statistics for linguistics students michaelmas 2004 week 5 bettina braun bettina
TRANSCRIPT
Statistics for Linguistics Students
Michaelmas 2004Week 5
Bettina Braunwww.phon.ox.ac.uk/~bettina
Overview
• P-values
• How can we tell that data are taken from a normal distribution?
• Speaker normalisation
• Data aggregation
• Practicals
• Non-parametric tests
p-values
• p-values for all tests tell us whether or not to reject the null hypothesis (and with what confidence)
• In linguistic research, a confidence level of 95% is often sufficient, some use 99%
• This decision is up to you. Note that the more stringent your confidence level, the more likely is a type II error (you don’t find a difference that is actually there)
p-values
• If you decide for a p-value of 0.05 (95% certainty that there indeed is a significant difference), then a value smaller than 0.05 indicates that you can reject the null-hypothesis
• Remember: the null-hypothesis generally predicts that there is no difference
• If we find an output saying p = 0.000, we cannot certainly say that it is not 0.00049; so we generally say p < 0.001
p-values
• So, in a t-test, if you have p = 0.07 means that you cannot reject the null hypothesis that there is no difference there is no significant difference between the two groups
• In the Levene test for homogenity of variances, if p = 0.001, then you have to reject the null-hypothesis that there is no difference so there is a difference in the variances for the two groups
Kolmogorov-Smirnov test
• Parametric tests assume that the data are taken from normal distributions
• Kolmogorov-Smirnov test can be used to compare actual data to normal distribution-- the cumulative probabilities of values in the
data are compared with the cumulative probabilities in a theoretical normal distribution
– Null-hypothesis: your sample is taken from a normal distribution
Kolmogorov-Smirnov test
• Non-parametric test• Kolmogorov-Smirnoff
statistic is the greatest difference in cumulative probabilities across range of values
• If its value exceeds a threshold, null-hypothesis is to be rejected
Kolmogorov-Smirnov test
• Kolmogorov test is not significant, i.e. the null-hypothesis that our sample is drawn from a normal distribution holds
• The distribution can therefore be assumed to be normal: Kolmogorov-Smirnov Z = 0.59; p = 0.9
Speaker normalisation
• We often collect data from different subjects but we are not interested in the speaker differences (e.g. mean pitch height, average speaking rate)
• We can convert the data to z-scores (which tell us how many sd away a given score is from the speaker mean)
Speaker normalisation in SPSS
• First, you have the split the file according to the speakers (Data -> split file)
Speaker normalisation in SPSS
• Then, Analyze -> Descriptive Statistics -> Descriptives
• This will create an output, but also a new column with z-values
Sorting data for within-subjects desings
Aggregating data
• One can easily build a mean for different categories, preserving the structure of the SPSS table
• Data -> Aggregate– Independent variables you want to preserve
are “break variables”– Dependent variables for which you’d like to
calculate the mean are “Aggregated variables”– Per default, new table will be stored as
aggr.sav
Aggregating data
• SPSS-dialogue-box
Non-parametric tests
• If assumptions for parametric tests are not met, you have to do non-parametric tests.
• They are statistically less powerful (i.e. they are more likely not to find a difference that is actually there – Type I error)
• On the other hand, if a non-parametric test shows a significant difference, you can draw strong conclusions
Mann-Whitney test
• Non-parametric equivalent to independent t-test
• Null-hypothesis: The two samples we are comparing are from the same distribution
• All data are ranked and calculations are done on the ranks
Wilcoxon Signed ranks test
• Non-parametric equivalent to paired t-test• The absolute differences in the two
conditions are ranked• Then the sign is added and the sum of the
negative and positive ranks is compared• Requires that the two samples are drawn
from populations with the same distribution shape (if this is not the case, use the Sign Test)
Examples
• English is closer to German than French is• A teacher compares the marks of a group
of German students who take English and French (according to the German system from 1 to 15)
• His research hypothesis is that pupils have better marks in English than in French
• One-tailed prediction!• File: language_marks.sav
Example
• For a one-tailed test divide the significance value bz 2
• Marks in English are better than in French (Z= -2.28, p = 0.011)
What are frequency data?
• Number of subjects/events in a given category
• You can then test whether the observed frequencies deviate from your expected frequencies
• E.g. In an election, there is an a priori change of 50-50 for each candidate.
• Note that you must determine your expected frequencies beforehand
X2-test
• Null-hypothesis: there is no difference between expected and observed frequency
• Data
• Calculation
Kerry supporter
Bushsupporter
observed 56 44
expected 50 50
X2-test example
• Null-hypothesis: there is no difference between expected and observed frequency
• Data
• Calculation
Kerry supporter
Bushsupporter
observed
expected
Looking up the p-value
Calculated value for X2
must be larger than the one found in the table
Degrees of freedom:
• If there is one independent variabledf = (a – 1)
• Iif there are two independent variables:df = (a-1)(b-1)
X2-test
• Limitations:– All raw data for X2 must be frequencies (not
percentages!)– Each subject or event is counted only once
(if we wish to find out whether boys or girls are more likely to pass or fail a test, we might observe the performance of 100 children on a test. We may not observe the performance of 25 children on 4 tests, however)
– The total number of observations should be greater than 20
– The expected frequency in any cell should be greater than 5
X2 as test of association
• Calculation of expected frequencies:
Cell freq =
Apect Past tense Present tense
total
Progressive 308 476 784Non-progressive
315 297 612
Total 623 773 1396
Row total x column total
Grand total