European Network to Develop Policy Relevant Models and Socio-Economic Analyses of Drug Use, Consequences and Interventions Final report: Part 2 – Prevalence of problem drug use at the national level

Abstract

The basic objective of the project work was to explore and to develop the multivariate indicator method. The method introduced by Person, Retka and Woodward (1977, 1978) and modified by Mariani (1999) estimates drug use by combining several population-standardized indicators directly corresponding to problematic drug use. With the use of principal component analysis, the complex information of the number of variables is reduced by extracting one single latent variable that is assumed to underlie all drug-related indicators, and that explains as much as possible of the variance of the original indicators. In a second step, the factor is used in a linear regression model with population-standardized prevalence estimates for at least two regions (the so-called anchor points). The linear regression results in population-standardized regional prevalence estimates. These are then used to calculate the national prevalence estimate.

Additionally, some variants of the method have emerged that differ in the way of transforming the indicator values (e.g. taking the logs, ranking, using the original values instead of the population- standardized ones) as well as in the method of reducing the information (principal component analysis, based on correlation matrix, summing up). Some of these variants were applied to existing data sets. Moreover, a cross-validation was conducted with an Austrian data set.

In the following, the results of these analyses are summarized:

1. At least three anchor points should be available, that should be from both sides of the continuum from low prevalence regions to high prevalence regions. The more anchor points are available, the more stable the method becomes towards other variations (such as choice of indicators, data weaknesses). Implication: Small scale studies are needed to provide a variety of independently obtained estimates. These studies should not be limited to areas with great drug problems, but also to areas with an assumed low prevalence.

2. The choice of indicators influences the model as well, however, this concerns mainly the rank of the regional prevalence estimates. Implication: Data collection should be organised na- tionally providing data collection and coding procedures that are comparable between the administrative regions. The choice of the drug-related indicators utilised for the study, how- ever, is not yet final.

3. The method is relatively robust towards systematic biases of the indicators, e.g. the use of event-based data instead of person-based data in some or all regions, the inclusion of previ- ous drug users or report not by area of residence. Implication: The method can be applied in spite of systematic biases.

4. The choice of the set of indicators should be theoretically based. Drug-related indicators representing consequences of problem drug use as e.g. treatment admissions or number of offences, cannot be easily replaced by social indicators. Aspects, such as face validity and basic assumptions, such as a monotonous relationship between drug prevalence and indica- tors should not be violated. Implications: Data on consequences of problem drug use should be made more easily available. If more indicators should be utilised, there should be empirical evidence that the indicator is drug-related.

5. Different variants of the method may result in a wide range of estimates. Implication:Different variants should be applied. In the case of rather different estimates it should be tried to find an explanation for the differences. At present, no recommendation for a certain variant can be given. The properties of the variants need further exploration.

6. As indicators are often not broken down by age group the choice of the age group is rather arbitrary. The choice of different age groups results in nearly the same regional and national prevalence estimates. Implication: To get prevalence estimates for the age groups recom- mended by EMCDDA a breakdown of the indicators and the anchor point estimates by age group is needed.

7. Overall the method seems to be appropriate for national prevalence estimation, but not for regional prevalence estimation. The choices of different sets of anchor points or indicators seem to effect more the regional prevalence rates than the national ones. In the sensitivity analysis and the cross-validation with capture-recapture estimates it turned out that changes of anchor points or indicators lead to high variations of the regional estimates – even if the national estimates are close to each other. Implication: Do not rely upon regional prevalence estimates obtained by the multivariate indicator method – especially if the regions are no anchor points.

Conclusions

From the effects above can be concluded, that the method works. The choice of the anchor point is crucial for the method but also the indicators should be selected carefully. The method is rather robust towards data flaws of the indicators, but it seems to be important that the indicators are consequences of problem drug use. However, there are still some properties of the method that could not be studied with the available data sets, such as the effects of anchor points estimates derived by different estimation methods and with different target groups or the effect of drug- related indicators not matching exactly to the target group of the anchor point estimates. It seems nearly impossible to analyse the latter problem as in practice no set of indicators will fit exactly to a the same, well-defined target group.

The influence of different methods for the anchor point estimates could, however, be analysed if at least two prevalence estimates derived with different estimations methods and/or different target groups were available for at least one of the anchor points. Even if the target groups are the same one method may be superior to the others, maybe due to obsolete multipliers or coverage errors.

Furthermore, at present we are unable to recommend the application of a certain variant of the multivariate indicator method. To create recommendations it would be to necessary to apply the different variants of the method to many appropriate data sets, to compare the results and to conduct sensitivity analyses. Because of the high correlation between indicators it was impossible to apply the correlation variants to the Austrian data set whereas the German data set is inappropriate since all anchor points are high prevalence regions. Unless enough appropriate data sets were unavailable simulation studies could be conducted. To enable the simulation of realistic situations, profound examination of the distribution properties of commonly used indicators in many empirical data sets is necessary.