Power of instructions for task implementation: superiority of explicitly instructed over inferred rules

Abstract

“Power of instructions” originally referred to automatic response activation associated with instructed rules, but previous examination of the power of instructed rules in actual task implementation has been limited. Typical tasks involve both explicit aspects (e.g., instructed stimulus–response mapping rules) and implied, yet easily inferred aspects (e.g., be ready, attend to error beeps) and it is unknown if inferred aspects also become readily executable like their explicitly instructed counterparts. In each mini-block of our paradigm we introduced a novel two-choice task. In the instructions phase, one stimulus was explicitly mapped to a response; whereas the other stimulus’ response mapping had to be inferred. Results show that, in most cases, explicitly instructed rules were implemented more efficiently than inferred rules, but this advantage was observed only in the first trial following instructions (though not in the first implementation of the rules), which suggests that the entire task set was implemented in the first trial. Theoretical implications are discussed.

Appendix

This appendix involves a few additional tests that were raised post hoc to try and clear the issues of (1) encoding during the instructions phase; and (2) learning throughout the experiment, given that the same inference is repeatedly needed. Importantly, we note that these analyses were raised in the review process and were not hypothesized in advance (this is especially important for the pre-registered Experiments 2 and 3) (Fig. 

Fig. 6

Study times in ms (including the 2 s delay) as a function of Rule-Type and Inference Difficulty in Experiments 1–3. Error bars represent Bayesian 95% credible intervals

6).

• 1) Examining study times.

  As has been mentioned, in the instructions phase each of the screens had a 2 s delay, after which participants could advance to the following screen via a spacebar self-paced response, thus introducing some inaccuracy in the measurement of study times.

  To test whether participants used the instructions phase to differentially encode the instructed/inferred rules, under the assumption that longer study times imply greater encoding efforts, we performed two sets of Bayesian ANOVAs in each of the three experiments. The first set of ANOVAs was performed for total study times (i.e., the sum of both instructions screens compared to one instruction screen including both instructed rules). This analysis was aimed as a first exploration, closer to our original hypotheses (that difficult inference should take longer than easy inference, and longer than when both rules were explicitly instructed). Thus, the independent variable was Block Type (3 levels: both rules explicitly instructed; one rule explicitly instructed and one inferred, with difficult vs. easy inference).

  In Experiment 1, The results supported a main effect for Block Type [F(2,78) = 9.56, p < 0.001, η²_p = 0.20, BF₁₀ = 123.39], showing support for our hypothesis (study times were 3.52 s for two explicit rules, 3.79 s for the easy inference condition, and 3.95 s for the difficult inference condition (these times include the 2 s delay). The same pattern was observed in Experiment 2 [F(2,78) = 6.79, p < 0.01, η²_p=0.15, BF₁₀ = 16.21], although the results indicated a smaller difference between the condition where both rules were explicitly instructed and the easy inference condition (5.77, 5.80, and 6.15 s in the both explicitly instructed, easy inference and difficult inference, respectively). Finally, Experiment 3 showed the same pattern [F(2,78) = 7.81, p < 0.001, η²_p = 0.17, BF₁₀ = 34.48], with results that more clearly demonstrate the difference between the three conditions (4.63, 4.82, and 5.08 s for the both explicitly instructed, easy inference and difficult inference, respectively).

  The second set of ANOVAs only involved the “one explicitly instructed S–R mapping and one inferred S–R mapping” condition. This analysis was performed with the within-subjects independent variables Rule-Type (instructed-inferred), Inference Difficulty (difficult-easy; in this analysis, this variable codes for whether the explicit instruction appeared first (easy inference) or second (difficult inference)). This analysis was meant to hopefully better differentiate the encoding processes within this condition.

  Unlike in the previous analysis, these results point to differences between the experiments. In Experiment 1, the results indicated a robust interaction between Rule-Type and Inference Difficulty [F(1,39) = 11.05, p < 0.01, η²_p = 0.22, BF₁₀ = 3,337.53], whereas both main effects were not significant (ps > 0.06, BF₁₀ < 0.38). The results demonstrate that participants took longer to study whichever stimulus appeared first, regardless of whether it was explicitly instructed or inferred (Figure 6): Both simple effects for Rule-Type were robust (BF₁₀ = 3.41 for difficult inference, and BF₁₀ = 40.56 for easy inference.

  A similar pattern was observed in Experiment 2 [F(1,39) = 5.80, p = 0.2, η²_p = 0.13, BF₁₀ = 68.76 for the interaction]. In this experiment, the simple effect of Rule-Type was only robust for easy inference mini-blocks (BF₁₀ = 0.66 for difficult inference, and BF₁₀ = 20.84 for easy inference) suggesting that when the inference was easy (i.e., the instructed rule presented first), participants took less time to study the inferred rule, but when the inference was difficult study times did not robustly differed between rule types.

  In Experiment 3 however, the results did not indicate a robust interaction (or any other main effects, BFs < 2.60) [for the interaction: F(1,39) = 0.10, p = 0.76, η²_p =  < 0.01, BF₁₀ = 0.25]. Here, the descriptive results indicate that participants learned explicitly instructed rules somewhat longer than inferred ones (main effect: F(1,39) = 4.25, p = 0.046, η²_p = 0.10, BF₁₀ = 2.53; which is considered an indecisive effect). We note that this experiment did not involve a NEXT phase, and while we do not completely understand this effect, which was not the focus of the current study, it could be that this (lack of) effect partly reflects the lower task demands in this experiment.

• 2) Learning to infer during the experiment

  The next set of analyses aims to test whether the superiority of instructions changes with the progress of the experiment, given that participants are required to repeat the same general inference. To perform this analysis, we added a Progress variable that divides mini-blocks into three equal parts with 50 mini-blocks each. To focus on the main result, we tested the first-trial RT instructions superiority effect above and beyond other variables. In Experiment 1, the results showed a main effect for Progress [F(2,78)=85.78, p < 0.001, η²_p=0.49, BF₁₀=1.44e⁺²⁹] demonstrating RT acceleration throughout the experiment. In addition, the results suggest a robust interaction between Rule-Type and Progress F(2,78)=9.33, p<0.001, η²_p= 0.19, BF₁₀=6.83], showing that the superiority effect decreased with progression in the experiment (47.4 ms in the first 50 mini-blocks, 19.3 ms in the next 50 mini-blocks, and 11.5 ms in the final 50 mini-blocks).

  In Experiments 2 and 3, although the Progress main effect was robust [F(2,78)=38.14, p<0.001, η²_p=0.49, BF₁₀=3.58^e+14 (Exp. 2); F(2,78)=43.60 p<0.001, η²_p=0.53, BF₁₀=2.04^e+16 (Exp. 3)], the interaction between Rule-Type and Progress was not significant [F(2,78)=2.52, p=0.09, η²_p=0.06, BF₁₀=0.31 (Exp. 2); F(2,78)=0.31, p=0.73, η²_p<.01, BF₁₀=0.09 (Exp. 3)] and allowed accepting H0.

  Therefore, the results do not support the hypothesis that participants have learned how to cope with required inference in the course of the experiment capitalizing on the constant abstract structure of the instructions. Specifically, while a supporting pattern was observed in some experiments, it was not robust across experiments.