... depression¹

Depression is defined differently for different studies. As individual studies are mentioned, the criteria used for the establishment of a diagnosis of depression for each study will be noted.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 1991).²

It is interesting to note that every study using a priming methodology found effects suggesting that depressed people were faster in naming negative words in a primed vs. nonprimed condition, presumably due to the increased activation of memory structures relevant to the stimulus, before the actual stimulus was presented. These results may provide information as to thresholds for activation of negative or positive constructs needed for results to be obtained in an unprimed condition, but this possibility is not explored in this paper.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... structures.³

Fundamental differences do exist between the notion of a semantic network and a more distributed conception of activations within the brain which are discussed elsewhere (Reggia, Berndt, & D'Autrechy, 1994; Siegle, Ingram, & Matt, 1995). Still, these distinctions will not, in general, compromise the parallels which are drawn between the representations in this thesis.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... decision.⁴

It makes sense to consider this proposition from a physiological perspective, since many structures implicated in the early processing of visual stimuli do project to the hippocampus. For example, the lateral geniculate nucleus, a structure widely implicated in vision, projects to the hypothalamus which is afferent to the hippocampus; the anterior thalamic nuclei which receive visual input also project to the hippocampus (Brodal, 1981).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... words.⁵

While researchers such as Watson and Clark (1984) discuss the considerable overlap in features of depression and anxiety, these disorders may have distinct etiologies (e.g., via overlearning of negative vs. threat stimuli) and relatively similar cognitive products.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... conditions).⁶

Confidence intervals around mean reaction times could not be computed because Mathews and Milroy (1994) did not provide relevant standard deviations. Also, Derryberry (1988) reports reaction times ranging from 478ms to 572ms, potentially weakening the statistical conclusion validity of this difference.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 150ms.⁷

Pilot studies determined that words were almost always identifiable at the 150ms condition in nondepressed people, and identifiable most the time in the 50ms condition. Strauss (1983) also reports finding almost no errors on an affective lexical decision task using a 50ms SD.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...$\eta^2$.⁸

t#tex2html_wrap_inline1346#is calculated as the quotient of the single relevant eigan value of HE^-1/(1+HE^-1), where H is the sum of squares and cross products matrix, and E is the analogous matrix for the error term, as recommended by the Scott Nichols (January, 1995, personal communication).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... condition⁹

Post-hoc contrasts investigating effects of valence within a task are computed using only the data from the 150ms condition since no interactions of depression with SD are present. This is because multiple SD's were included primarily in hopes of finding a single super-detection-threshold SD at which effects were present. Aggregation of data from this SD with shorter SD's may obscure true effects of valence-mediated rumination due to increased noise at these SD's, assumed to happen on data gathered between 50 and 150 ms.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... depression.¹⁰

SCID scores were determined based on a trained rater's initial scoring of the interview; no inter-rater reliability checks were done. This is due to the departure of two research assistants before the completion of the study.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... words.¹¹

This representation is not the only one which could be used. For example, in a previous neural network model of the same tasks (Siegle et al., 1994) I used a representation much closer to Bower's (1984) original semantic network representation, assuming that each word was represented by a single node. This representation became complex because as more words were added to the network, more nodes had to be added, weakening the analogy to biological systems.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... nodes).¹²

In an older model (Siegle, Ingram, & Matt, 1994) I included the direct perception of affective features from stimuli, as inputs to the network, feeding to nodes devoted to the affective valence of stimuli. This representational decision was debatable, since many models of affect (e.g., Williams and Oaksford, 1992) assume that emotive content is a property which emerges from training the an otherwise undifferentiated network on stimuli with affective content. In such a scheme, the emotionality of words is a property of nodes and connections internal to the network. While theorists such as LeDoux (1987) posit that affective content is attached to words at a very early stage of processing, and certainly before associations with the words in the context of an information processing task are accomplished, and while experiments demonstrating biases in processing affective information on preattentive tasks (e.g., Kitayama, 1990) support this assumption, how affect was initially associated with words could not be made explicit within the model. The current decision to attach affective valences to stimuli as part of the perception process alleviates this difficulty and does not prevent the affective valence of stimuli from being perceived before the semantic content, as described below.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... task.¹³

Importantly, this function of noise is an artifact of the use of an error-minimization learning algorithm. Were other weight-changing algorithms applied (e.g., a Hebb rule by which more frequently used connections were augmented) such noise would not be necessary.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.