Research Methods & Experimental Design 16.422 Human Supervisory Control April 2004 Research Methods ? Qualitative vs. quantitative ? Understanding the relationship between objectives (research question) and variables is critical ? Information ≠Data ? Information=data + analysis ? Planning in advance is a must ? To include how data will be analyzed Qualitative Research Methods ? Social & cultural phenomenon ? Case studies ? Focus groups ? Observations ? Usability testing ? Can be quantitative ? Interviews ? Questionnaires Quantitative Research Methods ? Natural phenomenon ? Mathematical modeling ? Experiments ? Optimization ? Game theory ? Surveys ? Bottom line – statistics are a must Project Assignment ? Design and conduct an experiment in which you explore some measure of human performance through testing, analyze the results, and discuss the broader implications. ? Design an actual display that uses automation for decision support… While formal experimental testing is not required, a small group of users should be used to identify problems with the design to include functionality evaluation as well as recommendation for future improvements and systems integration. The Experimental Design Process Research Question (Hypothesis) Design Experiment Collect Data Analyze Data Draw Conclusions Experimental Design ? Design of Experiments (DOE) defined: ? A theory concerning the minimum number of experiments necessary to develop an empirical model of a research question and a methodology for setting up the necessary experiments. ? A parsimony model ? Human subject vs. object experimentation ? Other DOE Constraints ? Time ? Money Experimental Design Basics ? Two kinds of data gathering methodologies ? Observation ? Can’t prove cause & effect but can establish associations. ? Hawthorne effect, social facilitation ? Experimental ? Cause & effect ? Variables of interest – factors vs. treatments ? Independent variable ? Treatment – manipulations of variables of interest ? Treatment vs. control group ? Dependent variable is what you are measuring More Basics ? Confounds ? Randomization Concerns ? Randomization prevents experimental bias ? Assignment by experimenter ? Counterbalancing ? Statistical assumptions ? A requirement for statistical tests of significance ? Why would you use the observation methodology instead of experiments? DOE Terminology ? Replications ? Independent observations of a single treatment. ? Variance ? The measuring stick that compares different treatments. ? Internal validity ? The extent to which an experiment accomplishes its goal(s). ? Reproducibility ? Given the appropriate information, the ability of others to replicate the experiment. DOE Terminology (cont.) ? External validity ? How representative of the target population is the sample? ? Can the results be generalized? ? Generalizations for field experiments are easier to justify than lab experiments because of artificialities. ? Medical Trials ? Placebo ? Double Blind ? If so, what is the population to which it can be generalized? ? Can the results be generalized to the real world? Data Analysis ? Data Types ? Variables ? Categorical ? Numerical ? Scales of Measurement ? Nominal ? Ordinal ? Interval ? Computer Programs ? Excel, SAS, S+, SPSS ANOVA Within group variance is noise and between group variance is information we seek. ANOVA separates these out. Basic Statistical Tests ? Assumptions for comparison of means ? Independent & random ? Normality ? Variances roughly equal ? t-tests ? One or two samples ? Chi-square tests ? NID(0,1) ? Categorical data, non-parametric Chi square important because any sum of squares in normal random variables divided by the variance is chi-square distributed Null Hypothesis: H o ? Defined: The difference in two different populations parameters is 0. ? H o : Always predicts absence of a relationship & assumed to be true. ? If the null hypothesis is NOT rejected, we CANNOT conclude that there is no difference, only that the method did not detect any difference. ? p < .05 ???? H o : μ 1 = u 2 H a : μ 1 ≠u 2 A Very Important Research Question ? Does drinking cappuccino one hour before a test improve results? ? What is the metric (dependent variable)? ? Experimental Design ? Treatment group vs. control group ? A single comparison ? Experimental efficiency ? Perhaps we want to look at who makes the cappuccino (Seattle’s, Starbucks, Pete’s) as well as the difference between coffee and cappuccino. ? 2X3 Factorial ? Interaction effects Caffeine/Performance Experiment GB SB ER Capp Coffee We now know the general layout of the experiment – but what is missing? Caffeine/Performance Experiment ? How many subjects do we need? ? Sample Size ? Related to power – the complement of a Type II error… Decision Ho True Ho False Reject Ho Type I error p = α Correct decision p = 1 - β = POWER Fail to reject Ho Correct decision p = 1 - α Type II error p = β Ask what Ho is? Null hypothesis – no significant difference exists between experimental groups. Don’t Panic… Caffeine/Performance Experiment ? So how do you determine sample size? ? http://members.aol.com/johnp71/javastat.html ? Sensitivity is an issue ? # of factors influences sample size ? Recruitment Issues ? Population selection ? How do we assign subjects to treatment categories? ? Confounds ? Experience ? Self-selection ? Control techniques Other Subject Considerations ? What is the most efficient way to use human subjects? ? Between subjects ? Within subjects ? Repeated measures ? Increases power but… ? Confounds – practice & fatigue ? Counterbalance ? Mixed subjects ? Pre-test/post-test ? Tests over time Pre/post Test Considerations Pre-Test Post-Test Intervention A Intervention B Between Subjects Within Subjects ? Ideally pre-test scores will be equivalent ? You want to see a difference between the experimental and control group. Statistical Tests (cont.) ? Analysis of variances (ANOVA) ? Testing the differences between two or more independent means (or groups) on one dependent measure (either a single or multiple independent variables). ? One way vs. factorial ? F test – ratio of variances ? MANOVA Other DOE considerations: ? Full Factorial ? Blocking ? More homogenous grouping ? Coffee of the day v. another kind ? Starbuck’s at the Marriott vs. Galleria ? Pairing ? Increases precision by eliminating the variation between experimental units ? Randomization still possible ? Many others… ? Full factorial – should be run twice ? Tennis shoe example – try to find out which sole is better for shoes so each boy wears two different shoes. Randomization comes in assigning which shoe to which foot. What test to use? Adapted from University of Maryland Psychology World. Yes Yes Yes Yes No No No No Yes No Includes a Categorical variable Only Between Subjects Variable Only Between Subjects Variable Only One Independent Variable Only Two Levels Pearson Correlation One-Way Analysis of Variance Mixed Two-Way ANOVA Between-Subjects 2-Way ANOVA Within-Subjects t-test Between Subjects t-test Example Experiment ? Are web-based case studies better than print versions. ? How can we test this? ? This question was tested with 2 classes with 2 different professors. ? What are the independent & dependent variables? ? Was it within/between/mixed? ? What statistical test should we use? Results Tests of Between-Subjects Effects Dependent Variable: GRADES 173.681 a 4 43.420 .986 .420 190832.489 1 190832.489 4333.757 .000 157.697 1 157.697 3.581 .062 26.217 2 13.109 .298 .743 11.840 1 11.840 .269 .605 3654.818 83 44.034 673001.300 88 3828.499 87 Source Corrected Model Intercept PROF TYPE PROF * TYPE Error Total Corrected Total Type III Sum of Squares df Mean Square F Sig. R Squared = .045 (Adjusted R Squared = -.001) a. Interactions ? Interaction effect: the response of one variable depends on effect of another variable ? No interaction – parallel lines ? Significant interaction: ? Which professor would you rather have? Estimated Marginal Means of GRADES PROF LB E s ti m a ted M a r g i n a l M e ans 89.5 89.0 88.5 88.0 87.5 87.0 86.5 86.0 85.5 TYPE D P Non-Parametric Tests ? Use when you have no good information about an underlying distribution ? Parametric tests: ? Parametric form - parameters either assumed to be known or estimated from the data ? The mean and variance of a normal distribution ? Null hypothesis can be stated in terms of parameters and the test statistic follows a known distribution. ? Non-parametric tests are still hypothesis tests, but they look at the overall distribution instead of a single parameter ? Particularly useful for small samples All data is not normal…. Parametric ? Correlation & Association ? Pearson ? T-tests ? Independent & dependent ? ANOVA ? Factorial ? Repeated measures ? MANOVA ? Linear regression Non-parametric ? Association ? Spearman ? Chi-Square ? Contingency tables ? Kruskal-Wallis test ? Sign-test ? Friedman ANOVA ? Logistic regression