? -1/2 Set H1, Ho, H1 is? Probability not ½ -gender, grades are the same, .5 -larger than or less than .5 New medicine. Healed faster? -H0 med not healed faster H1 healed same or slower Try again -H0 same or slower, H1 faster H0 always include same Want to include slower in H0 -what if want to prove same No way to prove, because not the same Nothing in this world is the same. Can decide if diff Is it small enough? Different kinds of analysis Can’t reject hypothesis same, but can’t prove can only fail to reject H1- this medication has no effect Test 1 medication compared to nothing, can’t have ‘not at all’ as H1, has to be H0 Reason: it is the only point where there is no effect, no difference. Another med, then can be part of H1 Important subtlety Cheating -H0 is that course 15 and 6 same probability to cheat, H1 is one is better or worse Called 2 sided hypothesis H1 both below and above. Many time when you test theories, they are not two-sided, but one-sided. That means H0 also includes the opposite of H1 Marriage over time. Give me a hypothesis for that. Marriage and happiness lets say -time increases, no correlation, Love marriages vs arranged marriages in India Love marriages start higher, arranged lower. Cross in year 3 Done H0s, H1s Why test hypothesis we don’t believe in? An odd way to think about life Now, what does it mean to reject an H0 hypothesis? Here’s what we do No difference between groups Find data and decide to reject this What it really means, what we have learned: if H0 is correct, problem of getting results we got so low, we reject H0. And therefore accept H1 Convoluted because we assume H0 is correct Question is: is the probability of getting this result high or low? Assume H0 is correct, this means we assume no difference between men, women If no difference, how likely will Women be 8 in niceness, men 7? If we say it is pretty likely it will happen, Cannot reject H0 Variance so high that P(experiment men 8, women 7) = 1/3. nothing makes us doubt H0 But if probability very low… What we mean when we say very small is 1/20, 5% If probability of getting this data or more extreme data is very low, we reject H0 If women less nice, I conclude H0, you H1 Regrettably, it will not help to spend 3 or 4 lectures on this difficult topic Ask grad students. Understand H0, H1? -yes -no I still want you to think about it. Could be H0 is correct. Cannot reject H0, then accept it (slide) Also could be wrong. Well, if we assume H0, unlikely to get result, so we reject it Can make 2 types of mistakes. Type 1 error – reject H1 by mistake. Do not disprove H0 Type 2 error – H1 is wrong. Cannot reject H0 Decide which is more important to us Together it is a constant sum Can be right or wrong, but if we are wrong, we can decide how we want to be wrong. Under H0, p(getting result) = 1/5, more likely type 1 error If it has to be 1/500, it is less likely to be a type 1 error. Maybe it turns out 1/400, and we can’t reject H0 Decision is a tradeoff between types of errors. If the result is so important, revolutionary, interesting, we might be willing to be more lenient Rejecting H0 is less stringent if the results are important, interesting, could be valuable Could be so damaging, we would want a very stringent type 1 error. Say we find gene difference between gay, non-gay people. Find peak value is .2 P(this or more extreme result) = 1/5. do we want to publish this or not? Hard to say If published, people start believing it, turns out wrong, then it might cause damage to a lot of people Standard is 5% Can adapt depending on how important things are What is the cost of the 2 types of mistakes? Not .05, but .07. - very close. Decide, what do you do? Publish? Not? Tell people? No magical solution to that Can run more subjects, experiments Not always possible Assume this distribution, this one (slide) H0 on left, H1 on right Some chance H0 is correct. Much bigger chance we did a good job. Can come all the way to the right, want to be absolutely sure Or left, can be more lenient. -p increases? As you move right, P is a smaller # -because p is divided in terms of H0? Right, exactly right. Definition of p? -smallest chance you would say can happen by coincidence. Smallest likelihood of something happening P is: assuming H0 correct, Probability of getting data we got or more extreme data under chance alone. If p very low, might decide H0 is wrong All kinds of Ps How about p = 0.11? Particularly for your projects Might discover some mistakes. Theory wrong? Data wrong? Data collection too small to be consistent? This is part of the interpretation of P If smaller P, do you have stronger confidence in H1? Say one p = .003, other .00003 In which do we have more confidence in H1? No, fact is we only can be confident in H0 being not correct. Generally, does not translate into H1 being correct Smaller P, effect size bigger? No, P not just about effect size. Some combination of mean variance and difference How about p and # of subjects More subjects we have, smaller p will be. That’s because of variance Talked before about effect sizes Remind you for your papers – giving percentages is not enough 8.1, 8 - is this a significant difference? Say all women rate themselves 8.1, all men 8. Is there a difference? In summary: Hypothesis testing, have H1, H0, set hypothesis to something we don’t believe in, hoping to reject it Meaning of P: probability of getting this or more extreme data, assuming H0 holds. If H0 doesn’t hold, we don’t know 2 types of error and tradeoffs Effect sizes Statistical tests For 1 sample, 2 samples, some about Anovas T test for 1 sample Say we had a medication, is it effective? Distribution of people in 1 dimension – how healthy they feel. Are they healthier than 0? Healthier that 5? Give me examples of other distributions, compare it to a single number -satisfied with customer service? -could ask how satisfied are you? Compare to specific # -satisfied with life at MIT? -drink more than 2 dozen beers a week? -fairness of coin bigger than .5? -age above or below average US age? -in this class? Statistical question never asks about individuals Gender proportion Here is the way T is calculated. (slide) Here is the logic of this test: Top: Difference between observation, mu, and hypothesized mean Hundred time flip coin, 40 times got tails N is hypothesized mean = .5 in this case Is this medication helpful? Medication helpful will just be mu Just ask about the mean differences Standard deviation should be familiar, we talked about this before Wider it is, the less likely it will be significant Square root of number of people that participated in the study The more people we take, the better estimate of what is really happening in reality If we’ve done research with thousands of people, we take mean difference more seriously 3 elements of the test Once you get it, there is a T table in the book Just compare it T becomes close to the normal distribution Say 7 observations Students in class Measured how aggressive they are Take difference between individual measure, overall mean, and square it, take sum of all of them Statements about H0 Aggressiveness -why using mu, not xbar? It is estimated I wouldn’t feel strongly about using xbar, that would be fine What do we have? T is the difference between my mean, hypothetical mean of what will happen. divide it, get 3.42 Go to table, figure out if it is significant Small sample, that will hurt us Probably significant but not sure Let me show example of how this would look Software called Statview. Can do it in Excel as well. How many good ideas does an MIT student have a year? What would be your null hypothesis? -7 I was generous What is the hypothesized mean? Ok, so 7 ideas Value at 1.69 is .1, doesn’t exactly make it, We can’t reject the hypothesis that students have 7 good ideas per year Can reject the hypothesis that people have 0 ideas Seems like students have somewhere below 10 We want to describe what is actually happening in the data Get mean, range, 95% And get outliers. This is called box plots What we’ve learned is that MIT students are creative, more than 0 good ideas/yr, but not 1x month 2 sample T test Basically, structure is very similar Basic structure for T test Compare for what we think will be the difference between 2 groups Relative to expected difference between 2 groups Here is the story 2 sample T test (slide) Real difference between men, women Hypothesized difference -parenthesis still under radical? Yes Example 1 slide Who eats more lollipops, M or F? We expected difference to be 0 Take #s, variances, number of people, get result We think it is close to being significant Example, does sun create freckles? H0, sunny side same or less freckles that non-sunny side Let me Show you how would look in some statistical software 2 columns, before, after [shows example] Want to cover Anova T test example - more general approach to life, Anova - analysis of variance First, if we make many comparisons, we are likely to get one by mistake By defining p = .5, we are saying for every 100 experiments, we will get 5 in which results might be wrong We have to worry about doing too many comparisons Second, Sometimes, even cells we are not using help us get a better estimation of error value IQ across MIT for all majors Take everybody into account, better estimation Story is the same, look at variance within cells and also across cells We ask, how does this affect our concept of variance? Here is a graphical attempt to illustrate this. (slide) Right now, looks like no order (red dots slide) what we ask is whether these barriers change the variance this case, answer is no can imagine things to cluster variance between groups note, across distinction makes difference, but not vertical distinction does it help us get smaller variance? In this example, everyone is 40 in some measure In this one, no variance within the group, but variance across groups Ideal case: Can have a situation, vary within a group, but no difference In the overall mean Most common - Can have variance in and variance between, can we make inferences about this? General formula the same Diff types of means w/ Anova Also look at other sum of squares To get sum of squares, extract from appropriate mean. Partition overall variance Once you have sum of squares, compare to sum of squares between and within. As it gets larger, the groups tell us more and more about these effects. Likely to have significant results One more thing, degrees of freedom Df stands for degrees of freedom If I told you all the data plus mean, I have replicated some information Take away information, you would still be able to calculate it Anovas can be 1 factor or 2 factor design Translates very nicely to Anovas We talked about hypothesis testing Once we reject it, we are willing to accept H1 How many think they took writing seriously? I want to point out, writing is important skill I don’t know how much time you spend on writing Now I am happy to produce a page in a day or two Make sure every sentence is one you like try and do better Writing I am certain will be useful for you