To control for the possibility of type II error in planned analyses performed in Greg Siegle’s dissertation, a series of exploratory analyses was also performed. These analyses examined interactions that qualified planned contrasts, effect size estimates when nuisance covariates were accounted for, and various indices of pupil dilation waveforms (e.g., mean dilation, peak dilation, time to maximum pupil dilation). They also included sensitivity analyses involving aggregation of behavioral and physiological indices a number of different ways. As these analyses did not directly address questions discussed in previous sections, their results were not included in the official document. Instead, they are included here.

The philosophy of this adjunct to the dissertation is that multiple indices and multiple ways of aggregating empirical data could be used to test any of the predictions for each of the domains examined within the dissertation. Examining all possible indices and methods for aggregating data would be cumbersome; highlighting just analyses that worked would likely capitalize on chance findings. At the same time, restricting analysis of the data to a single method, and not examining the robustness of findings could also compromise the depth of understanding gained from the data and the generalizability of findings. To strike a balance between these concerns the primary analyses in the dissertation involved only a priori contrasts. This more exploratory section was created as a way of understanding those data from those contrasts from a broader perspective that can stand as a springboard for future research.

To examine whether the model of semantic distance between representations of positivity and negativity in the model are consistent with actual semantic distances, Luce’s (1963) bias independent confusion coefficients (h) can be calculated for each cell of the confusion matrix as

h{i,j}=

The resulting matrix yields the observed semantic distance between valences for each group of individuals. As all terms in the formula are symmetric, semantic distances between any two valences are assumed to be symmetric.

PCA’s provide empirically derived summative indexes of cognitive activity. To make sure that any null results obtained using PCA reflect a true lack of relevant differences in cognitive activity, rather than poor summative indices, it was useful to explore other methods of aggregating pupil dilation. Thus, differences in various intuitive and common indices of late pupil dilation to personally relevant negative words and nonpersonally relevant positive words (the theoretically largest contrast on the valence identification task) between depressed and nondepressed individuals were examined. Similarly, to make sure that obtained null results were not a function of the method under which data were aggregated, various strategies for aggregating data were also explored. Because so many tests were conducted using these techniques, results obtained from them should be considered as sensitivity analyses; they are exploratory, and subject to replication.

For each analysis, indices of pupil dilation were examined under all combinations of the following conditions: inclusion of all data versus data only for words on which the valence rating at testing matched the normed valence or valence under which the words were generated (using the same criteria as for the reaction time analyses), inclusion or exclusion of valence confusions or incorrect lexical decisions, with and without simple reaction time waveforms subtracted from results to account for differences in motor responding, and with and without baseline dilations, defined as average pupil dilation in the one second preceding stimulus onset, subtracted from results.

To examine whether differences in characteristics of the entire waveform existed, the following variables were examined: total pupil area, amplitude of the peak (greatest point) in pupil dilation, and latency to peak amplitude. To examine whether characteristics of late pupil dilation, representing ruminative processes, differed between depressed and nondepressed individuals, two strategies were adopted. First, total area, peak amplitude, and the slope of pupil dilation occurring from three seconds after stimulus onset until the end of the trial was examined. Alternately, in case differences in reaction times greatly affected pupil dilation, the same variables (total area, peak amplitude, and the slope of pupil dilation) were examined in the window beginning 1.67 seconds after an individual’s reaction time on a given trial. 1.67 seconds was chosen based on visual inspection of a number of waveforms. This was the point in many non-depressed individuals at which pupil dilation returned to baseline after a response to the task.

Past research on pupil dilation has used each of these measures to reflect aspects of cognitive load (see Steinhauer, 1982 for a review). Peak amplitude is assumed to reflect the greatest amount of cognitive effort an individual pays during a time window. Peak latency is the time until that amount of effort is reached; if depressed individuals engage ruminative processes earlier than nondepressed individuals, their peak latency might be lower. Pupil area represents the average cognitive load expended during a time window (the whole task or after the "late" criterion). The slope of pupil dilation reflects the rate of change of pupil dilation in a time window. Based on the neural network model’s performance, it is expected that if depressed individuals ruminate on negative words in the late stages of attention, that their attention will increase throughout the task, even after their reaction. The slope of their pupil dilation, long after they react should therefore be increasingly positive. Individuals who do not ruminate are not expected to show this change in slope.

Two sets of analyses were performed on data collected in conjunction with Greg Siegle’s dissertation: planned contrasts and exploratory follow-up analyses. Error was controlled conservatively for all planned contrasts as described in the official document.

Because exploratory ANOVAs were interpreted as hypothesis generation tools, alpha was not as strictly controlled for these tests. Rather, to strike a balance between minimizing Type I and Type II errors, alpha was controlled at a level of .05 at each level of a hierarchical analysis. For example, interactions were tested at an alpha of .05, followed by simple effects tests, also evaluated with alpha set to .05.

A final, novel method of Type I error control was adopted for the analysis of late pupil dilations using multiple indices. A large number of tests were employed to determine which indices were robust to variations in conditions such as different baselines. Rather than controlling alpha for each test, a conservative approach is taken in which variables that show large differences between depressed and nondepressed individuals under large numbers of methods of aggregation are suggested to be the most robust indicators of a true difference. That is, it is unclear which method of aggregation should be considered "correct", but if results emerge from only one method of aggregation, they could be a function of the aggregation more than actual patterns in the data. If results emerge from multiple methods of aggregation it is more likely that they are truly present in the data. Tests of different variables are assumed to reflect different families of error. As such, each test was individually evaluated using an alpha of .05, but only indices that yielded significance levels below .05 under multiple conditions (over three) were interpreted as robust.

Not all planned contrasts were statistically significant. It is useful to explore whether obtained null results were due to a true lack of relationship, or possibly due to higher order qualifying interactions. Similarly, it is useful to explore the robustness of obtained results that were statistically significant through sensitivity analyses. In addition, many relationships in the data were suggested by the reported descriptive statistics, and may be informative. The following section examines these possibilities. Results from these analyses should be viewed as exploratory. Conclusions drawn based on them can be interpreted as promising directions for future research.

Planned analyses excluded trials for which valence ratings did not match normed valences, or valences for which words were generated, e.g., for which a person rated a word as negative that he or she had listed as a positive word. This technique had the consequence of retaining fewer matching trials for depressed than nondepressed individuals. Depressed people’s ratings of words matched their normed valence for an average of 46.5 of the 60 presented cases (M=77.3%, SD=.105) while nondepressed people’s ratings matched the normed valence for stimuli for an average of 50.3 of the 60 presented cases (M=83.3%, SD=.088), t(47)=-2.35, p=.023, D=6.49%. Potentially this restriction could thus have confounded other obtained results.

Table 1 presents descriptive statistics for match rates for each valence in depressed and nondepressed individuals. To examine whether the different elimination rates were primarily due to differences in match frequencies for a particular valence, a valence (3: positive, negative, neutral) x status (2: depressed, nondepressed) multivariate split plot ANOVA with match percentage was performed. This analysis revealed a statistically significant valence x status interaction, F(2,46)=5.41, p=.008. Simple effects analysis revealed that depressed individuals matched reliably less often for positive words, t(33.8)=-2.89, p=.005, D=-.17%, while differences were not statistically significant for negative words, t(47)=1.3, p=.20, D=.056%, and for neutral words, t(47)=-1.5, p=.119, D=-.08%.

A main effect of valence was found in the depressed group, F(2,22)=.3.91, p=.015$, h²=.26. Pairwise contrasts revealed that negative words were more often matched than positive words, F(1,23)=6.23, p=.019, h²=.22, and neutral words, F(1,23)=4.37, p=.048. h²=.16. No statistically significant main effect of valence was found in the nondepressed group, F(2,23)=1.89, p=.174, h²=.14.

Table 1

These findings suggest that depressed individuals reliably rate negative words as negative. In contrast, depressed individuals are not as likely to categorize positive or neutral words consistently. Nondepressed people do not present these biases.

Based on this data, it is unclear whether to have excluded stimuli for which normed or generated valences were not consistent with obtained ratings. For example, excluding trials on which positive words were categorized as somewhat negative ensured that measured reactions to positive words represent reactions to stimuli that are likely to be perceived as positive. From such analysis conclusions about depressed people’s reactions to information they perceive as positive could be made. By not excluding these trials it was possible to understand how depressed individuals perceive information that, for normed words, is labeled positive by other people, or for idiosyncratic words, that is sometimes labeled positive by them. This type of analysis could give an indication of the general type of reactions that depressed individuals may have to nominally positive information.

To examine whether non-matching trials might also provide valuable information, the same analyses for which planned contrasts were explored were re-examined without excluding non-matching trials. Results that differed are reported below. Table 2 shows the means and standard deviations of harmonic mean reaction times for the data set in which non-matching trials were not excluded.

Table 2

Significant differences between depressed and nondepressed individuals were not observed on the lexical decision task, in a manner consistent with the network’s performance. Potentially these null results were due to a true lack of group differences. Alternately, other features of the populations or the task could have obscured obtained findings. Additionally, the role of including personally relevant stimuli in the task is still unknown. As such, a number of subsequent analyses on the data were performed.

Analysis without personally relevant information. Previous experiments using affective lexical decision tasks have only examined non-personally relevant information. To make effects from the current study comparable to those observed in the other literature, a valence (positive, negative, neutral) x status (depressed, nondepressed) MANOVA on reaction time, using only non-personally relevant stimuli was performed. No planned contrasts from this analysis were statistically significant.

When all stimuli were used, the relative order of harmonic mean reaction times to each valence were the same as for the analyses using only stimuli for which valence ratings matched normed valences, but some effect sizes for relevant contrasts were higher.

Specifically, contrasts from this MANOVA suggested that depressed individuals were faster to respond to negative than neutral words, while nondepressed individuals were slightly slower F(1,46)=7.18, p=.01, h²=.12, though the difference in depressed and nondepressed individuals’ responses to positive versus negative words was not statistically significant, F(1,46)=1.66, p=.20, h²=.04. These results are not consistent with the network’s prediction that depressed individuals would be especially slow to respond to non-personally relevant negative information.

Omnibus test. To better understand the role of neutral and positive personally relevant information, a valence (3: positive, negative, neutral) x personal relevance (2: relevant, non-relevant) x status (2: depressed, nondepressed) MANOVA was performed. The test revealed a marginally significant valence x status interaction, F(2,46)=2.01, p=.147, h²=.087, and a personal-relevance x valence interaction, F(2,42)=4.41, p>.018, h²=.174. Simple effects analysis of the marginal valence x status interaction did not reveal a statistically significant effect of valence for depressed individuals.

Because differences in mean reaction times within groups appeared larger for the data set in which all words were used, the same analysis was conducted there. A valence x status interaction was apparent in this data set, F(2,21)=5.11, p=.016, h²=.327. Pairwise contrasts revealed that nondepressed individuals were reliably faster to respond to positive than neutral words, F(1,22)=10.3, p=.004, h²=.319. No other pairwise contrasts were statistically significant. The effect of valence was also statistically significant in the nondepressed group for the data set with all words, F(2,23)=7.46, p=.003, h²=394. Pairwise contrasts suggested that depressed individuals were reliably slower to respond to negative than either positive, F(1,24)=15.587, p=.001, h²=.394, or neutral words, F(1,24)=15.59, p=.001, h²=.194. Examination of the confidence intervals around the difference between depressed and nondepressed responses shows that depressed individuals responded reliably more slowly to neutral words, 95% CI=[25;225] ms, and positive words, 95% CI=[3;198] ms, but not to negative words, CI=[-23;166] ms.

Simple effects analysis of the personal-relevance x valence interaction revealed that individuals were slower to respond to personally relevant than non-personally-relevant positive information, t(46)=-3.41, p=.001, D=52ms, as well as neutral information, t(45)=-3.52, p>.001, D=67ms. In the data set in which matches were not required, individuals were faster to respond to personally relevant than non-relevant positive information, t(47)=5.41, p>.001, D=79ms, but were slower to respond to personally relevant than non-personally relevant neutral information, t(47)=-4.8, p>.001, D=89ms. This discrepancy could suggest that obtained results in either condition were artifactual, or that that qualitative differences in responses to matched versus unmatched positive information may be important to investigate in future studies.

To better illustrate the nature of the personal-relevance x valence interaction a number of complex contrasts were also performed on the data set in which valence matches were not required. These revealed that the difference in response times to personally relevant and nonpersonally relevant information was larger for negative than neutral, F(1,46)=12.79, p=.001, h²=.218, or positive words, F(1,46)=13.78, p=.001, h²=.231. Because this difference was so strong for all individuals, it may have tempered observed differences between depressed and nondepressed individuals.

Similarly, to better understand the role of variables such as personal relevance, additional analyses were performed on the valence identification task data.

Omnibus tests. To examine effects not revealed by planned contrasts, the omnibus test from the valence x status MANOVA with reaction times was analyzed. The MANOVA revealed a statistically significant valence x status interaction, F(3,42)=5.94, p=.002, h²=.298. Simple effects analysis revealed that depressed individuals were reliably slower than nondepressed individuals in responding to positive, t(33.36)=4.02, p>.001, 95% CI=.185-.563, negative personally-relevant, t(28.58)=3.07, p=.005, 95% CI=.120-.601, and neutral words, t(46)=3.69, p=.001, but not to negative words that were not personally relevant, t(46)=1.83, p=.074, 95% CI=-.015-.313. Simple effects analysis also revealed a statistically significant effect of valence within the depressed group, F(3,18)=5.08,p=.01, h²=.46, largely accounted for by slow responding to positive and personally relevant words. A main effect of valence was also present in the nondepressed group, F(3,22)=21.48, p>.001, h²=.75. Positive words were reliably more quickly responded to than words of all other valences, F(1,24)=26.5-51.4, p>.001.

Together, these results suggest that people who are not depressed are quick to say that positive words are positive. This same processing advantage for positive words was not present for depressed people. In addition, depressed individuals, are quick to say that negative words are negative, when they are not personally relevant, but are especially slow to respond to personally relevant negative information.

Omnibus tests for a valence x personal relevance x status ANOVA. To understand whether observed effects were specific to negative personally relevant words, or generalized to all personally relevant information, a valence (3: positive, negative, neutral) x personal relevance x status MANOVA was performed. This test revealed a statistically significant valence x personal-relevance x status interaction, Wilks l=.814, F(2,41)=4.699, p=.015, h²=.186.

Simple effects analysis revealed a valence x status effect for non-personally-relevant words, F(2,44)=6.787, p=.003, h²=.236. Simple-simple effects tests showed a significant effect of valence for non-personally relevant words for depressed individuals, F(2,20)=6.68, p=.006, h²=.401. As reported previously, depressed subjects responded to positive words reliably more slowly than negative words, F(1,21)=11.00, p=.003, h²=.34 and neutral words, F(1,21)=13.3, p=.002, h²=.383. Similar results held for nondepressed individuals.

In contrast, for personally relevant words, the valence x status interaction was not statistically significant, F(2,42)=.317, p=.737, h²=.015. A main effect of valence was present, though, F(2,42)=33.96, p>.001, h²=.678. Contrasts revealed that all valences were statistically significantly different from each other, F(1,44)>22.4, p>.001, h²>.34; both groups responded most quickly to positive words, followed by negative, and then neutral words.

All analyses were re-examined with all words, rather than just words for which the valence rating at testing matched the normed valence rating. This restriction did not change results qualitatively except as noted previously. All observed statistically significant effects reported above became larger. In no case did the inclusion of age, education, or gender as covariates change results qualitatively. Additionally, restricting the analysis to only responses that matched the normed valence of words (i.e., responding "neutral" to neutral words) did not change the results qualitatively. ANOVA summary tables of relevant effects tests and contrasts are presented in Appendix L for each of these analyses.

Planned analyses explored only the idea that depressed individuals would tend to label words as negative, more often than nondepressed individuals. A number of statistical analyses can be used to further explore the generality of this idea. First, implications for overall confusion rates, rather than confusions of one valence with another can be explored. Second, the obtained confusion rate data can be used to better understand the obtained reaction time data. Third, valence confusion data can be used to understand whether depressed and nondepressed individuals differ in the semantic distance between their representations of positivity, negativity, and neutrality.

Confusion matrices are reported in Table 3, in which each cell represents the average proportion of responses to a stimulus normed as having one valence that are confused with (i.e., responded to, as) another valence during the task. Mean confusion rates for non-personally-relevant words are presented in the table. In the table, stimuli are presented vertically, and responses are presented horizontally. Proportions of correct responses are reported along the diagonal. Because confusion rates could be confounded by different semantic distances between the valences, they are not interpreted without further analysis.

Table 3

Greater Numbers of Valence Confusions for Depressed Individuals. Due to low numbers of valence confusions, formal statistical tests could not be conducted on the bias parameters extracted using Luce’s response rule. A similar strategy does yield testable results though. Specifically, if depressed people label things as negative regardless of their actual valence, it is expected that their overall rate of confusions will be higher for non-negative words than for negative words. This result is not expected to hold for nondepressed individuals. Table 4 presents the proportion of valence confusions (e.g., the proportion of times positive words were labeled as negative or neutral) for each type of stimulus on the valence identification task.

Table 4

When personally relevant words were not included in a valence x status MANOVA, the status x valence interaction was not statistically significant, F(2,45)=1.65, p=.203, h²=.068. A main effect of valence accounted for a great deal of the variance in error rates, F(2,45)=24.47, p>.001, h²=.521; both depressed and non-depressed individuals made the fewest confusions on negative words and the most confusions on neutral words. No statistically significant differences between negative, and other words were observed, both before and after arcsine transformations to account for positive skew in the confusion proportions.

Because depressed people made many confusions on positive personally relevant words, while nondepressed individuals did not, contrasts from a valence x personal-relevance x status MANOVA suggested that depressed individuals displayed a greater difference in positive and negative confusions than nondepressed individuals when personally relevant and non-personally relevant responses were averaged, F(1,46)=52.27, p=.026, h²=.103. Interestingly, many individuals appeared to frequently classify personally relevant neutral words as either positive or negative, suggesting that individuals have a difficult time generating neutral words that are truly devoid of affect.

Omnibus tests. To fully assess the impact of personally relevant stimuli on information processing in depression, the omnibus tests from the valence x personal relevance x status MANOVA were examined. A marginally significant valence x status interaction was present, F(2,45)=3.02, p=.058, h²=.119, and was largely explained by the valence differences described above. In addition, a personal-relevance x valence interaction was present, F(2,45)=29.6, p>.001, h²=.568. Simple-effects analyses reveal statistically significant differences between personally relevant and non-personally relevant word types for each valence, as shown in Table 5. The differences in the table represent differences in proportions of confusions to personally relevant and non-personally relevant words, ranging from a possible minimum of zero to a possible maximum of one. Personally relevant negative and neutral information were confused more frequently than non-personally relevant information, while personally relevant positive information was confused less frequently than non-personally relevant information.

Table 5

Reinterpreting Valence Identification Reaction Time Data for Personally Relevant Information. Confusions during forced choice tasks may represent an individual thinking about one kind of information in response to the presentation of another. For example, if a depressed person says that a positive word is negative, it may be that the individual has seen a typically positive stimulus in a negative light. This type of confusion may be thought of as "affective interference." Because reaction times are often thought to index the same interference processes as confusion matrices (Podgorny & Gardner, 1979), confusion matrices can aid in the understanding of valence-identification reaction times.

Specifically, depressed individuals were slow to respond to personally relevant negative words on the valence-identification task. This finding was at odds with the expectation that depressed individuals would be especially fast to identify personally-relevant negative information, because they had learned these associations well. If the same processes led depressed individuals to be slow to identify personally relevant negative and positive information (e.g., a general lack of attention), error rates for each type of information should be similar, and confusions should be symmetric. Alternately, if different processes caused depressed individuals to react slowly to positive and negative words, confusion rates could be assymetric. Specifically, if depressed individuals were paid so much attention to negative words that they didn’t react to the task, but did not pay so much attention to positive words, it is expected that depressed people would have lower confusion rates for negative than positive words. Such "motor inhibition" on negative words might also lead to especially low correlations between reaction times and confusion rates for personally relevant negative words.

As discussed previously, confusion rates were lower for negative than positive words in depressed individuals, whereas this was not the case for nondepressed individuals. This observation is consistent with the idea that depressed individuals think of positive information in a negative light, and are slow to respond to negative information due to motor inhibition processes. Yet, personally relevant negative words were categorized as positive by depressed individuals in an average of 1.6% of cases, a virtually identical confusion rate to the 1.8% of cases in which positive non-personally-relevant words were categorized as negative. Still, all the confusion rates are quite low. Potentially the apparent symmetry is due to a floor effect.

Possibly the most compelling evidence that different processes operate in producing long reaction times to positive and personally relevant negative words can be seen by examining the relationship between reaction times and error rates in depressed individuals. In the depressed group, errors were strongly correlated with reaction times for personally relevant positive words, r=.581, p=.004, and non-personally relevant positive words, r=.420, p=.046, but not for personally relevant negative words, r=.247, p=.255, or non-personally relevant negative words, r=.007, p=.976. These correlations are consistent with the idea that slow responses to positive words involved interference by thoughts of other valences, while slow reaction times to negative words did not.

Semantic Distance Versus Bias. Williams et al.’s results suggested that positivity, negativity, and neutrality are not equidistant from each other in semantic space. This hypothesis can be tested within the confusion matrix data by showing that confusion rates are not similar between the valences. Bias-independent confusion rates derived using Luce’s response rule were generated. Using these parameters, a confusion (3: negative-positive, positive-neutral, negative-neutral) x status MANOVA revealed a main effect of valence, F(2,44)=11.21, p>.001, and no statistically significant interaction between depression status and valence, F(2,44)<1, p>.87. As such, biases, for depressed and non-depressed individuals counted together, are presented in Table 6. In the table, higher confusion rates are associated with smaller semantic distances. No pairwise differences in confusion rates were statistically significant on follow-ups to the omnibus MANOVA. As negative and positive words appear slightly closer to neutral words than to each other, as used in the network’s representation, larger samples could support the representation used in the network model.

Table 6

False alarm rates were not statistically significantly different for non-personally relevant words for depressed and nondepressed individuals on the task. To examine whether this result was due to true null effects, or other aspects of signal detection, overall signal detection rates (D) were calculated for each valence. Signal detection rates (D) for each valence are shown in Table 7. D values greater than .3 are considered adequate classification rates. As all D values in the table were relatively high, both depressed and nondepressed individuals could be considered good at classifying words on the lexical decision task.

Table 7

A valence x personal relevance x status MANOVA was performed on signal detection rates (D). This analysis was used specifically to determine whether systematic disturbances in signal detection were associated with some valence for either group. No statistically significant main effects or interactions with status were present, p>.33, h²<$.02 for all tests. There was a personally relevance x valence interaction, F(2,45)=5.89, p=.005, h²=.207. Simple effects analysis revealed that signal detection was generally lower for positive personally relevant words than non-personally relevant words, D=.034, t(47)=3.86, p>.001, while differences in detection rates between personally relevant and non-personally relevant negative, D=.005, and neutral words, D=.021, were not statistically significant. A statistically significant effect of valence was observed for both personally relevant words and non-personally relevant words, F(2,46)=7.49, p=.002, h²=.246. For non-personally relevant words, discrimination rates were significantly higher for negative words than both positive words F(1,47)=11.1, p=.002, h²=.192, and neutral words,F(1,47)=7.88, p=.007, h²=.144, but not did not differ statistically significantly between positive and neutral words, F(1,47)=.086, p=.770, h²=.002.

Planned analyses examined contrasts on five summative indices of pupil dilation, empirically derived through principal components analysis (PCA). These analyses did not generally detect expected differences in pupil dilations for different valences. Perhaps some differences in pupillary responses to different valences did exist, but were not tested for using planned contrasts. Alternately, other theoretically derived indices of pupil dilation may reveal differences in pupillary responses to stimuli of different valences that were not detected by the empirically derived indices. Each of these possibilities is discussed in the following section. Additionally, relationships between the first extracted (ruminative) factor and a self-report measure of ruminative coping are investigated.

To examine whether pupil responses differed between groups in ways not predicted by the network, as a function of valence or personal relevance, a valence (3: positive, negative, neutral) x personal relevance (2) x component (3: ruminative, cognitive, motor) x status (2: depressed, nondepressed) multivariate split-plot ANOVA on component loadings was performed on each task. On the valence identification task, the ANOVA revealed a personal relevance x component x valence x status interaction, F(4,39)=2.75, p=.042, h²=.22. Simple effects analysis of non-personally relevant words were largely accounted for by the differences between groups in responding to the different components, described through planned contrasts.

Depressed individuals appeared to respond remarkably similarly to words of all valences. Within the depressed group, no interactions were statistically significant (p>.24, h²<.27). The only statistically significant main effect was for component, F(2,18)=4.11, p=.03, h²=.31. This variation was primarily accounted for by a linear increase in factor loadings for later components, F(1,19)=6.25, p=.02, h²=.24. In contrast, in the nondepressed group, a statistically significant interaction of personal relevance with component was present, F(2,22)=9.66, p=.001, h²=.47, and a main effect of valence, F(2,22)=9.60, p=.001, h²=.47, accounted for by the fact that positive words generally provoked much less dilation than neutral words, F(1,23)=19.3, p>.001, h²=.456.

When the early attentive component (factor 4) was added into the ANOVA as an additional level of component, the four way interaction was not present. No interactions between valence and status accounted for more than 11% of the variation in factor loadings.

The first extracted component from the PCA peaked approximately four seconds after individuals reacted to stimuli on each task. It was thus labeled a ruminative component, representing cognitive activity occurring long after task-relevant behavior. Because the lexical decision task revealed higher factor scores on the ruminative factor for negative personally relevant information than other information, it is suggested that this ruminative factor may reflect the same processes thought to operate in depressive rumination.

To assess whether aspects of pupil dilation were predictably related to conventional assessments of ruminative coping processes, or distractive coping processes which are assumed to be negatively related to rumination, six hierarchical regressions were performed using rumination (factor 1) factor scores in response to positive, negative, and neutral personally relevant and nonpersonally relevant words as response variables. On the first step of the regressions, covariates including gender, age, and education were entered. On the second step, depression status was entered. On the third step, the ruminative and distractive coping scales from the Response Styles Questionnaire (RSQ) were entered. In no case did RSQ scores statistically significantly predict scores on the first factor above and beyond the other variables (maximum R²change<.03).

Bivariate correlations between RSQ scores and scores on the first factor were, in fact, uniformly low, as shown in Table 8. Based on these results, it is suggested that the late factor may represent ruminative coping processes, but not the same ruminative processes measured by the RSQ. Alternately, the late factor may not represent ruminative processes at all, but rather any thinking that occurs, about any topic, seconds after an experimental stimulus disappears.

Table 8

Principal components analysis did not reveal the expected differences in depressed and nondepressed individuals’ responses to personally relevant negative information and other types of information. This null result could be due to a true lack of differentiation in responses to different valences. Alternately, it could be because the PCA did not yield a summative index of late pupil dilation that adequately captured relevant differences. To examine whether other, more traditional indices of pupil dilation might reveal the expected differences in responses to different valences, complex contrasts were performed on traditional indices of pupil dilation, using various strategies for eliminating potentially irrelevant aspects of pupil dilation, as described in the Methods.

Appendix M presents the 198 relevant complex contrasts, along with t-tests of each contrast, for each pupil dilation index, analyzing whether the difference in pupil dilations to personally relevant negative words was greater than that for non-personally relevant positive words, for each task. Mean indices for negative personally relevant information were higher than for positive information on many of these variables. Still, this was the case for both depressed and nondepressed individuals, and thus few tests of differences between depressed and nondepressed individuals were significant at a level of p=.05. The following sections discuss patterns of statistically significant results from the tests reported in Appendix M.

Valence Identification Task. On the valence identification task, one variable consistently discriminated between depressed and nondepressed individuals. The difference in slope of pupil dilation to negative and positive words, three seconds after stimulus onset, was greater for depressed than nondepressed individuals. This phenomenon was apparent when the match between rated and normed valences was enforced, regardless of whether baselines were subtracted, or correct trials were removed. On examination, depressed individuals had nearly equivalent negative slopes for positive and negative words. Nondepressed individuals had slightly less negative slopes for positive words, and more negative slopes for negative words. This relationship suggests that depressed individuals appear to think equivalently about positive and negative information long after a stimulus is presented. Nondepressed individuals appear to think more after seeing positive information, and less after seeing negative information.

One other contrast was statistically significant. When all stimuli were used, and a baseline was subtracted, depressed individuals had a slightly higher peak in the window after three seconds post stimulus onset, in response to negative words than positive words. Nondepressed individuals had an even higher peak in response to negative words. Thus, the difference between peaks to negative and positive words was greater for nondepressed than depressed individuals. Because this difference was not significant under other methods of aggregation it is likely to be a chance finding, or due primarily to valence confusions or inconsistent ratings.

No other variables yielded statistically significant differences between depressed and nondepressed individuals. To support these results, a number of ANOVAs on the described variables including valence (positive, negative, neutral), personal relevance, and status (depressed, nondepressed) were performed, which yielded no statistically significant valence x status interactions.

Lexical Decision Task. Findings for the lexical decision task are more complex than for the valence identification task. As on the valence identification task, the variable that discriminated most consistently between depressed and nondepressed individuals was the slope of pupil dilation in the window three seconds after the onset of the stimulus. This was true whenever matching the rated valence to the normed valence was enforced, regardless of whether incorrect trials were eliminated, and regardless of whether baselines were subtracted. Yet, the direction of the results was reversed. Depressed individuals had a slightly negative slope for both positive and negative words throughout the window. Nondepressed individuals had a similarly negative slope for negative words, but a much more negative slope for positive words. The slope of pupil dilation after the reaction time followed the same pattern as the slope three seconds post-stimulus, when simple reaction time curves were subtracted with slopes for depressed individuals generally being positive, and the slope for nondepressed individuals, on positive words, being negative. This finding suggests that depressed individuals continued to think about positive and negative words long after they were presented, as on the valence identification task. Nondepressed individuals quickly ceased to think about positive words.

In addition, a number of other indices discriminated between nondepressed and depressed individuals. Specifically, the average pupil dilation after the onset of a stimulus was slightly greater for positive than negative words in both groups, but the difference was much larger in nondepressed individuals, in all conditions in which the baseline dilation was not subtracted. This finding suggests that nondepressed individuals thought much more about positive words in general than did depressed individuals.

The most complex finding regards peak amplitude. The difference in peak amplitude for personally relevant negative and nonpersonally relevant positive words was greater for depressed individuals when baseline values were not subtracted and when rating’s were forced to match normed ratings. In each case, depressed individuals reacted similarly to positive and negative words while nondepressed individuals had greater peaks to positive words. Yet, when baselines are subtracted and ratings are not forced to match, depressed individuals’ differences in peaks appear smaller than those for nondepressed individuals. This finding comes because, both depressed and nondepressed individuals show greater dilation to negative words, but the difference is especially pronounced for nondepressed individuals. Because these findings are rather inconsistent, no single interpretation can be made for the role of depressive information processing in affecting peak amplitude.

The analysis of alternative indices of pupil dilation suggests that depressed individuals may pay progressively more attention attention to personally relevant negative words than other negative words over time, as predicted by the computational neural network, but only when they are focussed on identifying the emotional aspects of information. Nondepressed individuals often appear to pay somewhat more attention to positive than negative information.

Exploratory results investigating changes in the slope of pupil dilation during late phases of attention suggest that depressed individuals may pay progressively more attention attention to personally relevant negative words than other negative words over time, as predicted by the computational neural network, but only when they are focussed on identifying the emotional aspects of information. Similarly, as the rumination component demonstrated the greatest factor scores for personally relevant negative words on the lexical decision task, it is suggested that depressed individuals may think primarily about personally relevant negative information on that task. In contrast, nondepressed individuals pay attention to the both the affective and semantic content of stimuli very quickly after they are presented. Nondepressed individuals thus respond differently to different types of stimuli based on their valence and personal relevance. They often appear to pay somewhat more attention to positive than negative information.

These findings are consistent with the idea that depressed individuals do not process presented stimuli very much. Rather, all incoming information reminds them of personally relevant negative information, and they think about this negative information rather than the presented information, in a ruminative fashion. Nondepressed individuals actually process presented information.

The analysis of pupil dilation for this project was problematic from its inception. There were at least five commonly used indices of cognitive load. There was no precedent for how to aggregate data. For months, my advisors and I agonized over exactly which effect size, under which method of aggregation would most reasonably represent my hypotheses. It was only after becoming thoroughly overwhelmed with the combinatorial mayhem engendered by picking and choosing combinations of ways of aggregating data that I decided to try everything (198 separate contrasts of what was nominally the same phenomenon), and to see what came out consistently. This method worked surprisingly well. Very little came out under one method of aggregation and not another. The few tests that were only significant under one method of aggregation were suspect for a variety of theoretical reasons.

While the generality of this technique has yet to be established, I believe it could be very useful for researchers working in domains in which clear methodological paths have not yet been paved. In the past, statistics were difficult to calculate, so researchers had to carefully pick and choose exactly which tests were performed. The ability to generate multiple tests easily on a computer has largely alleviated this difficulty; hundreds of contrasts took under 15 minutes for a Pentium II to perform, albeit after a few hours of programming.

Concerns over Type I error and chance findings have lead to a status quo in which conservation of tests is the norm. Each test is assumed to increase the chances of spurious findings and thus each test is evaluated more strictly. Such a philosophy does not account for patterns of results obtained through multiple tests. The current technique, while strictly evaluating any single test, is more lenient with robust results, emphasizing phenomena that hold up under various methodological manipulations. Were these results trusted, I believe the chances that conclusions will hold up on replication and meta-analysis will be increased.

The discipline of computationally intensive testing, exemplified by the, bootstrap, jackknife, and randomization test methods, appears to share much of the same theoretical basis as the current process. Each of these techniques encourages the user to rely most on conclusions that hold up under repeated resampling from a given data set. No one test, with a single sample of items is trusted, but rather, different combinations of data are repeatedly selected from a single pool. Such a technique resembles the current methodology in that a single data set is repeatedly attacked from multiple angles; if any contrast is repeatedly verified, its validity is supported.