Volume 4, Issue 1 -

Public policies promoting ‘screening’ for a cancer, while directly bearing on the cancer’s early diagnosis and treatment, are really aimed at reducing mortality from the cancer in the populations in question. In the adoption, or continuation, of such a policy, the policy-makers are informed by an estimate of the magnitude of the mortality reduction resulting from the policy. Even though this estimate has to do with mortality in a strictly epidemiological – community-medicine – meaning of the word, the relevant scientific input to it is about mortality in a very different, expressly clinical meaning of it. But a third, seriously malformed concept of mortality from the cancer is the focus in the type of research that has been, and still is, commonly viewed as the only relevant, and also solely sufficient, source of the requisite empirical inputs into the mortality-reduction estimates used in those policy decisions. I here outline the essence of the fundamental misguidedness of this still-orthodox original research – and of its correspondingly misguided reviews – for scientific inputs into policy decisions, with special reference to breast cancer. I also sketch the type of research that really is needed for the estimations at issue – and comment on the larger implications of the prevailing orthodoxy in both original and derivative research for the advancement the knowledge-base of those eminently important public-health policies.

From the writings in medical journals and newspapers, for many decades already, it has seemed uncertain whether programs of mammographic screening reduce mortality from breast cancer. Reviewers of the evidence, in the Cochrane Collaboration in particular, have repeatedly concluded that there is no clear evidence supporting such screening. But their results and conclusions have not impressed the advocates who continue to believe that it does reduce mortality from the cancer.

Our commentary of the article “‘Screening’ for Breast Cancer: Misguided Research Misinforming Public Policies” has two main parts. First we address some of the methodological points raised by Professor Miettinen. Then we review more specific aspects of the Swiss Medical Board statement on mammography screening for early detection of breast cancer.

I’d like to highlight some of the points about research on mammography screening that Dr. Miettinen and I agree on. But I disagree with the main point he tries to make.

Randomized controlled trials have been the mainstay of high-level evidence for or against cancer screening strategies. However, simple proof of reduction in cause-specific mortality is not enough for policymaking, particularly in the last 10 years. Today’s clinical and public health guidelines take into account the balance of risks to benefits from screening, costs, utilities, the political risk of inaction, the societal tolerance to risk, healthcare providers’ preferences, client choices, and other imponderables or subjective variables that cannot be captured or addressed via epidemiologic studies. This commentary uses the above perspective in discussing Miettinen’s arguments concerning the science of breast cancer screening.

Professor Miettinen offers a scathing critique of the criteria used by official bodies to decide for whom and how often breast cancer screening should be offered (Miettinen, 2015). He notes that these bodies often simply synthesize the results of prior randomized clinical trials with very little attention given to the question of whether individual trials constitute an adequate data source on which to make decisions concerning screening policy. He observes that most of these randomized trials suffer from many deficiencies. These include non-compliance with the assigned protocol and cross-over, problems with the randomization itself, the outmoded technologies used in older trials, inadequate length of follow-up, and restriction, by design, to just a few episodes of screening.

Individual-level models (ILMs) are a class of complex, statistical models that are often fitted within a Bayesian framework, and which can be suitable for modeling infectious disease spread. The deviance information criterion (DIC) is a model comparison tool that is appropriate for complex, Bayesian models, and since its development a number of variants have been proposed, including those for its application to missing data models. Here, we assessed five variants of the DIC and their application to ILMs, in particular a class of infectious disease models known as latent conditional LC-ILMs, which depend on a potentially unknown latent grouping variable for each individual in the population. The effectiveness of the traditionally defined DIC was compared to alternative DIC definitions through a simulation study, to assess which is most applicable for this class of models. Epidemic data was generated under an LC-ILM, to which both a spatial ILM (SILM) and the LC-ILM were fitted. Each variant of the DIC was then calculated for every fitted model, and the DIC values obtained for the LC-ILM were compared to those from the SILM. The results of the simulation study indicate that the DIC can be effective for model comparison within complex Bayesian models; however, the degree to which it is effective is dependent upon the variant of the DIC used and the amount of available information on the latent grouping variable.

A common goal of epidemiologic research is to study the association between a certain exposure and a certain outcome, while controlling for important covariates. This is often done by fitting a restricted mean model for the outcome, as in generalized linear models (GLMs) and in generalized estimating equations (GEEs). If the covariates are high-dimensional, then it may be difficult to well specify the model. This is an important concern, since model misspecification may lead to biased estimates. Doubly robust estimation is an estimation technique that offers some protection against model misspecification. It utilizes two models, one for the outcome and one for the exposure, and produces unbiased estimates of the exposure-outcome association if either model is correct, not necessarily both. Despite its obvious appeal, doubly robust estimation is not used on a regular basis in applied epidemiologic research. One reason for this could be the lack of up-to-date software. In this paper we describe a new R package, drgee , which carries out doubly robust estimation in restricted mean models. The package is constructed to be user-friendly and fast, to facilitate routine use of doubly robust estimation. The paper is structured into theory sections and example sections. The former are intended to serve as a brief but self-consistent tutorial in doubly robust estimation. The latter illustrate the use of the drgee package through practical examples. We have used publically available data throughout the paper, so that the reader can easily replicate all examples.

Separating the effects of ages, periods, and cohorts is a classic problem not only in epidemiology but also in demography and the social sciences in general. Frost provides a classic example in epidemiology that I use as an empirical example. In the classic age–period–cohort (APC) approach a single constraint is used to eliminate the linear dependency and identify the model. Among the infinite number of possible choices for a constraint, a particular constraint has been discovered and rediscovered in multiple guises. This constraint produces a solution to the APC model that is perpendicular to the null vector. Among these guises are the minimum norm solution, the Moore–Penrose solution, the principle components solution, the intrinsic estimator, the singular value decomposition solution, the partial least squares solution, and the maximum entropy solution. The results based on the perpendicular constraint provide a solution that has a smaller variance than any other constrained solution. This would be an important advantage, if there were some reason to believe that the perpendicular solutions were an unbiased estimate of the parameters that generated the dependent variable. To help explicate the common features of these solutions, this paper carefully examines the algebraic, geometric, and statistical characteristics of the perpendicular solution.