Quantitative Methods for Researchers Paul Cairns
Objectives Statistics for severe tests Designing for analysis A whizz through some stats Time for questions Quant for RS, Paul Cairns
But first… Does making game playing more sociable improve the player experience? Experiments, please Quant for RS, Paul Cairns
A little example Ten people played a game with AI on or with AI off. Get scores: Quant for RS, Paul Cairns AI OffAI On Mean
Quant for RS, Paul Cairns Calculate probabilities Statistical test – Transform data to a standard form – Calculate a statistic – Convert to probability “Unlikely” events (low p) are good Test depends on data
Mann-Whitney Non-parametric test Rank all scores (lowest =1, highest = 10) Sum ranks of each group Look up probability (in Howell) Let’s do that now Quant for RS, Paul Cairns
Significance What is unlikely? Conventions: 0.05, 0.01 Is AI harder? 1 in 20 experiments give an incorrect result!
Statistical Argument Inference is an argument form What is being severely tested? – Alternative hypothesis – “X causes Y” – Null hypothesis Statistics serve the test Quant for RS, Paul Cairns
Gold standard argument, mk 2 1.Specify your idea 2.Devise a severe test of hypothesis 3.Collect data rigorously 4.Statistics give probabilities of null 5.Unlikely predictions “prove” your case Quant for RS, Paul Cairns
Implications Must have an idea under test No multiple testing No post hoc analysis Need multiple experiments Quant for RS, Paul Cairns
Silver standard argument, mk 2 1.Specify some ideas 2.Devise an experiment to exhibit ideas 3.Collect data rigorously 4.Use statistics to suggest (unlikely) patterns 5.Follow up findings with gold standard work Quant for RS, Paul Cairns
Fishing: This is bad science 1.Collect lots of data – DVs and IVs 2.Data variations could be chance 3.Test until a significant result appears 4.Report the tests that were significant 5.Claim the result is important Quant for RS, Paul Cairns
Statistical pit… … is bottomless! Safe designs – One (or two) IV – Two (or three) conditions – One primary DV Other stuff is not severely tested Quant for RS, Paul Cairns
Descriptive statistics Central tendency (averages) Spread Correlations Quant for RS, Paul Cairns
Seeing location Boxplots Median, IQR, “Range” Outliers Quant for RS, Paul Cairns
Seeing correlation Scatterplots Linear relationships! Quant for RS, Paul Cairns
Distributions Theoretical stance – Starting model Must have this! Not inferred from samples Quant for RS, Paul Cairns
Choosing a test What’s the data type? Do you know the distribution? Within or between What are you looking for? Quant for RS, Paul Cairns
Parametric tests Normal distribution Two parameters Null = one underlying normal distribution Differences in location (mean) Quant for RS, Paul Cairns
t-test: null vs alternate Quant for RS, Paul Cairns
t-test Two samples Two means Are means showing natural variation? Compare difference to natural variation Quant for RS, Paul Cairns
Effect size How interesting is the difference? – 2s difference in timings – Significance is not same as importance Cohen’s d Quant for RS, Paul Cairns
ANOVA Parametric Multiple groups Why not do pairwise comparison? Get an F value Follow up tests Quant for RS, Paul Cairns
ANOVA++ Multiple IV – So more F values! Within and between Effect size, η 2 – Amount of variance predicted by IV Quant for RS, Paul Cairns
Non-parametric tests Unknown underlying distribution Heterogeneity of variance Non-interval data Usually test location Effect size is tricky! Quant for RS, Paul Cairns
Basic tests Mann-Whitney Wilcoxon Kruskal-Wallis Friedman No accepted two-way tests Quant for RS, Paul Cairns
Health warnings Craft skill Simpler is better – Doing it – Interpreting it – Communicating it Experiments as evidence Software packages are deceptively easy Quant for RS, Paul Cairns
Questions Specific problems Specific tests Other tests? Quant for RS, Paul Cairns
Useful Reading Cairns, Cox, Research Methods for HCI: chaps 6 Rowntree, Statistics Without Tears Howell, Fundamental Statistics for the Behavioural Sciences, 6 th edn. Abelson, Statistics as Principled Argument Silver, The Signal and the Noise Quant for RS, Paul Cairns
Multivariate Multiple DV Multivariate normal distribution – Normal no matter how you slice MANOVA Null = one underlying (mv) normal distribution Quant for RS, Paul Cairns
Issues Sample size Assumptions Interpretation Communication Quant for RS, Paul Cairns
Monte Carlo Process but not distribution Generate a really large sample Compare to your sample Still theoretically driven! Quant for RS, Paul Cairns
Example Event = 4 heads in a row from a set of 20 flips of a coin You have sample of 30 sets 18 events How likely? – Get flipping! Quant for RS, Paul Cairns