Interactions between Binary & Quantitative Predictors The purpose of the study was to examine the possible joint effects of the difficulty of the practice task and the amount of practice, upon the performance of the target task. Participants were randomly assigned to receive a practice that was either "" (easier than the target task), "" (same difficulty as the target task), or "" (er than the target task). Each participant was permitted to practice the appointed task as many times as they wanted, and then all were given the same target task and the performance recorded. Here's the plot of the data 0 8 6 TESTPERF 4 0 4 6 8 0 UMPRACT Here are the basic stats for the groups and the syntax to produce the dummy codes ( codes for 3 groups), centered covariate and the interaction terms ( -- each a product of the centered covariate and one of the group dummy codes) Report TESTPERF UMPRACT 5.500 5.0000 6 6.83.30940 6.50 6.3750 6 6.8063.6996 5.85 5.5000 6 6.9754.78089 5.79 5.650 48 48.56.58987 Below are three analyses of these data: Full model regression Hierarchical regression ACOVA using GLM
Full This involves including both dummy codes, the centered covariate and the two interaction terms in a single model. Summary R R Square.87 a.667 a. Predictors:,, IT, PRACT_D, PRACT_D, Sum of Squares AOVA b df 7.996 5 4.599 6.807.000 a 36.483 4.869 09.479 47 Square F Sig. a. Predictors:,, IT, PRACT_D, PRACT_D, b. PRACT_D PRACT_D IT Unstandardized Coefficients a. B Std. Error Coefficients a Standardized Coefficients Beta 5.90.33 5.300.000 -.847.336 -.64 -.50.06 -.4 -.36 -.4.58.337.08.766.448.85.7.068.7.087.807 8.9.000.334.785.73 -.04.35 -.898-7.558.000 -. -.759 -.673 -.757.6 -.758-6.0.000.000 -.680 -.535 t Sig. Correlations Zeroorder Partial Part Each term in the model are corrected for all other terms, which is part of interpreting each. a -- the constant tells us the mean of the comparison group, after correction for the covariate and the interaction. So, the practice group had an average performance of 5.9 PRACT_D -- tells the direction and extent of the difference between the mean of the comparison group and the target group for that dummy code, after correction for the covariate and the interaction. The t-test of the regression weight tests if that mean difference is statistically significant, after correction for the covariate and the interaction. So, the practice group had an average performance -.847 lower than the practice group or a mean of 5.9 -.847 = 5.053. The means of these two groups are significantly different, after correction for the covariate and the interaction. PRACT_D -- tells the direction and extent of the difference between the mean of the comparison group and the target group for that dummy code, after correction for the covariate and the interaction. So, the practice group had an average performance.58 higher than the practice group or a mean of 5.9 +.58 = 6.58. The means of these two groups are not significantly different, after correction for the covariate and the interaction. -- tells the slope of the COV-DV relationship, after correcting for the IV and the interaction IT -- tells about the difference in slope of the CIV-DV relationship of the target group for the related dummy code relative the slope for the comparison group, after correction for the covariate and the IV. So, the slope of simple regression line for the practice group is less positive than the slope of the simple regression line for the practice group -- tells about the difference in slope of the CIV-DV relationship of the target group for the related dummy code relative the slope for the comparison group, after correction for the covariate and the IV. So, the slope of simple regression line for the practice group is less positive than the slope of the simple regression line for the practice group Remember -- if any b of any code representing an effect is significant, that effect is significant!
We can use the IntPlot program to obtain the simple regressions for each group and the plotting points to portray this multivariate model. Select the one for 3 dummy-coded groups otice that the differences between the slopes of the simple regression lines match the regression weights for the interaction terms Slope for the practice group =.7 For the practice group =.7 + (-.04) = -.33 For the practice group =.77 + (-.757) = -.046 Test Performance 3 4 6 8 0 Medium >.58 > Hard Hard Practice Easy Practice Med Practice Easy <.847 < Hard -5. -.6 0.6 5. umber of Practices The relative slopes of the group s simple regression lines match the information from the various b-values. The corrected main effects are shown. These are made at the mean of the covariate, which is 0 because it has been centered, and also match the related b-values. While there is small main effect, clearly the important effect in this model is the interaction! Practice with items has no effect. Practice with items hinders performance. Practice with items improves performance,
Hierarchical The first model includes only the main effects -- the IV and the covariate. Then the interaction terms are added in the second step. Summary Change Statistics R Square R R Square Change F Change df df Sig. F Change.377 a.4.4.433 3 44.078.87 b.667.54 33.053 4.000 a. Predictors:,, PRACT_D, PRACT_D b. Predictors:,, PRACT_D, PRACT_D, IT, The main effects model is not significant. However there is a significant increase in R² when the interaction terms are added (showing the interaction is significant) and the resulting full mode is significant. Sum of Squares AOVA c Square F Sig. df 5.575 3 5.9.433.078 a 93.905 44.34 09.479 47 7.996 5 4.599 6.807.000 b 36.483 4.869 09.479 47 a. Predictors:,, PRACT_D, PRACT_D b. Predictors:,, PRACT_D, PRACT_D, IT c. PRACT_D PRACT_D PRACT_D PRACT_D IT a. Coefficients a Unstandardized Standardized Coefficients Coefficients B Std. Error Beta t Sig. 5.834.365 5.969.000 -.475.58 -.48 -.96.365.59.5.049.304.763.76.084.99.086.043 5.90.33 5.300.000 -.847.336 -.64 -.50.06.58.337.08.766.448.7.087.807 8.9.000 -.04.35 -.898-7.558.000 -.757.6 -.758-6.0.000 The full model is the same as was found above, because entry order does not influence the inclusive model.
ACOVA using SPSS GLM Analyze General Linear Univariate The results Tests of Between-Subjects Effects Type III Sum Source of Squares df Square F Sig. Corrected 7.996 a 5 4.599 6.807.000 Intercept 479.938 479.938 703.738.000 58.675 58.675 67.547.000 IT 49.68 49.68 57..000 3.383 3.383 36.9.000 9.930 4.965 5.76.006 Error 36.483 4.869 685.000 48 Corrected 09.479 47 a. R Squared =.667 (Adjusted R Squared =.67) The only difference between the AOVA table and the regression weights (below and from the regression analysis) is that the AOVA table includes a single F-test for the IV effect instead of a t-test for each dummy code of that effect. Parameter Intercept IT [=.00] [=.00] [=3.00] Parameter Estimates 95% Confidence Interval B Std. Error t Sig. Lower Bound Upper Bound 5.90.33 5.300.000 5.43 6.37.7.087 8.9.000.537.886 -.04.35-7.558.000 -.97 -.750 -.757.6-6.0.000 -.0 -.503 -.847.336 -.50.06 -.55 -.69.58.337.766.448 -.4.938 0 a..... a. This parameter is set to zero because it is redundant. 95% Confidence Interval Std. Error Lower Bound Upper Bound 5.079 a.38 4.598 5.559 6.83 a.38 5.70 6.665 5.95 a.35 5.45 6.400 a. Covariates appearing in the model are evaluated at the following values: =.0000, IT = -.083, =.500. The slight differences between the corrected group means from the regression weights and this table are simply rounding differences during the various calculations.