The two-sample procedures described last week only apply when the two samples are
independent. However, you may want to perform a hypothesis tests to data that are related.
Some examples of related (non-independent) data include: before and after, observed vs.
predicted, right vs. left, identical twins, etc. Pairing is a good idea when you expect greater
variation between the pairs when compared to variation within a pair.
t=
d
,
sd
where d = mean difference and s d = standard error of the mean difference. Thus, the paired-
sample t-test is essentially a one-sample t-test. Note, however, that each observation in one
sample must be correlated to one and only one observation in the second sample. The pairedsample t-test does not have the normality and homogeneity of variances assumptions as with the
two-sample t-test, but it does assume that the differences are normally distributed.
Example: We want to compare ground versus air-based temperature sensors to determine the
earths temperature, which is important for agricultural modeling, etc. Ground-based sensors are
expensive, and air-based (from satellites or air-planes) of infrared wavelengths may be biased.
We collected temperature data from ground and air-based sensors at ten locations, and we want
to test if they are different. We will test the following hypothesis:
Ho: d = 0
Ha: d 0
= 0.05
d=
Location
Ground (oC)
Air (oC)
Difference (di)
46.9
47.3
-0.4
45.4
48.1
-2.7
36.3
37.9
-1.6
31.0
32.7
-1.7
24.7
26.2
-1.5
22.3
23.3
-1.0
49.8
50.2
-0.4
40.5
42.6
-2.1
37.7
39.4
-1.7
10
35.5
37.9
-2.4
15.5
= 1.55 oC
10
sd =
sd
n
0.7706
10
= 0.24 oC
t-statistic: t =
d 1.55
=
= 6.458
sd
0.24
Example: Continuing with our earlier example of temperature sensors, we can obtain a 95%
confidence interval for the population mean difference:
We can also determine power of the test and necessary sample size for a specified level of
precision as we did for a one-sample t-test (chapter 7). Simply substitute d for X , and s d2 for s2.
When Ho: Population 1 Population 2 and Ha: Population 1 > Population 2, and if
T T (1),n , then reject Ho.
When Ho: Population 1 Population 2 and Ha: Population 1 < Population 2, and if
T+ T (1),n , then reject Ho.
Example: Lets redo our previous example using the non-parametric test:
Ho: Ground based sensors = Air based sensors
Ha: Ground based sensors Air based sensors
= 0.05
Signed Rank of d i
46.9
47.3
-0.4
1.5
-1.5
45.4
48.1
-2.7
10
-10
36.3
37.9
-1.6
-5
31.0
32.7
-1.7
6.5
-6.5
24.7
26.2
-1.5
-4
22.3
23.3
-1.0
-3
49.8
50.2
-0.4
1.5
-1.5
40.5
42.6
-2.1
-8
37.7
39.4
-1.7
6.5
-6.5
10
35.5
37.9
-2.4
-9
Sum of Ranks:
=0
= 55 .
Decision Rule: If T+ or T T0.05(2 ),10 = 8 , then reject Ho; otherwise, do not reject Ho.
Conclusion: Since T+ = 0 < 8 (P < 0.005), reject Ho and conclude that the mean difference
between ground and air-based sensors at these five locations is significantly different.