5.3 Results
5.3.4 Typing Errors
This section presents a fine grain analysis by categorizing the types of input errors: in- sertions - added characters; substitutions - incorrect characters; and omissions - omitted characters (MacKenzie and Soukoreff, 2002a). Figure 5.6 and Figure 5.7 shows the type of errors performed on mobile and tablet devices, respectively, for each participant.
Figure 5.6: Types of error for each participant in mobile condition.
Mobile device. Results show that on average 6% (sd=3%) of input errors are insertions.
Overall, this is the least common error type (although no significant differences were found to other types of errors). Moreover, generally speaking, correlations betweeninser- tionerrors and tremor measures (both objective and subjective) were weak (r<0.3). How- ever, there were some medium correlations: self-report tremor [Spearman rho=.392, n=15, p=.165], right-hand task specific tremor [Spearman rho=.312, n=15, p=.278], left-hand task specific tremor [Spearman rho=.401, n=9,p=.156], left-handoscillation X [Pearson r=.356, n=9,p=.348], and left-handoscillation XYZ [Spearman rho=.333, n=9, p=.381]
axes.
Regarding substitution errors, participants obtained a mean 7.8% (sd=7%) error rate.
This was the second most common error type and consisted in pressing incorrect keys.
Results show a high standard deviation, which reflects difficulties of some participants (see Figure 5.5). While eight participants had a substitution error rate between 0 and 5% (#1, #2, #4, #6, #7, #10, #13, #15), six achieved an error rates higher than 10% (#3, #5, #8, #11, #12, #14). Moreover, we found a large positive correlation between substitution error rate and both right [Spearman rho=.624, n=15, p<.05] and left [Spearman rho=.541, n=9,p<.05] handtask-specific tremor, which accounts for 39%
and 29% of shared variance, respectively. Indeed, visual inspection of participants’ profile revealed that only one of participants with average substitution error rate above 10% was absent of tremor; all remaining had slight or moderate task-specific tremor. Moreover, only one of the participants (#4) with (moderate) tremor did not obtain a mean substitution error rate above 10%; instead their main cause of errors were insertions (Figure 5.6).
Concerning omission error rates, results show an average of 12.65% (sd=16%). This error type is often described as a cognitive error, since it does not depend on motor abilities
Figure 5.7: Types of error for each participant in tablet condition.
(Kristensson, 2009). Instead, users forget to type the intended characters or misunderstand the required sentence. Although we did not account for cognitive differences and therefore cannot confirm this hypothesis, we did find a large positive correlation betweenomission error rate and left-handoscillation Y [Spearman rho=.733, n=9, p=.025], which suggest that this error type may (also) be related to tremor. Also, we should notice that the high variance (16%) is suggestive of a user-dependent measure. Indeed, most participants (nine) obtained an error rate below 10%, while participant #8 and #10 obtained a 57%
and 37% error rate, respectively (Figure 5.6). These results will be further investigated in Section 5.3.7 (Omission Errors in Detail).
Tablet device. Participants obtained an average of 3.8% insertion rate (sd=4.1%) in the tablet condition. As in other text-entry measurements, insertion error rate shows a high variance. Particularly, three participants achieved an error rate above 5%: participants #3 (16.89%), #10 (7.75%), and #11 (5.48%). A medium size positive correlation was found betweeninsertionerror rate andtask-specific tremor [Spearman rho=.398, n=15,p=.142].
All other tremor measures were weakly correlated with user performance, suggesting a weak relationship between insertion rates and tremor when using tablet devices.
Regarding substitution error rate, participants achieved a mean of 3.75%, however with high variance (sd=3.61%). While nine participants are below 3% error rate, six (#3,
#8, #9, #11, #12, #15) are between 3% and 13% (see Figure 5.7), resulting in a user- dependent measure. Indeed, we found a large positive correlation withtask-specific tremor [Spearman rho=.539, n=15,p=.038]; that is, participants with higher tremor had higher substitution error rates.
As for omission error rate, this is also the most common error type in the tablet condition,
accounting for an average of 9% (sd=10%). Again, results show high variance with nine participants below 10% error rate and four participants (#8, #11, #13, and #14) above 15%. No large correlations were found between omission error rate and both subjective and objective tremor measures in the tablet condition.
Summary. Overall, results for each type of error showed a high variance, which suggests that older adults text-entry performance is highly user-dependent. As for insertion errors, we found small relationships with all tremor measurements, thus its variance could not be explained by participants’ motor abilities. Nevertheless, we found significant differences betweendevice conditions [Z=-2.103, p<0.05, eta squared statistic=.25], in which tablet devices show a decrease of 1.72% of insertion errors. Regarding substitution error rate there was a large correlation with task-specific tremor for both device conditions, explain- ing 29-39% of users’ performance. Also, we found a statistically significant decrease of 4%
frommobiletotablet device [Z=-2.731,p<.01] with a medium effect (eta squared statistic
= .36). Concerning omission error rate results are inconclusive. This error type is often associated with cognitive errors, thus it would be expected that performance across de- vice conditions would be similar. Indeed, no statistical significant differences were found betweenmobile (m=12.65%, sd=16%) and tablet (m=9%, sd=10%) conditions [Z=-.722, p>.4]. On the other hand, we found a large correlation between omission error rate and hand oscillation on mobile condition, which suggests that this measure may be related to motor abilities. We will further explore these results on Section 5.3.7(Omission Errors in Detail), analyzing each participant individually.