We thank the reviewers. We begin by addressing R3’s concerns, as these
were highlighted by the AC. R1 also raised many important questions,
which we discuss.

Sample size: R3 was concerned that a sample size of 6 participants per
group was too small and that the older group was too diverse. To
clarify, there were 12 participants per age group (p.3, and in the N
values of the figure captions).  To alleviate any ambiguity, we will
adjust the wording on page 3 to say: “We recruited 12 participants
from each age group (AGE), for a total of 24 participants”. While we
agree with the AC that a larger sample would provide additional
evidence, our sample size is consistent with the field (see [13,27,28]
in the paper). In terms of diversity, we used the standard groupings
for age-comparative research on movement control [14]. Moreover, very
few significant differences were found between the 55-69 and 70+ age
groups in [19].

Importance: R3 was also concerned that pen interfaces might be
inappropriate for older adults; however, prior work has not yet
discounted pen interaction. When compared directly, older adults
perform better with a pen than a mouse [5]. While older adults do
experience problems with pen interaction, they experience at least as
many difficulties with a mouse (contrast the difficulties listed in
[25] with [19]). Touch interaction is also good for simple interfaces,
however, it has been shown to be more error prone than the mouse when
targets are small (9.8% vs. 2.1% in [ii]).

In contrast to R3, R1 was uncertain whether improvements to the
already low error rates for pen interfaces is worthy of additional
study. We argue that because of older adults’ cognitive impairments,
errors are of substantially greater significance in the real world
than reflected in experimental studies. Because there is no error
penalty in a Fitts task, errors are likely to be more acceptable to
users in that context than they would be in a real world situation.
This makes determining what is an acceptable error rate for computer
use challenging. There is a shortage of real-world data on selection
frequency (and none on error frequency), but one workplace study
(N=2146, mean age=42) found that in 1.3 hours of computer use (daily
average), participants made on average about 1300 selections [i]. We
extrapolate that for those users, a 5% error rate would result in 65
errors, which could have a substantial impact, depending on the exact
cost of the errors. While these figures may not map directly to older
users’ experiences, it is reasonable to expect that an older user
would encounter a comparable number of errors, from which they might
have additional difficulty recovering. An additional factor to
consider is the variability between participants. In [19], the overall
error rates were 4.2%, and 6.4% (for pre-old and old, respectively),
however the respective maximum error rates were 16.7%, and 22.2%,
suggesting that some users had substantial problems with errors.

We were unable to fully understand R1's 3rd point. We believe many of
the concerns could be addressed by a better breakdown of the overall
error data, including the following means (for Control, Steady,
Bubble, Steady-Bubble, respectively): older: 22.2%, 19.3%, 10.9%,
10.6% and younger: 16.0%, 13.0%, 7.5%, 6.8%. Similar data is currently
in the paper, but we agree that a summary table would help aid
interpretation. Further, we can clarify that most of the errors (and
largest effects) occurred with the smallest target sizes, which is
consistent with prior work [19] and representative of many real world
tasks.

Target-awareness: We agree with the excellent points made by R1 on the
limitations of target aware techniques, but note that the majority of
successful pointing techniques to date have been target-aware (as
discussed in [iii]), suggesting there is a tension between the
usefulness of target knowledge and the difficulty of obtaining it.
Moreover, one of the strengths of combining approaches is that support
can seamlessly shift to match different task constraints (potentially
using target knowledge when it is possible and relying on other
techniques when it not). Though our particular combination is not
perfect, it is the first example of this general approach, and
demonstrates potential that has not been sufficiently explored to
date.

Finally, there were many excellent recommendations for improving the
clarity, which we will incorporate (e.g., we will expand on the
description of the combined technique as recommended by R2).

[i] Andersen et al. Computer mouse use predicts acute pain but not
prolonged or chronic pain in the neck and shoulder. Occup Environ Med,
2008, 65: 126-131.
[ii] Sasangohar et al. Evaluation of mouse and touch input for a
tabletop display using Fitts’ reciprocal tapping task. Proc. HFES’09,
pp. 839-843.
[iii] Wobbrock et al. The angle mouse: target-agnostic dynamic gain
adjustment based on angular deviation. Proc. CHI’09, 1401-1410.