Abstract

This study was designed to further our understanding of the central role of motivational activation in mediated information processing and media choice. To do this, a dynamic model was developed to formalize the dynamic effects of three basic motivational input variables (arousing content, positivity, and negativity) on four physiological output measures (heart rate, skin conductance level, corrugator activity, and zygomatic activity) and a behavioral choice measure of television channel selection. The input and output variables were selected based on extensive theoretical and empirical research that has explicated static relationships among these variables. In general, the findings of the dynamic modeling approach were consistent with the previous literature using traditional static statistical methods. However, this study also theoretically extended the previous work. doi:10.1111/j. 1460-2466.2010.01527.x For almost half a century, it has been generally acknowledged that communication is a dynamic process occurring over time (Berlo, 1960). There has, however, been ''very little general, systematic examination of dynamic processes in the specific context of communication '' (VanLear & Watt, 1996, p. 3). This research seeks to help correct this situation by introducing and testing a set of dynamic models that provide an integrative conceptual framework of motivational processing and channel choice behavior in the context of television viewing. The dynamic models being proposed are built on psychological theories of motivational processing (e.g., Review of the theories Motivational processing Emotion plays a central role in our experience A dynamic approach to understanding motivational processing The primary theoretical importance of this study rests on its commitment to dynamic information processing models and methodology. Motivational Processing and Choice Behavior real-time psychophysiological data. It is based on explanations of information processing that are theorized to be modulated by motivational activation over time. However, its predictions are limited in the sense that only the directions of the information processing patterns, aggregated across time, can be proposed and tested. This limitation arises from the reliance on static methods, such as analysis of variance, for analyzing the data. The great potential to explicate the dynamic interactions between messages and individual responses cannot be realized unless formal (i.e., mathematical) dynamic models are employed. In other words, rich information about dynamic processes still remains hidden in real-time data due to the constraints of static analytical tools that aggregate the data over time The most general idea tested in this study is that key dependent variables (i.e., outputs) in information processing are causally influenced by motivational activation elicited by mediated emotional content (i.e., inputs). The DMA formalizes these multiple motivational inputs and outputs as a second-order linear stochastic system with lagged input effects and autoregressive residuals. Each of the outputs is influenced by three basic motivational input variables, their dynamic interactions, and feedback from its own previous responses. The feedback effect term is critical: It is responsible for the time it takes the physiological systems to be activated or deactivated with the changes in motivational inputs (e.g., as well as the cumulative effects of the inputs. This is discussed in detail in the section of Hypothesized models. To the authors' knowledge, this is the first attempt to build dynamic models explaining the causal influence of motivational inputs on physiological outputs across time. Previous research concerning time series models only involved analyses of the physiological dependent variables across time which is mostly descriptive. An integrative approach to the black box of motivational processing Emotional media content elicits motivational activation in individuals, which drives emotional experience. Although we cannot see directly into the black box of emotion and motivation, empirical research has developed a set of measures that we can use to infer what is happening in the box. P. Lang (1993) grouped measures of emotional experience into three output systems: (a) behavioral measures, including overt or functional behaviors and behavior modulations (e.g., fight, flight, and emotional modulation of task performance); (b) language measures, including expressive communication and evaluative report (e.g., verbal aggression, screams, and self-reported emotion); and (c) physiological measures, including reactions in the viscera (e.g., sweat glands, cardiovascular system, and tear ducts); somatic muscles (e.g., corrugator supercilii, zygomaticus major, and general action muscles); the respiration, endocrine, and immune system; and the brain. If motivation is a fundamental principle in emotional experience (M. This study follows this suggestion and simultaneously measures the physiological and behavioral outputs of message processing as a function of motivationally relevant media content. All the dependent variables are recorded continuously while viewing emotional television content and analyzed using time series models. Hypothesized models General hypotheses of DMA and specific models of physiological responses The primary hypothesis of this study is that the temporal variance in psychophysiological responses and channel-changing behavior during television viewing can be explained by the dynamic effects of three motivational inputs-the level of arousing content, positivity, and negativity in the media content (Hypothesis 1). In addition, this study emphasizes that the effects of the three motivational inputs will be dynamic. The effects of motivational inputs do not occur or cease instantaneously, but instead, build up and ramp down over time Motivational Processing and Choice Behavior course of physiological responses. A second-order system is proposed here for two major reasons. First, most time series analysis applications find two feedback terms to be sufficient to catch the dynamics in the systematic part of a time series model (c.f., higher order models used in modeling the errors; In addition, it is expected that there is a time delay between an onset or a change in a motivational input and the physiological responses that it elicits. This is because it takes time for the motivational inputs to reach and activate physiological systems. Therefore, it is proposed that there will be time lags between the onset of the motivational inputs and the physiological responses (Hypothesis 3). Dynamic theory-driven hypotheses are expressed formally as difference or differential equations. To systematically test the dynamic effects of the three motivational inputs on each physiological and behavior output, a second-order linear stochastic difference equation model with delayed input effects and with autoregressive error term is proposed. Basically, the proposed model is composed of two main parts: a systematic model and an error model. The systematic model models the effects of the three motivational inputs on the physiological response systems (i.e., Hypotheses 1 and 3) and the lagged feedback effects of the physiological system (i.e., the major origin of the dynamics of the system that moderates and cumulates the motivational input effects, expressed in Hypotheses 2.1 and 2.2). This is the focus of this study. To accurately estimate the parameters for the systematic model, correlations between errors should be modeled and removed; and in this study, an autoregressive model is used for this purpose. The formal equation of the heart rate (HR) model is given below to serve as an example for the four physiological models. Description and rationales for the model are explained immediately following the equation under three subtitles-the lagged feedback effects of the physiological systems, the delayed effects of motivational inputs, and the error. The lagged feedback effects of the physiological systems The term H(t) on the left-hand side of the equation is what we are trying to predict, HR at time point t. The first two terms on the right-hand side of the equation, a 1h H(t − 1) and a 2h H(t − 1), are the first-and second-order lagged feedback terms: H(t − 1) and H(t − 2) are HR at time points t − 1 and t − 2, and their coefficients are a 1h and a 2h . These are critical for explaining the time course of the system's response or output to a motivational input. In a dynamic system, the output of the system depends on several factors: first, the size and duration of the input effect, as the reader might expect; second, the lagged feedback effect/coefficient. The same input can produce dramatically different outputs simply by changing the lagged feedback coefficient of the system which determines the speed, strength, and duration of the output The delayed effects of motivational inputs The remaining terms in the HR model with coefficients symbolized by bs represent the causal effects of the motivational inputs on HR. This looks very much like a regression equation on its face, but the effects generated by this model do not behave like a regression model. As explained above, this is because the total effects generated by this model include the dynamic response to input effects caused by the feedback factors, a 1h H(t − 1) and a 2h H(t − 2). Based on a review of the literature, the proposed input effects include all the possibly interesting theoretical effects, including (a) main linear effects of the three motivational inputs (e.g. Motivational Processing and Choice Behavior Note that the input effects are delayed by a common amount d h . It represents the hypothesized time lag for the physiological systems to respond to the motivational inputs from the media content, as proposed in Hypothesis 3. This time delay will be estimated from the data. Lastly, in the HR model, there is also an intercept parameter, b 0h . Generally, for dynamic models, the intercept is meaningful as it can change the dynamic output curve. In this case, however, detrending the data eliminated the intercept effect in most of the data series. This is appropriate because this study focuses on explaining the causal effects of the motivational inputs on the physiological systems and not on modeling simple changes in the physiological systems over time (such as habituation or adaptation). The error Finally, the model includes an error term, e h . In time series modeling, the error in data also requires a dynamic model because there are statistical dependencies in the error across time. If the autocorrelated error is not removed from the total variance, the systematic model cannot be estimated correctly. Here, an autoregressive error model, AR(p), was used to model the autocorrelated residuals, where p is the highest order of the autoregressive terms. It was chosen for its flexibility and mathematical simplicity. Two-model comparison criteria were used to search for the best AR(p) error model: the Bayesian information criterion and the statistical significance of parameters for each lagged error term in competing models. An AR(5) model was found to be sufficient to capture most of the dependence in the over-time residuals-that is, the error model includes lags 1-5. The equations for the other three physiological variables follow the same form but the value of the parameters are estimated separately for each variable. This allows the models to accommodate the undoubtedly different feedback effects and time courses of the different physiological systems. Hypothesized dynamic model of channel choice behavior A somewhat different model is proposed for channel-changing behavior. Channel choice is viewed as a function of interest (I) in the channel which is theorized to be a function of the motivational inputs of that channel. The equation below shows the hypothesized model for interest as a function of the motivational inputs, with A(t) representing the arousing content input at time t, P(t), the positivity input at time t, and N(t), the negativity input at time t. Like the physiological response models, this model includes estimates of the linear and quadratic components of the main effects of the motivational inputs as well as their linear interactions. Also, it includes the time delay d. The rationales for including these terms are similar to those proposed for the physiological models. In addition, considering the explorative nature of this modeling effort and the focus on motivational input effects, we included these terms so that we could compare findings across all the dependent variables. Note that a decision was made to keep this model simpler by not including the feedback terms included in the physiological models. Although previous viewing experience influences current viewing experience (e.g., , the binary choice (to switch vs. not to switch) in this experiment was expected to be primarily influenced by the content being viewed at a given moment. The next step is to model how interest influences channel choice. The probabilities of the choices (e.g., stay on a channel or switch) are bounded by zero and one, and the probability of choosing an option (e.g., stay) has been empirically found to be an increasing S-shaped function of the strength of the option (i.e., interest in the channel) (e.g., Pr(S(t) = 1) = G(I(t − k)), where G is an increasing function bounded by zero and one and k is the time lag. For each participant, based on his or her own data, the best fitting time lag was selected from 11 competing models (using lags from 0 to 10). The probability function G, which is also called the link function in the generalized linear model, is a logistic function It is chosen here because it is the most commonly used and empirically supported model for choice data (e.g., Motivational Processing and Choice Behavior factorial design were selected for the pretest. In total, 125 undergraduate students (46.4% male, 76.8% White) with an average age of 20.44 (SD = 0.12) viewed and rated the clips using the continuous response measurement (CRM; Each participant viewed and rated 15 randomly assigned and randomly ordered clips. Across participants, all 30 clips were rated continuously on three scales: (a) how aroused do you feel? (the arousing content scale); (b) how positive do you feel? (the positivity scale); and (c) how negative do you feel? (the negativity scale). Each participant watched a given clip once and rated that clip on only one of the three CRM scales. The clips and scales were randomly assigned to participants in such a way that each participant watched two to three clips in each emotional category, rated five clips on each of the three CRM scales, and each clip was rated on all the three CRM scales by a similar number of people. Based on the average summative ratings for the clips, the final 24 messages were selected as follows. The 12 clips that were rated highest on the positivity scale (Ms > 5) and stayed below 3 on the negativity scale (Ms < 3) were selected as positive clips; the 12 that were rated highest on negativity (Ms > 5) but stayed below 3 on positivity (Ms < 3) were defined as negative clips. Then, within valence categories, the 12 clips were ranked on arousing content and divided into three levels (arousing, moderately arousing, and calm), with four in each level. A manipulation check confirmed that arousing content levels, positivity, and negativity were manipulated successfully (ps < .001 and Ms in the expected direction). After the final 24 clips were selected, CRM data series were processed for each clip to serve as dynamic motivational inputs for the DMA models. The median of CRM ratings at each time point was selected as motivational input for that time point. Finally, to test the reliability of CRM ratings, one CRM data point was randomly selected from every 25 seconds of each clip, generating 12 rating points per clip. Based on the 12 rating items, Cronbach's α was computed for each rating on each clip among the participants. The Cronbach's α indicated that the CRM ratings were reliable (M α = .94, SD α = .03). Experiment design In the main experiment, participants watched television for 30 minutes. The television had four available channels and participants were instructed to watch whatever they would like to on the channels. They were informed that they could change channels at will using a remote control and also practiced how to use the remote. In total, 6 of the 24 selected stimuli from the pretest were assigned to each channel, and they were edited together to form a coherent viewing session of 30 minutes (5 minutes × 6 clips). There were three different viewing orders. Within each viewing order, there were four different orders of presentation of the six clips on each of the four channels. The within and between channel orders were designed to counterbalance the position of clips with different valence and arousing content. Dependent variables Zygomatic and corrugator electromyography These are conceptualized as indices of positive and negative emotional responses which result from viewers' appetitive and aversive activations respectively Skin conductance level This is a measure of sympathetic nervous system activation that is theorized to be related to motivational activation (M. Bradley & P. Lang, 2000). Higher skin conductance level is attributed to increased activation in the sympathetic nervous system which indexes higher physiological arousal and suggests more intense motivational activation. The data were collected from the palm of the nondominant hand sampled at 20 Hz. HR During resource allocation to external stimuli such as television messages, activation of the parasympathetic nervous system increases, resulting in measurable decreases in HR Channel changing Channel changing is viewed as choice behavior which is a function of interest in the motivational content of television programming. VPM recorded both the time at which a channel change was made and the channel options (corresponding to certain motivational content). Then, a MATLAB program was created to generate the 1800-second long time series with one data point per second for each channel, dummy-coding 1 for a time point if the channel was watched at that time and 0 if not. Procedures and participants Participants completed the 1.5-hour experiment individually. After the experimenter demonstrated how to use the television remote control and attached electrodes to the participant, the experimenter left the room and closed the door to provide privacy Motivational Processing and Choice Behavior during viewing. Then, the four 30-minute stimulus tapes were simultaneously played on a wired set of four videocassette recorders (VCRs) outside the experiment room. The participant watched the content on a 25-inch television that was connected to the VCRs. While viewing, the participant could use the remote control to change channels at any time. Participants' physiological responses and channel-changing behaviors were collected continuously during viewing. In total, data were obtained from 67 participants. The average age was 21.13 (SD = 1.28, range 18-25); 41 (61.2%) were males; and the majority were White (80.3%), followed by Asian (8.5%), African American (4.2%), and Hispanic (2.8%). Modeling analysis and results Time series data sets For the dependent variables, time series were created for each participant using four physiological measures and channel choice behavior obtained at a rate of one observation per second for 1800 seconds. The physiological data were initially processed by removing linear trends of time using the general linear model procedure in SAS (PROC GLM) for each variable and each person. This detrending process is needed because when using time (1800 time points) as a predictor, linear regression tests found that time had a significant effect on each of the physiological variables (ps < .001); but this general, nonstimulus-specific trend in the physiological responses is not the focus of this study. After detrending, to put the physiological variables onto the same scale for easier interpretation of the model parameters, the data were transformed to standardized scores for each variable for each person. To create the independent variables, three 1800-second time series were created using the medians for arousing content, positivity, and negativity CRM data obtained in the pretest. Next, because participants were in control of their channel-changing behavior and therefore each participant viewed different video content at different times, three time series were created for each participant based on their personal channel viewing. A MATLAB program was created to align the CRM ratings in time (i.e., second by second) with the video content actually watched by each participant. For each participant, this data-matching procedure produced a series of 1800 seconds of ratings for the three motivational inputs over time based on the person's actual viewing experience. Thus, for each participant, we obtained a data matrix consisting of eight columns of variables (three inputs and five outputs) and 1800 rows of observations across time. This data set was used to estimate the parameters of the proposed DMA models (for a visual representation of the raw data of a single participant's actual inputs and outputs, http://wongzheng.web.officelive.com/dynamics.aspx ). Model fitting and model performance For each participant, the proposed four physiological models were estimated using maximum likelihood methods and PROC AUTOREG in the SAS software. To search for the best delay lags for the motivational input effects, for each model and for Note: HR = heart rate; EMG = electromyography. each person, 11 lagged models (using lags from 0 to 10) were estimated. Based on the regression R 2 predicted by the systematic model, the best lag model was selected for each physiological model for each individual. The average regression R 2 across all participants' data sets is relatively large. The descriptive statistics are shown in For the channel choice model, first, a chi-squared test was performed on the hypothesized full model (with 10 parameters) compared with a restricted null model (without parameters associated with the motivational input effects). The Bayesian information criterion was used to evaluate the complete model compared with the null model because its evaluation on models is based on both goodness of fit and model complexity Effects of motivational inputs After model fitting for each model for each viewer, we obtained a set of 12 systematic model parameters for each physiological variable from the best fit lag model and 10 for the channel choice model. First, multivariate analysis of variance (MANOVA) was used to test the significance of the motivational effects on the physiological responses and channel choice (i.e., parameters for A, P, N, A × P, A × N, P × N, A 2 , P 2 , and N 2 ). Each of the motivational input effects, as estimated by the model parameters, was tested. For example, to test the linear effect of arousing content (i.e., the parameter for A), each individual's parameters for A from all five of the models were entered into the MANOVA test simultaneously. Hotelling's T 2 was significant for the main effect of arousing content (both the linear and quadratic components), its interactions with positivity and with Motivational Processing and Choice Behavior negativity, and the quadratic component of the positivity main effect. For parameters with significant Hotelling's T 2 , Student's t tests were performed on each parameter for each model to determine: (a) whether the mean (across participants) of each parameter differs significantly from zero, and (b) the sign and size of each parameter. If the mean of a model parameter is significantly different from zero, this suggests that the effect associated with the model parameter is significant. The sign and size of the parameter tell us about the size and direction of the effect estimated by each parameter. The motivational effects on physiological variables are summarized in Lagged responses of physiological systems Next, we examined the feedback parameters that determine how quickly the motivational inputs activate or deactivate the physiological systems. Specifically, the first step is to determine how the physiological response feeds back on and influences itself. t tests were conducted on the two lagged feedback parameters in each model to determine the significance, direction, and size of the feedback effects. The results are reported in Delayed motivational input effects Next, to determine how quickly the motivational inputs reach and produce an effect on the physiological systems, the number of lags in the best-fit lagged model for each participant and each model was examined. Recall that the model for each physiological response uses the motivational inputs that occurred at a time d seconds earlier to predict the current physiological response (Hypothesis 3). The delay d is the lag parameter. Also recall that during our parameter estimation for each physiological model, we compared each model with 11 different lags (0-10 lags/seconds) and selected the best fit for each participant. The results support Hypothesis 3. Indeed, there were time lags between the onset of the motivational inputs and the physiological responses. The mean for the only unimodal distribution, HR lags, is 5.37 (SD = 2.99). For skin conductance level, the lags are bimodally distributed (M = 5.25, SD = 3.15), with the dominant mode being 7 (n = 10) and another peak at 2 (n = 8). For corrugator and zygomatic activities, the lags are also bimodal (M = 4.49, SD = 3.31 and M = 4.94, SD = 3.34, respectively), with modes at 0 (n = 10) and 5 (n = 9) for corrugator activity and 4 (n = 8) and 10 (n = 8) for zygomatic activity. These distribution patterns suggest that participants showed