Human performance in many visual and cognitive tasks declines with age, the rate of decline being task dependent. Here, we used a multiple-object tracking (MOT) task to provide a clear demonstration of a steep cognitive decline that begins relatively early in adult life. Stimuli consisted of 8 dots that moved along linear trajectories from left to right. At the midpoint of their trajectories, a certain number of dots, *D* (1, 2 or 3), deviated either clockwise or counter-clockwise by a certain magnitude (57°, 38° or 19°); the task for observers was to identify the direction of deviation. Percent correct responses were measured for 22 observers aged 18–62 years and were converted to *effective numbers of tracked trajectories* ( *E*) (S. P. Tripathy, S. Narasimhan, & B. T. Barrett, 2007). In 5 of the 7 conditions tested, there was a significant negative correlation between age and *E,* indicating an age-related decline in tracking ability. This decline was found to be equivalent to a mean performance drop of 16% per decade over the four decades of adulthood tested. Further analysis suggests that performance in this task starts to decline at around 30 years of age and falls off at the rate of approximately 20% every subsequent decade.

- The paradigm used by Tripathy et al. (2007) is useful whether the number of trajectories tracked is small, or large. Observers' performance can be measured and compared even when as many as 10 or 12 trajectories have to be tracked.

*E*) dropped very sharply with age over much of the age-range tested, showing that tracking performance in adults starts to deteriorate early in adulthood.

^{−2}. Chin and forehead rests were used to minimize head movements and maintain a viewing distance of 127 cm. At this distance, individual pixels subtended 1′ × 1′. Normal room illumination was used to ensure that the persistence of trajectories on the screen could not be used as a cue to the task.

^{−1}. Each dot subtended 5′ × 5′ and had a luminance of 61.9 cd m

^{−2}. The choice of dot speed and dot size ensured that the motion of the dots was perceived to be smooth at the frame rate used. At the vertical midline of the screen, indicated by two markers, the trajectories of a certain number of dots,

*D*(1, 2 or 3), deviated either clockwise (CW) or counter-clockwise (CCW) by a certain magnitude (57°, 38° or 19°). The

*D*= 3 condition was only tested for deviations of 19°. Tripathy et al. (2007; Figure 9) found that when the angle of deviation and the number of deviating trajectories (

*D*) were both large, the average orientation of the trajectories in the two halves of the screen could provide cues to the direction of deviation. In addition, for these conditions, performance can approach 100% correct; this ceiling is not ideal when one is interested in measuring changes. For these reasons the

*D*= 3 condition was not tested when the deviation was 57° or 38°. The magnitudes of deviation used in this study are all substantially supra-threshold, i.e. much larger than the threshold for detecting a deviation in a single trajectory (Tripathy & Barrett, 2004). When

*D*dots deviated, the trajectories of the remaining (8 −

*D*) dots did not deviate. In any trial, all deviations were in the same direction and had the same magnitude and, in any block, 50% of the trials showed CW deviations. Stimuli were presented for 51 frames (904 ms). All trajectories reached the midline of the screen at the same time (i.e. on frame 26). A typical stimulus presentation is depicted schematically in Figure 1, showing 2 out of 8 trajectories deviating CCW by 38°.

*D*) and magnitude of deviation were fixed within a block of trials, but varied between blocks. Observers were aware that deviations would occur only at the screen midline and their task was to indicate whether the deviations were CW or CCW. Viewing was binocular. Observers were not directed to fixate any particular point on the screen but were encouraged to adopt an eye-movement strategy that they felt might help them best perform the task. Feedback was given as to the correctness of each response in the form of one of two audible tones. Each block of data collection involved 100 trials. For each combination of deviation magnitude and

*D,*observers carried out a practice block. The seven practice blocks were followed by 21 further blocks of data collection. The data presented were obtained in three cycles of 7 blocks each, with each of the seven conditions tested in pseudo-random order within each cycle. The data collection was self-paced, with each cycle of seven blocks typically lasting slightly over an hour, but observers were permitted more time if needed. Results were expressed in terms of the percentage of trials on which deviation direction was correctly reported.

*E*) using the concept of a limited capacity hypothetical observer (LCHO). This process has been described in detail elsewhere (Tripathy et al., 2007), but the basic concept can be outlined as follows. Firstly, it is assumed that this hypothetical observer has a limited amount of tracking resources available and that these resources are distributed among the maximum number of trajectories that can be tracked perfectly (

*A*). Next, since all deviations are substantially supra-threshold, it is assumed that the direction of deviation will be correctly identified if any of the deviating trajectories (

*D*) are among those that are tracked. Given any total number of trajectories (

*T*), the predicted performance of the LCHO (

*P*) can then be described by the following equations based on simple probability:

*y*items from

*x*available items.

*D*> (

*T*−

*A*), as the

*A*trajectories allocated resources will contain at least one deviating trajectory. In Equation 2, the first expression within the double braces represents the probability that the

*A*tracked trajectories will contain at least one deviating trajectory, and the second expression represents the probability of correctly guessing the direction of deviation when there is no deviating trajectory among the

*A*that are tracked.

*D*for various values of

*A*when

*T*= 8 ( Figure 2). For example, if 2 trajectories can be tracked (

*A*= 2) and 2 of the 8 trajectories deviate (

*D*= 2) then the predicted performance would be 73%. The effective number of tracked trajectories (

*E*) is defined as the number of trajectories that must be tracked perfectly by the LCHO in order to match the performance of a human observer. This can be calculated from the human observer's proportion of correct responses for any value of

*D*by interpolating between the curves in Figure 2. The procedure used for interpolation has been explicitly described in Figure 5 of Tripathy et al. (2007). The advantage of this metric is that

*E*permits comparison of performance across different conditions since it takes into account the increased probability of detecting a deviation when there are many deviating trajectories. For example, 80% correct when

*D*= 2 may not be better performance than 75% correct when

*D*= 1. Indeed, Tripathy et al. (2007) showed empirically that while

*E*is an increasing function of the deviation angle, it is largely independent of

*D*.

*D,*and least-squares regression was used to fit a straight line to each data set. For each condition, values of Pearson correlation (

*r*) were calculated and were tested for deviations from zero (Bobko, 2001, pp. 44–47). Figures 3a and 3b show the relationship between percent correct responses and age for deviations of ±57° and ±38° respectively, when 1 (red circles, dashed lines) or 2 trajectories (blue squares, dotted lines) out of 8 deviated. Percent correct responses are generally higher for

*D*= 2 compared to

*D*= 1, but for all conditions performance decreases systematically with age. In each case, the correlations were significantly negative (57° deviations:

*r*= −0.465,

*p*= 0.0292 for

*D*= 1 and

*r*= −0.551,

*p*= 0.0078 for

*D*= 2; 38° deviations:

*r*= −0.666,

*p*= 0.0007 for

*D*= 1 and

*r*= −0.521,

*p*= 0.0129 for

*D*= 2), with a steep decline in performance with increasing age. Figure 3c shows percent correct responses and regression lines for deviations of 19°. Analysis shows that for

*D*= 1 (red circles, dashed line), there is a significant negative correlation between percent correct and age (

*r*= −0.671,

*p*= 0.0006). However, for

*D*= 2 (blue squares, dotted line) and

*D*= 3 (green crosses, dash-dot line), although the correlations are negative, they are not significantly different from zero (

*r*= −0.247,

*p*= 0.267 for

*D*= 2 and

*r*= −0.406,

*p*= 0.0608 for

*D*= 3).

*E*) and the data were re-plotted in terms of

*E*rather than percent correct ( Figure 4). Plotting the data in this way causes a clear overlap of the data for different values of

*D*. Values of

*E*range from 0 to 4.18 for deviations of ±57° ( Figure 4a), from 0 to 3.25 for ±38° ( Figure 4b) and from 0 to 1.44 for deviations of ±19° ( Figure 4c). Again, regression lines were fitted to the data and again, for deviations of 57° and 38° there is a significant negative correlation between age and

*E*for all four combinations of

*D*and deviation magnitude (57° deviations:

*r*= −0.465,

*p*= 0.0293 for

*D*= 1 and

*r*= −0.543,

*p*= 0.0090 for

*D*= 2; 38° deviations:

*r*= −0.661,

*p*= 0.0008 for

*D*= 1 and

*r*= −0.494,

*p*= 0.0196 for

*D*= 2). As before, for deviations of 19°, the correlations are significantly negative for

*D*= 1 (

*r*= −0.657,

*p*= 0.0009) but not for

*D*= 2 (

*r*= −0.236,

*p*= 0.290) or

*D*= 3 (

*r*= −0.388,

*p*= 0.0745). Potential reasons for this are discussed later.

*E*reduces at a rate of 0.36 (

*D*= 1) and 0.41 (

*D*= 2) trajectories per decade for deviations of ±57°, and 0.46 (

*D*= 1) and 0.27 (

*D*= 2) trajectories per decade for deviations of ±38°. Compared to mean performance levels of an 18 year old (as estimated from the regression lines), these figures are equivalent to reductions in performance of 13%, 14%, 20% and 15% per decade respectively, or an average performance drop of 16% per decade over the four decades of adult years tested.

*E*). This age-related decline in

*E*is, therefore, a robust, reliable effect. In the 2 conditions where the age-related correlations did not differ significantly from zero, there may have been a floor effect; performance for young observers is already poor and the performance for older observers cannot drop lower than chance, i.e.

*E*cannot fall below 0.

*D*= 1, for the 12 observers in the 18–31 age range, the mean value of

*E*was 2.62; to obtain the normalized

*E*for this condition, each value of

*E*was divided by 2.62.) Figure 5a shows the normalized effective number of tracked trajectories in all four conditions as a function of age for observers aged 18–31 years. Data for observers in the age-range 28–62 years were similarly normalized and plotted in Figure 5b. Regression lines were fitted to the two sets of data. For the younger set of observers the correlation with age was not significantly different from zero (

*r*= −0.01;

*p*= 0.94) and the slope of the regression-line was 0.006 per decade. For the older set of observers the correlation was significantly different from zero (

*r*= −0.473;

*p*= 0.00008) and the slope was 0.255 per decade. From the fits in Figures 5a and 5b, the normalized effective number of tracked trajectories for an 18-year-old observer is 1.004 and that for a 28-year-old observer is 1.332 (this is greater than for the 18-year-old observer because of separate normalization of the data for the two age-groups). Compared to these baselines the drop in performance is 0.6% per decade over the 18–31 year-range and 19.2% per decade for the 28–62 year-range. Our data suggest that the age-related decline in tracking performance is very small or non-existent in adults younger than 30 years, but drops off very steeply, at almost 20% per decade over the 30–60 year-range.

- Tripathy and Barrett (2004) measured deviation thresholds for single-trajectory stimuli as a function of the number of frames presented; deviation thresholds varied systematically with the number of frames in the stimulus even when the number of frames exceeded 15 (see their Experiment 1, Figure 2a). It is likely that at least 15–20 frames are utilized when tracking a single item and many more frames when multiple items are to be tracked. (Also see Hohnsbein & Mateeff, 1998, for the minimum time that a deviation must persist in order to be detected.)
- Using similar stimuli, Tripathy and Barrett (2006) showed that even when the central part of the trajectories are occluded by having the deviation occur within the blind spot of the observer, observers can report the deviation of the target trajectory as well as the perceived path of the moving target dot. Reporting the perceived path of the target dot requires tracking the target dot.
- Kanai, Sheth, and Shimojo (2007), using stimuli undergoing linear motion, showed that transients in the stimulus occurring within the first 200 ms of motion onset were more easily detected than transients occurring after more than 300 ms following motion onset, suggesting that longer sequences were “perceived as single, indivisible gestalt integrated over space as well as time” (p. 937). The duration of our stimuli was well in excess of 200 ms.

*E*generally increase with deviation angle (see Figure 4) is in agreement with the latter proposal, i.e. the more difficult the task, the fewer the number of trajectories that can be tracked.

*E*for each observer, averaged across different values of

*D,*were calculated for each magnitude of deviation. Regression lines were fitted to these values and are shown in Figure 6. That the lines are separated and non-overlapping highlights that the main factor affecting performance in this task is the angle of deviation. This, along with the fact that plotting the data in terms of

*E*rather than percent correct collapses together the data for different values of

*D*( Figure 4), indicates that, for each deviation magnitude, there is a distinct capacity limit to the number of trajectories that can be tracked, and this limit is independent of

*D*(Tripathy et al., 2007). Note, however, that this independence is an approximation, as can be seen from the slightly different slopes seen in the fits for the two values of

*D*in Figure 4b.

*E*to be, on average, only 15% lower in amblyopic eyes (i.e. equivalent to the reduction seen each decade in normal observers in the current study) compared to non-amblyopic, even though the acuity of the amblyopic eye of some observers was as much as 800% poorer.