Civil rights activist, litigator, scholar, and security and privacy sceptic.
Does the use of capital letters in English affect reading speed (compared to using all lowercase)? In March 2009, using the program contained in this repository, we carried out a study to attempt to answer this question.
The basic idea of this study was to administer reading comprehension tests where the sample text was either normally capitalised or in all-lowercase and to record how long it took to complete the test in order to measure reading speed. The null hypothesis was that the mean reading time would be the same regardless of whether the text was normally capitalised or in all-lowercase. The alternative hypothesis was that the mean reading time would be different depending on whether capitals were used.
The study was conducted using a reading comprehension web application written by Cathy J. Fitzpatrick and contained in this repository. The study web site explained the nature of the study and invited visitors to participate. Potential participants were advised that the study would require up to ten minutes of their time and that they would be timed, so they should focus exclusively on the reading comprehension test for the duration of those up to ten minutes.
When a visitor agreed to take a test, one of four pre-created reading comprehension tests was randomly selected, and it was also randomly chosen whether the text would be in all-lowercase or normally capitalised. Each visitor was allowed to take up to two tests, but could also stop after taking one. If a visitor decided to take a second test, it would not be the same reading comprehension sample.
Reading comprehension texts were displayed as images to prevent easy searching through the text (which might otherwise have been an alternative to reading the text). Together with the text, subjects were presented with four reading comprehension questions about the text in order to ensure they had read it. The questions were pre-created and specific to each individual test. They appeared in a random order for each subject.
Invitations to participate in this study were publicly posted on several popular internet forums of which the authors were members. After the study had been open for a few weeks, participation dried up, so the data collection phase of the study was deemed completed.
Overview of data:
For the remainder of this page, we consider only completed tests with acceptable scores. The other data is discarded.
All times are measured in minutes.
|Test||Lowercase mean time||A*2||Normal mean time||A*2||Mean difference||P-value|
|0||(2.4, 3.19)||0.609||(1.91, 2.59)||0.349||(0.03, 1.08)||3.98%|
|1||(2.19, 2.94)||0.425||(2.27, 2.86)||0.355||(-0.47, 0.47)||99.78%|
|2||(2.98, 4.48)||0.467||(2.85, 4.04)||1.018**||(-0.63, 1.2)||53.2%|
|3||(2.95, 3.85)||0.21||(2.82, 3.99)||0.71||(-0.71, 0.69)||97.72%|
The Anderson-Darling statistic (A*2) was calculated for each of the eight samples in order to determine whether each sample came from a normally distributed population. With the exception of normally capitalised test 2, there was no reason to believe that any of the populations were non-normal. Normally capitalised test 2 exhibited an unusual distribution, but that was ultimately not significant.
For each test, we used the two-tailed Student’s t-test to determine whether the population mean completion lowercase time was equal to the population mean normally capitalised time. One requirement for this test is that each of the populations are normally distributed, which appeared to be the case here (again with the possible exception of normally capitalised test 2). The P-value given in the chart is the chance of the sample means being this far apart or father if the population means are equal (that is, a low P-value would mean that the use of capitals has a statistically significant correlation with a change in reading speed).
The intervals shown in the chart are 95% confidence intervals; that is, the true population mean lies within the interval with 95% confidence. In the case of the column for the mean difference, if the true mean difference is zero, then it indicates that the mean reading time is the same regardless of whether capitals are employed. For tests 1 through 3, the interval does contain zero, but in the case of test 0, it does not contain zero, so with 95% confidence, the population means are not equal for test 0.
There was no reason to reject the null hypothesis for tests 1 through 3. The P-value was very high for tests 1 and 3. Although it was slightly lower for test 2, it was still very high, and the relatively low value was likely a result of the non-normal shape of the sample data for normally capitalised test 2. However, in the case of test 0, the P-value was much lower and provides evidence that the mean reading time was different for that test depending on whether it was in capitals or all-lowercase.
These results suggest that for most texts, whether a text is capitalised or not has no effect on reading speed, but for certain texts it may provide a marginal benefit. Assuming the effect was not merely caused by chance (of which there is a 3.98% probability), it was not clear from our study which properties of text 0 caused the use of capitals to provide a benefit.
This table shows how many tests of each kind were completed.
|Test||Total completed||Normally capitalised||All lowercase|