Analysis of Variance (ANOVA)

Size: px

Start display at page:

Download "Analysis of Variance (ANOVA)"

Tamsyn Bruce
6 years ago
Views:

Analysis of Variance (ANOVA) Two types of ANOVA tests: Independent measures and Repeated measures Comparing 2 means: X 1 = 20 t - test X 2 = 30 How can we Compare 3 means?

1 Analysis of Variance (ANOVA) Two types of ANOVA tests: Independent measures and Repeated measures Comparing 2 means: X 1 = 20 t - test X 2 = 30 How can we Compare 3 means?: X 1 = 20 X 2 = 30 X 3 = 35 ANOVA Test compares three or more groups or conditions For ANOVA we work with variances instead of means Why use ANOVA (3 advantages): (a) ANOVA can test for trends in our data. (b) ANOVA is preferable to performing many t-tests on the same data (avoids increasing the risk of Type 1 error). Effect of fertilizer on plant height (cm) (too much of a good thing) Suppose we have 3 groups. We will have to compare: Population of seeds Randomly select 40 seeds HEIGHT of plant (Cm) No Fertilizer 10 gm Fertilizer 20 gm Fertilizer 30 gm Fertilizer group 1 with group 2 group 1 with group 3 group 2 with group 3 Each time we perform a test there is a (small) probability of rejecting the true null hypothesis. These probabilities add up. So we want a single test. Which is ANOVA. 1

2 (c) ANOVA can be used to compare groups that differ on two, three or more independent variables, and can detect interactions between them. Age-differences in the effects of alcohol on motor coordination: Independent-Measures ANOVA: Each subject participates in only one condition in the experiment (which is why it is called independent measures). score (errors) An independent-measures ANOVA is equivalent to an independent-measures t-test, except that you have more than two groups of subjects. Alcohol dosage (number of drinks) Logic behind ANOVA: Example Effects of caffeine on memory: FOUR GROUPS: each group gets a different amount of caffeine, followed by a memory test (words remembered from a list) Variation in the set of scores comes from TWO sources: Random variation from the subjects themselves (due to individual variations in motivation, aptitude, mood, ability to understand instructions, etc.). Systematic variation produced by the experimental manipulation. 2

Random variation Systematic variation + Random variation ANOVA compares the amount of systematic variation to the amount of random variation, to produce an F-ratio: F = systematic variation random

3 Random variation Systematic variation + Random variation ANOVA compares the amount of systematic variation to the amount of random variation, to produce an F-ratio: F = systematic variation random variation ( error ) Large value of F: a lot of the overall variation in scores is due to the experimental manipulation, rather than to random variation between subjects. Small value of F: the variation in scores produced by the experimental manipulation is small, compared to random variation between subjects. Analysis of variance implies analyzing or breaking down variance. We start by breaking down Sum of Squares or SS. We saw these first when we did the t-test. 9 calculations go into doing an ANOVA test (read this figure from bottom to top with F as your final calculation) 2 sum of squares = #(X " X ) 1 Calculation 1 Calculation 1 Calculation We divide SS by the appropriate "degrees of freedom" (usually the number of groups or subjects minus 1) to get variance. 3 Calculations 3 Calculations 3

4 step 1 The null hypothesis: step 2 Total SS H 0 : µ 1 = µ 2 = µ 3 = µ 4 No treatment effect set alpha level: α =.05 steps 2, 3 & 4 Calculate 3 SS values: 1) Total 2) Between treatments 3) Within treatments Total SS " G ) 2 SS Total = 297 G = 9.5 step 3 Between treatments SS step 4 Within treatments SS X 1 = 4 X 2 = 9 X 3 =12 X 4 =13 n = number of participants SS between treatments = n [(X 1 " G ) 2 + (X 2 " G ) 2 + (X 3 " G ) 2 + (X 4 " G ) 2 ] SS between treatments = 245 X 1 = 4 X 2 = 9 X 3 =12 X 4 =13 SS 1 " X 1 ) 2 SS 3 " X 3 ) 2 SS 2 " X 2 ) 2 SS 4 " X 4 ) 2 SS within treatments = SS 1 + SS 2 +SS 3 +SS 4 = 52 4

step 5 Calculating df (total, between treatments & within treatments) step 6 Calculating the Between treatments and Within treatments Variance or Mean squares (MS) MS = SS df The ANOVA summary table:

(Within-treatments SS) (total df) = (between-treatments df ) + (within-treatments df ) step 7 Computing F and Assessing its significance: F = MS between treatments MS within treatments = 81.67 3.

5 step 5 Calculating df (total, between treatments & within treatments) step 6 Calculating the Between treatments and Within treatments Variance or Mean squares (MS) MS = SS df The ANOVA summary table: df total = All scores 1 = 19 df between treatments = Number of treatments 1 = 3 df within treatments = df 1 + df 2 + df 3 + df 4 = = 16 (Total SS) = (Between-treatments SS) + (Within-treatments SS) (total df) = (between-treatments df ) + (within-treatments df ) step 7 Computing F and Assessing its significance: F = MS between treatments MS within treatments = = The bigger the F-ratio, the less likely it is to have arisen merely by chance and the more likely it is that you can reject the null hypothesis. Use the between-treatments and within-treatments df to find the critical value of F. Your F is significant if it is equal to or larger than the critical value in the table. Here, look up the critical F-value for 3 and 16 df Columns correspond to between-treatments df; rows correspond to within-treatments df Here, go along 3 and down 16: critical F is at the intersection. Our obtained F = 25.13, is bigger than 3.24; it is therefore significant at p <

Interpreting the Results: A significant F-ratio merely tells us is that there is a statistically-significant difference between our experimental conditions; it does not say where the difference comes

6 Interpreting the Results: A significant F-ratio merely tells us is that there is a statistically-significant difference between our experimental conditions; it does not say where the difference comes from. In our example, it tells us that caffeine dosage does make a difference to memory performance. BUT the difference may be ONLY between: Caffeine VERSUS No-Caffeine To pinpoint the source of the difference we can do: (t tests) (a) planned comparisons - comparisons between (two) groups which you decide to make in advance of collecting the data. (b) post hoc tests - comparisons between (two) groups which you decide to make after collecting the data: Many different types - e.g. Newman-Keuls, Scheffé, Bonferroni. AND there might be NO difference between: Large dose of Caffeine VERSUS Small Dose of Caffeine Assumptions underlying ANOVA: ANOVA is a parametric test (like the t-test) It assumes: (a) data are interval or ratio measurements; (b) conditions show homogeneity of variance; (c) scores in each condition are roughly normally distributed. Using SPSS for a one-way independent-measures ANOVA on effects of alcohol on time taken on some task such as whack a mole Three groups (10 individuals in each) Treatments: Group 1: two drinks Group 2: one drink Group 3: no alcohol Dependent variable: Time taken to whack 20 ten moles 6

7 RUNNING SPSS (Analyze > compare means > One Way ANOVA) Data Entry 7

Linear trend: Quadratic trend: (one change in direction) Cubic trend: (two changes in direction) With two groups, you can only test for a

8 Click Options Then Click Boxes: Descriptive; Homogeneity of variance test; Means plot SPSS output Trend tests: (Makes sense only when levels of Independent Variable correspond to differing amounts of something - such as caffeine dosage - which can be meaningfully ordered). Linear trend: Quadratic trend: (one change in direction) Cubic trend: (two changes in direction) With two groups, you can only test for a linear trend. With three groups, you can test for linear and quadratic trends. With four groups, you can test for linear, quadratic and cubic trends. 8

9 Conclusions: One-way independent-measures ANOVA enables comparisons between 3 or more groups that represent different levels of one independent variable. A parametric test, so the data must be interval or ratio scores; be normally distributed; and show homogeneity of variance. ANOVA avoids increasing the risk of a Type 1 error. 9

Analysis of Variance: Repeated measures

Repeated-Measures ANOVA: Analysis of Variance: Repeated measures Each subject participates in all conditions in the experiment (which is why it is called repeated measures). A repeated-measures ANOVA is