标题: Select the Correct Statistical Test [打印本页] 作者: shiyiming 时间: 2004-6-29 18:44 标题: Select the Correct Statistical Test Imagine you are a researcher interested in sex differences in student's attitudes toward homosexuals. Specifically, you want to test the idea that women are more accepting of homosexuals and thus have more positive attitudes toward homosexuals than do men. You collect data from 10 men and 10 women using a scale that measures homophobia. Your data look like this (note: higher scores equal more positive attitudes).
[u:00088]Which of the following tests would you use to test your hypothesis? [/u:00088]
Pearsons Correlation Coefficient
Regression Analysis
t-test for Independent Means
Paired t-test
[u:00088]Pearson's Correlation Coefficient[/u:00088]
Incorrect. While correlation is a useful tool for describing the relationship between 2 paired variables it does not serve as a test of our hypothesis.
Correlation is used to describe to what degree pairs of data are related. For example, correlation could be used to describe the relationship between education level and attitudes toward homosexuals, but not to test a specific hypothesis about this relationship.
[u:00088]Regression Analysis[/u:00088]
Incorrect. Regression analysis can be used to derive an equation from which we can predict scores on one variable based on scores on another. Regression is however, inappropriate in this case for a variety of reasons. First, regression analysis is not usually used to test hypotheses about the differences between groups (though advanced uses of regression analysis do allow for this option). Additionally, regression is inappropriate for our data because we have 2 distinct groups rather than paired data.
[u:00088]t Test for Independent Means[/u:00088]
Correct! The t test for independent means is used to compare means derived from unrelated (uncorrelated) samples. We have 2 independent groups here, women and men. To test a hypothesis regarding the differences between the 2 groups, we can use a t-test for independent means.
[u:00088]Paired t-test[/u:00088]
Incorrect. A paired t-test is appropriate for data which is paired, matched, or before-after. In this case, we have 2 groups which we have no reason to believe constitute a "pair." Because our groups do not constitute a pair, they are termed independent.
If, for example, we were interested in how attitudes toward homosexuals might change after contact with a homosexual, and we had data that measured attitudes before (pre) and after (post) contact, then the paired t-test would be the statistical technique of choice.
It is extremely important to distinguish clearly between paired and independent means t-tests as each test rests on different assumptions. The paired test assumes that the variables of interest are correlated. In this manner each individual's score on 1 variable (e.g. pre-test) is compared to their own scores on the other variable (e.g. post-test). In this manner, the paired test takes into consideration each individual's change in score. Applying this test incorrectly to unpaired data can lead to errors as scores that are unrelated are incorrectly treated as pairs.
[u:00088]Conclusion[/u:00088]
When deciding on an appropriate statistical test, it is essential to pay attention to your data and your hypotheses.
Try asking yourself the following questions:
1. Do I have groups of data or continuous scores on two variables? If you have groups you need to use a test that looks at groups such as an independent t-test. If you want to compare the scores on two different variables, correlation/regression is your best bet.
2. Do I have a hypothesis? T tests or ANOVA can be used to test hypotheses when we have a categorical independent variable whereas correlation/regression is usually not used in this manner (these techniques can be used for categorical data but are more often used on continuous variables).
3. If you have groups, are they independent or paired? Independent groups are unrelated to each other. Paired scores are usually either before-after or matched pairs type designs.