Math 143 C/E, Spring 2001
IPS Reading Questions
Chapter 7, Section 2 (pp. 537-548)
The naming seems appropriate, since the 1-sample procedures require collecting only one sample, while the 2-sample procedures require two samples. (Note that this holds even in the case of a matched-pairs t test.) The 2-sample procedures are generally intended to compare two potentially-different populations, and the samples come from each of these populations.
The (population) distribution for x2 - x1 will be N(m2 - m1, (s12 + s22)(1/2). The other (sampling) distribution would be N(m2 - m1, (s12/n1 + s22/n2)(1/2)).
The center of any such confidence interval is `x2 -`x1. In the former case, you could say (with at least 95% confidence) that m1 is larger than m2. In the latter case, you cannot make this claim. There is a different level of confidence C (don't ask me what it is, but the value of C must be smaller than 95%; in fact, small enough so that the associated level C confidence interval comprises only negative numbers) for which you could say (with confidence C) that m1 is larger than m2.
In the case of level C confidence intervals, it means that, if off, our margins of error will be a little larger than necessary (so that, if anything, we are a little more than C% confident that the population parameter lies inside this interval). In the case of a test of significance, we are even less likely to errantly reject the null hypothesis (that is, when it is, in fact, true) because the P values, if anything, will be larger than they would be if no approximation were necessary.
If the two populations under consideration have similar distributions, the results should be trustworthy even for sample sizes as small as 5 (from both populations). In general, even if the two population distributions have shapes that are quite different, the 2-sample procedures are pretty safe to use (i.e., will yield fairly accurate and conservative results) so long as the two sample sizes are roughly equal and the total number of units (from both samples) is at least 40. Assuming that the researcher has no difficulty getting 40 units to study, the bulk of attention needs to be placed upon getting an SRS or something close enough to it, since a biased sample contaminates any conclusions drawn later using statistical inference no matter what the sample size.