QSCI 482/Prof. Conquest TAs
Kennedy, Malinick,
HW 6 - DUE
1. You have been hired as a statistial consultant to the Water Quality
Division of the Municipality of Metropolitan Seattle (METRO). (Even
though you only have one quarter of QSCI 381 and some weeks of QSCI 482, that
puts you way ahead of a lot of other people!) METRO is going to be comparing
two areas (one designated "polluted", one designated a
"reference control site") with respect to a number of measured responses
like dissolved oxygen, turbidity, pH, and a number of toxicants(polyaromatic hydrocarbons, DDT, polychlorinated biphenyls,
heavy metals, etc.) Since there are so many potential responses to consider,
rather than focus on any one particular response, one way to have a sample
size/power analysis discussion is to do a graph, with "delta/s" on
the y-axis and "n per treatment group (assuming equal n's)"
on the x-axis. Because of the expense
involved, the sample sizes per area are going to be rather small, perhaps as
low as 3 samples per area for either sediment samples or water samples! We're trying to show METRO the what happens
statistically in terms of how the Minimum Detectable Effect changes as the
number of samples per group changes. So, try out the values n = 3, 4, 5, 6, 8,
10, 15, 20 [that's the n *per group*], solve for the quantity
"delta/s", and plot delta/s against n. Use alpha(2) = .10 and power =
.90.
After you obtain your graph, what point would you emphasize to the METRO folks
if you were doing a sample-size-power-analysis presentation? (A
few sentences along with your graph. This means that you also have to
turn in your graph.)
n |
df |
t 0.10 (2), 2(n-1) |
t 0.10 (1), 2(n-1) |
Delta/s |
3 |
4 |
2.132 |
1.533 |
2.992 |
4 |
6 |
1.943 |
1.440 |
2.392 |
5 |
8 |
1.860 |
1.397 |
2.060 |
6 |
10 |
1.812 |
1.372 |
1.838 |
8 |
14 |
1.761 |
1.345 |
1.553 |
10 |
18 |
1.734 |
1.330 |
1.370 |
15 |
28 |
1.701 |
1.313 |
1.101 |
20 |
38 |
1.686 |
1.304 |
.946 |
Conclusions: Metro should be advised that initial
increases in sample size per area (n) reap cost-effective benefits in terms of decreasing
the standardized MDE. However, there are
diminishing returns, so it may not be worth investing in additional samples
beyond a certain n.
2. A soil scientist wants to see if the soil composition (having
something to do with the amount of carbon in the soil--the more carbon
in the soil, the darker the color) is different under
two different
Treatments; denoted "T" and "C". Pairs of plots are located
throughout a designated area (see notes p. 11-3 for a diagram). One member of
each
pair is randomly assigned the "T" treatment and the other member of
the
pair is assigned the "C" treatment. Now, the final measuring
instrument
for amount of carbon in the soil is simply a SCORE between 0 and 100,
basedly largely on "carbon color observation"
as described above.
Therefore, although a higher score represents "more carbon" than a
lower score, the scientist is not sure that the scores are sufficiently exact
to be treated numerically. Use a *nonparametric* test, then, to test whether
the effects of the two treatments are the same, or whether they are different.
You may use the .10 level of significance. The data are as follows:
Pair T C
1. 82 63
2. 69 42
3. 73 74
4. 43 37
5. 58 51
6. 56 43
7. 76 80
8. 65 82
What do you conclude?
We
will do a Wilcoxon Signed Rank Test, which yields a T
min of 9. The T crit.
is T 0.10 (2), 9 = 5. Since T min is
larger than the T critical, we fail to reject a null hypothesis (Ho) of no
difference between treatments.
3. A random sample consisting of 11 automobile drivers was selected to see if
alcohol affected time to complete a specific task. Each person's
response time was measured in a laboratory setting. Under one scenario, the person would drink a
beverage that contained NO alcohol; under another scenario, the person would
drink a beverage WITH alcohol.
Treatments were randomized, and neither the drivers *nor* the people
recording the test results knew which beverage actually contained the alcohol
(this is known as a "double-blind" study). It is expected that the response time will
*lengthen* when alcohol is consumed. The
data (response time in seconds) are as follows:
SUBJECT NO ALCOHOL WITH ALCOHOL Di
1 7.1
7.4 0.3
2 6.3 6.2 -0.1
3 6.8 6.6 -0.2
4 8.4 9.3 0.9
5 6.9 7.2 0.3
6 8.5 8.8
0.3
7 7.3 7.6
0.3
8 7.7 7.9
0.2
9 8.1 8.7
0.6
10 7.4 7.9
0.5
11 6.6 7.0
0.4
d-bar = .318
s of d = .303
variance of d = .092
At the .10 level of significance, do a *parametric* test to see whether
the consumption of alcohol *increases* reaction time. What do you
conclude?
H0: response time with alcohol is the same as with
no alcohol;
Ha: response time with alcohol > response time
without alcohol.
Data are normal and we perform a paired t- test.
t0.10(1),10 = 1.372; Reject if t crit < t obs
t = (d-bar)/(S of d-bar) = 0.318/0.09127 = 3.486
1.372 < 3.486, so reject H0. 0.0025 < p < 0.005
We have enough evidence to reject the null hypothesis and say that alcohol does increase the response time for automobile drivers doing specific tasks.