MP2 Experimental Design Review HCI W2014 What is experimental design? How do I plan an experiment? Acknowledgement: Much of the material in this lecture is based on material prepared for similar courses by Saul Greenberg (University of Calgary) as adapted by Joanna McGrenere 1 Experimental Planning Flowchart Stage 1 Stage 2 Stage 3 Stage 4 Stage 5 Problem definition Planning Conduct research Analysis Interpretation data reductions interpretation feedback research idea literature review statement of problem hypothesis development define variables pilot testing generalization controls data collection apparatus hypothesis testing procedures select subjects design feedback 2 statistics reporting What’s the goal? Overall research goals impact choice of study design – – The stage in the design process impacts the choice of study design – – 3 Exploratory research vs. hypothesis confirmation Ecological validity vs tightly controlled Formative evaluation (to get iterative feedback on initial design and/or design choices) Summative evaluation (to determine whether the design is better/stronger/faster than alternative approaches) What’s the research question? Study research questions impact choice of: – – – – Testable hypotheses impact – 4 Protocol, task Experimental conditions (factors) Constructs (effectiveness) Measures (task completion, error rate) choice of statistical analysis (also impacted by nature of the data and experimental design) Experimental Planning Flowchart Stage 1 Stage 2 Stage 3 Stage 4 Stage 5 Problem definition Planning Conduct research Analysis Interpretation data reductions interpretation feedback research idea literature review statement of problem hypothesis development define variables pilot testing generalization controls data collection apparatus statistics reporting hypothesis testing procedures select subjects design feedback 5 Reality check: does the final design support the research questions Quantitative system evaluation Quantitative: – – Methods – – Controlled Experiments Statistical Analysis Measures – – 6 precise measurement, numerical values bounds on how correct our statements are Objective: user performance (speed & accuracy) Subjective: user satisfaction Controlled experiments The traditional scientific method – – clear convincing result on specific issues in HCI: insights into cognitive process, human performance limitations, ... allows comparison of systems, fine-tuning of details ... Strive for – – – – – 7 – lucid and testable hypothesis (usually a causal inference) quantitative measurement measure of confidence in results obtained (inferential statistics) ability to replicate the experiment control of variables and conditions removal of experimenter bias The experimental method a) Begin with a lucid, testable hypothesis H0: there is no difference in user performance (time and error rate) when selecting a single item from a pop-up or a pull down menu, regardless of the subject’s previous expertise in using a mouse or using the different menu types File Edit Insert File Edit New Open View Close Insert Save 8 View New Open Close Save The experimental method b) Explicitly state the independent variables that are to be altered Independent variables – – the things you control (independent of how a subject behaves) two different kinds: 1. 2. treatment manipulated (can establish cause/effect, true experiment) subject individual differences (can never fully establish cause/effect) in menu experiment – – – 9 menu type: pop-up or pull-down menu length: 3, 6, 9, 12, 15 expertise: expert or novice (a subject variable – the researcher can not manipulate) The experimental method c) Carefully choose the dependent variables that will be measured Dependent variables – variables dependent on the subject’s behaviour / reaction to the independent variable – Make sure that what you measure actually represents the higher level concept! in menu experiment – – – 10 time to select an item selection errors made Higher level concept (user performance) The experimental method d) Judiciously select and assign subjects to groups Ways of controlling subject variability – – recognize classes and make them an independent variable minimize unaccounted anomalies in subject group superstars versus poor performers – use reasonable number of subjects and random assignment Novice 11 Expert The experimental method... e) Control for biasing factors – unbiased instructions + experimental protocols prepare ahead of time – – double-blind experiments, ... Potential confounding Now you get to do the pop-up menus. I think variables you will really like them... I designed them myself! – – – 12 Order effects Learning effects Counterbalancing (http://www.yorku.ca/mack/R N-Counterbalancing.html) The experimental method f) Apply statistical methods to data analysis – Confidence limits: the confidence that your conclusion is correct “The hypothesis that mouse experience makes no difference is rejected at the .05 level” (i.e., null hypothesis rejected) means: – a 95% chance that your finding is correct – a 5% chance you are wrong g) Interpret your results – – 13 what you believe the results mean, and their implications yes, there can be a subjective component to quantitative analysis Experimental designs Between subjects: Different participants - single group of participants is allocated randomly to the experimental conditions. Within subjects: Same participants - all participants appear in both conditions. Matched participants: participants are matched in pairs, e.g., based on expertise, gender, etc. Mixed: Some independent variables are within subjects, some are between subjects 14 www.id-book.com Within-subjects It solves the individual differences issues Allows participants to make comparisons between conditions But raises other problems: – Need to look at the impact of experiencing the two conditions Order Effects Changes in performance resulting from (ordinal) position in which a condition appears in an experiment (always first?) Arises from warm-up, learning, learning what they will be asked to reflect upon, fatigue, etc. Effect can be averaged and removed if all possible orders are presented in the experiment and there has been random assignment to orders Sequence effects Changes in performance resulting from interactions among conditions (e.g., if done first, condition 1 has an impact on performance in condition 2) Effects viewed may not be main effects of the IV, but interaction effects Can be controlled by arranging each condition to follow every other condition equally often Counterbalancing Controlling order and sequence effects by arranging subjects to experience the various conditions (levels of the IV) in different orders Self-directed learning: investigate the different counterbalancing methods – – – – – Randomization Block Randomization Reverse counter-balancing Latin squares and Greco squares (when you can’t fully counterbalance) http://www.experiment-resources.com/counterbalanced-measuresdesign.html Between, within, matched participant design Design Advantages Disadvantages Between No order effects Many subjects & individual differences a problem Within Few individuals, no individual differences Counter-balancing needed because of ordering effects Matched Same as different participants but individual differences reduced Cannot be sure of perfect matching on all differences 19 www.id-book.com Internal Validity the extent to which a causal conclusion based on a study is warranted Internal validity is reduced due to the presence of controlled/confounded variables – But not necessarily invalid It’s important for the researcher to evaluate the likelihood that there are alternative hypotheses for observed differences – Need to convince self and audience of the validity External validity The extent to which the results of a study can be generalized to other situations and to other people If the experimental setting more closely replicates the setting of interest, external validity can be higher than in a true experiment run in a controlled lab setting Often comes down to what is most important for the research question – Control or ecological validity? Control True experiment = complete control over the subject assignment to conditions and the presentation of conditions to subjects – Control of the who => random assignment to conditions – Control over the who, what, when, where, how Only by chance can other variables be confounded with IV Control of the what/when/where/how => control over the way the experiment is conducted Quasi-Experiment When you can’t achieve complete control – – Lack of complete control over conditions Subjects for different conditions come from potentially non-random pre-existing groups Experts vs novices Early adopters vs technophobes? It’s a matter of control True Experiment Random assignment of subjects to condition Manipulate the IV Control allows ruling out of alternative hypotheses Quasi Experiment Selection of subjects for the conditions Observe categories of subjects – If the subject variable is the IV, it’s a quasi experiment Don’t know whether differences are caused by the IV or differences in the subjects Other features In some instances cannot completely control the what, when, where, and how – – Need to collect data at a certain time or not at all Practical limitations to data collection, experimental protocol