12.3 Research hypothesis, alternative hypothesis, and nil hypothesis
The null hypothesis is central to significance testing. If the test is statistically significant, that is, if the p value is below the significance level, we reject the null hypothesis.
The alternative hypothesis covers all situations not covered by the null hypothesis. The null hypothesis stating that average media literacy in a population of children is 5.5, is paired with the alternative hypothesis stating that the average is not 5.5. In this way, we cover all possible outcomes.
If we reject the null hypothesis, we say that our data lend support to the alternative hypothesis. We doubt that the null hypothesis is true. Of course, we know that we can be mistaken. There is five per cent chance that we reject a null hypothesis that is actually true (Type I error, Section 12.2.2). Rejecting the null hypothesis does not mean that this hypothesis is false or that the alternative hypothesis is true. Please, never forget this.
The alternative hypothesis is of interest because it usually represents the research hypothesis (but not always as some statistics textbooks would have us believe). Most of the research hypotheses in social research are alternative hypotheses because our theories tell us to expect differences or changes but not the size of differences or changes.
Not knowing which precise difference or association to expect, we usually formulate the research hypothesis that there is a difference or association. Because a particular value for the difference or association cannot be specified, these research hypotheses are alternative hypotheses. The associated null hypothesis is that there is no difference or no association. It equates the population statistic to one value, namely zero. This type of null hypothesis is called a nil hypothesis or just plainly the nil.
If we expect that groups have different average scores on a dependent variable, for example, willingness to donate to a charity, but we do not know how different, we test the null hypothesis that the differences between the group averages are zero (no difference) in the population. If we expect a correlation between exposure and brand awareness in the population but we have no clue about the size of the correlation, we test the null hypothesis that the population correlation coefficient (Spearman’s rho or Pearson’s correlation coefficient) or regression coefficient (\(b\) or \(b^*\)) is zero. For all measures of association, zero means that there is no association. All of these are examples of nil hypotheses because the population value is hypothesized to be zero.
If the research hypothesis is the alternative hypothesis, we have to choose a value for the null hypothesis ourselves. This is very important, because the null hypothesis is actually tested. If statistical software does not report the null hypothesis that is being tested, you may assume that it equates the parameter of interest to zero (see Section 12.7 on null hypotheses in SPSS).