Following Topher's lead and googling for some info I've found the following article:
Conscious and non conscious emotional processes and the arrow of time
It looks like it's an original article by Radin and Bierman.
There I've found the following text:
Examining possible artifacts
Radin (1997) has adequately treated a number of potential normal explanations of the effect. The current replication, using completely different hardware and software, does strengthen the conclusion that the results are not due to technical artifacts.
The major (and maybe only) source of normal explanations remaining after Radin's original analyses is the hypothesis that subjects developed anticipatory strategies that resulted in different anticipatory physiology preceding calm or emotional pictures. At first glance this seems to be a real possibility. However, the current results do not support this idea because the presponse effect does not depend on the ratio between calm and emotional targets in any systematic way. We would expect some systematic relationship to exist if anticipation strategies were indeed based upon the Gambler's Fallacy.
There are three further arguments against an explanation in terms of normal anticipatory strategies. The first is that we find suggestive internal effects that can not easily be explained by this type of strategies. For example, the differences observed between erotic and violent emotional stimuli would require an anticipatory strategy (i.e., a probabilistic strategy) able to discriminate both between upcoming calm and emotional targets and between two types of emotional pictures. This seems most unlikely given that the subjects were blind both to the ratios between calm and emotional targets, and to the content of the emotional targets.
Secondly, sequential presentation histories starting with one emotional, followed by one, two or three calms always have a larger presponse before a final emotional in comparison with a final calm picture. Thus the effect is basically independent of presentation orders. Presentation orders starting with two or more consecutive emotional pictures and with larger lags are too infrequent to analyze.
The final argument is that computer simulations of anticipatory strategies, using the same emotional to calm target ratios and the same number of exposures used in the current studies, do not show the expected main calm vs. emotional effects. It does turn out that these simulations are sensitive for the type of randomization used. If we used a random selection with replacement of the targets, then the simulation effects were nil. However, if we used a random shuffling scheme without replacement then the effects ranged between 0% and 10%. This was a surprise because the reasoning as sketched in the introduction has such a direct appeal.
The following anticipatory strategies were tested:
a) Increase anticipation by 1 unit after each calm target, and reset anticipation to 1 after each emotional target.
b) Double the anticipation after each calm target (to a maximum of 500) and reset the anticipation after each emotional target to either half of the previous value, or 1.
The simulated effects in the open-deck situation were never larger than 2% while the observed experimental effects in Studies 1 and 2 were generally larger than 10%. However, these analyses are by no means exhaustive and there may be less plausible statistical anticipation models that may result in larger differences. The major point in favor of the psi hypothesis is that there are no indications in the real data that support any of these sequential strategy models.