6. Reliability Findings
6.4. Dependence of Age at First Use Reliability and Duration of Recall
The reliability of age at first use under the exact agreement criterion was mostly
(1977) for categorization of the kappa. When the approximate agreement criterion was used, allowing for up to a 1-year difference, the reliability improved to Landis and Koch's "substantial" range (0.61 to 0.80 kappas) for the age at first use of alcohol, cigarettes, and marijuana and hashish. Respondents' ability to accurately recall their age at first use may decline as more years have passed since their first use. Therefore, logistic regression models were fitted to the data to explore this possibility, where the age at first use responses at the first interview (T1) and at the reinterview (T2) were compared.
Two levels of consistency between these age at first use responses were considered: exact agreement and approximate agreement, allowing for up to a 1-year difference. The dependent variable in the models was defined as 1 if the T1 and T2 age at first use responses agreed (exactly or approximately), and 0 otherwise. The covariate of interest was the number of years since first use. However, additional covariates that may have had an effect on the outcome were included in the models. These covariates were "gender," "race/ethnicity" (four levels),
"education," and a covariate indicating whether both interviews were administered by the "same field interviewer" (FI). The current "age" (using T1 data) was not included in the models because of its high correlation with the number of years since first use, making it difficult to separate the effects of these two factors from each other. Indeed, the correlation between the number of years since first use and the current age was 0.97 for both cigarettes and for alcohol and 0.90 for marijuana and hashish. Because in the case of young individuals recall may be less of an issue, two age groups were modeled: 18 years old or older and 24 years old or older.
In this section, the length of time since first use is referred to as "duration of recall" (DOR) and is defined as the number of years since first use as reported at T1. Thus, DOR = AGE – AFU, where AGE is the current age and AFU is the age at first use, both calculated from T1 data.
6.4.1 General Observations on the Consistency of Age at First Use of Cigarettes, Alcohol, and Marijuana and Hashish and the Years Since First Use
Twelve models were fitted. These models were all combinations of the three substances (cigarettes, alcohol, and marijuana and hashish), two age groups (18 years old or older and 24 years old or older), and for exact agreement and approximate agreement (up to a 1-year difference). For the results, see Tables 6.51 to 6.57.
Eleven of the models showed the number of years since first use having a negative effect on the consistency between the two ages at first use, either exactly or approximately. In the case of approximate agreement for the marijuana and hashish age at first use, there was no effect. Although not all p values were significant, most were, and the uniformity in the direction of the effect of years since first use indicated that this phenomenon indeed exists.
6.4.2 Effect of the Duration of Recall on the Consistency of Age at First Use of Cigarettes, Alcohol, and Marijuana and Hashish and the Years Since First Use
Testing the effects of the duration of recall on exact agreement between the age at first use reported at T1 and that of T2 showed statistical significance. Similarly, when the agreement requirement was relaxed to allow up to a 1-year difference between the T1 and T2 age at first use
responses, the effects of the duration of recall were also significant. To determine the magnitude of these effects, predicted marginals were calculated using the model for the 18 or older age group for each of the three substances either for the outcome of exact agreement or with
approximate agreement (allowing a 1-year difference). This model is referred to as "Model 1" in Table 6.57. Additionally, predicted marginals were calculated using the 12 or older group (i.e., all ages). "Education" was not included in this model because it is an inappropriate covariate in the case of those aged 12 to 17. The two models gave very close predicted margins.
6.4.3 Predicted Marginals and Their Interpretation
To help to appreciate the impact of the duration of recall on the agreement of the age at first use, predicted marginals are presented in Table 6.57. Predicted marginals at x= , where x0
0
x is a certain value of a covariatex, are defined as follows. Suppose every respondent had his or her covariates as in the data, except for the covariate x, which is set at x= . Then the x0 predicted marginal is the mean of the predicted values of these modified observations of the dependent variable. For example, in the hypothetical case where the duration of recall was set to 5 years for all those who had valid age at first use for cigarettes at both T1 and T2, Model 1 predicted 66 percent giving same age at first use at T1 and T2, and 85 percent having a difference of no more than 1 year between the age at first use at T1 and that at T2.