&copy Copyright 1996 by Barbara Bruhns Frey & Keith Clayton
Presented to the Society for Chaos Theory in Psychology and the Life Sciences, Berkeley, CA; June 1996

Examining Sequence Effects Using Nonlinear Dynamics

Barbara Frey & Keith Clayton

FreyBA@ctrvax.vanderbilt.edu & ClaytoKN@ctrvax.vanderbilt.edu

Vanderbilt University
301 Wilson Hall, Nashville, TN, 37240
telephone: (615) 322-0060 fax: (615) 343-8349



Abstract:

In this paper, nonlinear dynamics is used as a means of describing and testing data sets exhibiting sequence effects. The term sequence effects describes the phenomena in which the dependent measure appears to be systematically influences by prior events such as the stimulus from the previous trial. For example, a particular type of sequence effect known as a repetition effect has faster response latencies to the letter X when the letter X was preceded by another letter X. Although sequence effects have been observed over a wide range of paradigms, the discussion will focus on sequence effects in choice reaction time experiments. There are a number of experimental variables known to systematically influence sequence effects. The question we are examining is whether or not these mean differences in the experimental variables reflect differences in the number of nonlinear dynamic variables (i.e., a change in the dimensionality).


Examining Sequence Effects Using Nonlinear Dynamics

Variability is typically assumed to be the consequence of many unknown independent factors. With the increased availability of nonlinear dynamical statistics, it is possible to test this basic assumption. From a nonlinear dynamical perspective, variability is the interaction of a limited number of non-independent nonlinear dynamical factors. This paper examines these two accounts of trial-to-trial variability. The first account assumes that the variability from trial to trial is independent and the result of random error. The nonlinear dynamics account views variability as deterministic and the result of nonlinear dynamics. Four different tests are used to examine both accounts.

To examine the plausibility of a nonlinear dynamical account of trial-to-trial variability, we chose a data set that appears to have trial-to-trial dynamics. This data set exhibits sequence effects (i.e., behavior that appears to be influenced by prior trial events). An example of a sequence effect is a speeded response when the stimulus is the same as the previous stimulus (e.g., responding to an X preceded by an X would be faster than an X preceded by an O).

Figure 1 depicts the sequence effect data from twelve participants who were instructed to make simple X-O classifications (i.e., when the stimulus X is presented, press the X key and when the stimulus O is presented, press the O key). Although neither the task instructions nor the task required the subjects to use past trial events, subjects appeared to use past trial events from at least four trials back. The sequence patterns (e.g., OXOXX) correspond to the temporal order in which the stimuli were presented. For the sequence pattern, OXOXX, the right most stimulus, X, corresponds to the stimulus most recently presented (see Table 1). As shown in Figure 1, response times (RT) decrease with increasing number of past repetitions (cf. OOOOO) and recent alternations admits a history of repetitions slows response latencies (cf. OOOXO). Sequence effects are not unique to this simple X-O classification task. Sequence effects have been observed in a wide range of paradigms and dependent measures (see for a review; Kirby, 1980; Kornblum, 1973; Luce, 1986).


     Figure 1.  Stimulus Sequence plot for O-X classification task (n=12).  Stimuli O 
     and X are collapsed.  For simplicity, the graph is labeled using sequence patterns 
     corresponding to the O stimulus. (data discussed in detail in Frey, 1995; Clayton & 
     Frey, 1995).  Order corresponds to the amount of past history used to 
     partition latencies (see Table 1).

Table 1. Sample of sequence patterns used in Figure 1 with corresponding order and stimuli history.
Sequence Patterntrial number (t)t-4t-3t-2t-1t
XO2nd orderXO
OXO3rd orderOXO
OOOXO5th orderOOOXO

Both trial-to-trial variability and the dependence on past history suggests that data sets exhibiting sequence effects may be good candidates for a nonlinear dynamical analysis. From a deterministic nonlinear dynamical perspective, trial-to-trial fluctuations such as the ones in Figure 2 are the result of the deterministic interaction of nonlinear dynamical variables and parameters.


     Figure 2.  Response Time (RT) - time series plot for subject 1, O-X classification 
     task.

From a linear perspective, these trial-to-trial fluctuations are typically accounted for by viewing reaction times as composed of a true score plus random error. In the case of the sequence effect data in Figure 1, the true score for a particular sequence pattern would consist of both the grand mean and the sequence effect mean. From this perspective, the repeated measures in the Figure 2 data would be viewed as distributions of responses centered around the true score. Figure 3 gives an example of two possible distributions, one for O responses and the other for X responses.


     Figure 3.  Response latencies to the stimuli O (left) and X (right).

These two general classes of models, linear & nonlinear, differ in their account of variability. From a linear perspective, if the sequence effects were removed linearly, then the resulting time series would exhibit all the characteristics of random error distributed around the grand mean. From a nonlinear dynamical perspective, the variability in the observed times series should exhibited characteristics suggesting low-dimensional nonlinear dynamics.

Methods & Data Sets

The data from participant 1 was obtained from Frey (1995). This data file (Obs. TS) was treated as a single time series of length 500. The participant was instructed to press the key that corresponded to the presented stimulus (an X or an O). The experiment started with 100 practice trials in which the participant received both accuracy and response time (RT) feedback. The practice trials were followed by 500 experimental trials in which the participant received no feedback. Each trial was initiated by the participant. Both the trial initiation times and the response times were recorded. However, only response times are discussed in this paper.

To examine the assumptions concerning the variability in the observed time series (Obs. TS), two other time series are analyzed: a shuffled version of the observed time series (Shuf. TS) and the observed times series with the sequence effects removed (Obs. TS-SEr). The observed time series was shuffled so that the dynamical structure in the time series was destroyed while other characteristics of the data set (e.g., mean and standard deviation) were maintained.

The sequence effects were removed from the observed time series in the following manner. Suppose the observed time series is describe by the following linear notation:

  (Response Time_t )= (Grand Mean) + (Sequence Effect_c) + (Random Error_t), 
          where t= index for a particular trial,
          c = index for a particular fifth-order condition (e.g., OOOOO, see Figure 1)

The observed time series with the sequence effects removed would be equivalent to the above equation with a linear removal of the Sequence Effect_c component. For example, if the grand mean for the data in Figure 1 equals 495 msec and the mean for sequence condition OOOOO equals 418 msec, then the contribution of the sequence effect condition would be a speed up of the response by 77 msec. To remove the sequence effect contribution from a response times (e.g., Response Time_t equals 403 msec), the sequence effect contribution needs to be subtracted out.

     For example:
          Grand mean = 495
          OOOOO condition mean = 418
          Adjustment that removes SE = 77 = 495 - 418
          RT  = 403
          RT SEr = RT  + (adjustment that removes SE)
          RT SEr = 480 = 403 + 77


Tests
The Shuf. TS is used to test the nonlinear dynamical perspective. If the time series was produced by a low-dimensional nonlinear dynamical system, there should be a difference between the shuffled and observed time series. The three data sets (Obs. TS, Obs. TS-SEr, and Shuf. TS) were submitted to a number of statistics designed to test the assumptions of both the linear and nonlinear accounts of variability. Besides testing the two different accounts of trial-to-trial variability, the results from statistics, listed in Table 2, should provide convergent support for either the linear or nonlinear account of the variability. Table 2 also lists the expected outcomes from the two different accounts of trial variability.

Table 2. Summary of statistics.

TestsExpected Outcome from a linear white noise perspectiveExpected Outcome from a deterministic nonlinear dynamical perspective
Spectral analysis: Fourier spectrum (FFT), linear analysis of dominant frequencies.broad frequency spectrum, zero or flat slope; all frequencies equally representedbroad frequency spectrum, could have zero or non-zero slope
Brock, Dechert, & Scheinkman (BDS) for IID: (Brock, Hseih, & LeBaron, 1991)tests the hypothesis of IID (independently & identically distributed). cannot reject the hypothesis of IIDreject the hypothesis of IID
Dimensionality estimates (Grassberger & Procaccia, 1983; Judd, 1992, 1994) estimates the number of dynamic variables needed to describe the data setthe estimate of dimensionality should be infinitethe estimate of dimensionality could be low
Sugihara & May (1990):Examines if predictability changes as prediction time increase - signature of chaotic dynamics predictability should be low & should not change as prediction time increasespredictability should be high & should decrease as prediction time increases

Spectral analysis

The graphical comparison between the FFTs for the three file types is displayed in Figure 4. The shuffled TS exhibits a flat or zero slope typical of white noise. Both the Obs. TS and the Obs. TS-SEr appear to have similar spectra. If the variability in the observed TS is the result of white noise, the Obs. TS-SEr should have a similar spectrum to the Shuf. TS, but this does not appear to be the case (see Table 3). Rather the Obs. TS and the Obs. TS-SEr both have non-zero slope near -0.25 (for similar findings see Gilden, et al, 1995).


     Figure 4.  Fast Fourier Transform for participant 1, X-O classification task.  See 
     Clayton & Frey (1995) for FFT data for each task collapsed across 12 subjects.

The BDS test for IID

The linear account of variability as white noise is the assumption of IID. To test this assumption the data sets were submitted to the BDS statistic (Brock, Hseih, & LeBaron, 1991). The results are presented in Table 3. In both the Obs. TS & the Obs. TS-SEr the assumption of IID can be rejected. This is not the case for the Shuf. TS. The results of both the FFT and the BDS analyses do not provide support for an account advocating that the variability is due to white noise.

Table 3. Results from the FFT analysis and the BDS statistic.

StatisticsObs. TSObs. TS - SErShuf. TS
FFT - flat or zero slope suggests - white noiseslope = -.28slope = -.25slope = -.02
BDS - tests if the hypothesis of IID can be rejectedz-score = 7.9; p<.05>z-score = 5.4; p<.05>z-score = -.53; p>.05

Estimates of Dimensionality

The previous two tests argue against the white noise account of variability. The next two tests examine the low-dimensional nonlinear dynamical account. The first test examine the number of dynamical variables needed to produce the time series (i.e., dimensionality). Two different estimates of dimensionality are used: Grassberger & Procaccia (1983) and Judd (1992, 1994). The Grassberger & Procaccia (1983) algorithm is widely used, but when the sample size is small as in this case it tends to under-estimate the dimensionality. JuddŐs estimate has been reported to give reasonable estimates with small data sets.

The dimensionality is estimated via the point at which the estimated dimensionality stops increasing as the embedding dimension increases. Since white noise is assumed to have an infinite number of factors, the estimated dimensionality should not level off, but rather continue to increase as embedding dimension increases. As seen in Figure 5, neither estimate suggests that the dimensionality is less than 6. For both dimensionality estimates, the Obs. TS & the Obs. TS-SEr appear to have similar estimates. Both of these time series also appear to differ from the Shuf. TS, particularly in JuddÕs estimate. The closeness of dimensionality estimate for the Shuf. TS to the other two time series and the ambiguity in whether the estimated dimensionality is leveling off, leaves these estimates of dimensionality tenuous.


     Figure 5.  Dimensionality Estimates using Grassberger & Procaccia (1983)Judd (1992, 1994) algorithm (right).  Thin solid line 
     corresponds to the expected dimensionality estimate of white noise.

Sugihara & May (1990) Test

The final test (Sugihara & May, 1990) provides a means of examining how predictability changes as prediction time increases. The predictability of white noise should not change as prediction time increases. A signature of chaotic systems, on the other hand, is high predictability that drops off as prediction time increases. As presented in Figure 6, the Shuf. TS behaves very similar to what is expected from white noise (i.e., low-predictability that remains constant as prediction time increase). Both the Obs. TS and the Obs. TS-SEr exhibit non-zero predictability that drops off as prediction time increases. This suggests that the Obs. TS and the Obs. TS-SEr might be the result of a chaotic system.


     Figure 6.  Predictability as prediction time increase, E=11, lag=1, Sugihara & May 
     (1990).

Conclusion

The random error account of trial-to-trial variance via independently distributed white noise is not supported by the statistics. The spectral analysis and the BDS test suggest that the shuffled time series exhibits characteristics of white noise and both the observed time series and the observed time series with the sequence effects removed appear to differ from the shuffled time series. The estimates of dimensionality are less clear. Two different estimates of dimensionality suggest that the dimensionality of the observed times series and the observed times series with the sequence effects removed might be less than the dimensionality of the shuffled time series. The estimates also suggest that the dimensionality for the observed time series and the observed time series with the sequence effects removed might be as low as 6-9 dynamic variables. The observed time series and the observed times series with the sequence effects removed have the signature characteristic of chaotic dynamics in which the prediction accuracy decreases as prediction time increases. The convergent results provide support for the nonlinear dynamical account and these results suggest that other data sets exhibiting sequence effects might be excellent candidates for nonlinear analysis.

Table 4. Summary of Results.

StatisticsObs. TSObs. TS - SErShuf. TS
Spectral analysis non-zero slope non-zero slope, similar to Obs. TS zero slope
BDS test for IID reject hypothesis of IID reject hypothesis of IID cannot reject hypothesis of IID
Dimensionality estimates unclear, greater than 6, maybe differ from white noise unclear, greater than 6, maybe differ from white noise, similar to Obs. TS unclear, greater than 6, maybe differ from Obs. TS & Obs. TS-SEr
Sugihara & May moderately high predictability that decreases with increasing prediction time moderately high predictability that decreases with increasing prediction time, similar to Obs. TS low predictability that does not decreases with increasing prediction time

Notes

During the intervening time, I have had the opportunity to apply the discussed tests to the remaining 11 participants' data sets. All but the dimensionality conclusions appear to be supported. The results from the dimensionality tests do not suggest a difference between the three file types and neither observed time series appears to have a dimensionality less than 7. Click here to see the corresponding data figures for n=12.


References

Brock, W.A. Hseih, D.A., & LeBaron, B. (1991). Nonlinear dynamics, chaos, and 
     instability: Statistical theory and economic evidence. Cambridge, MA: MIT 
     Press.

Clayton, K.& Frey, B. (August, 1995). Studies of Mental 'Noise'. Paper presented to the annual convention of the Society for Chaos in Psychology and the Life Sciences, Garden City, NY.

Frey, B. (Feb, 1995). Examining the Nonlinear Dynamics of Short-Term Memory. Master's Thesis. Vanderbilt University.

Gilden, D. L., Thornton, T., & Mallon, M. W. (1995). 1/f noise in human cognition. Science, 267, 1837-1839.

Grassberger, P., & Procaccia, I. (1983). Measuring the strangeness of strange attractors, Physica 9D, 189-208.

Judd, K. (1992). An improved estimator of dimension and some comments on providing confidence intervals. Physica D, 56, 216-228.

Judd, K. (1994). Estimating dimension from small samples. Physica D, 71, 421-29.

Kirby, N. (1980). Sequential effects in choice RT. In A. T. Welford (Ed.), Reaction times (pp 129-172). London: Academic Press.

Kornblum, S. (1973). Sequential effects in choice RT: a tutorial review. In S. Kornblum (Ed.), Attention & Performance IV (pp. 259-288).

Luce, R. D. (1986). Response Times: Their Role in Inferring Elementary Mental Organization. New York, NY: Oxford University Press.

Sugihara, G. & May, R. M. (1990). Nonlinear forecasting as a way of distinguishing chaos from measurement error in time series. Nature, 344, 734-741.


URL:http://www.vanderbilt.edu/AnS/psychology/cogsci/clayton/papers/C haos96/Chaos96-seq_effects.html