Review. Number of variables. Standard Scores. Anecdotal / Clinical. Bivariate relationships. Ch. 3: Correlation & Linear Regression

Transcription

1 Ch. 3: Correlation & Relationships between variables Scatterplots Exercise Correlation Race / DNA Review Why numbers? Distribution & Graphs : Histogram Central Tendency Mean (SD) The Central Limit Theorem Logic and Logical Fallacies Precision vs. Accuracy Population vs. Sample- Law of Large Numbers Measurement Scales Norms and Standard Scores: Z, IQ, T 1 57 Standard Scores Number of variables One variable, one dimension Number Line Frequency Distribution / Histogram dimensional graph of 1D data Percent of cases within each range Standard Deviations Percentile rank Z score T score Standard Scores Difference Score 1 dimension dimensions IQ score Bivariate relationships is factor A related to factor B? Methods of analysis: Anecdotal / Clinical - before systematic research Numerically -- check extremes Visually -- scatterplots see relationships and problems w/data can t test hypothesis Statistically -- correlation & regression hard to detect problems w/data easy to test hypothesis Anecdotal / Clinical Many interesting findings began from nonscientific approaches Intuition that something is related through experiencing multiple situations Pattern recognition - Good and Bad Problems -- faulty memory, confirmation biases, prejudice, etc Next step after a gut feeling : design experiment and collect data. 1

2 Simple numerical analysis Simplify the situation by using Categorical variables (or reducing Continuous variables to Categorical variables) Use extreme cases to maximize effect Compute percentages in a x matrix Do the results suggest an effect? Compute Chi-square statistic to judge significance Example I think there is brain dysfunction in HIV disease as measured by neuropsychological testing Medical status: control vs. HIV+ symptomatic NP test results: normal vs. impaired NP Status Medical Status Control HIV+ Normal 5% 5% Impaired 15% % 3 Standard Scores Cases below -1.0 SD are about 15% of the total x Analysis Pro: easy to understand Con: dividing continuous variables into binary reduces power Percent of cases within each range Standard Deviations Percentile rank Z score T score Standard Scores Graphical and Statistical methods should be used as well. IQ score Scatterplots Graph two variables in relation to each other on two-dimensional, axis Easy to see relations problems Can t prove relationship is significant Difficult to interpret clinically or in common sense terms x y Scatterplots 9

3 Assume that two variables are related, and that this relationship is linear -- model the data by a simple straight line for the data. For any given data set, we pick the line that best fits our data Similar terms: linear regression, fitting a line, finding the trend, creating a trendline, best fit line, etc. Residuals = difference between prediction and actual value minimizes the square of the residuals, often called Ordinary Least Squares Why Regression Frances Galton Height of children vs parents. Tall parents have tall children (and vice versa) But children are closer to the mean than their parents (by a factor of ~/3) Galton called this Regression to the Mean His paper fit** straight lines to data points. The technique has been called regression ever since ** He never calculated the lines, he just eyeballed them Anscombe s Quartet I Equation: y = x Correlation rx,y = 0. Which line fits the best? 0 x y Anscombe s Quartet II Anscombe s Quartet III x y x y

4 Anscombe s Quartet IV Anscombe s Quartet x y Anscombe s Quartet Summary Each series has the same Quantitative stats: linear regression equations correlations Each one is Qualitatively different Each series needs special handling Lesson Graph our Data Equation = a + b = predicted = actual b = slope D/D ( rise over run) a = intercept value when =0 0 run rise 79 0 Residuals in : independent variable : dependent variable Model: predict from : ( prime) : predicted = a + b Prediction is imperfect. Difference between predicted ( ) and actual () is called a Residual = ( - ) Calculation of best fit line minimizes the sum of the squared residuals Σ(- ) Residuals in Residual is difference between actual and predicted ( - ) Graphically it is equal to how far away (vertically) a point is from the linear regression line Residual 5

5 Residuals and Error Residuals (error) are greater when values are further from prediction. More Error 0 Less Error 0 Residuals d i = y i y i Difference between predicted and actual y value 9 Sum of Squares SST = N i=1 (y i ȳ) SSR = N i=1 d i SSR = N i=1 (y i y i ) Sum of Squared Residuals Residual = ( - ) Squared residual = (- ) SSR: Sum of squared residuals Linear regression minimizes this value SSR is hard to interpret R R = 1 SSR SST R = 1 - (SSR/SST) Ranges from 0 to 1 (0% to 0%) R Terminology Coefficient of Determination Explained Variance Shared Variance Meaning what % of variation in the values can we predict if we are given values Correlation: not causation 9 93

6 Interactive Correlation Demo Standard Error of Estimate Residual = ( - ) Standard Deviation of residuals measure of average error aka Standard Error of Estimate In Prism: Sy.x 9 57 Family Tree Charles Darwin (09-) Francis Galton (-1911) Karl Pearson (57-193) Correlation (r) Pearson s r Pearson s Product-Moment Correlation Measures the strength of the linear relationship between two variables Ranges between -1.0 and +1.0 Is a special case of linear regression, when both and have been turned into Z scores. r is transitive commutative (correlation between and is same as correlation between and ) R = explained variance is the proportion of variation in the data explained by the model. R ranges from 0 to 1.0 (0% to 0%) 5 59 Correlations Regression vs. Correlation Linear Regression Correlation Scores Raw Z R = 0% R = 9% R = 5% Mean, Std Dev sample means sample Std Dev Equation = a + b = r 0 1 Slope b = change in per change in r = correlation coefficient Slope meaningless R = % variance explained R = 9% R = 1% R = 9% Commutative? no yes, Rxy = Ryx

7 Other Correlation Coefficients Continuous (interval & ratio): Pearson s r Ordinal (Ranked): A B C D 1st, nd, 3rd... Spearman s Rho: correlation between two ordinal / ranked variables. Dichotomous (yes/no, one/zero, T/F, Male/ Female, Pass/Fail...) True vs. Artificial? Continuous vs. Dichotomous Type of / Type of Continuous Artificial Dichotomous True Dichotomous Continuous Pearson r Biserial r Point biserial r Artificial Dichotomous True Dichotomous Biserial r Tetrachroic r Phi Point biserial r Phi Phi Correlation : Issues Technical / Calculation : Non-normal distribution Non-linear data and relationships Outliers, data errors Restricted Range Interpretation: Correlation =? Causation Third variable explanations Non-linearity & Correlation assume a linear relationship between and When it s not linear: Restrict the range of Transform (log, square root, etc.) other statistical analyses (Spearman s Rho ) Life expectancy / national income Restrict range of

8 log transform (or ) Outliers & Data Errors? Correlation = Causation? A relationship (linear or otherwise) between and tells us nothing about whether causes Lack of correlation between and does not mean that doesn t cause Ice cream sales are positively related to increases in drowning deaths Hypothesis Testing All parameters (equations) we estimate from data have inherent error How do we know if a given estimate is correct? How big is the error likely to be (confidence intervals)? Inferential Statistics - covered later Formulas to calculate probability, confidence intervals. Higher N is better statistical significance not the same as clinical significance 5 5 Statistical vs Clinical Significance Regarding the change in the Dependent Variable (DV) Statistical Significance: Could the change be due to chance? P value (p <.05 : less than 5% probability) Clinical Significance Was the change big enough to matter? Effect Size (R ) Depends on context Significance vs. Effect Size Two coin flips 0% heads big effect, not statistically significant 00 coin flips 9.% heads small effect, statistically significant 00 coin flips 35% heads big effect, statistically significant 5 5

9 Lies, damned lies, and statistics Statistical significance (P) is a function of Errors of measurement (E) Effect Size (D) Sample Size (N) Reporting Results Men had higher IQ than women. Results were statistically significant p <.001 P-value : yes Effect Size :? p ~ E / (D x N) Review : Is race real? Pre-DNA Gold, Silver, Brass, Iron -- Plato Pre-DNA theory Post-DNA theory There is a physical difference between the white and black races which I believe will for ever forbid the two races living together on terms of social and political equality. -Abraham Lincoln Genetics : DNA Genetics Human genome contains about billion pairs of deoxyribonucleic acid (DNA) DNA is Transcribed into RNA RNA is Translated into Proteins Proteins serve as structural components function as enzymes to catalyze biochemical reactions Human DNA is grouped into chromosomes 3 pairs, one of each pair comes from each parent pairs in both males and females (autosomes) 1 pair determines sex: either (females) or (males)

10 Post-DNA theory Variance variation between individuals aka variation within groups variation between groups Variance variation between individuals : 3mbp / person variation within groups : 5% variation between groups: 15% about 5% - within races about % - between races Genetic Differences Fst = % of subpopulation variance DNA Differences Identical Twins 0.0% Human vs. Human 0.1% Humans vs Gorillas 1.% Humans vs Chimps:.0% Humans vs. Cats.0% Twins Strangers Gorillas Chimp Cat 01 Variance: Genetic Variation Within local populations Within race Between race 5% For example: 5% within Japanese 5% between Japanese & Korean % between Asian and Caucasian 0 % 5% Prehistorical Migration 0