Theory-Based Behavioral Indicators for Children’s Purchasing Self-Control in a Computer-Based Simulated Supermarket

Self-Controlled Behavior in the Purchasing Process

Market demands with their multitude of stimuli and information can be particularly overwhelming for children, whose cognitive abilities and skills are not yet fully developed and who lack market experience and knowledge (Mau et al., 2014). To better understand children’s purchase decisions, associated behaviors, and deficits, Mau et al. (2016) analyzed children’s behaviors in a simulated supermarket environment. They showed that children often behave differently at the point of sale than they intended and expected to when making their purchase decision. Slightly more than half of the children indicated that they would primarily look for low prices when shopping. Although the children in the subsequent observation of their shopping behavior had a limited budget and were tasked with buying the cheapest products, it was found that they clearly tended to select products more based on package design or brand. These findings indicate that children have difficulty implementing a basic requirement of goal-oriented consumer behavior, namely, taking the right actions to achieve a set goal (Bagozzi & Dholakia, 1999).

Regarding the question of how actions in purchasing processes could be implemented in a goal-oriented manner, we draw on basic theoretical frameworks of action regulation that include feedback loops, such as the cybernetic TOTE model (Powers, 1973). Carver and Scheier (1981) assume that the successful realization of a goal state (e.g., fulfillment of the shopping list) occurs by passing through loops in which the actual and target states (e.g., contents of the shopping cart vs. the shopping list) are repeatedly compared with each other and a deviation is successively reduced by operations until the loop is exited. Consequently, the execution of the action is expressed as a sequence of corresponding operations and is always in the interplay between the goals of the agent and the situational requirements. According to Gillebaart (2018), setting standards or goals as well as monitoring deviations are aspects of self-regulation, while successful self-control comes into play within the feedback loop in the ‘operate’ phase. Self-control has been described by Baumeister et al. (2007) as the mental processes that enable people to control their thoughts, emotions, and behaviors to achieve higher-level goals. While operating, various aspects of self-control can be observed, such as suppressing the impulse to be tempted by alluring stimuli that are not in line with the defined long-term goals (e.g., completing the shopping list), avoiding situations that might lead one into temptation (e.g., forgo the candy shelf), or even delaying gratification with an immediate, smaller reward in order to obtain a larger, delayed reward. According to Inzlicht et al. (2014), as a result of repeated self-control efforts, there may be a change in the degree of self-control displayed. This is attributed to a change in task priorities, a shift in motivation away from so-called “have-to” to “want-to” goals that provide more pleasure and satisfaction. Therefore, the process of action regulation is always under the influence of changing motivations and the attendant changes in emotions and attention.

The Present Study

Although self-control has been highlighted in its importance for the successful implementation of consumer goals (e.g., by avoiding impulsive purchases; Baumeister, 2002), there is still no study that specifically captures children’s operations in the purchasing process and relates the extent of self-controlled behavior to the successful implementation of a purchase intention. To fill this gap, we used a computer-based supermarket simulation to assess children’s purchasing literacy based on their shopping behavior at the point of sale. In this task, children were asked to complete a shopping trip based on a shopping list, which is not an uncommon task in children’s daily life (Mau et al., 2016). The simulation, which recreates an essential and routine form of consumer behavior, imitates key elements of grocery shopping while simplifying or missing non-essential aspects of real-world shopping situations. Most importantly, it is designed so that at the behavioral level, children’s attention behavior can be derived from the log data (Silberer, 2009). Attentional behavior comprises observable attention to objects in the store environment, that is, how often or how long the children look at individual products. The supermarket simulation was purposefully designed to include elements that are not required for the performance task. These include irrelevant shelves and products, as well as products that in themselves (e.g., toys) or in their presentation (e.g., packaging design) are intended to be particularly appealing to children. The extent to which children engage with these irrelevant elements can be seen in their attention behavior and, at the behavioral level, enables differentiation between actions that are more conducive to have-to goals (according to the task) or want-to goals in the sense of Inzlicht et al. (2014). Possible behavioral indicators for the pursuit of want-to-goals and thus (low) self-control are the number of irrelevant products purchased or the time spent with irrelevant shelves throughout the main task.

The aim of the present study is to investigate self-control as a subdomain of purchasing literacy that relates to manifestations in concrete actions. Therefore, the study investigates whether children’s self-control can be measured using behavioral indicators from the simulation and how this hypothesized construct relates to task performance and other variables in the study. In order to capture the lack of self-control not only via the situational behavioral data from the supermarket simulation, children’s self-reported impulsivity is also captured as a cross-situational facet of self-control, that is, the predisposition to react rapidly and unplanned to internal or external stimuli without regard to the negative consequences (Moeller et al., 2001). Such a cross-situational disposition could affect the situational behavior in the supermarket simulation and thus also have consequences for task performance. Specifically, the following hypotheses will be tested in the course of the study:

In the final step of our study, we explore how task performance, impulsivity, and the generated factor of self-control are related to demographic and socioeconomic background, distrust of advertising, prior shopping experience, brand imprinting, financial monitoring, and children’s ratings of the extent of task fulfillment.

Supermarket Simulation

The simulation involved a bird’s eye view of a two-dimensional supermarket environment on a tablet (Figure 1). In a training run, the children were able to try out the navigation options through the supermarket and its functions in a comparable environment. Based on a shopping list, the children were asked to buy eight different products for their parents, of which they were instructed to buy only the cheapest option with their budget of €20.00. To complete the task, the children were free to move around the supermarket and check the shopping list as needed. They could complete the task at any time by proceeding to the checkout.

Process Data

Students’ interactions during testing were recorded in log files, consisting of time-stamped sequences of events and enabling the retrospective tracing of the solution path, acting as a basis for the scoring procedure and the creation of behavioral indicators. First, we defined the events in our log files, which defined the start of task processing (the first relevant product is placed in the shopping cart) and the two possible end situations (the last relevant product is placed in the shopping cart or the purchase is completed without full task completion). Due to the particular chronological structure of the log data, the subsequent extraction of behavioral indicators required some preparation of the data. In log files, events can be named the same, but take on different meanings due to their position in the chronological sequence of events. A tool that facilitates the creation of process indicators from log data, taking into account their chronological sequence, is the R-package LogFSM (Kroehne, 2021). The method uses the concept of finite-state machines (FSM), in which the entire process of task processing is decomposed into a finite number of states (Figure 2) and events are defined that cause the transition from one state to the other (Kroehne & Goldhammer, 2018). Subsequently, it is possible to determine how long and how often certain states or the transitions between them have occurred. Using this method, we extracted four indicators, three of which characterize behavior during task processing (S1, S2, S3) and one of which presents one outcome of the entire simulation interaction (S4, see Table 1).

Table 1 Behavioral indicators extracted from the log data

Name	Indicator	Description	Hypothesized factor	Hypothesized direction of factor loading	Scale	Mean (SD) or frequency
Note. Time is reported in seconds.
S1	time_irr_shelf	Time spent on irrelevant shelves.	Self-control	Negative	Metric	4.95 (10.41)
S2	time_irr_prod	Time spent with irrelevant products.		Negative	Metric	0.73 (1.99)
S3	freq_irr_shelf	Frequency with which irrelevant shelves were visited.		Negative	Recoded as dichotomous	76 vs. 39 (at least one irrelevant shelf visited)
S4	freq_irr_prod	Frequency with which products not on the shopping list were purchased.		Negative	Recoded as dichotomous	68 vs. 47 (at least one irrelevant product purchased)

Table 1 Behavioral indicators extracted from the log data

View as image HTML

Task Performance

Students’ task performance was derived from the computer-generated log data. Their performance was evaluated with a partial credit solution, in which partial steps relevant to the solution were extracted from the process data. The products on the shopping list were treated as ordinally scored test items, with one point awarded in each case for finding the correct shelf, the correct product category, the cheapest product, and (for two test items) the best price-quantity ratio. We fitted a partial credit model with the R package TAM (Robitzsch et al., 2021) and obtained Weighted Likelihood Estimates (WLEs) as an indicator of students’ performance (WLE reliability = .72).

Questionnaires

In addition to questions on the demographic background (age, school year), the number of books in the household was used as an indicator to capture educational and socioeconomic background (e.g., in TIMSS 2015; Wendt et al., 2017). Data collection also included questionnaires on children’s impulsivity (Wendt et al., 2017, e.g., “I often do and say something without having thought about it.”) and distrust of advertising (self-developed, e.g., “Advertising is meant to persuade me to buy things.”). In addition, the children were asked whether they had ever made a similar purchase in a real supermarket (Q1), how sure they were that they had bought all the products on the shopping list (Q2), how much they had bought the brands they knew from home (Q3), and how much they had been careful to spend little money (Q4). Agreement or disagreement with the statements was recorded using an illustrated 5-point Likert scale. After shopping, the children were asked to estimate how much they had spent. The inaccuracy of these estimates was calculated using the difference from the actual remaining amounts.

Sample and Procedure

The children were recruited from two different elementary schools in Germany and participated voluntarily. The sample consisted of N = 136 children in their third (n = 73) or fourth (n = 63) year of elementary school. This sample size was planned to obtain robust estimates of simple item statistics, for example, difficulty. Originally, a larger sample size of more than 1,500 students was planned to allow the analysis of rare behaviors and accurate estimates of extreme item difficulties. The schools for this larger sample had already been recruited, but data collection was canceled due to the COVID-19 pandemic. Testing took place on the schools’ premises. For the questionnaires, the children were assisted by an interviewer who read out the questions and entered the answers into a computerized scoring environment on a tablet. The supermarket simulation, on the other hand, was completed by the children themselves.

For one participant, the test had to be terminated prematurely due to external circumstances, and once the test was terminated due to a device error. Because this was the first time this version of the supermarket simulation was used, no specific exclusion criteria were established prior to data collection. Screening of the behavioral data and univariate and multivariate outliers identified 16 children who had not looked at the shopping list and therefore could only solve the task by chance. Two other children purchased so many products (40, 58) that the budget was overdrawn (4 and 7 times the budget) and relevant products may have ended up in the shopping basket purely by chance. These exclusions reduced the final sample to N = 115 children (59% female) aged 7–12 years (M = 9.09, SD = 0.82), who were included in the subsequent analyses.

Data Preparation

A special challenge was posed by the count variables (frequencies), which show an excess of zero counts. Since the behaviors we consider to be indicative of lower self-control, such as buying irrelevant products, occur independently of the shopping task solution process, they are not exhibited by all children. For the two behavioral indicators showing this high proportion of zero (S3, S4), categories were formed, each represented by at least 20% of the children. This resulted in dichotomous coding for both variables. Because of the skewness of the metric variables S1 and S2 (up to 2.83), we normalized the distributions by using the logarithms of the indicators. Since the variables contained zero values, the variables were increased by the value +1 to allow the logarithmic transformation. Thereafter, all metric variables were centered around their respective grand mean.

Statistical Analyses

To statistically test the hypothesized relationships, a structural equation model (SEM) was specified. This consists of two parts: In the measurement model, the four extracted behavioral indicators (S1–S4) from the supermarket simulation represent the exogenous latent variable “self-control”. The path model adds another exogenous “impulsivity” as well as an endogenous manifest variable “task performance” and describes the hypothesized mediated relationship (Figure 3). It should be noted that the relationships between the variables of the path model are represented as single-headed arrows, which indicate regression effects and thus causal effects from one variable to another. Conceptually, we assume such causal relationships between the variables, even though their causal effects cannot be tested using our study design. Within the measurement model, two pairs of variables (1) S1 and S3 and (2) S2 and S4 are associated with the same behavior. Due to the small number of degrees of freedom, a residual correlation could only be included in the measurement model for the first of these pairs.

Because of the dichotomous variables, diagonally weighted least squares (WLSMV) were used to estimate the model parameters of the self-control factor. Model fit was evaluated using χ² goodness of fit test and standard model fit indices such as the root mean square error of the Error of Approximation (RMSEA), the comparative fit index (CFI), the Tucker Lewis index (TLI), and as a residual-based fit index the standardized root mean square residual (SRMR). The interpretation of these model fit indices and the use of cut-off criteria are based on Schreiber (2017), who classifies a ratio between χ² and degrees of freedom of < 3, TLI ≥ .95, CFI ≥ .95, SRMR ≤ .08, and RMSEA < .05 as an acceptable fit. We consider the CFI and the TLI to be particularly central to our evaluation since they have a low sample size sensitivity compared to the other criteria (Schreiber, 2017). Factor scores were estimated from the measurement model to examine the correlations of the hypothesized factors with the test score and other student variables. Analyses were conducted using Mplus 8.4 (Muthén & Muthén, 2017).

Measurement Model

The measurement model for self-control (SC) included the variables time_irr_shelf, time_irr_prod, freq_irr_prod, and freq_irr_shelf (Table 1). The loading for time_irr_shelf was fixed to −1 to define the factor as SC and not a lack thereof. For time_irr_shelf and freq_irr_shelf residual correlation was included in the model as both variables refer to the same behavior. The model showed a largely good fit (χ²(1) = 2.276, p = .131, RMSEA = 0.105, 90% RMSEA CI [.000, .294], CFI = 0.993, TLI = 0.956, SRMR = .04), only the RMSEA exceeded the cut-off criterion. The measurement model and factor loadings are shown graphically in Figure 3. McDonald’s Omega (ω = 0.95) was calculated as a reliability coefficient for this model.

Path Model

The model fit indices for the full SEM (χ²(7) = 15.850, p = .027, RMSEA = 0.105, 90% RMSEA CI [.034, .174], CFI = 0.959, TLI = 0.913, SRMR = .104) indicate a poor-fitting model. Only the direct path (p < .001) of SC on task performance was significant (see Figure 3 for standardized path coefficients). Neither the direct (p = .727) nor the indirect (p = .862) path of impulsivity on task performance was significant.

Correlations

The results of the Pearson correlations are listed with their confidence intervals and significance level in Table 2.

Table 2 Means, standard deviations, and correlations with p-values and confidence intervals

Variable	M	SD	1. Self-control	2. Performance (WLEs)	3. Impulsivity
Note. WLE = Weighted Likelihood Estimates; M = mean; SD = standard deviation. Values in square brackets indicate the 95% confidence interval for each correlation. Values printed in bold are significant correlations with p-values < .05.
1. Self-control	−0.01	0.72
2. Performance (WLEs)	−0.05	1.19	.44 (p < .001)
			[.28, .58]
3. Impulsivity	11.20	4.18	.09 (p = .353)	.01 (p = .943)
			[−.10, .27]	[−.18, .19]
4. Age	9.12	0.81	−.11 (p = .262)	−.07 (p = .426)	.01 (p = .962)
			[−.28, .08]	[−.25, .11]	[−.16, .17]
5. School year	1.46	0.50	.13 (p = .165)	.22 (p = .019)	.00 (p = .913)
			[−.05, .31]	[.04, .38]	[−.17, .17]
6. Number of books	3.49	1.18	.02 (p = .841)	.06 (p = .525)	−.14 (p = .116)
			[−.16, .20]	[−.12, .24]	[−.30, .03]
7. Distrust of advertising	12.24	3.86	.22 (p = .018)	.20 (p = .028)	.29 (p < .001)
			[.04, .39]	[.02, .37]	[.14, .44]
8. Experience_Q1	1.68	0.74	.07 (p = .491)	.01 (p = .911)	−.04 (p = .532)
			[−.12, .25]	[−.17, .19]	[−.21, .13]
9. Sureness_Q2	4.57	0.98	.40 (p < .001)	.47 (p < .001)	−.19 (p = .061)
			[.21, .53]	[.31, .60]	[−.35, −.02]
10. Brands_Q3	3.42	1.32	−.31 (p < .001)	−.27 (p = .004)	−.03 (p = .711)
			[−.49, −.16]	[−.44, −.09]	[−.20, .14]
11. Spending_Q4	4.48	1.03	.03 (p = .444)	.23 (p = .015)	−.18 (p = .037)
			[−.16, .21]	[.04, .39]	[−.34, −.01]
12. Inaccuracy of estimates	9.74	25.99	−.35 (p < .001)	−.25 (p = .007)	.06 (p = .548)
			[−.43, −.08]	[−.42, −.07]	[−.13, .24]

Table 2 Means, standard deviations, and correlations with p-values and confidence intervals

View as image HTML

Limitations and Future Research

Two limitations should be considered when interpreting the results of this study. First, our sample is so small, N = 115, that differences in the population may not be representatively covered. Second, it remains to be investigated to what extent the results from the supermarket simulation can be generalized to real decision-making situations and SC at the point of sale.

In addition to the representativeness and generalizability of our results, there is also the question of the extent to which evidence and theory support the interpretation of the factor as a measure of SC. Our development of arguments supporting the interpretation of the extracted behavioral indicators and their aggregation into a common factor was based on a theory that defines the construct of SC and could explain behavioral differences captured by the process data. Future research could focus on empirical validation of the construct interpretation. By examining the relationship between the extracted factor of SC and standardized measures purported to measure a similar construct, convergent evidence could further support the interpretation of the factor.

Continuous assessments allow us to characterize the work process in the form of appropriate indicators and to collect diagnostic evidence beyond the achievement of a correct solution. The present study has used this potential on the one hand to enable a more sophisticated assessment of the extent of competence and on the other hand to map another aspect of consumer behavior with a factor of SC. The untapped potential of the process data lies in the possibility to analyze changes in different behavioral indicators during the course of task processing. This implies for future research that process data could be used to model the relationship between SC and task performance over time, including different stages of task processing. In this context, it might also be possible to address core questions about SC, as different theories exist about how individual changes in SC might be accounted for. Very influential, also in the field of consumer behavior research (see, e.g., Hofmann et al., 2008), is the belief that self-control is based on limited resources that are consumed in the course of repeated acts of restraint (i.e. ego depletion). However, there are more recent non-resource-based considerations of SC that attribute declines in one’s SC more to motivational changes in task processing (Inzlicht et al., 2014).

In the spectrum between conventional formats in competence assessment with distinct, predefined test items and observation of performance in the real world, we opted for a simulation-based assessment in which we can capture facets of purchasing literacy through situational representation in simulated situations and examine them in relation to each other. Although it is important to gather further arguments for the interpretations of performance and the factor of SC, they are probably more valid in terms of a person’s actual purchasing literacy than they could have been on the basis of individual items, and they exhibit greater operationalization, standardization, and objectivity due to the automatic recording of behavioral data in the log files as opposed to behavioral observations in a real supermarket.

With the successful theory-driven extraction of construct-relevant facets based on log data from a computer-based task, our study not only extended our understanding of the construct purchasing literacy but also demonstrated the diagnostic potential inherent in such data. This potential can be used in the future to improve the quality and richness of psychological assessments.