use had a very small negative effect (b = - 0.029), which means that students with medium computer use performed a bit lower on the Reading Literacy Test than students who never used a computer. High computer use exerted a large negative effect (b = -0.33): Students who frequently used a computer performed much lower on the PIRLS Reading Literacy Test than students who never did. The random effects of high computer use were significant at the levels of the country and the class; the effects could vary between -0.55 to -0.11 across countries and between - 0.73 and 0.07 across classes. There may be classes in industrialized countries where the effect of high computer use is positive.
With respect to the student’s home, five variables appeared to be important for the prediction of Reading Literacy. The first significant home predictor was the frequency with which the parents undertook literacy activities with their children prior to
elementary school entrance (Preschool Activities, b = 0.032). When the parents undertook such activities more frequently, their children performed better on the PIRLS Reading Literacy Test. The effect of Preschool Activities on Reading Literacy could vary between –0.03 and 0.09 across schools, moreover, which means that there may be schools where the effect of preschool activities is negative.
The second significant home predictor was the Number of Books in the Home (b = 0.14). Those students with more books in the home performed better on the PIRLS Reading Literacy Test. The random effects of this variable were not significant at any level.
The third significant home predictor of Reading Literacy was Parental Attitudes towards Reading (b = 0.074). Students of parents with a more positive attitude towards reading performed better on the PIRLS Reading Literacy Test. The effect could vary between -0.06 and 0.19 across schools, which means that there may be schools where students who have parents with a negative attitude towards reading actually show higher Reading Literacy skills. The covariance between the effect of parental attitudes towards reading (A) and the intercept at the level of the school (B) was significant (cov(A,B) = -0.010). The negative covariance shows that if the mean reading literacy level for one school is higher than the mean reading literacy level for a second school, the effect of parental attitudes towards reading will be smaller for the second school.
The final two home predictors related negatively to Reading Literacy. The frequency of Parental Reading for Information purposes related negatively to Reading Literacy (b = -0.046) with the students of parents who read more frequently for information performing lower on the PIRLS Reading Literacy Test. The frequency of Parental Visitation of the Library or Bookstore with their children was also negatively related to Reading Literacy (b = -0.02) with the students of parents who did this more frequently performing lower on the PIRLS Reading Literacy Test. The random effects of these variables were not significant at any level.
At the level of the class, only one factor was entered into the model. Class Size was found to positively relate to Reading Literacy (b = 0.07). Students in larger classes performed better on the PIRLS Reading Literacy Test. The random effects of class size were significant at the level of the country, moreover. The effect of class size may thus vary between –0.08 and 0.23 across countries, which means that there may be countries where students in smaller classes perform better than students in larger classes.
At the level of the school, two predictor variables were entered into the model and found to relate significantly to Reading Literacy. The first significant school predictor was the Percentage of students in the school coming from Economically
Disadvantaged Homes (b = -0.09). Students in schools with fewer students coming from economically disadvantaged homes performed better on the PIRLS Reading Literacy Test. The second significant school predictor was the Percentage of first through fourth grade students in the school with Another Home Language than the test language (b = -0.03). Students in schools with fewer students speaking another language at home performed better on the PIRLS Reading Literacy Test.
Comparison of Models 1 and 2
In order to compare the variances for Model 1 (i.e., the model including the control variables) with the variances for Model 2 (i.e., the model including both the control and predictor variables), the results of the analyses allowing random intercepts but no random effects for the predictor variables are presented in the right side of Table 5.5 (Model 1A) and Table 5.6 (Model 2A). In Model 1a, the random intercept at the level of the country was 0.037; in Model 2a, the random intercept was 0.038.
Comparison of these coefficients leads to the conclusion that the variance in Reading Literacy at the level of the country is not explained by the predictor variables. This was also expected because no country variables were included in the models. At the level of the school, almost half (47.5%) of the variance was explained by the predictor variables in both models (
V
2 0.12 in Model 1a andV
2 0.063 in Model 2a). At the level of the class, virtually none of the variance was explained (V
2 0.036 in Model 1 andV
2 0.034 in Model 2). At the level of the student, 21.5% of the variance in Reading Literacy was explained by the predictor variables (V
2 0.62 in Model 1 andV
2 0.49 in Model 2). The differences in the fixed effects for the two models were negligible. Given the above information and given that only a small part of the total variance in Reading Literacy occurs at the level of the country, it can be concluded that the differences between countries can best be interpreted in terms of differences between schools and students. The difference in the fit provided by the two models was significant (p < .001), which shows the twelve predictor variables to significantly contribute to the explanation of the variance in Reading Literacy.12
131Influence of the Missing Replacement Procedure on Results
The procedure followed for the replacement of missing values was repeated four times in order to assess the influence of such replacement. Appendix C shows the results for the original model and three variants with replacement values. Only the fixed effect of Parental Visitation of Library or Bookstore with Children and the random effects of Parental Education were found to differ across the four versions of the model. Given that the effects for all of the other variables were equal across the four versions, there is no reason to believe that the replacement procedure for missing values influenced the results of the multilevel analyses.
Model 3: Country-Level Variable
The school-level variable of Percentage of Students in School coming from
Economically Disadvantaged Homes was aggregated onto the country-level variable of Percentage of children in the Country coming from Economically Disadvantaged Homes. Appendix D shows the complete results for Model 3 with 12 predictor variables and the country-level-variable. When the country-level variable was added to the model, the fixed effects of the other variables remained unchanged. The random intercept at the level of the country in the model with the fixed and random effects of the intercept only (i.e., Model 3a) was found to decrease from 0.038 to 0.020, which shows almost 50% of the variance at the level of the country to be explained by the addition of this country-level variable while the percentage of the total variance explained increased from 25% to 32%.
Conclusions and Discussion
The results of the present study clearly show the variance in Reading Literacy across industrialized countries to occur mainly at the level of the student. The school also accounts for a significant amount of the variance in Reading Literacy. The class and country accounted for virtually none of the variance in Reading Literacy, however. Although a complicated procedure for the selection of variables, replacement of missing data, and adjustment of the PIRLS sample weights was necessary to conduct the present multilevel analyses, it was possible to identify twelve predictor variables that explained 25% of the variance in reading literacy. These variables explained one fifth of the variance in reading literacy at the student level and almost half of the variance at the school level. Most predictions were in line with earlier studies as described in the introduction.
At first, a student’s reading self-concept was found to play the most important role in the explanation of the variance in reading literacy scores. Next, students who read for fun or watched television outside of school more often had higher reading achievement. Even though time spent watching television may reduce the amount of time available to read books, television watching appears to improve children’s reasoning and text interpretation. The association of computer use with reading literacy was more complicated: The reading literacy of students who sometimes used a computer was best followed by students who never used a computer with the
performance of students who frequently or very frequently used a computer being lowest. It is certainly possible that students who spend too much time on the computer perform lower simply because they do not have time to read. Interestingly, reading motivation did not contribute to the explanation of the variance in reading literacy in our model.
Within the homes of the students, the number of books in the home was found to influence reading literacy the most. The effect cannot be attributed to its relation to parental education, moreover, as parental education was controlled for. Further, parents with a positive attitude towards reading had children with better reading literacy skills, and the influence of this variable was higher in schools with lower reading literacy performances. Contrary to what we expected (e.g., Baker et al 1997; Wigfield & Asher, 2002), parents who read more frequently for information had children with lower reading achievement. It is possible that students who see their parents read for information are less stimulated to read themselves as their parents are then less interactive and may exude less pleasure in reading than parents who do not read frequently for information purposes. It is also possible that parents who read frequently for information purposes at home do not want to spend more time reading with their children, but prefer to undertake other non-literary activities with their children. Furthermore, parents who undertook such literacy activities with their children as reading books, telling stories, singing songs, playing with alphabet toys, word games, writing letters/words, and reading signs and labels aloud before their children entered elementary school were found to have children with better reading literacy skills in fourth grade. Nevertheless, fourth graders who visited the library or a bookstore with their parents on a regular basis performed lower on the Reading Literacy Test.
Within the schools of the student, class size was found to play an important role: Students in larger classes performed better on the Reading Literacy Test. One possible explanation may be that larger classes generally occur in larger schools with many facilities like remedial teachers and extra reading programs. Another
explanation may lay in the fact that small classes generally occur in rural schools with less financial possibilities. In addition, the economic background and the home language of school population were found to play an important role. Students in schools with many students coming from economically disadvantaged homes or speaking another language at home performed lower. Given that parental education and home language were controlled for, this effect cannot be attributed to the individual backgrounds of the students. The number of instruction hours per school year did not significantly contribute to the explanation of the variance in reading literacy. This finding suggests that the amount of instruction may be less important than the content of the instruction.
Knowing that the most important part of the variance in reading literacy occurs at the level of the student and the level of the school, the question of whether it makes sense to seek predictors at the level of the country arises. In order to shed more light