Les expriences contrles Plusieurs mthodes dvaluation vues en

Les expériences contrôlées

Plusieurs méthodes d’évaluation vues en LOG 350 … • • • Sondages Évaluation heuristique Tests d’utilisabilité Expériences Etc.

Les expériences • Une partie fondamentale de la méthode scientifique • Permettent de trouver des relations causales entres des conditions et leurs effets • En IHM, permettent de trouver si une interface A est plus rapide/cause moins d’erreurs/etc. qu’une interface B

Les expériences • On varie (manipule) au moins une variable (exemple: l’interface à utiliser). C’est la variable indépendante. Chaqu’une de ses valeurs correspond à une condition. • On mesure au moins une variable (exemples: le temps, le nombre d’erreurs, la satisfaction subjective). C’est la variable dépendante. • On analyse les résultats pour voir s’il y a des différences significatives.

Exemple d’expérience • Les « expanding targets » Référence: M. Mc. Guffin, R. Balakrishnan (2002). Acquisition of Expanding Targets. Proceedings of ACM Conference on Human Factors in Computing Systems (CHI) 2002, pages 5764, http: //doi. acm. org/10. 1145/503376. 503388

Example: Mac OS X • Does this really make acquisition easier ?

Additional motivation Furnas Generalized fisheye views CHI 1986 Mackinlay, Robertson, Card The Perspective Wall CHI 1991 Bederson Fisheye Menus UIST 2000

Fitts’ Law A Target Cursor W

Fitts’ Law Target 1 Target 2 Same ID → Same Difficulty

Fitts’ Law Target 1 Target 2 Smaller ID → Easier

Fitts’ Law Target 1 Target 2 Larger ID → Harder

Fitts’ Law W Open-loop Closed-loop Speed Overshoot Undershoot Distance

Expanding Targets Basic Idea: • Big targets can be acquired faster, but take up more screen space • So: keep targets small until user heads toward them Click Me ! Okay Cancel

Experimental Setup W Target Start Position A

Experimental Setup Expansion: • How ? Animated Expansion

Experimental Setup Expansion: • How ? Fade-in Expansion

Experimental Setup Expansion: • How ? • When ? P = 0. 25

Experimental Setup Expansion: • How ? • When ? P = 0. 5

Experimental Setup Expansion: • How ? • When ? P = 0. 75

Pilot Study 7 conditions: • No expansion (to establish a, b values) • Expanding targets – Either animated growth or fade-in – P is one of 0. 25, 0. 75 (Expansion was always by a factor of 2)

Pilot Study 7 conditions x 16 (A, W) values x 5 repetitions x 2 blocks x 3 participants = 3360 trials

Pilot Study: Results Time (seconds) ID (index of difficulty)

Pilot Study: Results Time (seconds) ID (index of difficulty)

Pilot Study: Results Time (seconds) ID (index of difficulty)

Pilot Study: Results Time (seconds) P = 0. 25 ID (index of difficulty)

Pilot Study: Results Time (seconds) P = 0. 5 ID (index of difficulty)

Pilot Study: Results Time (seconds) P = 0. 75 ID (index of difficulty)

Implications • Pilot Study suggests the advantage of expansion doesn’t depend on P • So, set P = 0. 9 and perform a more rigorous study

Full Study 2 conditions: • No expansion (to establish a, b values) • Expanding targets, with – Animated growth – P = 0. 9 – Expansion factor of 2

Full Study 2 conditions x 13 (A, W) values x 5 repetitions x 5 blocks x 12 participants = 7800 trials

Results Time (seconds) A, W values

Results Time (seconds) ID (index of difficulty)

Results Time (seconds) ID (index of difficulty)

Results Time (seconds) ID (index of difficulty)

Results Time (seconds) P = 0. 9 ID (index of difficulty)

Implications • For single-target selection task, – Expansion yields a significant advantage, even when P=0. 9 • What about multiple targets ?

(Fin des diapos sur les « expanding targets » )

Les variables dans une expérience • Variables indépendantes: celles qu’on manipule (on les appelle aussi les facteurs); correspondent aux conditions (ou traitements ou niveaux) • Variables dépendantes: celles qu’on mesure • Variables de contrôle: celles qu’on contrôle, c. -à-d. qu’on essaie de garder constantes entre les conditions • Variables aléatoires: celles qu’on laisse varier, de manière le plus aléatoire possible. – Exemples: âge, sexe, profil socio-économique, etc. – Comment assurer une variation aléatoire entre les conditions ? • Assignation aléatoire des participants aux conditions – Désavantage: Ces variables vont introduire plus de variabilité dans nos résultats – Avantage: Nos résultats seront plus généraux; nos conclusions vont s’appliquer à plus de situations • Variables confondantes: celles qui varient de manière systématique entre les conditions. On veut éliminer ces variables!

Régression linéaire Y X • Sortie: pente, intersection, et coéfficient de corrélation de Pearson r qui est dans l’intervalle [-1, 1]

Un lien causal … • Dans une expérience bien contrôlée, s’il n’y a pas de variables confondantes, et on trouve que les variable dépendantes changent lorsqu’on change les variables indépendantes, on peut conclure qu’il y a un lien causal: le changements dans les variables indépendantes cause le changement dans les variables dépendantes. Dans ce cas, une corrélation impliquerait un lien causal.

… versus une corrélation simple • Par contre, si on ne fait qu’observer une corrélation entre deux variables X et Y, sans contrôler les conditions, cela n’implique pas un lien causal entre eux. Il se pourrait que – X a un effet sur Y – Y a un effet sur X – Une troisième variable, Z, a un effet sur X et Y • C’est pour ça qu’on essaie d’éliminer les variables confondantes dans les expériences

Exemple • Des chercheurs voulait savoir quelle variable pourrait prédire les chances qu’un conducteur de motocyclette ait un accident de moto. Ils ont cherché des corrélations entre le nombre d’accidents, et l’âge, le niveau socioéconomique, etc. • Ils ont trouvé que la plus forte corrélation était avec le nombre de tatous du conducteur. • Évidemment, les tatous ne causent pas les accidents, ni l’inverse.

Examples of Questions to Answer in an Experiment • Of 3 interfaces, A, B, C, which enables fastest performance at a given task? • Does prozac have an effect on performance at tying shoe laces? • How does frequency of advertisements on television affect voting behaivour? • Can casting a spell on a pair of dice affect what numbers appear on them?

Elements of an Experiment • Population – Set of all possible subjects / observations • Sample – Subset of the population chosen for study; a set of subjects / observations • Subjects – People/users under study. The more politically correct term within HCI is “participants”. • Observations / Dependent variable(s) – Individual data points that are measured/collected/recorded • E. g. time to complete a task, errors, etc. • Condition / Treatment / Independent variables(s) – Something done to the samples that distinguishes them (e. g. giving a drug vs placebo, or using interface A vs B) – Goal of experiment is often to determine whether the conditions have an effect on observations, and what the effect is

Tasks to Design and Run an Experiment • Design – – – Choose independent variables Choose dependent variables Develop hypothesis Choose design paradigm Choose control procedures Choose a sample size • Pilot experiment – Often more exploratory, varying a greater number of variables to get a “feel” for where the effect(s) might be • Run experiment – Focuses in on the suspected effect; tries to gather lots of data under key or optimal conditions to result in a strong conclusion • Analyze data – Using statistical tests such as ANOVA • Interpret results

Hypothesis • Statement, to be tested, of relationship between independent and dependent variables • The null hypothesis is that the independent variables have no effect on the dependent variables

Experimental Design Paradigms • Between subjects or within subjects manipulation (entre participants vs à travers tous les participants) • Example: designs with one independent variable – Between subjects design • One independent variable with 2 or more levels • Subjects randomly assigned to groups • Each subject tested under only 1 condition – Within subject design • One independent variable with 2 or more levels • Each subject tested under all conditions • Order of conditions randomized or counterbalanced (why? )

What To Control • Subject characteristics – Gender, handedness, etc. – Ability – Experience • Task variables – Instructions – Materials used • Environmental variables – Setting – Noise, light, etc. • Order effects – Practice – Fatigue

How to Control for Order Effects • Counterbalancing – Factorial Design – Latin Square

Data Analysis and Hypothesis Testing • Describe data – Descriptive statistics (means, medians, standard deviations) – Graphs and tables • Perform statistical analysis of results – Are results due to chance? (That is, with what probability)

ANOVA • “Analysis of Variance” • A statistical test that compares the distributions of multiple samples, and determines the probability that differences in the distributions are due to chance • In other words, it determines the probability that the null hypothesis is correct • If probability is below 0. 05 (i. e. 5 %), then we reject the null hypothesis, and we say that we have a (statistically) significant result – Why 0. 05 ? Dangers of using this value ?

Techniques for Making Experiment more “Powerful” (i. e. able to detect effects) • Reduce noise (i. e. reduce variance) – Increase sample size – Control for confounding variables • E. g. psychologists often use in-bred rats for experiments ! • Increase the magnitude of the effect – E. g. give a larger dosage of the drug

Uses of Controlled Experiments within HCI • Evaluate or compare existing systems/features/interfaces • Discover and test useful scientific principles – Examples ? • Establish benchmarks/standards/guidelines – Examples ?

Exemple d’un plan d’expérience … • Pour chaque participant … – Pour chaque condition majeure. . . * • On fait des essais de réchauffement • On a un certain nombre de blocs, séparés par des pauses • Pour chaque bloc … • On répète chaque condition mineure un certain nombre de fois * • * Comment ordonner ces choses ?