At a workshop last week, we were asked to bring along examples of good and bad academic writing. Several of the bad examples were papers where the number of significance tests was so large that the workshop participants felt that they couldn’t make sense of the Results section. It’s not that they didn’t understand each test separately but rather that they couldn’t see the forest for the trees. I, too, wish researchers would stop inundating their readers with t, F and p-values (especially in the running text), but until such time readers need to learn how to survive the ANOVA onslaught. Below I present a list of guidelines to help them with that.

  1. Identify the central, genuine research questions and the corresponding hypotheses.
    Research papers surprisingly often contain ‘padding’ research questions (RQs) that are largely unrelated to the core goal of the study. When scanning the Results section, you can usually leave aside the paragraphs about these uninteresting RQs. For example, in a report on a pretest/posttest experiment where participants were randomly assigned to conditions, you may find RQs such as Do participants in the treatment condition have different pretest scores from those in the control condition? or Do participants have higher scores on the posttest than on the pretest? Both questions are uninteresting as they don’t tell you whether the treatment actually worked.

  2. Draw a graph of the predictions.
    Having identified the key RQs and hypotheses, I often find it useful to sketch how the data would look like if the researchers’ predictions panned out and what kind of data would, within reason, falsify their hypotheses. These graphs are usually simple hand-drawn line charts that illustrate the expected group differences, but I find that they help me to better understand the logic behind the study and to focus on the important analyses in the Results section.

  3. Look for a graph in the paper.
    Ideally, the paper will contain a graph of the main results that you can then compare with the graphs you drew yourself. Do the results seem to confirm or disconfirm the researchers’ predictions? Sometimes, a good graph will allow you to carry out the easiest of significance tests yourself: the ‘inter-ocular trauma test’–if the conclusion hits you between the eyes, it’s significant. If the results are less clear cut, you’ll need to scan the Results section for the more details, but by now, you should have a clearer idea of what you’re looking for–and what you can ignore for now. If the paper doesn’t contain a graph, you can often draw one yourself on the basis of the data provided in the tables.

  4. Ignore tests unrelated to the central research questions.
    Results sections sometimes contain significance tests that are unrelated to the research questions the authors formulated up front. Such tests include include balance tests in randomised experiments (e.g. The control and intervention group did not differ in terms of SES (t(36) = 0.8, p = 0.43).), tautological tests (e.g. A one-way ANOVA confirmed that participants categorised as young, middle-aged and old differed in age (F(2, 57) = 168.2, p < 0.001).) as well as some less obvious cases. By and large, these tests tend to add little to nothing to the study and can be ignored the first time round. In non-randomised experiments, systematic differences on background variables between the groups may represent confounds, but these can be assessed based on the descriptive statistics and don’t need to be rubber-stamped with a significance test.

  5. Main effect or interaction?
    Complex designs often target the interaction between two independent variables rather than each independent variable separately. If researchers write that they’re expecting an interaction between variable A and variable B, then what they mean is that they expect A to have a different effect on the outcome variable depending on what value B has. When running the test for the interaction between A and B, statistical software also returns the significance tests for A and B separately, and researchers often report these, too. But if you’re not really interested in the effects of A and B separately, then you can ignore the tests for the main effects.
    Two further references: interpreting main effects in the presence of an interaction; interpreting ‘removable’ interactions.

  6. Don’t get distracted by F1 and F2 analyses.
    Until fairly recently, articles in psycholinguistics journals would often contain two F-tests, labelled F1 and F2, for each comparison. The goal of these double analyses was to find out whether the results would generalise not only to different participants but also to different stimuli from the ones used in the experiment. To this end, researchers calculated average scores, reaction times etc., for each participant within each ‘cell’ (combination of independent variables) and analysed these averages in a ‘subject ANOVA’ (F1). Then, again within each cell, the data were aggregated for each stimulus, and the by-stimulus averages were analysed in an ‘item ANOVA’ (F2). Findings were deemed to be generalisable when both analyses returned significance.
    F1 and F2 analyses were stopgap solutions and have now been eclipsed by more suitable methods of analysis.
    So in case you were wondering what these F1 and F2 numbers mean in that pre-2012 paper you were reading: it’s to see whether the finding would generalise to both new participants and new stimuli.

  7. Ask yourself some critical questions.
    a. Is it the means you’re really interested in?
    ANOVA compares group means, but when the data are strongly skewed or bimodal (= double-humped), means may be poor indicators of the tendencies in the data. Ideally, you’d be able to tell from the graphs how well the group means capture the tendencies in the data, but often all you get is a barplot with the group means.

    b. Are there dependencies in the data that the analysis doesn’t account for?
    If the paper reports on a cluster-randomised experiment (e.g. an intervention study in which whole classes were assigned to the same condition), then these clusters (e.g. classes) need to be somehow taken into consideration in the analysis. Not doing so can dramatically affect the results. Clustering could also occur when the intervention involves students interacting with each other, for instance when learning vocabulary in a communicative task.

    c. Are the groups real groups or are they the result of arbitrarily cutting up a continuous variable?
    Cutting up some continuous variable such as age so that participants can be categorised into ANOVA-friendly groups has its problems on the statistical side. Furthermore, when interpreting the results, discretising continuous variables may sometimes tempt the researchers or readers to infer threshold effects where none exist.

    d. Is the outcome variable really continuous or was is it binary variable converted to percentages?
    ANOVAs and t-tests work for continuous outcome variables; binary outcome variables (e.g. accuracy) call for different statistical approaches.

    e. How generous are the researchers when interpreting their results?
    Running a test is one thing, interpreting its result is another. Common shortcuts include interpreting ‘no significant difference’ to mean ‘equal’, interpreting differences in significance and reading a lot into post-hoc analyses (see ‘researcher degrees of freedom’ and ‘the garden of forking paths’).

This may not cover all the bases, but I reckon it covers enough of them to get started.


18 May 2016


comments powered by Disqus