Cognitive Ability Measures
To reduce the 21 measures of component abilities in the CREATE battery for subsequent regression analysis, seven factors were initially hypothesized: Perceptual Speed, Fluid Intelligence, Spatial Visualization, Working Memory, Episodic Memory, Crystallized Intelligence, and Psychomotor Speed. The Digit Span Test was not included in the analysis because of problems in group test administration at two of the sites. These factors are broadly consistent with taxonomies of human abilities promoted by Horn (1989)
and Carroll (1993)
(see McGrew, 1997
An initial confirmatory factor analysis, with each variable loading only on its hypothesized single factor, was performed with a structural equation model, estimated with AMOS 5.0 (Arbuckle, 2003
), that included the 1,110 people with complete data on the relevant cognitive measures. Because the four measures that were based on reaction times (simple reaction time, choice reaction time, Stroop Color–Word Test, and the difference between Trails B and Trails A) generally showed skewed (right-tailed) distributions, with several individuals having very long responses, we applied the natural log transformation to these variables. A freely estimated covariance between the residuals of California Verbal Learning Test–Immediate and California Verbal Learning Test–Delayed was specified because they use the same items for immediate and delayed memory tests. Likewise, a freely estimated covariance between residuals of digit symbol substitution, and digit symbol recall was specified given the likely relation between memorizing the pairings and speed of standard digit symbol substitution performance. Age was included in the model as a separate latent variable so that we could estimate correlations of age with each of the factors.
The initial model did not fit the data well, χ2
= 161) = 2160.45, root-mean-square error of approximation (RMSEA) = .106, comparative factor index (CFI) = .866. Inspection of fit diagnostics suggested two possible sources of misfit: (a) Age had a different relationship to individual measures within a factor than was allowed by our approach, and (b) variables loaded on factors other than those specified. To address the first issue, Shipley Vocabulary, Digit Symbol Recall, Wechsler Adult Intelligence Scale–III (WAIS–III) Information, and Stroop Color–Word were regressed on age. To accommodate possible misspecification, Reading Comprehension was allowed to load on Memory, and Inferences was allowed to load on Crystallized Intelligence. The latter two relationships are consistent with what is known about these specific constructs and how they are measured. Reading Comprehension is tested with memory for text passages, which is known to be more broadly related to episodic memory (e.g., Hultsch, Hertzog, Dixon, & Small, 1998
). The inferences test requires that inferences be made regarding propositions about the world and is probably, for that reason, influenced both by reasoning abilities (a primary marker for fluid intelligence) and by world knowledge (a primary marker for crystallized intelligence; see Horn, 1989
, for further discussion of why reasoning tests may load on knowledge factors).
Results from this model showed extremely high factor correlations, leading us to an alternative model in which we reduced the number of factors. In particular, the Working Memory and Spatial Visualization factors were absorbed into the Fluid Intelligence factor. To preserve fit to the correlations of variables that measured the narrower abilities of working memory and spatial visualization, we modelled the Computation Span/Alphabet Span and the Cube Comparison/Paper Folding relations as residual covariances. In addition, because the associations of age with Paper Folding (main indicator for Spatial Ability) and Letter Sets (main indicator for Fluid Intelligence) differed, we included a regression of Paper Folding on age. This five-factor model fit the data well, χ2(N = 163) = 714.124, RMSEA = .055, CFI = .963. reports the standardized factor loadings and factor correlations for the final model. Although age was included in the model, it is not depicted in for purposes of clarity. The correlations of age with the five factors were −.12 with crystallized intelligence, −.06 with fluid intelligence, −.54 with memory, −.48 with psychomotor speed, and −.71 with perceptual speed.
Standardized factor loadings and factor correlations for the final five-factor ability model. WAIS–III = Wechsler Adult Intelligence Scale–III.
We computed composite ability variables on the basis of the factor analysis results, with unit weighting of standardized variables that loaded uniquely on each factor. For instance, the crystallized intelligence composite was formed by summing z scores for Shipley Vocabulary, Multidimensional Aptitude Battery (MAB) Information, and WAIS–III Information. Reading Comprehension was excluded from the composite because of its loading on perceptual speed. Because Inferences loaded somewhat weakly on both crystallized intelligence and fluid intelligence, it was not included in either of those composite variables.
One-way analyses of variance (ANOVAs) were used to examine age differences on these composite variables. Post hoc comparisons were performed with Scheffé’s test (α = .05). Age-group differences were found for crystallized intelligence, F(2, 1132) = 19.42, p < .001, η2 = .03; fluid intelligence, F(2, 1132) = 301.68, p < .001, η2 = .35; memory, F(2, 1132) = 203.23, p < .001, η2 = .27; psychomotor speed, F(2, 1132) = 131.22, p < .001, η2 = .19; and perceptual speed, F(2, 1132) = 317.46, p < .001, η2 = .36. The younger adults performed better than did the middle-aged adults, who performed better than did the older adults, on the perceptual speed, memory, fluid intelligence, and psychomotor speed factors. For the crystallized intelligence factor, older adults outperformed both younger adults and middle-aged adults, who did not differ from each other (see ). (The means and standard deviations on the individual ability tests as a function of age are presented in Appendix B.)
Age-Group Differences in Ability Factor Scores
Attitudes Toward Computers
All participants completed the Attitudes Toward Computers Questionnaire (ATCQ; Jay & Willis, 1992
), a 35-item multidimensional scale assessing seven dimensions of attitudes toward computers: comfort (feelings of comfort with computers and their use), efficacy (feelings of competence with computers), gender equality (the belief that computers are important to both men and women), control (the belief that people control computers), interest (the extent to which one is interested in learning about and using computers), dehumanization (the belief that computers are dehumanizing), and utility (the belief that computers are useful). Each dimension is assessed with five or six items and scored on a 5-point Likert-type scale with anchors of strongly agree
to strongly disagree
. After inspection of the items, we decided to drop the gender/equality scale from further analysis because of the possible dated item content. Participants also completed a 10-item Computer Anxiety Scale (Loyd & Gressard, 1984
) that assessed feelings of comfort/ease with computers. Participants were required to indicate the degree to which they agreed with the 10 statements (e.g., “Computers make me feel uncomfortable”) on a 4-point Likert-type scale with anchors of strongly agree
to strongly disagree
Given that there was content overlap between the ATCQ scales and the Computer Anxiety Scale, we computed correlations among the six remaining ATCQ scales and the Computer Anxiety Scale. Correlations were in the expected direction and were generally moderate to large in magnitude. Hence, we conducted an exploratory factor analysis on the seven scales to determine whether composite scales should be constructed. An unweighted least squares factor analysis revealed two initial factors with eigenvalues greater than 1.0, accounting for 66% of the variance of the data. After promax rotation, the correlation between the two factors was .57. The first factor consisted of the comfort, anxiety, and self-efficacy scales, with factor pattern weights of .92, −.87, and .64, respectively. Clearly, the factor was primarily marked by the two affect-related scales, but the association of self-efficacy and anxiety constructs was also expected from the self-efficacy literature (Bandura, 1997
). The second factor was defined by dehumanization, utility, and control, with loadings of −.62, .90, and .41, respectively. The interest scale was split approximately evenly between the two factors. On the basis of these outcomes, we then rescaled the seven scales to standard scores (M
= 0, SD
= 1) with the means and standard deviations of the total sample. We then formed a composite variable of general computer attitudes as the average of the (reversed-scored) dehumanization, utility, and control scales. The comfort and anxiety scales were also averaged (after reverse scoring of the comfort scale) to define an anxiety scale. We used a unit-weighted method to form these composites, as generally, these methods are less affected by differences in sample size and are preferred for oblique factors (Grice, 2001
). Given the importance of computer self-efficacy as a construct and the conceptual distinction of anxiety as a derivative of self-efficacy, we opted to leave the self-efficacy scale as a separate variable. Likewise, interest was left as a separate variable.
Age-group and gender differences in computer attitudes, computer anxiety, computer self-efficacy, interest, and general attitudes were examined with univariate 2 (gender) × 3 (age group) ANOVAs. Interactions were tested with contrasts through the general linear model procedure. The results of these analyses are presented in .
Analysis of Variance for Computer Attitude Scales by Age Group and Gender
Significant Age × Gender interactions were found for the anxiety scale, F(2, 1183) = 3.05, p < .05, η2 = .01, and the general attitude scale, F(2, 1183) = 3.67, p < .05, η2 = .01. The older women reported more anxiety and less positive general attitudes about computers relative to older men. Neither the difference between the middle-aged women and the middle-aged men nor the difference between the younger women and the younger men was significant for these variables. In general, the older adults (both men and women) indicated more computer anxiety and lower computer self-efficacy than did younger and middle-aged adults. The middle-aged adults were also significantly different on these constructs than were the younger adults. The older adults also reported less interest in computers than did the younger and middle-aged adults. Women, overall, also reported higher computer anxiety, lower computer self-efficacy, lower general computer attitudes, and less interest in computers than did men.
Use of Technology and Experience With Computers and the World Wide Web
General use of technology was measured by having the participants indicate on a 17-item list whether they had used common, everyday technology (e.g., cellular phone, automated teller machine, fax machine, microwave oven, videocassette recorder). The list did not include computer equipment. A composite score was obtained by summing the responses (0–17). Participants were also asked whether they had experience with computers. Those who reported having experience with computers then responded to six questions that assessed frequency of use and breadth of computer experience and knowledge. A breadth-of-computer-use variable was computed by summing responses to questions regarding experience with input devices (e.g., keyboard, mouse [0–7]); proficiency with basic computer operations (e.g., insert a disk, save a file [0–6]); and proficiency with computer applications (e.g., computer graphics, e-mail, spreadsheets [0–12]). Given that the number of items for each of the breadth questions varied, z scores were computed for each question and averaged to generate a composite score.
Participants who reported computer experience were asked to respond to five questions regarding their experience with the World Wide Web. The questions pertained to frequency of Web use, training, and activities performed on the Web (e.g., e-mail, games, news information, shopping). A breadth-of-Web-experience variable was computed by summing the number of activities for which participants reported experience (0–33). The activities were also grouped into the following categories: communication, news and weather, information gathering (legal information), community resources, health, travel, leisure/entertainment, and shopping.
The data indicated a significant Age Group × Gender interaction for use of technology, F(2, 1197) = 6.70, p < .001, η2 = .01. As shown in , older women reported using fewer types of technology than did older men. Gender differences for the younger and middle-aged groups were not significant. There were also significant differences in use of technology as a function of age, F(2, 1197) = 206.96, p < .001, η2 = .26. In general, older people reported less use of technology than did middle-aged people, and older people and middle-aged people reported less use of technology than did younger people.
General use of technology according to age group and gender.
The age groups also differed in terms of having computer experience, χ2(2, N = 1204) = 77.50, p < .001. As expected 99% of the younger participants reported experience with computers, as compared with 90% of the middle-aged participants and 84% of the older participants. Overall, there was no difference in experience with computers between men (91%) and women (92%). However, when examining age differences within each gender, younger men (99%) were more likely than middle-aged (84%) and older (86%) men to report having had experience with computers, χ2(2, N = 454) = 26.58, p < .001. With respect to women, both younger women (100%) and middle-aged women (94%) were more likely than older women (82%) to have experience with computers, χ2(2, N = 750) = 61.87, p < .001.
There was a significant Age Group × Gender interaction for breadth of computer experience, F(2, 1087) = 6.63, p < .001, η2 = .01. For younger and older adults, breadth of computer experience was lower among women as compared with men. There was no difference in breadth of computer experience between middle-aged men and women. In general, younger people reported more breadth of computer experience than did middle-aged people, and middle-aged people reported more experience than did older adults, F(2, 1087) = 350.92, p < .001, η2 = .39. Overall, women reported less breadth of computer experience than did men, F(2, 1087) = 55.60, p < .001, η2 = .05.
There was also a significant age-group difference in experience with the Web, χ2(2, N = 1203) = 182.68, p < .001. As expected, experience with the Web was greater among the younger participants (97%) as compared with the middle-aged (75%) and older (61%) participants. The difference in Web experience between the middle-aged and older adults was also significant. These age differences in Web experience were found for both men, χ2(2, N = 454) = 159.03, p < .001, and women, χ2(2, N = 749) = 133.89, p < .001. Younger men (93%) were more likely than middle-aged men (69%) and older men (69%) to report having experience with the Web; however, there was no difference in Web experience between older and middle-aged men. With respect to women, younger women (97%) were more likely than middle-aged (79%) and older women (56%) to have Web experience. The difference in Web experience between middle-aged and older women was also significant. There was no overall difference in Web experience between men (80%) and women (77%).
Similar to the findings regarding breadth of computer experience, there was also a significant difference among the age groups in breadth of Web use, F(2, 938) = 125.78, p < .001, η2 = .21. The younger participants (M = 21.90, SD = 7.47) reported that they used the Web for more activities than the did participants in the middle-aged (M = 16.88, SD = 8.61) and older age (M = 13.63, SD = 7.70) groups, and the middle-aged participants reported that they used the Web for more activities than did the older participants. There was also a difference among men and women in use of the Web, F(2, 938) = 22.32, p < .001, η2 = .02. Women (M = 16.00, SD = 8.42) reported less breadth of experience with the Web than did men (M = 18.16, SD = 8.60).
As shown in , there were also differences among the age groups in Web activities. The two most frequent Web activities for the younger and middle-aged participants were communicating and searching for information. For the older people, the two top activities were communication and entertainment, followed by searching for information and travel-related activities. The older people used the Web much less for shopping and finding information about community resources than did younger and middle-aged people.
Participants were asked to indicate how they learned to use the Web. Overall, the most common methods were learning by trial and error (self-taught; 75%), following directions on the Web (25%), attending a class (19%), and reading books (14%). (Note that the question asked participants to indicate all the methods that they used.) It is interesting that 87% of the younger participants indicated that they learned by trial and error, as compared with 70% of the middle-aged participants and 61% of the older participants, χ2(2, N = 717) = 67.62, p < .001. A greater number of the older adults (27%), as compared with the younger adults(4%), indicated that they learned by reading books, χ2(2, N = 717) = 80.93, p < .001, and attending classes (25% vs. 14%), χ2(2, N = 717) = 18.11, p < .01.
Predictors of Use of Technology, Computers, and the World Wide Web
One of the primary goals of this study was to determine whether and how individual differences in demographic characteristics, attitudes, and abilities predict technology and computer use patterns. The first question we addressed was whether these variables predicted general technology use and experience with computers. The next set of analyses focused only on those participants who reported experience with computers. People may use computers minimally or for a wide range of tasks. Our goal was to determine whether breadth of computer experience was predicted by demographic variables, attitudinal variables, and abilities. Finally, we examined factors that predicted having experience with the World Wide Web. Again, the analysis of breadth of Web use was restricted to those with Web experience.
Multiple regression analyses were performed to examine these relationships. Given the high correlations among all of the ability factors (see ), the two most broadly defined ability factors, fluid and crystallized intelligence, were chosen as ability predictors for these analyses. Also, given the high correlation among all the health variables, we used functional health as the health predictor, given our interest in functional impact and the fact that this measure had the highest correlation with the technology use variables. For these analyses, we restricted the ethnic group variable to include participants from three ethnic groups: White/European Americans, Black/African Americans, and Hispanic/Latino Americans. These ethnic groups represented 91% of the sample, and the remaining 9% consisted mostly of Asian students (41%) and a few individuals from other ethnic groups such as Native Americans (n = 8) and multiracial (n = 23).
For both gender and ethnicity, we used weighted effects coding. For the ethnicity variable, the reference group was White/European Americans. We chose effects coding because we wanted to interpret the first-order effects of gender and ethnicity as average effects; we used weighted coding because the sample sizes were unequal for the levels of the variables (Cohen, Cohen, West, & Aiken, 2003
For each of the outcomes measures, we conducted an initial hierarchical regression analysis. We entered all of the main effects in the first step, all of the two-way interactions in the second step, and the three-way interactions in the third step (Cohen et al., 2003
). The following variables were entered into the model in the first step: functional health, gender, education, ethnicity, fluid and crystallized intelligence, the four computer attitude measures, and age. The following two-way interactions were entered in the second step: Age × Education, Age × Crystallized Intelligence, Age × Fluid Intelligence, Age × Computer Anxiety, Age × Computer Efficacy, Ethnicity × Education, Ethnicity × Computer Anxiety, Ethnicity × Computer Efficacy, and the interactions among each of the ability measures and computer anxiety and computer efficacy. The three-way interactions among age, each of the ability measures, and computer anxiety and computer efficacy were entered in the third step. Interaction terms were retained in the model on the basis of the significance of the F
statistic for increments to R2
. Nonsignificant predictor variables entered in Step 1 were trimmed from the model if there were no significant interaction terms associated with that variable. Separate equations were computed for each of the significant interaction terms, and tests of simple slopes were performed at α = .05 (Aiken & West, 1991
). Given that experience with computers and the World Wide Web were dichotomous variables, logistic regression was used for those variables.
We then used the trimmed models to conduct hierarchical multiple regressions for each of the dependent measures. We were interested in determining (a) excluding age, how much variance could be explained by the remaining exogenous sociodemographic variables (gender, education, ethnicity) and then by basic cognitive abilities (crystallized and fluid intelligence) and attitudinal measures (computer self-efficacy and computer anxiety); (b) how much age-related variance remains after the other independent variables have been entered; and (c) how much variance was explained by the remaining two-way and three-way interactions. Given the large sample size, we considered only interaction effects that accounted for at least 1% of the variance as being meaningful for interpretation.
General use of technology
Results of the hierarchical regression analysis for general use of technology are summarized in . After accounting for the exogenous social/demographic variables, the cognitive variables resulted in a significant increment in R2 of .29. Adding age to the model after accounting for the exogenous social/demographic variables, cognitive ability variables, and computer anxiety and self-efficacy variables resulted in a significant increment in R2 of .06. The interaction terms also resulted in significant changes in R2.
Hierarchical Regressions With R2 and Increment in R2: Use of Technology
As shown in , general use of technology was predicted by education, age, ethnicity, fluid and crystallized intelligence, computer anxiety, and computer self-efficacy. In general, people who were better educated, were younger, had higher levels of crystallized and fluid intelligence and computer self-efficacy, and had lower levels of computer anxiety used more types of technology. Black/African Americans used less types of technology than did White/European Americans and Hispanic/Latino Americans.
Summary of the Final Regression Model for Use of Technology
The two-way interactions accounted for 2% of the variance. As shown in , this was the combined effect of six different interaction terms. The squared semipartial correlations for these terms were examined, and the only interaction that satisfied our criterion of 1% was the interaction between fluid intelligence and computer efficacy. In general, for people with lower fluid intelligence, t1081 = 5.91, p < .01, greater computer self-efficacy was associated with more use of technology. This relationship was not found for people with high fluid intelligence, t1081 = −.88, p < .05 (see ).
Interaction of Fluid × Computer efficacy for use of technology.
With respect to experience with computers, as shown in , having experience with computers was predicted by age, fluid intelligence, ethnicity, computer anxiety, and education. In general, people who were younger, had higher levels of fluid intelligence and education, and had lower levels of anxiety about computers were more likely to have experience with computers. Black/African Americans and Hispanic/Latinos were also less likely than White/European Americans to have experience with computers. This set of predictors accounted for about 34% of the null deviance (Cohen et al., 2003
). The data also indicated that ethnicity, cognitive abilities, and computer anxiety had the strongest effect on experience with computers (see ). After accounting for gender and education, ethnicity resulted in an increment in Nagelkerke R2
of .03, and after accounting for the exogenous social/demographic variables, the cognitive ability variables resulted in a significant increment in Nagelkerke R2
of .19. Finally, adding the computer anxiety variables to the model, after accounting for the exogenous social/demographic variables and the cognitive ability variables, resulted in a significant increment in Nagelkerke R2
Final Hierarchical Logistic Regression Model: Computer Experience
Breadth of computer use
The results of the hierarchical regression analysis () indicate that, after accounting for the exogenous demographic/social variables, cognitive abilities resulted in a significant increment in R2 of .27. The addition of the computer attitude variables to the model accounted for a significant increment in R2 of .144, and the addition of age to the model accounted for a significant increment in R2 of .13.
Hierarchical Regressions With R2 and Increment in R2: Breadth of Computer Use
Breadth of computer use was predicted by gender, education, fluid and crystallized intelligence, computer anxiety, and age (see ). Ethnicity was not a reliable predictor of breadth of computer use. In general, people who were less educated, were older, and had lower fluid and crystallized intelligence had less breadth of computer experience. Women also had less breadth of computer experience than did men. Breadth of computer use was also lower among people with high computer anxiety as compared with those with low computer anxiety. Although there were also significant interactions for this outcome variable, none met the criterion of accounting for at least 1% of the variance.
Summary of the Final Regression Model for Breadth of Computer Use
Use of the World Wide Web
Having experience with the World Wide Web was predicted by education, ethnicity, fluid and crystallized intelligence, computer anxiety, and age (see ). People who were less educated, were older, had lower fluid and crystallized intelligence, and had more anxiety about using computers were less likely to have experience with the World Wide Web. Black/African Americans were also less likely than White/European Americans to have experience with the Web. As shown in , the strongest predictors of having experience with the Web were ethnicity, fluid intelligence, and computer anxiety. After accounting for gender, education, and ethnicity, cognitive abilities resulted in an increment in Nagelkerke R2
of .31, and after controlling for the exogenous social/demographic variables, cognitive ability variables, and computer anxiety and self-efficacy variables, age resulted in a significant increment in Nagelkerke of R2
of .07. The Age × Fluid Intelligence interaction, although significant, resulted in an increment in Nagelkerke R2
of only .01; thus, we constrained the model to include only main effects, given the tradeoff between the improved goodness of fit and the complexity of interpreting interactions in logistic regression. As the inclusion of the interaction resulted in only a small improvement in the model, we determined that the increased complexity associated with interpretation of the interaction was not warranted (Glantz & Slinker, 1990
Final Hierarchical Logistic Regression Model: Web Experience
Breadth of World Wide Web experience
The results of the hierarchical regression analysis () indicate that cognitive abilities resulted in a significant increment in R2 of .17. The addition of the computer attitude variables to the model accounted for a significant increment in R2 of .11, and the addition of age to the model accounted for a significant increment in R2 of .10.
Hierarchical Regressions With R2 and Increment in R2: Breadth of Web Use
Breadth of Web experience was predicted by age, crystallized intelligence, computer anxiety, and computer self-efficacy (). In general, people who were older, had lower crystallized intelligence, had lower computer anxiety, and had higher computer self-efficacy had more breadth of web experience. Gender and ethnicity were not reliable predictors of breadth of Web experience. The two-way interactions, although significant, accounted for less than 1% of the variance. It is interesting that there was a significant three-way interaction between age, fluid intelligence, and computer anxiety. As shown in , computer anxiety had a more debilitating effect for older adults with high fluid intelligence than for older adults with low fluid intelligence. Computer anxiety had less impact on younger adults.
Final Multiple Regression Model for Breadth of Web Use
Interaction of Age × Computer anxiety for high fluid intelligence and low fluid intelligence for breadth of World Wide Web use.
Structural Regression Models
To test our hypothesized model and to determine the extent to which computer attitude variables and cognitive abilities mediate the effects of age and education on breadth of computer experience and breadth of web experience, we used structural equation modeling with AMOS 5.0 (Arbuckle, 2003
). We restricted our analysis to these variables because having experience with computers and having experience with the World Wide Web were categorical variables with two response categories (Byrne, 1998
). General use of technology was included as a potential mediator that, in principle, indirectly reflected prior domain-specific knowledge and experience with technology that is relevant as a predictor of computer use (e.g., Beier & Ackerman, 2005
). Although our previous multiple regression analyses revealed some significant interaction effects, we did not attempt to include interaction effects in the model; instead, we chose to evaluate mediated paths averaged over interaction effects. Bootstrapping techniques were used to test for indirect effects.
We used single indicators in the structural models with the procedure recommended by Liang, Lawrence, Bennett, and Whitelaw (1990)
. This procedure allowed us to use the same composites that were used in the regression analysis and still disattenuate regression coefficients for measurement error. Cronbach’s alpha reliability estimates were calculated for each indicator variable. The factor loadings for each indicator variable were fixed at √α(σ), and their error variances were fixed at (1 − α)σ2
, where α is the estimated reliability and σ2
is the variance of the indicator. This correction was not applied to the use of technology variable because coefficient alpha would not have been an appropriate estimate of reliability for that index. This procedure was used because we did not have multiple scales measuring both computer self-efficacy and computer anxiety and because we wanted to maintain similar correction procedures for both ability and attitude predictors. We checked the results by using multiple indicator models for abilities and found good agreement in estimated structural regression coefficients, justifying the present approach.
Separate structural equation models evaluated the two target criteria of breadth of computer use and breadth of Web use. The initial model was the hypothesized model in which the effects of age and education were mediated by the computer attitude variables and cognitive abilities (see ). We added general use of technology as a mediator to examine whether experience with technology predicted computer use independently of attitude and ability variables. We specified use of technology as an influence on computer self-efficacy and computer anxiety and evaluated whether it had a direct effect on breadth of computer use and breadth of Web use independent of these attitude variables. Our models also specified a specific order of flow of influence for the attitude and ability variables. We specified that computer efficacy influenced computer anxiety and that fluid intelligence influenced crystallized intelligence. The latter direction of influence is justified by Cattell’s investment theory (see Beier & Ackerman, 2005
); the former is justified by Bandura’s (1997)
theoretical claim that anxiety is typically an outcome of low self-efficacy.
We adjusted models in two ways. First, we trimmed nonsignificant regression coefficients ( p > .01), unless retaining the coefficient was needed to ensure statistical control on other predictors. Second, we used modification indices to assess whether additional parameters would improve the fit of the model to the data.
With respect to breadth of computer use, the final model fit was excellent, χ2(N = 11) = 20.44, RMSEA = .028, CFI = .997. The standardized effects for the final models are presented in . The model accounted for 64% of the variance in breadth of computer use, with direct paths from age, computer anxiety, crystallized intelligence, and general use of technology (see ). In addition, age was indirectly linked to breadth of computer use through fluid intelligence, crystallized intelligence, computer efficacy, and computer anxiety ( p < .01). An important feature of the model is that the direct effect of age was larger than the indirect effects of age, suggesting that our study variables did not account for all the pathways by which age differences in computer use are realized. Only 37% of the total effects of age on breadth of computer use were indirect; the rest were direct. On the other hand, it should be noted that the final model contained no direct effects of education on breadth of computer use. We trimmed a small path from education to breadth of computer use that was reliable at the .05 level (but not the .01 level), but the standardized effect of education was only .04. Also, there was no direct effect of fluid intelligence on breadth of computer use. Instead, effects of fluid intelligence were fully mediated by use of technology and crystallized intelligence.
Standardized Effects of Variables on Breadth of Computer Use and Breadth of Web Use
The final structural model for general breadth of computer use, with standardized regression weights.
We evaluated a few alternative models to test the robustness of our solution and the theories that generated it. Some alternatives matter conceptually but are not empirically testable. For example, our assumption that computer efficacy influences computer anxiety, versus the alternative assumption that anxiety influences efficacy, is not empirically testable. The alternative assumptions generate equivalent models fits and hence cannot be evaluated on the basis of differential empirical fit of the models to the data (MacCallum, Wegener, Uchino, & Fabrigar, 1993
). However, we identified two alternative specifications that are not empirically equivalent and hence represent meaningful alternatives to the assumptions specified in our accepted model. In one model, we allowed computer self-efficacy to directly influence breadth of computer use, eliminating the effect of computer anxiety. This model stipulates that self-efficacy is the primary influence on computer use, with anxiety being an inert derivative of self-efficacy. Despite the strong relationship of self-efficacy to anxiety, the locus of the effect was not interchangeable. This model fit the data considerably worse than did our preferred model, even though the overall level of fit was good, χ2
= 11) = 89.80, RMSEA = .081, CFI = .977. Thus, it seemed more plausible to model the effect of computer efficacy on breadth of computer use as being fully mediated by computer anxiety. In the second model, we attempted a model in which use of technology had no direct effect on breadth of PC use, instead being mediated through computer efficacy and computer anxiety. This model also specified a direct effect of fluid intelligence on breadth of computer use (in the accepted model, this had been modeled as an indirect effect of fluid intelligence mediated through use of technology). This alternative model also fit more poorly than did our accepted model, χ2
= 11) = 147.20, RMSEA = .107, CFI = .961. Further, the direct effect of fluid intelligence on breadth of computer use was not reliable (p
> .01). Thus, we concluded that prior use of technology was an important, independent influence on breadth of computer use.
A similar pattern of results was also found for the breadth of Web use variable (see ). The regression model fit extremely well, χ2(N = 11) = 12.80, RMSEA = .013, CFI = .999. The final model accounted for 46% of the variance in breadth of Web use, and 29% of the age effect was mediated by the attitudinal and ability variables (see ). Given that the pattern of indirect and direct effects was similar to the breadth of computer use variable, we do not elaborate further on these outcomes.
The final structural model for general breadth of Web use, with standardized regression weights.
We did test the same two alternative models considered earlier for breadth of computer use. Forcing computer efficacy, instead of computer anxiety, to predict breadth of Web use degraded the fit of the model, χ2(N = 11) = 38.84, RMSEA = .052, CFI = .989. Removing the direct effect of use of technology on breadth of Web use likewise resulted in poorer fit, χ2(N = 11) = 82.66, RMSEA = .083, CFI = .971. We concluded that neither alternative model was preferable to our accepted model.