24.
963
Linguistic Phonetics
Basic statistics
2.5%
-3S -2S -S X +S +2S +3S
68%
95%
99.7%
Image by MIT OCW.
Adapted from Kachigan, S. K. Multivariate
Statistical Analysis. 2nd ed. New York, NY: Radius, 1991.
1
Assignments:
• Send me a paragraph on your final project (due 11/24)
• Write up the affricates experiment (due 12/1)
2
Writing up an experiment
The report on an experiment usually consists
of four basic parts:
1. Introduction
2. Procedure
3. Results
4. Discussion
3
Writing up an experiment
1. Introduction
• Outline of the purpose of the experiment
• state hypotheses tested etc
• provide background information (possibly including
descriptions of relevant previous results, theoretical
issues etc).
2. Procedure - what was done and how.
• instructions for replication, e.g.
– Experimental materials
– Subjects
– Recording procedure
– Measurement procedures (especially measurement
criteria).
4
Writing up an experiment
3. Results
• Presentation of results, including descriptive statistics
(means etc) and statistical tests of hypotheses.
4. Discussion
• Discuss the interpretation and significance of the results
5
Some Statistics
Two uses of statistics in experiments:
• Summarize properties of the results (descriptive statistics).
• Test the significance of results (hypothesis testing).
6
Descriptive statistics
Measures of central tendency:
• Mean: Σxi
M= N
– M is used for sample mean, μ for population mean.
• Median: The value that separates the lower half of
a set of observations from the higher half.
– Arrange the values from low to high. The median is in
the middle of this list.
7
Descriptive statistics
A measure of dispersion:
• Variance: mean of the squared deviations from the mean
Σ(xi-μ)2
σ2 = N
• Standard deviation: σ (square root of the variance).
8
9
10
11
Hypothesis Testing
Which properties differentiate /tS/ and /dZ/?
mean mean mean
VOT intensity F1 onset
(ms) (dB) (Hz)
/tS/ 112 67 531
/dZ/ 58 64 423
12
Hypothesis Testing
• Is peak intensity of frication in /tS/ higher than in /dZ/? I.e.
is 67 dB significantly higher than 64 dB?
• Could the apparent differences have arisen by chance,
although the true (population) means of frication
intensities in the two affricates are the same?
• I.e. given that intensity of /tS/ and /dZ/ frications varies, we
might happen to sample most of our /tS/ tokens from the
high end of the distribution, and most of our /dZ/ tokens
from the low end.
• Statistical tests allow us to assess the probability that this is
the case.
13
• Could the apparent differences have arisen by chance, although the true
(population) means of /tS/ and /dZ/ are the same?
dZ/tS
dZ tS
?
• Statistical tests allow
us to assess the
probability that this is
the case.
dZ tS
14
Hypothesis Testing: t-test
• The t-test allows us to test hypotheses concerning means
and differences between means.
– The mean frication intensity of /tS/ differs from the
mean frication intensity of /dZ/ (Mean difference ≠ 0)
• We actually evaluate two exhaustive and mutually
exclusive hypotheses, a null hypothesis that the mean
has a particular value, and the alternative hypothesis that
the mean does not have that value.
– The mean frication intensity of /tS/ = the mean
frication intensity of /dZ/ (Mean difference = 0).
• Statistical tests allow us to assess the probability of
obtaining the observed data if the null hypothesis were
true.
15
Hypothesis Testing: t-test
• Basic concept: If we know what the distribution
of sample means would be if the null hypothesis
were true, then we can calculate the probability
of obtaining the observed mean, given the null
hypothesis.
• We arrive at the parameters of the distribution of
sample means through assumptions and
estimation.
16
Hypothesis Testing
Distribution of sample means
σ = 10 ms
…it is unlikely that we
would get a sample mean
if this were the
population mean… of this value
110 120 130 140 150 160 170 180 190
17
Hypothesis Testing
• Basic assumption: The samples are drawn from normal populations.
2.5%
-3S -2S -S X +S +2S +3S
68%
95%
99.7%
Image by MIT OCW.
Adapted from Kachigan, S. K. Multivariate
Statistical Analysis. 2nd ed. New York, NY: Radius, 1991.
18
Distribution of Sample Means
• Basic assumption: The distribution of sample
means is normal.
• This is guaranteed to be true if the population
from which the sample is taken is normally
distributed.
• But the distribution of sample means is
approximately normal even if the population is
non-normal, as long as the samples are large
enough.
• http://onlinestatbook.com/stat_sim/sampling_dist/index.html
19
Hypothesis Testing
• Properties of distribution of means of samples of size N
drawn from a normal population:
– The sample means are normally distributed.
– Mean is the same as the population mean.
– The variance is less than the population variance:
σ2
σM2 = N
20
(a) Parent population
x
300 350 400 450 500 550 600
(b) Sampling distribution of
the mean based on samples
of size n = 4
x
300 350 400 450 500 550 600
(b) Sampling distribution of
the mean based on samples
of size n = 16
x
300 350 400 450 500 550 600
(b) Sampling distribution of
the mean based on samples
of size n = 64
x
300 350 400 450 500 550 600
Image by MIT OCW.
Adapted from Kachigan, S. K. Multivariate
Statistical Analysis. 2nd ed. New York, NY: Radius, 1991.
21
Hypothesis Testing
• The mean of the distribution is determined by hypothesis.
– E.g. mean = 1500 Hz or mean difference = 0.
• Population variance is estimated from the sample
variance. Unbiased estimate of the population variance:
Σ(xi-Μ)2
S2 = N-1
– N-1 is the number of degrees of freedom of the
sample.
• So estimated variance of distribution of sample means,
SM2 = S2/N
=
M-μ
• t score: t S
M
22
Hypothesis Testing
• t scores follow a t-distribution - similar to a normal distribution, but
with slightly fatter tails (more extreme values) because S may
underestimate σ.
• t-distribution is actually a family of distributions, one for each
number of degrees of freedom.
• Calculate t-score then consult relevant t distribution to determine the
probability of obtaining that t-score or greater (more extreme).
23
t test for independent means
• When we compare means, we are actually sampling a population of
differences (e.g. differences in intensity of frication of /tS/ and /dZ/).
• If the null hypothesis is correct, then the mean difference is 0.
• Variance of the distribution of mean differences is estimated based
on the variances and sizes of the two samples.
– Or, if the observations are paired, based on the variance of the
differences (‘paired t-test’)
• Using a paired t-test (ignoring repeated measures):
– SM = 3.34/√252 = 0.21
– M = M1-M2 = 66.7-64.3 = 2.4
(difference between sample means)
– μ = 0 (null hypothesis is that the difference between pop.
means is 0).
M − μ 2.4 − 0
t(22) = = = 11.4
SM 0.21
p( t ( 252 ) ≥ 11.4) = 2.56 ×10 −6
24
Hypothesis testing
• Statistical tests like the t test give us the probability of obtaining the
observed results if the null hypothesis were correct - the p value.
E.g. p < 0.01, p = 0.334.
• We reject the null hypothesis if the experimental results would be
very unlikely to have arisen if the null hypothesis were true.
• How should we set the threshold for rejecting the null hypothesis?
– Choosing a lower threshold increases the chance of incorrectly
accepting the null hypothesis.
– Choosing a higher threshold increases the chance of incorrectly
rejecting the null hypothesis.
– A common compromise is to reject the null hypothesis if p < 0.05,
but there is nothing magical about this number.
25
Hypothesis testing
• In most experiments we need more complex statistical analyses than
the t test (e.g. ANOVA), but the logic is the same: Given certain
assumptions, the test allows us to determine the probability that our
results could have arisen by chance in the absence of the
hypothesized effect (i.e. if the null hypothesis were true).
26
Fitting models
• Statistical analyses generally involve fitting a model to the
experimental data.
• The model in a t-test is fairly trivial, e.g.
affricate VOT = μ + voice (voice is ‘voiced’ or ‘voiceless’)
27
Fitting models
• Statistical analyses generally involve fitting a model to the
experimental data.
• The model in a t-test is fairly trivial, e.g.
affricate VOT = μ + voice (voice is ‘voiced’ or ‘voiceless’)
VOTij = μ + voicei + errorij
• Analysis of Variance (ANOVA) involves more complex models, e.g.
VOTijk = voicei + contextj + errorijk
(context takes different values for /stop_, /vowel_)
VOTijk = voicei + contextj + voice*contextij + errorijk
• Model fitting involves finding values for the model parameters that
yield the best fit between model and data (e.g. minimize the squared
errors, maximize the probability of the observed data).
• Hypothesis testing generally involves testing whether some term or
coefficient in the model is significantly different from zero.
28
MIT OpenCourseWare
https://ocw.mit.edu
24.915 / 24.963 Linguistic Phonetics
Fall 2015
For information about citing these materials or our Terms of Use, visit: https://ocw.mit.edu/terms.
