Different expressions of emotion can be recognized by unique combinations of facial features. Studies usually ask observers to either inspect full faces, measuring which face parts are being explored, or they use masked faces and measure how performance changes when different face parts are visible.
We linked these two approaches by first asking 202 observers to highlight the most informative parts of a full face for each expression. Then, we generated masked faces, where only these highlighted parts were visible. A new group of 201 observers had to recognize each expression in the masked faces, with masks being either matched (e.g. happy face/happy mask) or mismatched (e.g. disgusted face/angry mask).
When mask and expression matched, recognition accuracies for all expressions remained high (e.g. 99% for both full and masked happy face). When expression and mask were mismatched, performance dropped sharply. When using a happy mask on other emotions, they were systematically mistaken for neutral, with e.g. 86% neutral ratings for surprised expressions. For the other emotions, confusions were often biased towards the mask (e.g. disgust was mistaken for anger in 75% of cases when using an anger mask).
These results are in line with happiness engaging unique features of the face, while the other emotions engage sets of partly overlapping features, causing the masking manipulation to amplify confusions which observers already make in fully visible faces. Using the results of one study to generate stimuli for another study allows us to integrate two main approaches of face expression research.