The Problem of Auto-Correlation in Parasitology

Summary

Explaining the contribution of host and pathogen factors in driving infection dynamics is a major ambition in parasitology. There is increasing recognition that analyses based on single summary measures of an infection (e.g., peak parasitaemia) do not adequately capture infection dynamics and so, the appropriate use of statistical techniques to analyse dynamics is necessary to understand infections and, ultimately, control parasites. However, the complexities of within-host environments mean that tracking and analysing pathogen dynamics within infections and among hosts poses considerable statistical challenges. Simple statistical models make assumptions that will rarely be satisfied in data collected on host and parasite parameters. In particular, model residuals (unexplained variance in the data) should not be correlated in time or space. Here we demonstrate how failure to account for such correlations can result in incorrect biological inference from statistical analysis. We then show how mixed effects models can be used as a powerful tool to analyse such repeated measures data in the hope that this will encourage better statistical practices in parasitology.

Mixed Effects Models as Important Tools for Parasitologists

Parasitologists aim to understand the factors that determine the outcome of infections (e.g., host and pathogen genetic effects), and how these factors change in response to a new intervention or other environmental variables. However, infections are complex and dynamic: multiple interacting factors shape parasite traits, and within-host environments vary over time and between different niches [1]–[5]. Studies examining infections as a snapshot in time are consequently likely to miss a large degree of complexity and subtle—yet relevant—variation in patterns, and risk missing confounding effects or reporting misleading results. For example, in genetically mixed infections of human malaria parasites, different strains circulate in the blood at different times during the infection [2], and in the lungs of patients with cystic fibrosis, bacterial species compete, which results in varying relative frequencies over time [1]; therefore, single time points are unlikely to be representative of the infection as a whole. While in some cases this problem could be overcome using traditional statistical tests with summary statistics (e.g., the slope of the relationship between parasite density and time), relationships will often not be linear and will co-vary with multiple variables. These dynamical differences have important consequences for disease severity and, as within-host dynamics also determine transmission, are essential to understanding epidemiology [6]. Tracking and analysing pathogen dynamics across and between infections is consequently an important goal, but the statistical complexities of dealing with such data offer many traps for the unwary.

Recent years have seen the increasing use of statistical tools, such as mixed effects models, which allow researchers to analyse pathogen dynamics within infections while controlling for issues of pseudo-replication arising from repeated measurements on the same host (i.e., time-series data [7]). In addition to enabling statistically rigorous tests of biological hypotheses relating to infection dynamics, these approaches cut the number of animals needed in experiments, reducing financial costs and ethical concerns. Mixed effects models work by fitting fixed effects, random effects, and error terms into the model. Fixed effect terms explain the variation in a response variable (e.g., parasite density) that is due to the treatment or predictor variable of interest (e.g., drug versus no drug, competitor versus no competitor). Random effect terms, which are specific to a particular group of observations (e.g., all the measurements made on one individual or group of individuals), describe the constant deviation from the mean of that individual or group. Finally, error terms describe the variation (the residuals) remaining in the data that is not explained by either the fixed or the random terms [8]–[10]. The use of random effect terms allows researchers to make use of all of the available data points, whilst removing many of the statistical problems associated with repeatedly measuring the same individual. See [7], [11], [12] for discussion of additional benefits of this approach.

The Problem of Auto-Correlation

At the heart of most statistical tests is the “independence assumption”, which states that model residuals should not be correlated in time or space. Studies where individual subjects are measured on multiple occasions (repeated measures studies) contain potential sources of non-independence, which are not present when individuals are only measured once. In particular for time-series data, unmeasured factors can produce correlations in the data (temporal auto-correlation) over days, weeks, or months (i.e., data points adjacent to one another in a time series are more likely to be similar than those further apart). These correlations can be strong and, importantly, may lead to the appearance of spurious patterns in the data (Figures 1 and 2). Failure to account for this temporal non-independence when carrying out an analysis will incorrectly inflate test statistics and can dramatically increase the likelihood of false positives (type 1 errors [13]), where a significant difference between treatment groups is wrongly concluded. Indeed, simulations based on levels of auto-correlation found in real data show that failing to account for it in analyses has the potential to double or even triple the number of false positives (Figure 2).

There are straightforward solutions to the problems caused by temporal auto-correlation that are routinely used in other biological disciplines, but remain rarely implemented in parasitology. While previous reviews have highlighted the need to track infection dynamics (e.g., [2]) and encouraged the use of statistical models as a valuable tool for parasitologists [7], the importance of meeting the assumptions of these tests has been largely ignored. In a literature search of papers published in the last three years we found that, of 76 papers using mixed effects models to analyse infection data in seven high impact parasitology journals, only 25% of publications explicitly checked and/or controlled for temporal auto-correlation (Figure 3). This indicates a worrying trend and potentially a major problem with the validity of reported findings.

Controlling for Auto-Correlation

There are various ways in which violations of the auto-correlation assumption can be dealt with. One approach is to make tests more conservative by reducing the p-value for which a significant difference is accepted [14], [15]. However, a more elegant (and less arbitrary) alternative is to fit error structures that account for auto-correlation as additional terms in mixed effects models. These error structures describe, and therefore control for, the correlations between residuals at different time points [9]. The simplest form is an auto-regressive error structure, which assumes that the correlation decreases with lag time, e.g., measurements on day 4 of an infection are more similar to those on day 5 than to those on day 20 [9], [13]. Auto-regressive error structures are likely to be common in time-course data [13], are straightforward to understand and implement (see Box 1), and can restore confidence in statistical inference (Figure 2). However, as with all analyses, it is important to consider if this error structure is appropriate for one's own data and whether more complex error structures could potentially provide a better fit (see [9] for examples).

Advances in statistical methodology should provide important and useful tools for understanding infections and disease in just the same way as do advances in genetic, molecular, and immunological methods. Investing in learning how to effectively use tools, such as mixed effects models, pays by providing robust and novel insight into the roles of hosts and parasites in shaping patterns of disease. However, as with other methodological advances, the improvements to biological understanding they provide depend crucially on them being applied and interpreted correctly. Temporal correlation in time-course data can compromise statistical analyses by increasing the likelihood of false positives, yet this problem has been largely overlooked in parasitology. We strongly support the implementation of more sophisticated statistical analyses in which the assumptions underlying models are fulfilled to safeguard against inaccurate or misleading results and provide a solid foundation from which to progress understanding of disease.