定量數(shù)據(jù)分析_Quantitative_data_analysis

1、Will G HopkinsAuckland University of TechnologyAuckland NZQuantitative Data AnalysisSummarizing Data: variables; simple statistics; effect statistics and statistical models; complex models.Generalizing from Sample to Population: precision of estimate, confidence limits, statistical significance, p v

2、alue, errors.Reference: Hopkins WG (2002). Quantitative data analysis (Slideshow). Sportscience 6, /jour/0201/Quantitative_analysis.ppt (2046 words)Summarizing Data Data are a bunch of values of one or more variables. A variable is something that has different values. Values can be numbe

3、rs or names, depending on the variable:Numeric, e.g. weightCounting, e.g. number of injuriesOrdinal, e.g. competitive level (values are numbers/names)Nominal, e.g. sex (values are names When values are numbers, visualize the distribution of all values in stem and leaf plots or in a frequency histogr

4、am.Can also use normal probability plots to visualize how well the values fit a normal distribution. When values are names, visualize the frequency of each value with a pie chart or a just a list of values and frequencies. A statistic is a number summarizing a bunch of values. Simple or univariate s

5、tatistics summarize values of one variable. Effect or outcome statistics summarize the relationship between values of two or more variables. Simple statistics for numeric variables Mean: the average Standard deviation: the typical variation Standard error of the mean: the typical variation in the me

6、an with repeated samplingMultiply by (sample size) to convert to standard deviation. Use these also for counting and ordinal variables. Use median (middle value or 50th percentile) and quartiles (25th and 75th percentiles) for grossly non-normally distributed data. Summarize these and other simple s

7、tatistics visually with box and whisker plots. Simple statistics for nominal variables Frequencies, proportions, or odds. Can also use these for ordinal variables. Effect statistics Derived from statistical model (equation) of the form Y (dependent) vs X (predictor or independent). Depend on type of

8、 Y and X . Main ones:YXEffect statisticsModel/Testnumericnumericslope, intercept, correlation regressionnumericnominalnominalnominalnominalnumericmean differencefrequency difference or ratiofrequency ratio per t test, ANOVA chi-squarecategorical Model: numeric vs numerice.g. body fat vs sum of skinf

9、olds Model or test: linear regression Effect statistics: slope and intercept= parameterscorrelation coefficient or variance explained (= 100correlation2)= measures of goodness of fit Other statistics:typical or standard error of the estimate= residual error= best measure of validity (with criterion

10、variable on the Y axis) Model: numeric vs nominale.g. strength vs sex Model or test: t test (2 groups)1-way ANOVA (2 groups) Effect statistics: difference between meansexpressed as raw difference, percent difference, or fraction of the root mean square error (Cohens effect-size statistic) variance e

11、xplained or better (variance explained/100)= measures of goodness of fit Other statistics:root mean square error= average standard deviation of the two groups More on expressing the magnitude of the effect What often matters is the difference between means relative to the standard deviation: strengt

12、hfemalesmalesTrivial effect: strengthfemalesmalesVery large effect: Fraction or multiple of a standard deviation is known as the effect-size statistic (or Cohens d). Cohen suggested thresholds for correlations and effect sizes. Hopkins agrees with the thresholds for correlations but suggests others

13、for the effect size:CorrelationsEffect Sizes For studies of athletic performance, percent differences or changes in the mean are better than Cohen effect sizes. Model: numeric vs nominal (repeated measures)e.g. strength vs trial Model or test: paired t test (2 trials)repeated-measures ANOVA withone

14、within-subject factor (2 trials) Effect statistics: change in mean expressed as raw change, percent change, or fraction of the pre standard deviation Other statistics:within-subject standard deviation (not visible on above plot) = typical error: conveys error of measurement useful to gauge reliabili

15、ty, individual responses, and magnitude of effects (for measures of athletic performance).preposttrial Model: nominal vs nominale.g. sport vs sex Model or test: chi-squared test or contingency table Effect statistics:Relative frequencies, expressed as a difference in frequencies, ratio of frequencie

16、s (relative risk), or ratio of odds (odds ratio)Relative risk is appropriate for cross-sectional or prospective designs. risk of having rugby disease for males relative to females is (75/100)/(30/100) = 2.5Odds ratio is appropriate for case-control designs. calculated as (75/25)/(30/70) = 7.030%75%

17、Model: nominal vs numerice.g. heart disease vs age Model or test: categorical modeling Effect statistics:relative risk or odds ratioper unit of the numeric variable(e.g., 2.3 per decade) Model: ordinal or counts vs whatever Can sometimes be analyzed as numeric variables using regression or t tests O

18、therwise logistic regression or generalized linear modeling Complex models Most reducible to t tests, regression, or relative frequencies. Example Model: controlled trial (numeric vs 2 nominals)e.g. strength vs trial vs group Model or test: unpaired t test of change scores (2 trials, 2 groups)repeat

19、ed-measures ANOVA withwithin- and between-subject factors (2 trials or groups)Note: use line diagram, not bar graph, for repeated measures. Effect statistics: difference in change in mean expressed as raw difference, percent difference, or fraction of the pre standard deviation Other statistics:stan

20、dard deviation representing individual responses (derived from within-subject standard deviations in the two groups)preposttrialdrugplacebo Model: extra predictor variable to control for somethinge.g. heart disease vs physical activity vs age Cant reduce to anything simpler. Model or test:multiple l

21、inear regression or analysis of covariance (ANCOVA)Equivalent to the effect of physical activity with everyone at the same age.Reduction in the effect of physical activity on disease when age is included implies age is at least partly the reason or mechanism for the effect.Same analysis gives the ef

22、fect of age with everyone at same level of physical activity. Can use special analysis (mixed modeling) to include a mechanism variable in a repeated-measures model. See separate presentation at . Problem: some models dont fit uniformly for different subjects That is, between- or within-

23、subject standard deviations differ between some subjects. Equivalently, the residuals are non-uniform (have different standard deviations for different subjects). Determine by examining standard deviations or plots of residuals vs predicteds. Non-uniformity makes p values and confidence limits wrong

24、. How to fixUse unpaired t test for groups with unequal variances, orTry taking log of dependent variable before analyzing, orFind some other transformation. As a last resort Use rank transformation: convert dependent variable to ranks before analyzing (= non-parametric analysissame as Wilcoxon, Kru

25、skal-Wallis and other tests).Generalizing from a Sample to a Population You study a sample to find out about the population. The value of a statistic for a sample is only an estimate of the true (population) value. Express precision or uncertainty in true value using 95% confidence limits. Confidenc

26、e limits represent likely range of the true value. They do NOT represent a range of values in different subjects. Theres a 5% chance the true value is outside the 95% confidence interval: the Type 0 error rate. Interpret the observed value and the confidence limits as clinically or practically benef

27、icial, trivial, or harmful. Even better, work out the probability that the effect is clinically or practically beneficial/trivial/harmful. See . Statistical significance is an old-fashioned way of generalizing, based on testing whether the true value could be zero or null. Assume the nul

28、l hypothesis: that the true value is zero (null). If your observed value falls in a region of extreme values that would occur only 5% of the time, you reject the null hypothesis. That is, you decide that the true value is unlikely to be zero; you can state that the result is statistically significan

29、t at the 5% level. If the observed value does not fall in the 5% unlikely region, most people mistakenly accept the null hypothesis: they conclude that the true value is zero or null! The p value helps you decide whether your result falls in the unlikely region. If p0.05, your result is in the unlik

30、ely region. One meaning of the p value: the probability of a more extreme observed value (positive or negative) when true value is zero. Better meaning of the p value: if you observe a positive effect, 1 - p/2 is the chance the true value is positive, and p/2 is the chance the true value is negative

31、. Ditto for a negative effect.Example: you observe a 1.5% enhancement of performance (p=0.08). Therefore there is a 96% chance that the true effect is any enhancement and a 4% chance that the true effect is any impairment.This interpretation does not take into account trivial enhancements and impair

32、ments. Therefore, if you must use p values, show exact values, not p0.05.Meta-analysts also need the exact p value (or confidence limits). If the true value is zero, theres a 5% chance of getting statistical significance: the Type I error rate, or rate of false positives or false alarms. Theres also

33、 a chance that the smallest worthwhile true value will produce an observed value that is not statistically significant: the Type II error rate, or rate of false negatives or failed alarms.In the old-fashioned approach to research design, you are supposed to have enough subjects to make a Type II err

34、or rate of 20%: that is, your study is supposed to have a power of 80% to detect the smallest worthwhile effect. If you look at lots of effects in a study, theres an increased chance being wrong about at least one of them.Old-fashioned statisticians like to control this inflation of the Type I error rate within an