CONCLUSIONES Y RECOMENDACIONES

The present study marks a natural and necessary next step in probing and understanding the structure of cognitive processes. It combines the logic and rationale of mental chronometry with the most recent and auspicious models of decision making to the questions of architecture for multisignal integration and test quality. Answers to the questions will directly improve the theoretical understanding of both coactivation and the RMI test. In addition, the practical value and efficiency of the RMI test will be improved, by solidifying results that have already been or will be obtained in the future.

3

MODELING COACTIVATION IN BIMODAL

DETECTION TASKS

The redundant signals effect is a shortening in mean reaction times (RT), when observers process a redundantly defined stimulus compared to processing a target, defined in only one quality (i.e. feature, dimension or modality; see chapter 2.2 for a detailed

description of the redundant signals paradigm). The race model inequality test allows inferring the cognitive architecture by putting a distributional constraint on this speed-up. Whenever this RMI bound is violated, this is taken as evidence against the whole class of race models and in favor of integrative or coactivation models (see chapter 2.3 for a formal derivation of the RMI test and its statistical profile).

Even when researchers find violations of the RMI, it is still unclear at what stage(s) of the decision process an integration occurred. Decisional accounts assume that the RSE is at least to some degree due to a speed-up at the level of early visual stimulus analysis. This is the stage, where target attributes are coded and compared with those of non-target

elements, without necessarily involving stages following attentional selection (Koene & Zhaoping, 2007; Krummenacher et al., 2001, 2002; Töllner et al., 2011; Zehetleitner et al., 2009b; for a review see Zehetleitner et al., 2008b). There already exist two explicit models of a decisional account of the RSE (Diederich, 1995; Schwarz, 1994), however both only show a good fit on the level of means, while being unable to account for the spread of the data.

Contrary decisional accounts, there are researchers who believe that the RSE is a nondecisional effect and post-selective in nature (Corballis, 1998; Feintuch & Cohen, 2002; Iacoboni & Zaidel, 2003; Miller, 2007; Miller et al., 2009; for a review see Miller & Reynolds, 2003 ). From mean level only, those coactivation accounts cannot be discriminated as only the overall reaction time is measurable and not its latent subcomponents (see however, motor related measures in Ulrich & Giray, 1986; Ulrich & Stapf, 1984). Part of the reason is that the used reaction time models were of a descriptive and atheoretic nature.

Sequential sampling models in contrast, by design allow for an implementation of both decisional and nondecisional accounts to the RSE (see chapter 2.4 for an elaborate description of this model class). However, so far only decisional accounts of the RSE have been empirically tested, as both Diederich’s (1995) and Schwarz’s (1994) models assume a summation in the rate of evidence accumulation for redundant signals over single signals. There are no studies which have tried to fit nondecisional or combined coactivation accounts (where both decision and nondecision parameters may vary) in a comparative fashion.

Apart from this conceptual lack, all of the models so far only look at fits to the means. As reaction times are usually positively skewed, there are distinct patterns, which can be missed, when focusing only on the central moment. Balota and Yap (2011) showed that effects in means can be produced by either shifts of RT distributions, or stretching of slow tails of RT distributions, or a combination of the two. In the context of sequential sampling models, Ratcliff, Thapar, Gomez and McKoon (2004) have argued, that a complete

explanation of processing, requires accounting for all aspects of the experimental data. This encompasses the distributions for correct and incorrect responses as well as the proportion of correct and incorrect responses.

As the redundant reaction times are not only mean shifted but also differ in their skewness (mean-variance relation, Wagenmakers, Grasman, & Molenaar, 2005) an analysis on the mean level is in principle not able to distinguish subtle parameter shifts, which do not exclusively manifest in a change in means (for the problem of model mimicry, see chapter 2.1). Also, response times are known to show a positive correlation between mean and variance: higher mean response times are afflicted with a higher variance and vice versa. Diffusion accounts intrinsically produce this relation of means and variances for response times. Due to this correlation, a nondecisional model alone is unlikely to fit the data

adequately, as there a change in the base time parameter is only capable to shift the whole distributions horizontally. It is unclear whether a combined (decision and nondecision) model can outperform a decision-alone measured in its goodness-of-fit.

This study contributes to the debate on the source of RMI violations and coactivation both conceptually and methodically. A diffusion model analysis was performed to fit quantile proportions of the response times in two bimodal audiovisual RSP experiments to three model variants reflecting different sources of coactivation: (a) a decisional model (drift rate may vary), (b) a nondecisional model (base time may vary), and (c) a combined model (both drift rates and base times may vary). This way, the question of the source(s) of RMI

violations (and the RSE in consequence) can be addressed: Does coactivation occur at a decisional stage, a nondecisional or at both stages and to what degree?

On a methodical level, the present study promotes the application of sequential sampling models (the Ratcliff diffusion model in particular) to the redundant signals

paradigm, as it enables researchers to utilize all of their empirical data (RT distributions for both correct and incorrect response, and error rates) to plausibly map it to latent

psychological variables and generate theory-driven hypothesis to further shed light on the redundant signals paradigm.

To answer this question satisfactorily however, diffusion model analyses of a large class of redundant signals experiments with varying tasks and stimuli have to be performed. In this study, the endeavor is commenced by analyzing to bimodal RSP tasks: Experiment 1 utilizes a simple reaction time task and Experiment 2 a two-alternatives forced choice task.