advertisement

Introduction to Statistics Lecture 3 1 Covered so far Lecture 1: Terminology, distributions, mean/median/mode, dispersion – range/SD/variance, box plots and outliers, scatterplots, clustering methods e.g. UPGMA Lecture 2: Statistical inference, describing populations, distributions & their shapes, normal distribution & its curve, central limit theorem (sample mean is always normal), confidence intervals & Student’s t distribution, hypothesis testing procedure (e.g. what’s the null hypothesis), P values, one and two-tail tests 2 Lecture outline Examples of some commonly used tests: Two-Sample Inferences t-test & Mann-Whitney test chi-squared and Fisher’s exact test Correlation Paired t-test Two-sample t-test Inferences for more than two samples One-way ANOVA Two-way ANOVA Interactions in two-way ANOVA 3 t-test & Mann-Whitney test (1) t-test test whether a sample mean (of a normally distributed interval variable) significantly differs from a hypothesised value 4 t-test & Mann-Whitney test (2) Mann-Whitney test non-parametric analogue to the independent samples t-test and can be used when you do not assume that the dependent variable is a normally distributed 5 Chi-squared and Fisher’s exact test (1) Chi-squared test See if there is a relationship between two categorical variables. Note, need to confirm directionality by e.g. looking at means. 6 Chi-squared and Fisher’s exact test (2) Fisher’s exact test Same as chi-square test, but one or more of your cells has an expected frequency of five or less 7 Correlation Correlation Non-parametric . pwcorr price mpg , sig . spearman price mpg | price mpg -------------+-----------------price | 1.0000 | | mpg | -0.4686 1.0000 | 0.0000 | Number of obs = Spearman's rho = 74 -0.5419 Test of Ho: price and mpg are independent Prob > |t| = 0.0000 8 Two-Sample Inferences So far, we have dealt with inferences about µ for a single population using a single sample. Many studies are undertaken with the objective of comparing the characteristics of two populations. In such cases we need two samples, one for each population The two samples will be independent or dependent (paired) according to how they are selected 9 Example Animal studies to compare toxicities of two drugs 2 independent samples: 2 paired samples: Select sample of rats for drug 1 and another sample of rats for drug 2 Select a number of pairs of litter mates and use one of each pair for drug 1 and drug 2 10 Two Sample t-test Consider inferences on 2 independent samples We are interested in testing whether a difference exists in the population means, µ1 and µ2 Formulate hypotheses H 0 : 2 1 0 H a : 2 1 0 11 Two Sample t-Test x x It is natural to consider the statistic 2 1 and its sampling distribution The distribution is centred at µ2-µ1, with standard error 12 22 n1 n2 If the two populations are normal, the sampling distribution is normal For large sample sizes (n1 and n2 > 30), the sampling distribution is approximately normal even if the two populations are not normal (CLT) 12 Two Sample t-Test The two-sample t-statistic is defined as ( x2 x1 ) ( 2 1 ) t 1 1 , whe re sp n1 n2 (n1 1) s (n2 1) s s n1 n2 2 2 p 2 1 2 2 The two sample standard deviations are combined to give a pooled estimate of the population standard deviation σ 13 Two-sample Inference The t statistic has n1+n2-2 degrees of freedom Calculate critical value & p value as per usual The 95% confidence interval for µ2-µ1 is ( x2 x1 ) t 0.025s p 1 1 n1 n2 14 Example Population Drug 1 Drug 2 n 20 38 mean 35.9 36.6 s 11.9 12.3 2 2 ( n 1 ) s ( n 1 ) s 1 2 2 s 2p 1 n1 n2 2 (19)(141 .61) (37 )(151 .29) 56 148 .01 15 Example (contd) ( x2 x1 ) 0 t 1 1 2 sp n1 n2 -0.21 Two-tailed test with 56 df and α=0.05 therefore we reject the null hypthesis if t>2 or t<-2 Fail to reject - there is insufficient evidence of a difference in mean between the two drug populations Confidence interval is -7.42 to 6.02 16 Paired t-test Methods for independent samples are not appropriate for paired data. Two related observations (i.e. two observations per subject) and you want to see if the means on these two normally distributed interval variables differ from one another. Calculation of the t-statistic, 95% confidence intervals for the mean difference and P-values are estimated as presented previously for one-sample testing. 17 Example 14 cardiac patients were placed on a special diet to lose weight. Their weights (kg) were recorded before starting the diet and after one month on the diet Question: Do the data provide evidence that the diet is effective? 18 Patient Before After Difference 1 62 59 3 2 62 60 2 3 65 63 2 4 88 78 10 5 76 75 1 6 57 58 -1 7 60 60 0 8 59 52 7 9 54 52 2 10 68 65 3 11 65 66 -1 12 63 59 4 13 60 58 2 14 56 55 1 19 Example H 0 : d 0 H a : d 0 xd 2.5 sd 2.98 n 14 xd 0 2.5 t 3.14 sd 2.98 14 n 20 Example (contd) Critical Region (1 tailed) t > 1.771 Reject H0 in favour of Ha P value is the area to the right of 3.14 = 1-0.9961=0.0039 95% Confidence Interval for 2.5 ± 2.17 (2.98/√14) = 2.5 ±1.72 =0.78 to 4.22 d 1 2 21 Example (cont) Suppose these data were (incorrectly) analysed as if the two samples were independent… t=0.80 22 Example (contd) We calculate t=0.80 This is an upper tailed test with 26 df and α=0.05 (5% level of significance) therefore we reject H0 if t>1.706 Fail to reject - there is not sufficient evidence of a difference in mean between ‘before’ and ‘after’ weights 23 Wrong Conclusions By ignoring the paired structure of the data, we incorrectly conclude that there was no evidence of diet effectiveness. When pairing is ignored, the variability is inflated by the subject-to-subject variation. The paired analysis eliminates this source of variability from the calculations, whereas the unpaired analysis includes it. Take home message: NB to use the right test for your data. If data is paired, use a test that accounts for this. 24 50% of slides complete! 25 Analysis of Variance (ANOVA) Many investigations involved a comparison of more than two population means Need to be able to extend our two sample methods to situations involving more than two samples i.e. equivalent of the paired samples t-test, but allows for two or more levels of the categorical variable Tests whether the mean of the dependent variable differs by the categorical variable Such methods are known collectively as the analysis of variance 26 Completely Randomised Design/one-way ANOVA Equivalent to independent samples design for two populations A completely randomised design is frequently referred to as a one-way ANOVA Used when you have a categorical independent variable (with two or more categories) and a normally distributed interval dependent variable (e.g. $10,000,$15,000,$20,000) and you wish to test for differences in the means of the dependent variable broken down by the levels of the independent variable e.g. compare three methods for measuring tablet hardness. 15 tablets are randomly assigned to three groups of 5 and each group is measured by one of these methods 27 ANOVA example See that the students in the academic program have the highest mean writing score, while students in the vocational program have the lowest. Mean of the dependent variable differs significantly among the levels of program type. However, we do not know if the difference is between only two of the levels or all three of the levels. 28 Example Compare three methods for measuring tablet hardness. 15 tablets are randomly assigned to three groups of 5 Method A Method B Method C 102 99 103 101 100 100 101 99 99 100 101 104 102 98 102 29 Hypothesis Tests: One-way ANOVA K populations H 0 : 1 2 ... k H A : at least one is different 30 Do the samples come from different populations? Two-sample (t-test) YES NO Ho DATA Ha A B 31 Do the samples come from different populations? One-way ANOVA (F-test) A Ho DATA B AB C A AC C Ha BC B 32 F-test The ANOVA extension of the t-test is called the F-test Basis: We can decompose the total variation in the study into sums of squares Tabulate in an ANOVA table 33 Decomposition of total variability (sum of squares) Assign subscripts to the data i is for treatment (or method in this case) j are the observations made within treatment e.g. y11= first observation for Method A i.e. 102 y1. = average for Method A Using algebra Total Sum of Squares (SST)=Treatment Sum of Squares (SSX) + Error Sum of Squares (SSE) 2 ( y y ) ij 2 2 ( y y ) ( y y ) i. ij i. 34 ANOVA table df SS MS F P-value Treatment (between groups) df (X) SSX SSX df (X) MSX MSE } Look up ! Error (within groups) Total df (E) SSE SSE df (E) } df (T) SST 35 Example (Contd) Are any of the methods different? P-value=0.0735 At the 5% level of significance, there is no evidence that the 3 methods differ 36 Two-Way ANOVA Often, we wish to study 2 (or more) independent variables (factors) in a single experiment An ANOVA of observations each of which can be classified in two ways is called a two-way ANOVA 37 Randomised Block Design This is an extension of the paired samples situation to more than two populations A block consists of homogenous items and is equivalent to a pair in the paired samples design The randomised block design is generally more powerful than the completely randomised design (/one way anova) because the variation between blocks is removed from the test statistic 38 Decomposition of sums of squares 2 2 2 2 ( y y ) ( y y ) ( y y ) ( y y y y ) ij i. .j ij i. j. Total SS = Between Blocks SS + Between Treatments SS + Error SS Similar to the one-way ANOVA, we can decompose the overall variability in the data (total SS) into components describing variation relating to the factors (block, treatment) & the error (what’s left over) We compare Block SS and Treatment SS with the Error SS (a signal-to-noise ratio) to form Fstatistics, from which we get a p-value 39 Example An experiment was conducted to compare the mean bioavailabilty (as measured by AUC) of three drug products from laboratory rats. Eight litters (each consisting of three rats) were used for the experiment. Each litter constitutes a block and the rats within each litter are randomly allocated to the three drug products 40 Example (cont’d) Litter Product A Product B Product C 1 89 83 94 2 93 75 78 3 87 75 89 4 80 76 85 5 80 77 84 6 87 73 84 7 82 80 75 8 68 77 75 41 Example (cont’d): ANOVA table Source df SS MS F-ratio P-value Product Litter Error Total 2 7 14 23 200.333 391.833 405.667 997.833 100.167 55.9762 28.9762 3.4569 1.9318 0.0602 0.1394 42 Interactions The previous tests for block and treatment are called tests for main effects Interaction effects happen when the effects of one factor are different depending on the level (category) of the other factor 43 Example 24 patients in total randomised to either Placebo or Prozac Happiness score recorded Also, patients gender may be of interest & recorded There are two factors in the experiment: treatment & gender Two-way ANOVA 44 Example Tests for Main effects: Treatment: are patients happier on placebo or prozac? Gender: do males and females differ in score? Tests for Interaction: Treatment x Gender: Males may be happier on prozac than placebo, but females not be happier on prozac than placebo. Also vice versa. Is there any evidence for these scenarios? Include interaction in the model, along with the two factors treatment & gender 45 More jargon: factors, levels & cells Happiness score Factor 2 Treatment Levels Placebo Prozac Male 3 4 2 3 4 3 7 7 6 5 6 6 Female 4 5 4 6 6 4.5 5 5 5 4 6 6 Factor 1 Gender Cells 46 What do interactions looks like? H a p p i n e s s No H a p p i n e s s Placebo Prozac NO INTERACTION! H a p p i n e s s Yes Placebo Prozac Yes Placebo H a p p i n e s s Prozac Yes Placebo Prozac 47 Results Tests of Between-Subjects Effects Dependent Variable: Happiness Source Corrected Model Intercept Drug Gender Drug * Gender Error Total Corrected Total Type III Sum of Squares 28.031a 565.510 15.844 .844 11.344 12.708 606.250 40.740 df 3 1 1 1 1 20 24 23 Mean Square 9.344 565.510 15.844 .844 11.344 .635 F 14.705 889.984 24.934 1.328 17.852 Sig. .000 .000 .000 .263 .000 a. R Squared = .688 (Adjusted R Squared = .641) 48 Interaction? Plot the means Estimated Marginal Means of Happiness Gender 1.0 2.0 Estimated Marginal Means 6.0 5.0 4.0 3.0 1.0 2.0 Drug 49 Example: Conclusions Significant evidence that drug treatment affects happiness in depressed patients (p<0.001) Prozac is effective, placebo is not No significant evidence that gender affects happiness (p=0.263) Significant evidence of an interaction between gender and treatment (p<0.001) Prozac is effective in men but not in women!!* 50 After the break… Regression Correlation in more detail Multiple Regression ANCOVA Normality Checks Non-parametrics Sample Size Calculations 51