IS 4800 Empirical Research Methods for Information Science Class Notes February 15, 2012

1. IS 4800 Empirical Research Methods for Information Science Class Notes February 15, 2012 Instructor: Prof. Carole Hafner, 446 WVH hafner@ccs.neu.edu Tel: 617-373-5116 Course Web site: www.ccs.neu.edu/course/is4800sp12/

2. Outline • Review/finish reliability and validity techniques for composite measures • Sampling and volunteer bias

3. Validating a Composite Measure

4. Has reliability Has validity For psychological measures, these are collectively referred to as a measure’s “psychometrics”. What is a validated measure?

5. A reliable measure produces similar results when repeated measurements are made under identical conditions Reliability can be established in several ways Test-retest reliability: Administer the same test twice Parallel-forms reliability: Alternate forms of the same test used Split-half reliability: Parallel forms are included on one test and later separated for comparison Measure Reliability

6. For surveys, this also encompasses internal consistency: Do all of the questions address the same underlying construct of interest? That is, do scores covary? A standard measure is Cronbach’s alpha 0 = no correlation 1 = scores always covary in the same way 0.7 used as conventional threshold Reliability

7. Correlation coefficient • A measure of association between two numeric variables X and Y (Pearson’s R) • When will R be positive ? • When Xi and Yi are both larger than the mean • When Xi and Yi are both smaller than the mean • Represents a general tendency for X and Y to vary in the same way • Normalized to range from -1 to +1

8. Interpreting the correlation coefficient • -1.0 to -0.7 strong negative association. • -0.7 to -0.3 weak negative association. • -0.3 to +0.3 little or no association. • +0.3 to +0.7 weak positive association. • +0.7 to +1.0 strong positive association.

9. Cronback’s Alpha • A test for composite measure reliability • K = the number of test items • r is the mean of K(K-1) non-redundant correlation coefficients • Can take on negative numbers but they are not meaningful • Therefore considered to run from 0 to 1

10. Check to be sure the items on your questionnaire are clearly written and appropriate for those who will complete your questionnaire Increase the number of items on your questionnaire Standardize the conditions under which the test is administered (e.g., timing procedures, lighting, ventilation, instructions) Make sure you score your questionnaire carefully, eliminating scoring errors Increasing the Reliability of a Questionnaire

11. A valid measure measures what you intend it to measure Very important when using psychological tests (e.g., intelligence, aptitude, (un)favorable attitude) Validity can be established in a variety of ways Face validity: Is a measure clearly related to the construct. Least powerful method. Content validity: How adequately does a measure sample the full range of behavior it is intended to measure? Measure Validity

12. The degree to which a measure corresponds to what happens in the real world. Example: Assessing productivity/day in the lab vs. Assessing productivity/day in the office Ecological Validity

13. Criterion-related validity: How adequately does a test score match some criterion score? Takes two forms Concurrent validity: Does test score correlate highly with score from a measure with known validity? Predictive validity: Does test predict behavior known to be associated with the behavior being measured? Measure Validity

14. Construct validity: Do the results of a test correlate with what is theoretically known about the construct being evaluated? Convergent validity (subtype): measures of constructs that should be related to each other are related. (conservative, religious ??) Discriminant validity (subtype): measures of constructs that should not be related are not Measure Validity

15. MonitorSize Productivity Example • Assume we have good evidence for this model of the world.. • We now propose a new measure for Productivity • What would be evidence for convergent validity? • What would be evidence for discriminant validity? Seniority

16. Sensitivity Is a dependent measure sensitive enough to detect behavior change? An insensitive measure will not detect subtle behaviors Range Effects Occur when a dependent measure has an upper or lower limit Ceiling effect: When a dependent measure has an upper limit Floor effect: When a dependent measure has a lower limit. More Concerns with Measures

17. You want to assess the effect of TV viewing on whether people like large computer monitors or not (yes/no). You run an experiment in which participants are randomized to watch either 2 hrs or 0 hrs of TV per day for a week, then answer your question. What’s going on? Example Participant Condition LikesLargeMonitors 1 TV Yes 2 No TV Yes 3 TV Yes 4 No TV Yes

18. Say you decide you need a new survey measure, “attitude towards large computer monitors” (ATLCM) I like big monitors. Big monitors make me nervous. I prefer small monitors, even if they cost more. 7-pt Likert scales How would you validate this measure? Developing a New Measure

19. You want to assess the effect of TV viewing on attitude towards large computer monitors (ATLCM). You run an experiment in which participants are randomized to watch either 2 hrs or 0 hrs of TV per day for a week, then fill out the ATLCM. What’s going on? Example Participant Condition ATLCM 1 TV 7.0 2 No TV 6.7 3 TV 6.9 4 No TV 7.0

20. Reliability Test-retest/parallel forms/split-half Internal consistency Validity Face Content Criterion-related Concurrent Predictive Construct Convergent Discriminant Validation - Summary

21. Sometimes you really can measure the entire population (e.g., workgroup, company), but this is rare… “Convenience sample” Cases are selected only on the basis of feasibility or ease of data collection. Sampling

22. You should obtain a representative sample The sample closely matches the characteristics of the population A biased sample occurs when your sample characteristics don’t match population characteristics Biased samples often produce misleading or inaccurate results Usually stem from inadequate sampling procedures Convenience samples are not representative – they are subject to “volunteer bias” !! Acquiring A Survey Sample

23. Volunteer Bias • How can it affect external validity? • Characteristics of volunteers? • How do you address volunteer bias?

24. Characteristics of Individuals Who Volunteer for Research Maximum Confidence 1. tend to be more highly educated than nonvolunteers 2. tend to come from a higher social class than nonvolunteers 3. are of a higher intelligence in general, but not when volunteers for atypical research (such as hypnosis, sex research) 4. have a higher need for approval than nonvolunteers 5. are more social than nonvolunteers

25. Considerable Confidence Volunteers are more “arousal seeking” than nonvolunteers (especially when the research involves stress) Individuals who volunteer for sex research are more unconventional than nonvolunteers Females are more likely to volunteer than males, except when the research involves physical or emotional stress Volunteers are less authoritarian than nonvolunteers Jews are more likely to volunteer than Protestants; however, Protestants are more likely to volunteer than Catholics Volunteers have a tendency to be less conforming than nonvolunteers, except when the volunteers are female and the research is clinically oriented Source: Adapted from Rosenthal & Rosnow, 1975.

26. Remedies for Volunteer Bias • Make your appeal very interesting • Make your appeal as nonthreatening as possible • Explicitly state the theoretical and practical importance of your research • Explicitly state why the target population is relevant to your research • Offer a small reward for participation

27. Remedies for Volunteer Bias (cont.) • Have a high-status person make the appeal for participants • Avoid research that is physically or psychologically stressful • Have someone known to participants make the appeal • Use public or private commitment to volunteering when appropriate

28. Simple Random Sampling Randomly select a sample from the population Random digit dialing is a variant used with telephone surveys Reduces systematic bias, but does not guarantee a representative sample Some segments of the population may be over- or underrepresented Scientific Sampling Techniques

29. Systematic Sampling Every kth element is sampled after a randomly selected starting point Sample every fifth name in the telephone book after a random page and starting point selected, for example Empirically equivalent to random sampling (usually) May still result in a non-representative sample Easier than random sampling Scientific Sampling Techniques

30. Stratified Sampling Used to obtain a representative sample Population is divided into (demographic) strata Focus also on variables that are related to other variables of interest in your study (e.g., relationship between age and computer literacy) A random sample of a fixed size is drawn from each stratum May still lead to over- or underrepresentation of certain segments of the population Proportionate Sampling Same as stratified sampling except that the proportions of different groups in the population are reflected in the samples from the strata Scientific Sampling Techniques

31. You want to conduct a survey of job satisfaction of all employees but can only afford to contact 100 of them. Personnel breakdown: 50% Engineering 25% Sales & Marketing 15% Admin 10% Management Examples of Stratified sampling? Proportionate sampling? Sampling Example:

32. Cluster Sampling Used when populations are very large The unit of sampling is a group (e.g., a class in a school) rather than individuals Groups are randomly sampled from the population (e.g., ten classes from a particular school) Sampling Techniques

33. Multistage Sampling Variant of cluster sampling First, identify large clusters (e.g., school districts) and randomly sample from that population Second, sample individuals from randomly selected clusters Can be used along with stratified sampling to ensure a representative sample Scientific Sampling Techniques

34. Sampling and Statistics • If you select a random sample, the mean of that sample will (in general) not be exactly the same as the population mean. However, it represents an estimate of the population mean • If you take two samples, one of males and one of females, and compute the two sample means (let’s say, of hourly pay), the difference between these is an estimate of the difference between the population means. • This is the basis of inferential statistics based on samples

35. Sampling and Statistics (cont.) • If larger the sample, the better estimate (more likely it is close to the population mean) • The variance/SD of the sample means is related to the variance/SD of the population. However, it is likely to be LESS (!) than the population variance.

36. Inference with a Single Observation ? Population Parameter:  Sampling Inference Observation Xi • Each observation Xi in a random sample is a representative of unobserved variables in population • How different would this observation be if we took a different random sample? June 9, 2008 36 36

37. Normal Distribution • The normal distribution is a model for our overall population • Can calculate the probability of getting observations greater than or less than any value • Usually don’t have a single observation, but instead the mean of a set of observations June 9, 2008 37

38. Inference with Sample Mean ? Population Parameter:  Sampling Inference Estimation Sample Statistic: x • Sample mean is our estimate of population mean • How much would the sample mean change if we took a different sample? • Key to this question: Sampling Distribution of x June 9, 2008 38

39. Sampling Distribution of Sample Mean • Distribution of values taken by statistic in all possible samples of size n from the same population • Model assumption: our observations xi are sampled from a population with mean and variance 2 Sample 1 of size n x Sample 2 of size n x Sample 3 of size n x Sample 4 of size n x Sample 5 of size n x Sample 6 of size n x Sample 7 of size n x Sample 8 of size n x . . . Distribution of these values? Population Unknown Parameter:  June 9, 2008 39

40. Mean of Sample Mean • First, we examine the center of the sampling distribution of the sample mean. • Center of the sampling distribution of the sample mean is the unknown population mean: • mean( X ) = μ • Over repeated samples, the sample mean will, on average, be equal to the population mean • no guarantees for any one sample! June 9, 2008 40

41. Variance of Sample Mean • Next, we examine the spread of the sampling distribution of the sample mean • The variance of the sampling distribution of the sample mean is • variance( X ) = 2/n • As sample size increases, variance of the sample mean decreases! • Averaging over many observations is more accurate than just looking at one or two observations June 9, 2008 41

42. Comparing the sampling distribution of the sample mean when n = 1 vs. n = 10 June 9, 2008 42

43. Law of Large Numbers • Remember the Law of Large Numbers: • If one draws independent samples from a population with mean μ, then as the sample size (n) increases, the sample mean x gets closer and closer to the population mean μ • This is easier to see now since we know that • mean(x) = μ • variance(x) = 2/n 0 as n gets large June 9, 2008 43

44. Example • Population: seasonal home-run totals for 7032 baseball players from 1901 to 1996 • Take different samples from this population and compare the sample mean we get each time • In real life, we can’t do this because we don’t usually have the entire population! June 9, 2008 44

45. Distribution of Sample Mean • We now know the center and spread of the sampling distribution for the sample mean. • What about the shape of the distribution? • If our data x1,x2,…, xn follow a Normal distribution, then the sample mean x will also follow a Normal distribution! June 9, 2008 45

46. Example • Mortality in US cities (deaths/100,000 people) • This variable seems to approximately follow a Normal distribution, so the sample mean will also approximately follow a Normal distribution June 9, 2008 46

47. Central Limit Theorem • What if the original data doesn’t follow a Normal distribution? • HR/Season for sample of baseball players • If the sample is large enough, it doesn’t matter! June 9, 2008 47

48. Central Limit Theorem • If the sample size is large enough, then the sample mean x has an approximately Normal distribution • This is true no matter what the shape of the distribution of the original data!  June 9, 2008 48

49. Example: Home Runs per Season • Take many different samples from the seasonal HR totals for a population of 7032 players • Calculate sample mean for each sample n = 1 n = 10 n = 100 June 9, 2008 49