Presentation is loading. Please wait.

Presentation is loading. Please wait.

1 COMM 301: Empirical Research in Communication Lecture 15 – Hypothesis Testing Kwan M Lee.

Similar presentations


Presentation on theme: "1 COMM 301: Empirical Research in Communication Lecture 15 – Hypothesis Testing Kwan M Lee."— Presentation transcript:

1 1 COMM 301: Empirical Research in Communication Lecture 15 – Hypothesis Testing Kwan M Lee

2 2 Things you should know by the end of the lecture Structure and logic of hypothesis testing –Differentiate means and standard deviations of population, sample, and sampling distributions. –What is a null hypothesis? What is an alternative hypothesis? Know how to set up these hypotheses. –Know the meaning of significance level –Know the decision rules for rejecting the null hypothesis: Comparing observed value versus critical value (check z statistic example) Comparing p-value versus significance level. –Know the limitations of statistical significance. –Know Type 1 error, Type 2 error, power, and the factors that affect them.

3 3 Structure and logic of hypothesis testing Research question Hypothesis: a testable statement regarding the difference between/among groups or about the relationship between/among variables  we ONLY talked Hr (research H) [aka Ha (alternative H)] Data Gathering (you have done!) Analysis –Descriptive Measure of Central Tendency Measure of Dispersion –Inferential We will test H0 (null hypothesis) in order to test Hr

4 4 Cf. Normal Distribution A continuous random variable Y has a normal distribution if its probability density function is –Don’t worry about this formula! The normal probability density function has two parameters – mean (mu) and standard deviation (sigma) Mu and sigma changes particular shapes of a normal distribution (see next graphs) –Remember standard deviation is how far away “on average” scores are away from the mean of a set of data

5 5

6 6 Cf. Standard Normal Distribution Standard normal distribution is a normal distribution with mu = 0 and sigma = 1 z distribution = distribution of a standard normal distribution z transformation 

7 7 Population Distribution: Sample Distribution: & Sampling Distribution See SPSS example! (lect 15_2) Very important to know the difference! Central limit theorem –With large sample size and limitless sampling, sampling distribution will always show normal distribution.

8 8 Two Hypotheses in Data Analyses Null hypothesis, H 0 Alternative hypothesis, H a (this is your Research H) H 0 and H a are logical alternative to each other H a is considered false until you have strong evidence to refute H 0

9 9 Null hypothesis –A statement saying that there is no difference between/among groups or no systematic relationship between/among variables –Example: Humor in teaching affects learning. RM1XM2 RM3M4 Learning measured by standardized tests H0 : M2 = M4, meaning that …? Equivalent way of stating the null hypothesis: – H0 : M2 - M4 = 0

10 10 Alternative hypothesis Alternative H (H a ) = Research H (Hr) –A statement saying that the value specified H 0 is not true. –Usually, alternative hypotheses are your hypotheses of interest in your research project An alternative hypothesis can be bi-directional (non-directional) or uni-directional (directional). Bi-directional H 1 (two-tailed test) H 1 : M2 - M4  0 (same as M2  M4) –I am not sure if humor in teaching improves learning or hampers learning. So I set the above alternative hypothesis.

11 11 Alternative hypothesis (cont.) Uni-directional or Directional Ha (one-tailed test) You set a uni-directional Ha when you are sure which way the independent variable will affect the dependent variable. Example: Humor in teaching affects learning. RM1XM2 RM3M4 I expect humor in teaching improves learning. So I set the alternative hypothesis as –M2 > M4 –Then, H 0 becomes: M2 “< or =“ M4

12 12 Comparing two values Okay, so we have set up H 0 and Ha. The key questions then become: Is H 0 correct or is Ha correct? How do we know? We know by comparing 2 values against each other (when using stat tables): –observed value (calculated from the data you have collected), against –critical value (a value set by you, the researcher) Or simply by looking alpha value in SPSS output

13 13 Finding observed values How do we find the observed value? Observed values are calculated from the data you have collected, using statistical formulas –E.g. z; t; r; chi-square etc. Do you have to know these formulas? –Yes and No: for the current class, we will test only chi- square (  we will learn it later) –Most of the time, you need to know where to look for the observed value in a SPSS output, or to recognize it when given in the examination question.

14 14 Determining the critical value How to determine the critical value? four factors –Type of distribution used as testing framework –Significance levels –Uni-directional or bi-directional H 1 (one-tailed or two-tailed test) –Degree of freedom

15 15 Determining critical values Type of distribution –Recall that data form distributions –Certain common distributions are used as a framework to test hypotheses: z distribution t distribution chi-square distribution F distribution –Key skill: reading the correct critical values off printed tables of critical values. –Table example t-distribution  see next slide Also compare it with z distribution (what’s the key difference? z is when you know population parameters; t is when you don’t know population parameters)

16 16 z distribution Remember the standard normal distribution!

17 17 t distribution (next lecture) Why t instead of z? Relationship between t and z?

18 18 p df ……… chi-square distribution (next lecture)

19 19 Determining critical values Significance level (alpha level) –A percentage number set by you, the researcher. –Typically 5% or 1%. –The smaller the significance level, the stricter the test.

20 20 Determining critical values One-tailed or two-tailed tests –One-tailed tests (uni-directional H 1, e.g. M2 - M4 > 0) use the whole significance level. –Two-tailed tests (bi-directional H 1, e.g. M2 - M4  0) use the half the significance level. –Applies to certain distributions and tests only, in our case, the t-distributions and t-tests

21 21 Determining critical values Degree of freedom –How many scores are free to vary in a group of scores in order to obtain the observed mean –Df = N – 1 –In two sample t-test, it’s N (total sample numbers)-2. Why?

22 22 Accept or reject H 0 ? Next, you compare the observed value (calculated from the data) against the critical value (determined by the researcher), to find if H0 is true or H1 is true. The key decision to make is: Do we reject H0 or not? We reject H0 when the observed value (z: t; r; chi- square; etc.) is more extreme than the critical value. If we reject H0, it means, we accept H1. H1 is likely to be true at 5% (or 1%) significance level. We are 95% (or 99%) sure that H1 is true.

23 23 Cf. Hypothesis testing with z When you know mu and sigma of population (this is a rare case), you conduct z test Example. After teaching Comm301 to all USC undergrads (say, 20000), I know that the population distribution of Comm301 scores of all USC students is a normal distribution with mu =82 and sigma = 6 In my current class (say 36 students), the mean for the final is 86. RQ: Is the students in my current class (i.e., my current sample) significantly different from the whole USC students (the population)?

24 24 Cf. Hypothesis testing with z (cont) Step 1: State H 0 = “mean = 82” and Ha Step 2: Calculate z statistic of sampling distribution (distribution of sample means: you are testing whether this sample is drawn from the same population) = = 4 –Notice that for the z statistic here, we are using the sigma of the sampling distribution, rather than the sigma of the population distribution. –It’s because we are testing whether the current mean score is the result of a population difference (i.e., the current students are from other population: Ha) or by chance (i.e., students in my current class happened to be best USC students; difference due to chances caused by sampling errors or other systematic and non-systematic errors)

25 25 Cf. Hypothesis testing with z (cont) Step 3: Compare test statistic with a critical value set by you (if you set alpha at 0.05, the critical value is 2 [more precisely it’s 1.96])  see the previous standard normal distribution graph (z- distribution graph) Step 4: Accept or Reject H0 –Since the test statistic (observed value) is lager than the critical value, you reject H0 and accept Ha Step 5: Make a conclusion –My current students (my current samples) are significantly different from other USC students (the population) with less than 5% chance of being wrong in this decision. But still there is 5% of chance that your decision is wrong

26 26 Another way to decide: Simply look at p-value in SPSS output Area under the curve of a distribution represents probabilities. This gives us another way to decide whether to reject H 0. The way is to look at the p-value. There is a very tight relationship between the p-value and the observed value: the larger the observed value, the smaller the p-value. The p-value is calculated by SPSS. You need to know where to look for it in the output.

27 27 Another way to decide: p-value So what is the decision rule for the p-value to see if we reject H 0 or not? The decision rule is this: –We reject H 0, if the p-value is smaller than the significance level set by us (5% or 1% significance level). Caution: P-values in SPSS output are denoted as “Sig.” or “Sig. level”.

28 28 The meaning of statistical significance –Example: Humor in teaching affects learning. RM1XM2 RM3M4 –H 1 : M2 - M4  0; H 0 : M2 - M4 = 0. –Assume that p-value calculated from our data is 0.002, i.e. 0.2%  meaning that, assuming H 0 is true (there is no relationship between humor in teaching and learning), the chance that M2 - M4 = 0 is very, very small, less than 1% (actually 0.2%). –If the chance that H 0 is true is very very small, we have more probability that H1 is true (actually, 99.8%). –Key point: The conclusion is a statement based on probability!

29 29 The meaning of statistical significance (cont.) When we are able to reject H 0, we say that the test is statistically significant. It means that there is very likely a relationship between the independent and dependent variables; or two groups that are being compared are significantly different. However, a statistically significant test does not tell us whether that relationship is important.

30 30 The meaning of statistical significance (cont) Go back to the previous example. –So, we have a p-value of The test is significant, the chance that M2 - M4 = 0 is very very small. –M2 - M4 > 0 is very likely true. But it could mean M2 - M4 = 5 point (5 > 0), or M2 - M4 = 0.5 points (0.5 > 0). One key problem with statistical significance is that it is affected by sample size. –The larger the sample, the more significant the result. –So, I could have M2 - M4 = 0.5 (0.5 > 0, meaning that my treatment group on average performs 0.5 point better than my control group), and have statistical significance if I run many many subjects.

31 31 Need to consider “Effect Size” Then, should I take extra time to deliver jokes during the class for the 0.5 point improvement? So, beyond statistical significance, we need to see if the difference (or the relationship) is substantive. You can think of it this way: an independent variable having a large impact on a dependent variable is substantive. The idea of a substantive impact is called effect size. Effect size is measured in several ways (omega square, eta square, r square [coefficient of determination]). You will meet one later: r 2

32 32 Type 1 error At anytime we reject the null hypothesis H 0, there is a possibility that we are wrong: H 0 is actually true, and should not be rejected. By random chance, the observed value calculated from the data is large enough to reject H 0. = By random chance, the p-value calculated from the data is small enough to reject H 0. This is the Type 1 error: wrongly rejecting H 0 when it is actually true.

33 33 Type 1 error (cont) The probability of committing a Type 1 error is equal to the significance level set by you, 5% or 1%.  Type 1 error = alpha As a researcher, you control the chance of a Type 1 error. So, if we want to lower the chance of committing a Type 1 error, what can we do?

34 34 Type 1, Type 2 errors, and Power

35 35 Type 2 error and Power When we lower the chance of committing a Type 1 error, we increase the chance of committing another type of error, a Type 2 error (holding other factors constant). A Type 2 error occurs when we fail to reject H 0 when it is false. Type 2 error is also known as beta. From Type 2 error, we get an important concept: How well can a test reject H 0 when it should be rejected? This is the power of a test. Power of a test is calculated by (1 - beta). You don’t have to know how to calculate beta; it will be given or set by you.

36 36 Factors affecting power of a test –Effect size: The larger the effect size, the smaller beta, and hence the larger the power (holding alpha (significance level) and sample size constant). –Sample size: The larger the sample size, the smaller the beta, and hence the larger the power (holding alpha and effect size constant) –Measurement errors The less measurement errors  the more power


Download ppt "1 COMM 301: Empirical Research in Communication Lecture 15 – Hypothesis Testing Kwan M Lee."

Similar presentations


Ads by Google