Post on 14-Jan-2016
One Sample and Two Sample T-tests Introduce t test Hypothesis testing using a t test Paired t test Independent samples t test
Recap Single score compared to a known
distribution
Sample mean compared to a known sampling distribution (central limit theorem)
z X
X
Xz
Example test of a hypothesis Say the 8:00 stats students average 6.85
hours of sleep on weeknights. UNT Daily reports that students as a whole
average 7.5 hours of sleep per weeknight (standard deviation of 2.5).
Do the stats students get less sleep on average?
Example test of a hypothesis Hypothesis: UNT students average 7.5
hours of sleep (s = 2.5)
Data from stat students: 6.85 65X N
X
Xz
)65/5.2(
5.785.6 z
Example test of a hypothesis Z = -2.10 Consult table: what is the probability of
coming up with a z value of -2.10 (which is just a transformation of our mean of 6.85) or more extreme ?
p = .018
Example test of a hypothesis Another way to phrase our answer is that
the probability of getting a mean of 6.85 or less hours slept when we were expecting 7.7 is .018.
Question: are 8:00am stats students typical in their sleeping habits?
Is their mean sleeping average significantly different than what we’d expect?
Example test of a hypothesis Formally
Ho : μ = 7.7 H1 : μ ≠ 7.7
If the probability of the result is less than some criterion we’ve set up (e.g. p = .10), then we reject the null hypothesis (Ho ), and believe that our sample mean comes from a population with a mean that is different from the hypothesized population mean. The difference between the sample mean ( ) and the
hypothesized population mean (μ) is statistically significant.
Not just due to sampling error
X
One and two-tailed tests There are two ways to state our
alternative hypothesis For example
Ho : μ = 7.7 H1 : μ ≠ 7.7
However, seeing as we were dealing with an 8:00am class, it wouldn’t have seemed unlikely to have expected these folks to sleep less than typical UNT students. In other words Ho : μ = 7.7 H1 : μ < 7.7
One and two-tailed tests In the first case we have a non-directional
hypothesis, also called a two-tailed test We are expecting extreme results that are some
distance greater than or less than the hypothesized population mean (but we’re not sure which)
In the second case we had a directional hypothesis, one-tailed test We expect our result to be greater than the
hypothesized mean Or we expect our result to be greater than the
hypothesized mean We only test one of these two outcomes
One and two-tailed tests What difference does it make? Say I was going to state that any outcome
with a probability of < .05 would be deemed statistically significant
What z-score am I hoping to find for a one-tailed test vs. a two-tailed test in order to claim significance?
One-tailed test
H1: µ < some value H1: µ > some value
Two-tailed test
H1: µ ≠ some value
We don’t know if it’ll be greater than or less than
One and two-tailed tests Note that the critical z-value that we’d need to
reach the .05 level is different for each test. In the one-tailed situation, we only need a z score
of +1.645 or –1.645 (depending on whether we our alternative hypothesis states that the result will be greater than or less than the hypothesized population mean) Gist: Easier to find a result that qualifies as significant
(more statistical power) In the two-tailed situation, we need a z score of
1.96 Tougher
Two-tailed rejection Question: can you claim a directional
result from a two-tailed test?
One and two-tailed tests Why don’t we just do a one-tailed test all
the time? Sometimes we don’t know one way or
another and/or are just looking for a difference of some kind
Even when we have a good idea, our theory could be wrong Cover our ignorance
Limitation to the z-test We need to know (population standard
deviation) to compare things using our normal distribution tables
Most situations we do not know
Solution Remember that the standard deviation has
properties that make it a very good estimate of the population value
We can use our sample standard deviation to estimate the population standard deviation
T test
Which leads to:
where
and
remember degrees of freedom (n-1)
t X
sX
X
ss
n
1
2
n
XXs
Look familiar?
t distribution
Gossett, who worked crunching numbers for the Guinness brewing company, had a few too many one day and stumbled upon the t distribution. Having enough wits about him to figure he might regret his actions the next day, he sent off his work to the journal under the pseudonym “Student”.
However the work, like Guinness itself, has stood the test of time as a quality product Student’s t distribution
We need to use a variety of distributions for t depending on df (sample size)
What’s the difference?
Why a “t” now not a “z”? The difference involves using our
sample standard deviation to estimate the population standard deviation
Standard deviation is positively skewed, and so slightly underestimates the population value. Our standard error part of the formula
will also be smaller than it should, which would lead to a larger value of z than should be X
s
n
How is t different? Since s likely to be smaller than the
appropriate value of more likely to produce a larger z than if we had used the actual
So instead we use the t-distribution which accounts for this
Because we are trying to estimate , how well s does depends on the sample size
The t-distribution is normal for an infinitely large sample size With larger samples our ‘important’ t-scores will be
close to the ‘important’ z-scores (1.645, 1.96, 2.58)
What about when n is not large? Most of the time we do not have a very
large n With smaller samples, s is more likely to
underestimate positively skewed sampling distribution of s
As n gets larger and larger the t distribution more closely approximates the normal distribution
So now that we have something we call a t, how do we know what the critical value of t will be to reject the null hypothesis?
Looking up t distributions Table A.2 in book Use both df and to determine the critical value
I mean, what the heck is alpha?
The alpha level is simply going to be whatever our criterion for statistical significance is going to be Our alpha level is associated with a particular t-value
(critical value) It is also associated with making a type of error (but
we’ll get to that later)
Looking up t distributions Normally if the printout of our results shows the p-value to
be less than some criterion we’ve set up (e.g. .05), we reject the null hypothesis
But I don’t have a printout!
You calculate the t- score. If the t-score is beyond the critical t-value listed in the table that is associated with the .05 or .01 alpha levels (you get to pick your cutoff) for a one or two-tailed test, you reject the null hypothesis
Saves having to write a t-score for every p-value possible like we did for the normal distribution
Note: you would get a p-value from a stats program and should report it More info is better
Important point We will talk more about this later, but
alpha level and the p-value are not the same thing
We are entering dangerous territory!
Comparing t to z When = .05, then z (two tails) = 1.96 What are the comparable values of t with:
1 df, 3 df, 5 df, 15 df, 30 df, 60 df, 120 df?
Note your table: As with the z-test we performed, t-tests can be one-tailed or two
One vs Two-tailed test
For which is it easier to obtain a ‘significant’ result- .05 one-tailed or .05 two-tailed test?
Previous Example: New version The UNT Daily claims in its recruiting
literature that its students get an average of 8 hours of sleep a night (no s reported)
Collected sleep data from 25 grad students, this sample has a mean of 7.2 hours sleep, s = 1.5
Plug in the numbersFormula
where
t=
= = -2.667
What else do we need to know?
t X
sX
X
ss
n
7.2 8
1.5 / 25
.08
.3
Critical value of t Degrees of freedom (df) = n-1 = 25 - 1 = 24 Critical value at the .05 alpha level
t.05(24) = +2.064 In this situation we are looking for a value greater than
+2.064 or less than –2.064 (a two-tailed test) The t score observed from our data (-2.667) falls
beyond the critical value Therefore p < .05 Conclusion?
Summary of hypothesis testing (or at least one way to go about it)
1) State the research question2) State the null hypothesis, which simply gives
us a value to test, and alternative hypothesis3) Construct a sampling distribution based on the
null hypothesis and locate region of rejection (i.e. find the critical value on your table)
4) Calculate the test statistic (t) and see where it falls along the distribution in relation to the critical value (tcv)
5) Reach a decision and state your conclusion
Your turn...The average grade from year to year in statistics courses at UNT is an 81. This year the stats students (200 thus far) have an average of 83 w/ s = 10. Is this unusual?tcv = ?p = ?t = ?Conclusion?