The dual-system model’s predictions with respect to the effects of target pattern and learning condition on participants’ learning strategies were also of interest. To test the predictions of hypothesis (ii), two more GLM’s were fit to experiment data. Unlike GLM-1 and GLM-2 which modeled response accuracy, these GLM’s predicted participants’ learning strategies with the following predictor variables:
Table 5.11: GLM-3 and GLM-4 Predictor Variables
Predictor/Covariate Code Levels Reference
Target Pattern target.pattern c=1/ncs=2/ncns=3 c
Learning Condition learning.condition E=0/I=1 E
*Covariates areitalicized
5.4.1 GLM-3, Implicit Learning
The third GLM (GLM-3) predicted the binomial outcome(ER) Training-Intuition=true/false (0/1), with a logit link and the predictor variables given in Table 5.11. The model equation as defined in R was:
GLM-3 Equation
(ER)Training−Intuition∼target.pattern∗learning.condition
Model parameters were estimated using GEE. As we are only interested in the effects of target pattern and learning condition on the likelihood of using intuition in training, there were no repeated measures. Coefficient estimates are given in Table 5.12 below:
Table 5.12: GEE Coefficient Estimates for GLM-3 Estimate Std.err Wald Pr(>|W|) (Intercept) 0.9163 0.4830 3.60 0.0578 NCNS 0.3365 0.7448 0.20 0.6515 NCS -0.6286 0.7246 0.75 0.3856 I -0.5108 0.6646 0.59 0.4421 NCNS:I -0.3365 0.9856 0.12 0.7328 NCS:I 1.1676 0.9653 1.46 0.2264
Given the estimates in Table 5.12, it seems that none of the modeled predictor variables signifi- cantly affected participants’ learning strategies. However, if we look past the non-significance of the effects, we see the strange result that whereas theNCSpattern and theImplicitlearning task on their own seem todiscouragethe use of intuition (both having negative coefficients), we see that their interaction,NCS:I, has a positive coefficient, suggesting that participants in the I-NCS condition were slightly more likely to implement implicit learning than the average across all participants.
Comparing this result with the prediction of hypothesis (ii), namely that the dual-system is sensitive to the types of patterns being acquired and allocates learning respectively, GLM-3 does not seem to support such an account. Instead, we find that the likelihood of using intuition in training was not significantly affected by any of the provided predictor variables.
Using Least-Squares Means and back-transformations from the logit scale, the marginal mean probabilities of using intuition by pattern and condition were derived (Table 5.13 and Figure 5.2).
Table 5.13: GLM-3, Least-Squares Means Probability Estimates target.pattern learning.condition prob SE asymp.LCL asymp.UCL
C E 0.7143 0.0986 0.4924 0.8657 NCNS E 0.7778 0.0980 0.5353 0.9140 NCS E 0.5714 0.1323 0.3163 0.7935 C I 0.6000 0.1095 0.3801 0.7858 NCNS I 0.6000 0.1095 0.3801 0.7858 NCS I 0.7200 0.0898 0.5178 0.8603
We can see that, overall, participants in all conditions were more likely than not to use intuition in training in all conditions . One likely explanation for this is that the pattern of metathesis was too difficult/confusing for participants, and many reported “giving up” on finding a rule and switching to using their intuition instead.
In the end, GLM-3 shows no clear evidence of participants use of intuition being sensitive to the target pattern (or learning condition). This goes against hypothesis (ii) and suggests that, if there is a dual-system grammar, that the use of a particular mechanism is not explained by the target pattern. However, this would deserve further testing given the problem of participants struggling to find a rule and using their intuitions instead.
5.4.2 GLM-4, Explicit Learning
The fourth GLM (GLM-4) predicted the binomial outcome(ER) Sought-Rule=true/false (0/1), with a logit link and the predictor variables given in Table 5.11. The model equation as defined in R was:
GLM-4 Equation
(ER)Sought−Rule∼target.pattern∗learning.condition
Model parameters were estimated using GEE. As we are only interested in the effects of target pattern and learning condition on the likelihood of seeking a rule in training, there were no repeated measures. Coefficient estimates are given in Table 5.14 below:
Table 5.14: GEE Coefficient Estimates for GLM-4
Estimate Std.err Wald Pr(>|W|) (Intercept) -1.7918 0.6236 8.26 0.0041 NCNS 0.5390 0.8428 0.41 0.5225 NCS 0.8755 0.8596 1.04 0.3085 I 1.3863 0.7728 3.22 0.0728 NCNS:I -0.9808 1.0755 0.83 0.3618 NCS:I -0.7112 1.0534 0.46 0.4996
As with GLM-3, it seems that most of the predictor variables for GLM-4 had no significant effect on participants’ likelihood of seeking a rule in training. The one exception to this is the marginally significant positive effect of being in the Implicit learning condition (p= 0.0728). This is strange, as it was expected that the Explicit learning condition should foster explicit learning strategies while the Implicit learning condition would foster implicit learning strategies.
One possible explanation for the positive effect of Implicit learning condition could be that the Implicit task may have required less effort since participants weren’t actively having to compare two choices and could just focus on the given words and their combination. Whereas the explicit learning task may have been too much for participants and led them to give up on seeking a rule, the implicit task may have been simple enough to encourage rule seeking.
Using Least-Squares Means and back-transformations from the logit scale, the marginal mean probabilities of seeking a rule according to pattern and condition were derived (Table 5.15 and Figure 5.3).
Table 5.15: GLM-4, Least-Squares Means Probability Estimates target.pattern learning.condition prob SE asymp.LCL asymp.UCL
C E 0.1429 0.0764 0.0468 0.3614 NCNS E 0.2222 0.0980 0.0860 0.4647 NCS E 0.2857 0.1207 0.1115 0.5605 C I 0.4000 0.1095 0.2142 0.6199 NCNS I 0.3000 0.1025 0.1414 0.5272 NCS I 0.4400 0.0993 0.2629 0.6338
Figure 5.3: GLM-4, Least-Squares Means Probability Estimates
Table 5.15 and Figure 5.3 show that participants were generally unlikely to seek rules across all conditions. Estimated marginal means confirm that participants in the Implicit learning condition were overall more likely report seeking a rule. Furthermore, we see no significant difference in the
likelihood of seeking a rule betweenI-CandI-NCSparticipants as would have been expected by hypothesis (ii).
CHAPTER 6
DISCUSSION & CONCLUSIONS
We can now begin the undertaking of synthesizing the results of the AGL experiment. Of course, we should note the concerning fact that all experimental conditions should test accuracy at or below chance levels, suggesting that the tested pattern of metathesis was too difficult/unnatural. However, in-depth analysis does reveal some interesting results that bear on the considered hypotheses (copied below for convenience).
(10) Experiment Hypotheses
(i) Positive effect of explicit learning strategies onC-pattern test-phase accuracy, and a null effect of implicit learning onC-pattern test-phase accuracy
(ii) TheC-pattern has a positive effect on the likelihood of carrying out explicit learning, and this effect is significantly greater forCthanNCSandNCNS
(iii) Significant contrast between the test-phase accuracy of implicit learners in theNCS-pattern condition and in theNCNS-pattern condition. No significant contrast for explicit learners across these two pattern conditions.