Exploring Effects of Age on Conflict Processing in the Light of Practice in a Large-Scale Dataset

ABSTRACT

Introduction

The possible decline of cognitive functions with age has been in the focus of cognitive research in the last decades. The present study investigated effects of aging on conflict processing in a big dataset of a Stroop-inspired online training task.

Methods

We focused on the temporal dynamics of conflict processing in the light of task practice by means of inspecting delta plots and Lorenz-interference curves to gain insights on a process level.

Results

The results indicate a relatively constant increase of cognitive conflict over the course of adulthood and a decrease with practice. Furthermore, the latency of the automatic processing of conflicting information relative to the controlled processing of task-relevant information decreases relatively constantly with age. This effect is moderated by practice, that is, the relative latency of the automatic processing decreases less with age at high practice levels.

Conclusion

As such, practice seems to be able to partially counteract age-related differences in conflict processing, on a process level.

Increasing life spans due to medical progress in combination with reduced birth rates have lead to an age shift in western societies over the last decades. For example, the percentage of persons aged 65 or older in the US was 8.14 in 1950 (U.S. Bureau of Census, 1952) and 16.40 in 2019 (U.S. Bureau of Census, 2019). Consequently, different aspects of human development over later stages of the life span have moved more and more into focus. Specifically, research efforts on the development of cognitive functioning over adulthood were intensified, for example with respect to cognitive control (e.g., Andrés, Guerrini, Phillips, & Perfect, 2008; Erb, Touron, & Marcovitch, 2020; Lampit, Hallock, Valenzuela, & Gandy, 2014; Zanto & Gazzaley, 2017).

Cognitive control is the basic ability to focus on and functionally act upon relevant information in our environment while inhibiting the influence of distracting irrelevant information (e.g., Cohen, 2017). Cognitive control is, therefore, crucial for managing our daily life in situations ranging from conversations, over complex tasks at the work place to maneuvering safely through traffic. This wide range makes apparent the importance of understanding how cognitive control develops with older age in aging societies.

Cognitive control is commonly experimentally investigated using so-called conflict tasks, such as the Stroop (Stroop, 1935), Eriksen flanker (Eriksen & Eriksen, 1974) or Simon task (Simon & Small, 1969). These tasks have in common that participants are instructed to make a speeded decision to a specific aspect of a presented stimulus with a binary response choice, for example by pressing one of two keys, while ignoring a second salient yet task-irrelevant aspect of the stimulus. For instance, in the Stroop task, the stimulus is a color word displayed in an independent color. Participants are instructed to respond based on the display color of the word while ignoring the word’s meaning. Thus, there are two types of trials: Trials in which the word’s meaning and its display color correspond are called congruent trials and trials in which the meaning and the display color of the word differ, and thus interfere, are called incongruent trials. Although the word’s meaning is irrelevant to the task, it is automatically processed and thus influences the task performance. It can both facilitate responding in congruent trials and interfere with responding in incongruent trials (e.g., Parris, Hasshim, Wadsley, Augustinova, & Ferrand, 2022). The influence of the irrelevant information in the Stroop task and other conflict tasks is measured as the congruency effect, that is, the difference in response times (RTs) and accuracy of responding between congruent and incongruent trials. The size of this difference depends, among other factors, on the ability of the individual to selectively attend relevant information while suppressing the automatic activation of irrelevant information, that is, their level of cognitive control. Therefore, conflict tasks, such as the Stroop task, can be employed to investigate interindividual differences in cognitive control and, consequently, also its development over the life span.

One of the most prominent findings in research on cognitive aging is a slowing of RTs in decision making tasks (see Salthouse, 1996). With respect to cognitive control, there is some evidence indicating that there is indeed an increase in the congruency effect with age, for example in the Stroop task (e.g., Andrés, Guerrini, Phillips, & Perfect, 2008; Cohn, Dustman, & Bradford, 1984; Comalli, Wapner, & Werner, 1962; Li & Bosman, 1996; Panek, Rush, & Slade, 1984; Spieler, Balota, & Faust, 1996; West & Alain, 2000; West & Baylis, 1998; Wolf et al., 2014; see also Servant & Evans, 2020). However, it is necessary to differentiate an increase of the RT difference between conditions due to reduced cognitive control from an increase that is proportional to the general increase of RT with age. A meta-analysis employing Brinley and state-trace analyses concluded that the increase of the congruency effect is proportional to the general RT increase in the Stroop task, indicating general cognitive slowing instead of reduced cognitive control with age (Rey-Mermet & Gade, 2018; this is in line with findings of an earlier meta-analysis by Verhaeghen & De Meersman, 1998). However, some studies specifically reported age effects on the congruency effect that were overproportional to the general slowing of RT (e.g., Aschenbrenner & Balota, 2015; Bugg, DeLosh, Davalos, & Davis, 2007; Jackson & Balota, 2013; Li & Bosman, 1996; Nicosia & Balota, 2020; Spieler, Balota, & Faust, 1996). Other studies showed that only specific conflict processes are affected by aging (e.g., Augustinova, Clarys, Spatola, & Ferrand, 2018; Burca, Chausse, Ferrand, Parris, & Augustinova, 2022) and that a single factor of general slowing cannot fully account for age differences in conflict tasks (Allen et al., 2001). Thus, in some contexts aging seems to affect cognitive control beyond a general slowing of RT.

Servant and Evans (2020) further investigated the cognitive processes underlying these aging effects by modeling data from a flanker task assessed in young and older adults. The model that fitted their data best was the diffusion model for conflict tasks (DMC; Ulrich, Schröter, Leuthold, & Birngruber, 2015), which is an extension of the general decision diffusion model (DDM; Ratcliff, 1978). In the DDM, decisions are conceived as the result of a noisy information accumulation process. The DMC extends the DDM by a second process, to accommodate the parallel processing of the irrelevant in addition to the relevant information in conflict tasks. The DMC parameters estimated from the age data suggest a slowing in non-decision time, that is, the sensoric and motor execution component of RT, and higher response thresholds, that is, generally more careful responding, in older adults (Servant & Evans, 2020). This is in line with the results of studies investigating response processes applying the DDM in different decision tasks (e.g., Ratcliff & McKoon, 2015). They, too, conclude that slower RTs in older adults are mainly due to slower sensorimotor processing and more cautious decision making rather than slowing of information accumulation. Consequently, despite apparent age effects on the congruency effect, process-focused investigations are crucial to disentangle changes in cognitive control over adulthood.

One approach to a deeper understanding of the underlying mechanisms of cognitive control is by investigating the temporal dynamics of the conflict. This can be achieved by inspecting the distributions of RTs in congruent and incongruent trials. The difference between these distributions is often captured in so-called delta plots (De Jong, Liang, & Lauber, 1994). They depict, for certain quantiles of each distribution, the difference between conditions (congruent vs. incongruent) as a function of the mean of both conditions. Therefore, they depict how big the congruency effect is for faster and slower responses and, consequently, how the influence of the conflict evolves with RT. For instance, a delta plot with a positive slope, as it is often observed in Stroop data (e.g., Pratte, Rouder, Morey, & Feng, 2010; Spieler, Balota, & Faust, 1996), indicates that there is a stronger congruency effect in longer RTs. In other words, it means that the RT difference between congruent and incongruent trials is bigger in trials, in which the responses are on average slower.¹ Importantly, the shape of the delta plot is often interpreted as indicative of cognitive processes (e.g., Mittelstädt & Miller, 2020; Ridderinkhof, Wildenberg, Wijnen, & Burle, 2004; for a review of processing accounts of delta plots see Schwarz & Miller, 2012). Specifically, they can provide insight with respect to the relative latencies of cognitive processes. That is, a delta plot with a positive slope, also referred to as positive going, is interpreted to indicate an influence of the automatic processing of the irrelevant information, which is relatively slow compared to the processing of the relevant information. Correspondingly, a delta plot with a negative slope, also referred to as negative going, is interpreted to indicate a relatively fast influence of the automatic processing of the irrelevant information. The activation of the irrelevant information is often considered an automatic process, which needs to be inhibited (e.g., Wildenberg et al., 2010). Both the activation time and inhibition latency are subject to characteristics of the task and inter- and intraindividual processes. Thus, investigating the temporal dynamics of the conflict provides insights into the underlying cognitive processes and differences therein.

The goal of the present study was to explore aging effects on cognitive control and the underlying processes over the course of adulthood in the light of practice in a large-scale dataset of users of the online cognitive training platform Lumosity (www.lumosity.com). To this end, we investigated the temporal dynamics of conflict processing in responses to the Stroop inspired color match game by assessing how age in combination with practice modulates the shape of delta plots in this task.

Methods

Transparency and Openness

All data and analysis code are available on the Open Science Framework (OSF, https://osf.io/yb5au/). The data are owned by Lumos Labs, Inc and are made available with the company’s permission. Due to its exploratory nature, we did not preregister this study nor conduct power analyses.

Dataset

The dataset was derived from the online cognitive training platform Lumosity. Lumosity offers so-called ``brain training built on science,” that is, a range of cognitive tasks conventionally applied in research, assessment or treatment, which are gamified and embedded in an online training program. For example, the color match game is a complex version of the Stroop task. In this game, users are instructed to match the meaning of a reference color word, which is always displayed in black on the left side of the screen, with the display color of a second color word, which is, like in the Stroop task, displayed in an unrelated color on the right side of the screen. Since responses are indicated by key presses, the color match game rather resembles the manual Stroop task, for which commonly smaller congruency effects are found compared to the vocal Stroop task (e.g., Fennell & Ratcliff, 2019; White, 1969). Figure 1 depicts a trial of the color match game.

Figure 1. Trial of the Color Match Game. The meaning of the left word (always displayed in black) and both meaning and display color of the right word (displayed in red in this example) change independently between trials. A color version of this figure is available in the online version.

The deidentified dataset contains trial by trial data of the color match game of 1200 Lumosity users. These were drawn from six age groups, ranging from 21 to 80, with equal sizes of 200. We treated age as a categorical variable to reduce the complexity of our data analysis. However, these six age groups, equally spaced over adulthood, still allow for a much more fine-grained analysis than studies that rely on comparing a group of young adults, often university students in their twenties, with a group of older adults aged 60 to 80. Additional demographic variables contained in the dataset are gender and level of education.

All included users had played the color match game at least 60 times and the first 60 games are in the dataset. There was a high variability in the temporal spacing of the games. Specifically, some games were started immediately after the last and others up to 1406 days later. One color match game lasts exactly 45 seconds. Within this period, users completed on average 41.41 ($\mathit{SD}=$ 17.90) trials. Due to the high variability in the number of trials per game there was also a high variability in the total number of trials between users, $\overline{\mathit{trials}}=$ 2,313.79, $\mathit{SD}=$ 771.50.

Since we had data from the first 60 games of each user, we could relate the investigated aging effects with practice effects. We defined practice based on the number of completed trials, binned into the three categories which are defined by the average number of completed trials after the first, 20th, 40th and 60th game, that is the 22nd to 604th trial, the 605th to 1415th trial, and the 1416th to 2304th trial. The average number of trials in the first game (i.e., trial 1 to 21) were regarded training trials and thus excluded in the analysis. All trials above the average trial number after the 60th game (i.e., trials above 2304) were also excluded to increase comparability between overall slower and faster users. An alternative to the categorization of practice based on the number of completed trials would have been to categorize by the numbers of games played, thus defining practice as the amount of time spent with the task. An advantage of this alternative approach is that no data would have to be excluded to increase comparability. The main results of a sensitivity analysis following this approach are in the online supplemental material in Section C.

On a trial level, the data contain information on the stimuli, the RT, the correctness of the response given, and the scoring used for Lumosity training feedback. The information on the stimuli comprises the meaning of the reference word displayed on the left side of the screen, the meaning of the target word displayed on the right side of the screen and the color of this target word. Based on these, four to five distinct trial types can be differentiated (i.e., three to four more types than in the standard Stroop task), as is depicted in Table 1. Two attributes of the trial define the trial type: The first is whether the color of the target word matches the meaning of the reference word, that is, the information relevant to responding. The second is whether the meaning of the target word matches the meaning of the reference word, that is, the irrelevant information expected to interfere with responding. Both of these attributes can be either true or false, which leads to four distinct trial types: (1) Both true, (2) first true but second false, (3) first false but second true, and (4) none true. The fourth type can be further differentiated by extracting the special cases in which the meaning and the color of the target word are the same and thus match each other, while not matching the meaning of the reference word (Type 4b in Table 1). Table 1 contains examples for all trial types.

Table 1. Trial Types in the Color Match Game.

	Example			Match
Trial type	Reference meaning	Target color	Target meaning	Color	Meaning	Correct response
1	Red	Red	Red	+	+	“Yes”
2	Red	Red	Blue	+	-	“Yes
3	Red	Red	Red	-	+	“No”
4a	Red	Blue	Yellow	-	-	“No”
4b	Red	Blue	Blue	-	-	“No”

Note. Trial types in the color match game are defined by the meaning of the reference word and the meaning and color of the target word. The meaning of the reference word can match the meaning of the target word and/or the color of the target word. Responses refer to the matching of the meaning of the reference word and the color of the target word.

Analysis

We used R (Version 4.2.2; R Core Team, 2021),² for all our analyses.

Data Exclusion

In the pre-processing of the data, we had to exclude the duplicate data of one game, which was in the dataset twice, and the data of 4943 games for which no trial level data were available. Additionally, we excluded the first 21 trials of each user (i.e., 25200 trials in total) from the analysis, since we regarded the respective trials as practice trials, and all trials above 2304 (i.e., 383585 trials in total), to increase comparability between slower and faster respondents. Furthermore, we excluded 23,775 trials with an RT below the 0.50% quantile or above the 99.50% quantile per age group and practice level. Table 2 depicts number of users, games, and trials in the final dataset.

Table 2. Sample Size.

Age	Users	Practice	Games	Trials
(20,30]	200	low	3085	115371
		intermediate	3185	157905
		high	2911	162244
(30,40]	200	low	3313	115331
		intermediate	3299	158326
		high	2946	160921
(40,50]	200	low	3852	115427
		intermediate	3734	157785
		high	2888	144660
(50,60]	200	low	4010	115434
		intermediate	3971	159477
		high	2835	137133
(60,70]	200	low	4897	115406
		intermediate	4409	147588
		high	1985	87185
(70,80]	200	low	5846	115208
		intermediate	4501	132218
		high	1159	46365
Total	1200		60655	2343984

Note. Number of users, games, and trials per age and practice group in the final dataset.

To investigate the congruency effect in the color match game, trial types need to be categorized as congruent or incongruent. However, in the color match game this is not as straightforward as in the standard Stroop task, since there are more attributes that can be congruent or incongruent. Specifically, (1) the meaning and color of the target word can be congruent, (2) the meaning of the reference word and the color of the target word can be congruent, (3) the meaning of the reference word and the meaning of the target word can be congruent, and/or (4) the congruency with the reference word can be congruent for the meaning and the color of the target word. Since it is impossible to foresee which type of congruency would influence responding (to which degree), we decided to focus on two trial types, for which a congruency effect similar to that of the standard Stroop task was to be expected. Specifically, we focused on the two trial types, in which the color of the target word matches the meaning of the reference word, that is, the two trial types, in which the correct response is “Yes” (i.e., Types 1 and 2 in Table 1). The trial type, in which both the color and the meaning of the target word match the reference word, is congruent in all above mentioned ways. The trial type, in which only the color but not the meaning of the target word matches the reference word, is congruent with respect to the matching of the color of the criterion word and the reference word, but incongruent in every other way. Most importantly, these two trial types differ in the congruency of meaning and color of the target word, which is the type of congruency that most closely resembles that in a standard Stroop task. Thus, we based all analyses exclusively on these two trial types.³

Descriptive Analysis

For a summary of the RTs, we computed the mean and median per age group, practice level and trial type. Additionally, we calculated the accuracy, that is, the proportion of correct responses per user, practice level, and trial type, and computed the mean per age group, practice level, and trial type.

Analysis of Congruency

First, to describe the congruency effect, we computed summary statistics per age group and practice level. We plotted individual congruency effects on RT, using only correct responses, and on accuracy separately for all age groups and practice levels. We investigated whether the congruency effect differences between age groups and practice levels were proportional to general speed differences.

To then investigate the temporal dynamics of the congruency effect, we created delta plots for each age group and practice level based on vincentized percentiles (Ratcliff, 1979). Specifically, we computed percentiles of the RT distributions of correct congruent and incongruent trials of each game and aggregated them per age group and practice level.⁴ We then computed, for each percentile (1 to 100), the mean RT and the difference between RTs of the two trial types (i.e., delta). Finally, we plotted delta as a function of the mean RTs separately for age groups and practice levels. To test differences in slope between age groups and practice levels, we conducted a linear regression. In this regression, delta was the dependent variable. The independent variables were the mean RTs, the age group, and the practice level. We centered the mean RTs around their respective group means (per age group and practice level). This way, the main effects of age group and practice level refer to the respective average RTs. To be able to assess linear trends over age groups and practice levels, we transformed these two variables to integer values ranging from 0 to 5 and 0 to 2, respectively. Importantly, the interaction effect of mean RT and age group (practice level) captures the influence of age (practice) on the slope of the delta plot.

Additionally, to control for the influence of the overall size of the congruency effect, due to, for instance, general slowing, and to increase the comparability across groups with respect to the shape of the delta function, we computed so-called Lorenz-interference plots (Gajdos, Servant, Hasbroucq, & Davranche, 2020). Conceptually derived from classical Lorenz curves, these plots depict the cumulative proportion of the total congruency effect as a function of the RT quantile. Thus, the curve always starts at the origin, with the 0% RT quantile and 0% of the conflict, and ends with the 100% RT quantile and 100% of the conflict. A very early conflict (i.e., negative going delta plot) results in a concave function, mainly above the diagonal. For example, if 80% of the total congruency effect can be attributed to the fastest 20% of the trials, one would observe a very steep increase in the Lorenz curve, which then flattens out. On the other hand, a late conflict (i.e., positive going delta plot) results in a convex function, mainly below the diagonal. For example, if only 20% of the total congruency effect, can be attributed to the fastest 80% of the trials, one would observe a very flat Lorenz curve, which then steepens strongly toward the end. The area under the curve yields the interference distribution index (IDI; Gajdos, Servant, Hasbroucq, & Davranche, 2020). An IDI greater than .5 indicates an early conflict and an IDI smaller than .5 a late conflict. We applied this standardized index to compare congruency effect distributions across the age groups and practice levels with differing mean RTs. We used a bootstrapping procedure with 1000 bootstrap samples to compute confidence intervals for the IDI estimates and reduce potential bias in estimation (Efron, 1979).

Results

Response Times

Figure 2 depicts the mean and median RT aggregated across trials as a function of age group separately for practice level and trial type (out of Type 1 and Type 2). RTs increase with age and decrease with practice level. They are generally lower in trials, in which both the color and the meaning of the target word match the meaning of the reference word (i.e., congruent trials). Descriptively, the age differences become smaller with higher practice levels. However, these differences might be proportional to the mean practice effects. The medians, which are less susceptible to the skew in RT distributions, are overall lower but reflect the same general pattern with respect to age group, practice level, and trial type. Raw RT distributions and summary statistics for all trial types are in the online supplemental material in Section A.

Figure 2. Summary Statistics for Response Times Across Age Groups, Practice Levels and Trial Types. Error bars indicate 95 percent confidence intervals.

Accuracy

Figure 3 depicts the mean accuracy per user and trial type aggregated across all users as a function of the age group separately for practice level and trial type (out of Type 1 and Type 2). Clearly, there is a strong decrease in accuracy with age in the trials, in which the color of the target word matches the meaning of the reference word, but the meaning of the target word does not (i.e., incongruent trials), when the level of practice is low. This effect is less pronounced at intermediate practice level, and nearly absent (except in the oldest group) at high level. Accuracy is generally very high in trials in which both the color and the meaning of the target word match the meaning of the reference word (i.e., congruent trials). This was to be expected, since both the relevant and the irrelvant stimulus information trigger the same response (``Yes’’) in these trials and, therefore, even an accidental automatic response based on the irrelevant information is a correct response. Neither practice level nor age seem to strongly influence accuracy in these trials. Raw accuracy distributions and summary statistics for all trial types are in the online supplemental material in Section B.

Figure 3. Summary Statistics for Accuracies Across Age Groups, Practice Levels and Trial Types. Error bars indicate 95 percent confidence intervals.

Analysis of Congruency

Mean Congruency Effect

Figure 4 depicts the mean congruency effect on RTs per user and practice level, that is, the mean difference between RTs in trials in which both the color and the meaning of the target word match the reference word (i.e., congruent trials) and trials in which only the color of the target word matches the reference word (i.e., incongruent trials) using only correct trials. The data are split by age group and practice level. Virtually all users show a positive congruency effect at all practice levels, with a wider spread in older age groups and at lower practice levels. Table 3 displays the mean congruency effect per age group and practice level. It reflects that the mean and standard deviation of the congruency effect increase with age and decrease with practice. To indicate whether the differences are overproportional, Table 3 additionally contains the mean RT and the congruency effects proportional to this mean RT. The differences concerning the congruency effect are not proportional. With respect to the mean RT, there is an overproportional increase of the congruency effect with age and an overproportional decrease with practice. The proportional increase of the congruency effect with age is smaller at high practice level (0.09 to 0.18) than at low practice level (0.17 to 0.32).

Figure 4. Individual Congruency Effects per Age Group and Practice Level.

Table 3. Mean RT, Congruency Effect (SD in Parentheses) and Proportional Congruency Effect by Practice Level and Age Group.

	Mean RT			Congruency effect			Proportional congruency effect
Age	low	intermediate	high	low	intermediate	high	low	intermediate	high
(20,30]	1131 (395)	853 (212)	769 (152)	196 (181)	97 (104)	68 (75)	0.17	0.11	0.09
(30,40]	1236 (369)	888 (232)	776 (147)	243 (165)	110 (102)	66 (60)	0.20	0.12	0.09
(40,50]	1466 (458)	1019 (252)	871 (150)	342 (240)	162 (136)	98 (75)	0.23	0.16	0.11
(50,60]	1547 (423)	1086 (257)	931 (190)	396 (260)	195 (143)	122 (110)	0.26	0.18	0.13
(60,70]	1971 (652)	1392 (459)	1060 (211)	567 (387)	319 (287)	169 (121)	0.29	0.23	0.16
(70,80]	2428 (810)	1686 (564)	1180 (241)	765 (494)	461 (349)	210 (160)	0.32	0.27	0.18

Note. RT-(difference)s are reported in ms. “Low,” “intermediate,” and “high” refer to the practice level.

Figure 5 shows the individual congruency effects on accuracies. Table 4 displays the respective mean congruency effect on accuracy per age group and practice level. The data pattern basically echoes the one referring to RTs. That is, virtually all users show a positive congruency effect at all practice levels both in RT and accuracy.

Figure 5. Individual Congruency Effects on Accuracies per Age Group and Practice Level.

Table 4. Mean Congruency Effect on Accuracy ($\mathit{SD}$ in Parentheses) by Age Group and Practice Level.

	Practice level
Age	low	intermediate	high
(20,30]	0.11 (0.11)	0.08 (0.08)	0.07 (0.08)
(30,40]	0.11 (0.09)	0.07 (0.07)	0.06 (0.05)
(40,50]	0.17 (0.13)	0.09 (0.09)	0.07 (0.07)
(50,60]	0.18 (0.13)	0.09 (0.10)	0.07 (0.09)
(60,70]	0.26 (0.18)	0.14 (0.14)	0.07 (0.07)
(70,80]	0.37 (0.21)	0.22 (0.19)	0.12 (0.14)

Delta Plots

To investigate the temporal dynamics of the conflict, we created delta plots based on vincentized RT distributions per age group and practice level. These delta plots are in Figure 6. All plots are positive going, reflecting the typical shape of delta plots in the Stroop task.

Figure 6. Delta Plots per Age Group and Practice Level.

To investigate differences in the shape of the delta plots, we fit a linear function to the delta plots with age group and practice level as covariates. The results are in Table 5. In the reference category (age 21 to 30, low practice), the congruency effect is predicted to be 168 ms at an average RT and to increase with RT by 33 ms per 100 ms (i.e., positive going delta plot). The congruency effect is predicted to increase with age by 104 ms per age group and decrease with practice by −67 ms per practice level. The increase in the congruency effect with age is predicted to be smaller with increasing practice by −42 ms per practice level. These main effects and interaction term reflect the differences between the mean congruency effects described above. Additionally, one can see that there is a negative interaction of the linear slope of the delta plots with age and practice. Thus, the delta plots are predicted to be less steep with increasing age, at low practice level (see first row in Figure 6). They are also predicted to be less steep with more practice, at youngest age (see first column in Figure 6). However, there is a positive three-way interaction of the linear slope, age, and practice level. Thus, with increasing age, the negative influence of practice on the linear slope diminishes, and, as one can see in Figure 6, even becomes positive in the older age groups (see, e.g., last column in Figure 6). Or, put differently, with increasing practice, the negative influence of aging on the linear slope diminishes, and, as one can see in Figure 6, seems to disappear at high practice level (see last row in Figure 6).

Table 5. Results of Fitting the Delta Plot with a Linear Function with Age Group and Practice Level as Covariates.

Predictor	$\mathit{b}$	95% CI	$\mathit{t}$	$\mathit{df}$	$\mathit{p}$
Intercept	167.94	[162.42, 173.46]	59.69	1792	<.001
Mean RT gmc	0.33	[0.31, 0.36]	24.99	1792	<.001
Age num	103.91	[102.09, 105.74]	111.82	1792	<.001
Practice num	−67.18	[−71.46, −62.91]	−30.83	1792	<.001
Mean RT gmc $\times$ Age num	−0.03	[−0.04, −0.03]	−9.15	1792	<.001
Mean RT gmc $\times$ Practice num	−0.06	[−0.08, −0.03]	−3.77	1792	<.001
Age num $\times$ Practice num	−42.41	[−43.83, −41.00]	−58.92	1792	<.001
Mean RT gmc $\times$ Age num $\times$ Practice num	0.04	[0.03, 0.05]	9.94	1792	<.001

Note. Mean RTs are group mean centered (“gmc”). Age group and practice level are treated as numeric variables (“num”) with 0 indicating the lowest group, that is, “low” and “(20, 30],” respectively.

Lorenz Interference Plots

Since the mean RTs and mean size of the congruency effect differ strongly between different age groups and practice levels, we applied the procedure by Gajdos, Servant, Hasbroucq, and Davranche (2020) to increase comparability between groups. The resulting Lorenz-interference plots are depicted in Figure 7 separately for age groups and practice levels. All curves are convex, indicating a stronger congruency effect in longer RTs. The curvature is more pronounced for higher levels of practice and, at low practice level, at younger age. The age differences decrease with higher practice level, at the highest practice level they nearly disappear. This demonstrates a stronger shift of the congruency effect toward longer RTs with higher levels of practice and younger age, when practice is low.

Figure 7. Lorenz-Interference Plots by Age Group and Practice Level. The dashed line marks the hypothetical uniform distribution of the interference over the RT distribution. A color version of this figure is available in the online version.

From the Lorenz-interference plots, the IDI can be computed as a general standardized measure of the latency of the congruency effect relative to RTs. Figure 8 depicts the IDI as a function of age group and practice level. Nearly all IDIs are below .5, indicating a relatively late congruency effect. An exception is visible in the age group 70 to 80 with little practice, where the IDI approximates .5. Again, the pattern indicates a clear ordering by practice level, with decreasing IDIs with increasing level of practice. The ordering by age group, however, is not so consistent. At a low practice level, there is a general trend of increasing IDIs with increasing age. At intermediate practice level, there is still a slight trend of increasing IDIs with increasing age. However, at a high practice level the IDIs do not differ with age (if anything there seems to be a U-shaped relationship).

Figure 8. Interference Distribution Index (IDI) as a Function of Age Group and Practice Level. The IDI indicates the latency of the conflict on a scale from 0 to 1. Smaller IDIs indicate a larger congruency effect at longer RTs. Depicted are the means of 1000 bootstrap samples and the respective 95 percent confidence intervals.

Discussion

The present study investigated age effects in the light of practice on the temporal dynamics of cognitive conflict in a big Lumosity color match game dataset. As expected, we observed an increase of RT and a decrease of accuracy with age (and a decrease of RT and an increase of accuracy with practice). Likewise, the congruency effect both on RT and accuracy increases with age and decreases with practice. These differences seem to be overproportional with respect to the general speed differences, indicating an effect that goes beyond the general slowing account. The observed age effects are in line with the literature indicating a decrease of cognitive control with age (e.g., Andrés, Guerrini, Phillips, & Perfect, 2008; Aschenbrenner & Balota, 2015; Bugg, DeLosh, Davalos, & Davis, 2007; Jackson & Balota, 2013; Li & Bosman, 1996; Nicosia & Balota, 2020; Spieler, Balota, & Faust, 1996; Zanto & Gazzaley, 2017). Moreover, the changes seem to be rather constant over the course of adulthood.

Furthermore, this study specifically focused on the temporal dynamics of the conflict to gain insight in differences in the underlying processes. The observed positive going delta plots across all age groups and practice levels are typical for the Stroop task (e.g., Pratte, Rouder, Morey, & Feng, 2010; Spieler, Balota, & Faust, 1996). They indicate that the automatic activation of the task-irrelevant information (i.e., word meaning) is relatively slow compared to the controlled activation of the task relevant information (i.e., display color). Modeling the delta plots with a linear function, with age and practice as covariates, revealed that, at low practice level, the delta plots become less positive going with age. This means that, at low practice level, relative to the controlled processing of the relevant information, the automatic processing becomes faster with age. However, the effect of aging on the linear slope interacts with practice level, such that, at high practice level, the influence of age is dissolved.

To increase comparability across the age groups and practice levels with differences in the overall RT and, thus, control for general slowing, we additionally created Lorenz-interference plots. From these, the same interpretation follows: The curves in all age groups and practice levels are convex. At low practice level the curvature decreases with age. However, this trend diminishes with practice and at high practice level there are hardly any visible age differences in the curves. This pattern is reflected by the IDIs, which are mainly below .5. They increase with age, only at a low practice level (and slightly at intermediate practice level). This also indicates a long relative latency of the automatic processes, which is reduced with age, when there is little practice.

To summarize, there is a long relative latency of the automatic activation, which decreases with age at low practice level. However, since from the present observations we can only draw conclusions about relative latencies, we cannot infer the source of these changes. The decrease with aging at low practice level could either result from a slowing of the controlled processes or a faster or stronger influence of the automatic processes, for example due to reduced inhibition strength. To further disentangle these sources, cognitive models have been developed, such as the dual-stage two-phase model of visual attention (DSTP; Hübner, Steinhauser, & Lehle, 2010), the diffusion model for conflict tasks (DMC; Ulrich, Schröter, Leuthold, & Birngruber, 2015), and the shrinking spotlight model (SSP; White, Ratcliff, & Starns, 2011). In the study mentioned in the introduction of this paper, Servant and Evans (2020) fit these models to aging data from a flanker task. The DMC fit the data best and the resulting parameter estimates indicated slower sensorimotor processing with age, and higher decision thresholds (i.e., more careful responses). Unfortunately, efforts to fit the DMC to the present Lumosity dataset were not successful. The parameter estimates were not able to reasonably predict the data pattern. This is not too surprising since the task is much more complex than the standard Stroop task. Thus, in the present dataset we cannot disentangle if the age and practice differences in the relative speed of the processing of the relevant and the irrelevant information are due to differences in the speed of controlled processing or of automatic processing (possibly because of differences in inhibition strength).

Other studies in the literature can shed some light on possible sources. As mentioned before, Servant and Evans (2020) found that aging was related to slower sensorimotor processing and higher decision thresholds in the Flanker task. However, it is unclear, whether this conclusion generalizes to the Stroop task. Other studies investigated whether aging affects conflict processing on the level of semantic or response processing and obtained contrasting results (e.g., Augustinova, Clarys, Spatola, & Ferrand, 2018; Burca, Chausse, Ferrand, Parris, & Augustinova, 2022). Lien et al. (2006) observed an increased efficiency in parallel processing in a lexical task with age. They interpreted this finding to indicate that the automaticity of lexical processing increases with age, possibly due to experience (see also Allen et al., 2002). This interpretation is in line with our observation that the latency of the automatic processing of the irrelevant information decreases relative to the processing of the relevant stimulus with age. Nevertheless, importantly, it is not possible to disentangle processes at this level based on our analyses.

However, what the results of investigating the temporal dynamics show is that the effects of aging are moderated by practice, even at the level of the relative speed of controlled and automatic processes. This indicates that practice may help counteracting aging effects on a process level and that, therefore, practice might be especially beneficial in older age groups.

Limitations

This study is subject to certain limitations, which are mainly due to the non-experimental mode of data administration. First, as mentioned above, the mode of data administration differed strongly from that of a standard experiment. There is less control over the users’ response behavior in an online, self-administered training and arguably the motivators for engaging in the task differ substantially. The short blocks, with a high variance of inter-block intervals, and the lack of training trials presumably produced a lot of additional random noise, which might have been the reason we were not able to fit the DMC to the data. However, the present Lumosity dataset has apparent advantages. The size of the dataset increased the reliability of estimates reflected in the narrow confidence intervals and, thus, made inference statistics nearly dismissible.

Second, the data are not directly comparable to those from other Stroop studies due to the high complexity of the color match game. As explained in the methods section, there are various types of congruency that could have evoked different types of conflict. Despite our focus on two trial types, in which the congruency effect was relatively predictable, it is difficult to evaluate how conflict was evoked even in these trials. Additionally, the response format of the color match game differs from that of the standard Stroop task. Instead of indicating the color of a word, a rather neutral response, responses are given in a “Yes/No” format. “Yes” and “No” responses are not neutral and often produce differences in RT. Although these factors of task design decrease the comparability of the present results to those of standard Stroop experiments, they bare the potential of additional insights. Sections A and B of the online supplemental material contain summary statistics of RT and accuracy with respect to all trial types. The observed differences between trial types encourage reflecting on the possible influences of more complex conflict schemata. This, however, is beyond the scope of this paper. On the other hand, it is another indicator for the robustness of the Stroop effect that we found a strong and consistent congruency effect even in this complex task.

Third, we investigated practice effects within one specific task. This does not allow us to draw any conclusions about the effects of training on cognitive control in other tasks or even everyday life situations. In the literature, there is a controversy on such general training effects on cognitive functioning, to which we cannot contribute, here (see, e.g., Lampit, Hallock, Valenzuela, & Gandy, 2014; Sala & Gobet, 2019; Simons et al., 2016). However, we can show that, within the same task, even on the process level, practice can have effects on conflict processing, especially in older adults.

Despite these limitations, we could replicate standard Stroop task findings in this more complex task. We observed a consistent congruency effect, which increases with age and decreases with practice. The delta plots are mainly positive going, as expected, and they decrease in slope with age. Additionally, the fine-grained age groups allowed us to investigate the development of these differences over the course of adulthood. The data indicate a rather constant process with respect to all age effects. Furthermore, we could relate these age effects with practice effects. The data indicate that even on the level of the temporal dynamics of conflict processing, practice effects can moderate aging effects and, thus, practice seems to be able to partially counteract the age-related deterioration of mental speed.

Conclusion

We investigated the influence of aging in the light of practice on conflict processing by means of inspecting the temporal dynamics of the conflict as indicated by delta and Lorenz-interference plots in a large scale dataset of the Lumosity color match game. In addition to the expected increase of interference effects, the data indicate an increase in speed of the automatic activation of irrelevant information relative to the controlled activation of the relevant information with age, which can be partially counteracted by practice.

Data Availability

The data owned by Lumos Labs, Inc and made available with the company’s permission and all analysis scripts are available on the Open Science Framework (https://osf.io/yb5au/)

Author Contributions

The authors made the following contributions. Fabiola Reiber: Conceptualization, Data analysis, Writing – Original Draft Preparation; Rolf Ulrich: Conceptualization, Writing – Review & Editing.

Acknowledgments

We thank Lumos Labs, Inc for contributing the deidentified dataset free of charge through the Human Cognition Project without restricting analyses and interpretations and the research team for helpful comments and feedback. We thank Victor Mittelstädt for helpful discussions and suggestions.

Disclosure Statement

No potential conflict of interest was reported by the authors.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/0361073X.2023.2214051

Additional information

Funding

This research was funded by the Deutsche Forschungsgemeinschaft (DFG), grant 2277, Research Training Group “Statistical Modeling in Psychology” (SMiP).

Notes

1. A bigger congruency effect in longer RTs might seem counterintuitive on first sight, since one might assume that longer RTs are associated with more deliberation. However, importantly, also longer RTs reflect speeded responses in the Stroop task and, thus, do not necessarily involve more deliberation.

2. We, furthermore, used the R-packages papaja (Version 0.1.1; Aust & Barth, 2020), and tidyverse (Version 1.3.2; Wickham et al., 2019).

3. Nevertheless, Sections A and B of the online supplemental material contain summary statistics of the remaining trial types.

4. Since the number of trials was relatively low per game, we tested different quantile algorithms using continuous functions of p, which are available in the quantile function provided by the stats R-package. The interpretation of the results was not influenced by the choice of the algorithm type.