Invited Commentary: Challenges of using Contact Data to Understand Acute Respiratory Disease Transmission

Wallinga et al. (1) have done an excellent job of demonstrating how simple data can be used to improve our estimation of transmission parameters for infectious disease models. Their analysis involves several steps illustrating a unifying framework, from collecting data on social contacts to model-fitting with relevant infectious disease data. First, there is estimation of the contact matrix from age-specific data on conversations. Second, assumptions are made to estimate the age-specific transmission parameters of a particular transmission model. Third, comparison is made with infectious disease outcome data for goodness of fit and model choice. The problem is very important. Contact patterns are crucial determinants in both the spread of an infectious disease and the decision on which interventions would be most effective. Sometimes a group of investigators pulls together several ideas, in preliminary form, showing the way for future research. Such is the case with this paper by Wallinga et al. I would like to comment on four areas that deserve additional research: 1) data structure, 2) model-dependent transmission parameters, 3) statistical inference, and 4) infectious disease data for model-fitting. In the paper by Wallinga et al. (1), the data on which the contact matrix analysis was based were remarkably simple. A random sample of people in Utrecht, the Netherlands, were asked about the number of conversations they had with people of different age groups during a typical week. These simple data were adequate to estimate an age-structured contact matrix. The simple age-structured matrix was in turn adequate to estimate transmission parameters for an agestructured model. In this type of model, the mixing groups are mutually exclusive; that is, a person can belong to only one age group. The transmission parameters that are estimated in this paper are specific to the type of model being used. However, many current models being used to study the effects of interventions in populations, such as those for pandemic influenza (2–5) and smallpox (6), have more complex population structures. In these models, people can mix in several different places, including households, schools, and workplaces, as well as have age-specific components to the mixing within the different mixing groups. These more complex patterns are required to analyze the effect of household interventions, such as targeted antiviral prophylaxis, school closures, or quarantines. I would like to see more empirical studies like the one presented in this paper with these more complex models in mind (7). A natural extension of the data structure would be to ask people not only about the age groups of people with whom they have had conversations but also where they had the conversations. Such data would allow estimation of transmission parameters for models with more complex mixing structures. Since the interpretation of transmission parameters is model-specific, such data for estimation of transmission parameters would be very important. The third area, statistical inference, also needs further development to take honest account of the uncertainty in the estimates. The current analysis probably underestimates the uncertainty in the mixing matrix, which then carries over to an underestimate of uncertainty in the estimates for the transmission model. First, the bootstrap confidence intervals presented in Wallinga et al.’s table 1 are only for the mean values of the negative binomial distributions that were fitted to the data. There is no mention of k, the shape parameter, so we do not know the full uncertainty of the entire distribution. Second, to estimate the age-specific transmission parameters, Wallinga et al. keep the estimated means fixed at the maximum likelihood value. Heterogeneity in the number of contacts within age groups and the variability of the data