4.3.5 Choosing statistical tests

This section is optional in that it provides more detail on the statistical tests that can be applied to analyse the data. If you have time and/or such analysis is important to the type of work you do, you should study this section.

Many different statistical tests have been developed, and we will only look at the most common ones in this course. Firstly, statistical tests can be classified as parametric or non-parametric. Parametric tests assume a specific distribution of data (e.g. a normal distribution for numeric data). Examples of parametric tests include the chi-squared test, linear regression and logistic regression. Non-parametric tests do not assume a specific distribution of the data. A commonly used non-parametric test is the Mann-Whitney U test for ordinal data. Non-parametric tests are generally not as statistically powerful as parametric tests, but have the advantage of needing fewer assumptions.

We will look at two categories of tests in this section. First, let us consider the tests to use when the exposure variable is a categorical variable. In this case, statistical tests allow us to assess differences between counts or means of the outcome variable at different levels of the exposure variable (e.g. assess the differences in average healing time (outcome variable) for different treatments A, B and C). The choice of such a statistical test depends on the type and characteristics of the data being analysed. One way to determine what test is most appropriate is to identify the type of outcome variable, the number of groups being compared (i.e. in our example – there are three levels of the variable ‘treatment’: A, B and C) and whether the assumptions for the test are met (this step may be informed by the descriptive analysis). Table 6 lists the most important assumptions for each test. For example, a non-parametric Mann-Whitney test might be used to test whether there are differences in the presence of AMR genes in bacterial isolates sampled from patients in hospitals, compared to patients in the community.

Statistical tests in a second category are those used when the exposure variable is numeric. In this case, tests allow users to assess the existence of a significant relationship between the (numeric) outcome and exposure variables. When there are no other variables to consider, such a test is called a correlation or collinearity test. The most commonly used correlation tests are Pearson’s r (parametric, normally distributed variables) and Spearman’s ρ rank correlation (non-parametric). For example, researchers conducting a multi-country study have data on the prevalence of resistance to a particular antimicrobial for two target organisms for 20 countries. If they want to know whether resistance to a particular antimicrobial in one target organism is related (correlated) to resistance in the second target organism, they could use a test for correlation. If additional variables need to be considered in the analysis, multivariable regression models can be constructed, though this is beyond the scope of this course.

Table 6 provides an overview of the types of statistical tests you can conduct according to the type of data you have. If you’re just starting out in learning statistics, some of this information might seem very complex, but you don’t need to understand it all now. Keep a copy of this table for future reference for when you plan and conduct your own analyses of AMR data. An important point to note for now is that results obtained from statistical tests are only valid when all the assumptions for the test are met. This important point is not always easy to assess in practice. Two important assumptions common to the seven tests presented in Table 6 are that the observations are based on a random sample and that they are independent from each other. When observations are not independent, for example, when several observations are obtained from the same individual at different points in time (repeated observations) or from a matched study design, then other tests should be used, such as paired sample t-test or mixed-effect regression models. Such tests are out of the scope of this course.