Basic Statistics and Data Analysis for Health Researchers from Foreign Countries

Transcription

1 Basic Statistics and Data Analysis for Health Researchers from Foreign Countries Volkert Siersma The Research Unit for General Practice in Copenhagen Dias 1

2 Content Quantifying association between continuous variables. In particular: Correlation (Simple) regression Dias 2

3 Example Newly diagnosed Type 2 Diabetes pt glucose bmi sex age A data set with 729 newly diagnosed Type 2 diabetes patients. pt: Patient ID glucose: Diagnostic plasma glucose (mmol/l) bmi: sex: age: Body Mass Index (kg/m2) sex (1=male, 0=female) age (years) Dias 3

4 Research question Do fat people have a more severe diabetes when the diabetes is discovered? Or in a more statistical language: Is diagnostic plasma glucose (positively) associated with the body mass index at the time of diagnosis? Dias 4

5 Scatter-plot When investigating a potential association between only two variables (like diagnostic plasma glucose and BMI) a scatterplot is an important part of the analysis. It gives insight in the nature of the association. It shows problems in the data, e.g. outliers, strange or impossible values. Dias 5

6 Scatter-plot Dias 6

7 Scatter-plot There is no apparent tendency, specifically not one that would support our research question and if we have to point out a tendency, it would be that high BMI associates with lower diagnostic glucose (why is this not so strange if we think about the diagnosis of diabetes?). There seem to be some very large values, especially for diagnostic plasma glucose. These are valid measurements. Maybe a log transformation of glucose would make associations more apparent? Dias 7

8 Scatter-plot R code plot(diabetes$bmi,diabetes$glucose, frame=true, main=null, xlab= BMI (kg/m2), ylab= Glucose (mmol/l), col= green, pch=19) Dias 8

9 Scatter-plot log transformation Dias 9

10 Measures of association We want to capture the association between two variables in a single number: a correlation coefficient, a measure of association. Suppose that Y i is the diagnostic plasma glucose of patient i and X i the BMI for the same person. Then we want our measure of association to have the following characteristics: A positive association indicates that if X i is large (relative to the rest of the sample) then Y i is likely to be large as well. A negative association indicates that if X i is large then Y i is likely to be small. Dias 10

11 Measures of association between -1 and 1 0 : No association 1 : perfect positive association -1 : Perfect negative association Dias 11

12 Measures of association for the diabetes data r = ρ = τ = Dias 12

13 Measures of association for the diabetes data and log transformed r = ρ = τ = Only the first one changes! Dias 13

14 Pearson s correlation coefficient Pearson s correlation coefficient is computed from the data set (X i, Y i ), i = 1,,N as: X Y r = N i= 1 ( X where and are the respective means and SD x and SD y the respective standard deviations. i X )( Y ( N 1) SD SD x i Y ) y Dias 14

15 Characteristics of Pearson s correlation coefficient Pearson s correlation coefficient has the following properties: It measures the degree of linear association. It is invariant to linear change of scale for the variables. It is not robust to outliers. Coefficient values that are comparable between different data sets, and moreover a valid confidence interval and p-value, require that both X i and Y i are normally distributed. Dias 15

16 Pearson s correlation coefficient R code > cor(diabetes$bmi,diabetes$glucose,use= complete.obs ) [1] Gives only the correlation coefficient. > cor.test(diabetes$bmi,diabetes$glucose) Pearson's product-moment correlation data: diabetes$bmi and diabetes$glucose t = , df = 723, p-value = alternative hypothesis: true correlation is not equal to 0 95 percent confidence interval: sample estimates: cor Also performs a statistical test to see whether the coefficient is different from zero. Dias 16

17 Normally distributed? BMI Glucose A Normal distribution for comparison. Dias 17

18 Normally distributed? BMI Log(Glucose) Dias 18

19 Normally distributed? Dias 19

20 Normally distributed? Dias 20

21 R code A histogram of BMI: hist(diabetes$bmi,main= BMI,xlab= BMI (kg/m2),col= green ) A Normal Q-Q plot of BMI: qqnorm(diabetes$bmi,main= BMI,col= green ) qqline(diabetes$bmi,col= red ) And how do we get all these works of art in some decent format? jpeg(file= D:\mydirectory\mypicture.jpg,width=500,height=500) # # put here the code that generates the picture # dev.off() Dias 21

22 Rank correlation Spearman s ρ If data does not appear to be Normally distributed, or when there are outliers, one may instead compute the correlation between the ranks of the X i values and the ranks of the Y i values. This gives a nonparametric correlation coefficient called Spearman s ρ. It measures monotone association. It is invariant to monotone transformations (like a log transformation). It is robust to outliers. It has an odd interpretation. Dias 22

23 Spearman s rank correlation coefficient R code > cor.test(diabetes$bmi,diabetes$glucose,method= spearman ) Spearman's rank correlation rho data: diabetes$bmi and diabetes$glucose S = , p-value = alternative hypothesis: true rho is not equal to 0 sample estimates: rho Warning message: In cor.test.default(diabetes$bmi, diabetes$glucose, method = "spearman") : Cannot compute exact p-values with ties Dias 23

24 Rank correlation Kendall s τ A measure of monotone association with a more intuitive interpretation than Spearman s ρ is Kendall s τ. The observations from a pair of subjects i, j are and concordant if X i < X j and Y i < Y j or X i > X j and Y i > Y j discordant if X i < X j and Y i > Y j or X i > X j and Y i < Y j Kendall s τ is the difference between the probability for a concordant pair and the probability for a discordant pair. There are various versions of Kendall s τ depending on how ties are treated. Dias 24

25 Characteristics of Kendall s tau It measures monotone association. It is invariant to monotone transformations (like a log transformation). It is robust to outliers. It has a more straightforward interpretation than Spearman s rho. Dias 25

26 Kendall s rank correlation coefficient R code > cor.test(diabetes$bmi,diabetes$glucose,method= kendall ) Kendall's rank correlation tau data: diabetes$bmi and diabetes$glucose z = , p-value = alternative hypothesis: true tau is not equal to 0 sample estimates: tau Dias 26

27 Correlation in the diabetes data r = (p = 0.110) ρ = (p = 0.180) τ = (p = 0.169) Dias 27

28 Correlation in the diabetes data and log transformed r = (p = 0.154) ρ = (p = 0.180) τ = (p = 0.169) Dias 28

29 Limitations of correlation coefficients While it is (relatively) clear what a correlation coefficient of 0 means, and also 1 or -1, it is often unclear what a highly significant correlation of, say, 0.5 means Correlation rarely answers the research question to a sufficient extend; because it is not easily interpretable. Coefficients of correlation depend on the sample selection and therefore we cannot compare values of the coefficients found in different data. Dias 29

30 Dias 30 Department of Biostatistics

31 Regression analysis An (intuitively interpretable) way to describe a (linear) association between two continuous type variables. It models a response Y (the dependent variable, the exogenous variable, the output) as a function of a predictor X (the independent variable, the exogenous variable, the explanatory variable, the covariate) and a term representing random other influences (error, noise). Dias 31

32 Regression model formulation We say: To regress Y on X or: To regress glucose on BMI Mathematically: Y i = α + βx i + ε i Where ε i are independently Normal distributed noise terms with mean 0 and standard deviation σ. Dias 32

33 Regression model The mean of Y is modelled with a linear function of X; a line in the X-Y plane. For each X, Y is a random variable Normally distributed around the modelled mean of Y, with standard deviation σ Dias 33

34 Scatter-plot with regression line Dias 34

35 Interpretation of the parameters We have variation due to a systematic part, the explanatory variable, and a random part, the noise. The systematic part of the model is defined by the regression line. α = the intercept: mean level for Y i when X i = 0 β = the slope: mean increase for Y i when X i is increased 1 unit. Dias 35

36 Research question Do fat people have a more severe diabetes when the diabetes is discovered? Or in a more statistical language: Is diagnostic plasma glucose (positively) associated with the body mass index at the time of diagnosis? In a (simple) linear regression analysis, is the slope β different from 0 (or more pertinently, larger than 0)? Dias 36

37 How does the model answer the research question? Interest may focus on making a simple hypothesis about the two parameters: Null hypothesis : β = 0 Null hypothesis : α = 0 The second hypothesis often has no (clinical) meaning. Dias 37

38 Linear regression R code > mymodel <- lm(diabetes$glucose~diabetes$bmi) > summary(mymodel) Call: lm(formula = diabetes$glucose ~ diabetes$bmi) Residuals: Min 1Q Median 3Q Max Estimate of the slope P-value of the test for the null hypothesis β = 0. Coefficients: Estimate Std. Error t value Pr(> t ) Table with (Intercept) <2e-16 *** parameter diabetes$bmi estimates --- Signif. codes: 0 *** ** 0.01 * Residual standard error: on 723 degrees of freedom (4 observations deleted due to missingness) Multiple R-squared: , Adjusted R-squared: F-statistic: on 1 and 723 DF, p-value: Dias 38

39 Plot of regression line R code The lm() function can be used to plot the regression line in the scatter-plot: > plot(diabetes$bmi,diabetes$glucose) > mymodel <- lm(diabetes$glucose~diabetes$bmi) > abline(mymodel) Dias 39

40 Scatter-plot with regression line log transformed glucose Coefficients: Estimate Std. Error t value Pr(> t ) (Intercept) <2e-16 *** diabetes$bmi Dias 40

41 How are the parameters estimated? The estimated parameters of the linear model define the line (found among all possible lines) which minimizes the squared distance between the data-points and the line in the scatter-plot. The estimation method is called ordinary least-squares (maximum likelihood gives the same answer). Dias 41

42 Least squares fit Dias 42

43 Does the model fit the data? Dias 43

44 Diagnostic plots Dias 44

45 Diagnostic plots R produces some diagnostic plots (of varying usefulness). The residuals (the error or noise) was supposed to be Normal distributed, this can be studied in the Q-Q plot (top right) More importantly, the residuals should have a single standard deviation, i.e. the variance should not increase with, for example, BMI. This can be studied in the residuals vs. fitted plot (top left) > mymodel <- lm(diabetes$glucose~diabetes$bmi) > opar <- par(mfrow = c(2,2), oma = c(0,0,1.1,0)) > plot(mymodel) > par(opar) Dias 45

46 Data transformations If the residuals are not Normal, or (and this is more serious because the central limit theorem deals with much of the non- Normality issue) if variance seems to increase with level, it may be a good idea to transform one or both variables. This is the real reason to investigate log(glucose) instead of glucose. Dias 46

47 Data transformations log transform Dias 47

48 The influence of one outlier Dias 48

49 Simpson s paradox Florida death penalty verdicts for homicide relative to defendant s race White Black 11% (53/430) 8% (15/176) Dias 49

50 Simpson s paradox Victim white Victim black White Black 11% 23% (53/414) (11/37) 0% 3% (0/16) (4/139) Blacks tend to murder blacks and whites tend to murder whites and the murder of a white person has a higher probability of death penalty. For any victim the probability for a black person to get death penalty is about 2 times higher. Dias 50

51 Confounding Victim s race We are interested in the green highlighted association, but there is a correlation with the victim s race both with the defendant s race and the outcome of the trial. Defendant s race Death penalty Dias 51

52 Confounding A confounder influences both exposure and outcome Confounder When confounding is present we cannot interpret the green highlighted association as causal Exposure Outcome Dias 52

53 Randomization Exposure randomised Confounder Outcome Often there are many factors that may influence both exposure and outcome, some of them may not be observed or are unknown. If exposure is randomised, then there is no confounding. The green highlighted association can be interpreted causal. Dias 53

54 Two regressions The blue points denote patients with SBP>140 mmhg; the blue line the corresponding regression line. The red points denote patients with SBP < 140 mmhg; the red line the corresponding regression line. The black line is the general regression line. The slopes from the stratified analyses are less steep than the slope of the general line. Dias 54

55 Multiple regression > mymodel <- lm(log(diabetes$glucose)~diabetes$bmi+diabetes$sbp) > summary(mymodel) Coefficients: Estimate Std. Error t value Pr(> t ) (Intercept) <2e-16 *** diabetes$bmi diabetes$sbp * --- Signif. codes: 0 *** ** 0.01 * The adjusted slope (association) of bmi is less pronounced than before. SBP is related to both glucose and bmi and is a confounder. Dias 55

56 Multiple regression Adjusting a statistical analysis means to include other predictor variables into the model formula. Intuitively, a slope for BMI is determined for each level of the SBP variable separately and these are then averaged. including SBP in the analysis removes the confounding effect of SBP from the relationship between log(glucose) and BMI. Dias 56

57 Take home message Association between two continuous variables may be measured by correlation coefficients or in (simple) linear regression analysis. The latter provides arguably the best interpretable results. Moreover, it is straightforwardly extended to be able to deal with confounding, and more Dias 57