Title: Common statistical Test Problems:
1Common statistical Test Problems
Tests dealing with the mean of data samples
Tests dealing with the variance of the samples
Tests dealing with correlation coefficients
Tests dealing with regression parameters
Testing sample mean Is it equal/ larger/
smaller a prescribed value? Comparing two sample
sets Are the mean values different? Comparing
paired samples Are the differences equal/
larger/smaller a certain value?
Testing the correlation coefficient obtained from
two paired samples Is correlation equal 0,
larger 0, or smaller 0?
Testing a single sample variance Is the
variance equal/greater/smaller a prescribed
value? Testing the ratio between the estimated
variances from two sample sets Are the variances
equal? Is the ratio between the variances equal 1
greater 1 or smaller 1 ?
- Testing a simple linear regression model
- Is the regression coefficient different from 0,
greater 0 or smaller 0. - With multiple predictors Are all regression
coefficients as a whole significantly different
from 0?Which individual regression parameters
are different from 0?
2Testing the significance of the differences in
the speed(of the Starling bird flying through a
corridor with striped walls)
Experiment Sample size n Standard deviations (guessed)
Horizontal stripes 16.5ft/s 10 1.5
Vertical stripes 15.3ft/s 10 1
Step 1 Identifying the type of statistical test
We want to test the difference in the two mean
values The test compares two estimated means.
Both are random variables with an underlying
Probability Density Function (PDF) The
variance of samples (and the variance of the
means) are also unknown and must be estimated
from the data The samples are not paired (the
experiments were all done independent)
3Testing the significance of the differences in
the speed(of the Starling bird flying through a
corridor with striped walls)
Experiment Sample size n Standard Deviations (guessed)
Horizontal stripes 16.5ft/s 10 1.5
Vertical stripes 15.3ft/s 10 1
The appropriate test is A test for the
differences of means under independence (or
Comparing two independent population means with
unknown population standard deviations) The
null hypothesis is H0 The average speed is the
same in both experiments
If H0 is true then the random variable z is a
realization from a population with approximate
standard Gaussian distribution.
Note Only for large sample sizes n1 and n2
4The classical Student t-test
Testing if Albany temperatures anomalies from
1950-1980 were different from 0
January 1950-1980 anomalies with respect to the
1981-2010 climatological mean
Dashed line Theoretical probability density
function of our test variable. If H0 was true
then our test value should be a random sample
from this distribution. That means we would
expect it to be close to zero. The more our test
value lies in the tails of the distribution, the
more unlikely it is to be part of the
distribution.
Student' (1908a). The probable error of a mean.
Biometrika, 6, 1-25. William S. Gosset He
received a degree from Oxford University in
Chemistry and went to work as a brewer'' in
1899 at Arthur Guinness Son and Co. Ltd. in
Dublin, Ireland (Steve Fienberg. "William Sealy
Gosset" (version 4). StatProb The Encyclopedia
Sponsored by Statistics and Probability
Societies. Freely available at http//statprob.com
/encyclopedia/WilliamSealyGOSSET.html)
5The classical Student t-test
Testing if Albany temperatures anomalies from
1950-1980 were different from zero
Annual mean 1950-1980 anomalies with respect to
the 1981-2010 climatological mean
Test variable
The test value calculated from the sample.
Â
The test variable t is calculated from a
random sample. As any other quantity
estimated from random samples, it is a random
variable drawn from a theoretical population
with
6The classical Student t-test
Testing H0 Albany (New York Central Park)
temperatures anomalies from 1950-1980 not
different from 0.
Solid lines Cumulative density function (for the
test variable if H0 is true)
NYC 1950-1980 Jan
Albany 1950-1980 Jan
7The classical Student t-test
Testing H0 Albany (New York Central Park)
temperatures anomalies from 1950-1980 not
different from 0. Alternative hypothesis the
mean anomaly was less than 0! (i.e. it was colder
1950-1980 than 1981-2010)
Solid lines Choose a significance test level 5
one sided t-test
NYC 1950-1980 Jan
Albany 1950-1980 Jan
0.05
0.05
8The classical Student t-test
Testing H0 Albany (New York Central Park)
temperatures anomalies from 1950-1980 not
different from 0. Alternative hypothesis the
mean anomaly was less than 0! (i.e. it was colder
1950-1980 than 1981-2010)
Solid lines Choose a significance test level 5
one sided t-test
NYC 1950-1980 Jan
Albany 1950-1980 Jan
0.05
0.05
Accept H0!
Reject H0! Accept alternative!
9The single sided t-test
Null Hypothesis H0 Albany temperatures
anomalies from 1950-1980 not different from 0.
Alternative Hypothesis Ha Temperature anomalies
were negative
t
tcrit
0
Area under the curve gives the probability P(tlt
tcrit)
Note that we formed anomalies with respect to
the 1981-2010 climatology. Thus we test if
1950-1980 was significantly cooler than the
1981-2010.
10The single sided t-test
Null Hypothesis H0 Albany temperatures
anomalies from 1950-1980 not different from 0.
Alternative Hypothesis Ha Temperature anomalies
were negative
We reject the null hypothesis if the calculated
t-value falls into the tail of the distribution.
The p-value is chosen usually chosen to be
small 0.1 0.05 0.01 are typical p-values. We
then say We reject the null-hypothesis at the
level of significance of 10 (5) (1)
t
tcrit
0
Area under the curve gives the probability p(tlt
tcrit)
Calculated t
Note that we formed anomalies with respect to
the 1981-2010 climatology. Thus we test if
1950-1980 was significantly cooler than the
1981-2010.
11The two-sided t-test
Null Hypothesis H0 Albany temperatures
anomalies from 1950-1980 not different from 0.
Alternative Hypothesis Ha Temperature anomalies
were different from zero
t
-tcrit
tcrit
0
Area under the curve gives the probability P(t gt
tcrit)
Area under the curve gives the probability P(tlt
-tcrit)
Note that we formed anomalies with respect to
the 1981-2010 climatology. Thus we test if
1950-1980 was significantly cooler than the
1981-2010.
12The two-sided t-test
Null Hypothesis H0 Albany temperatures
anomalies from 1950-1980 not different from 0.
Alternative Hypothesis Ha Temperature anomalies
were different from zero
We cannot reject H0 at the two-sided significance
level of p-percent (e.g. 5)
t
tcrit
0
Calculated t
13Testing a Null Hypothesis
Hypothesis/Conclusion Null hypothesis H0 true Null hypothesis H0 false
Null hypothesis accepted Correct decision False decision (Type II error)
Null hypothesis rejected False decision (Type I error) Correct decision
14Test for differences in the MEAN
- H0 Here we would reject H0 for the given
p-value (a 0.05)
Calculated test value
Figure 5.1 from Wilks Statistical Methods in
Atmospheric Sciences (2006)
15Test for differences in the MEAN
- H0 Here we would accept H0 for the given
p-value (a 0.05)
Calculated test value
Figure 5.1 from Wilks Statistical Methods in
Atmospheric Sciences (2006)
16Testing a Null Hypothesis
Hypothesis/Conclusion Null hypothesis H0 true Null hypothesis H0 false
Null hypothesis accepted Correct decision False decision (Type II error) Probability of this type of error is usually hard to quantify ( ßbeta)
Null hypothesis rejected False decision (Type I error) Probability of this error is given by the p-value ( a alpha) Correct decision