Common biases in observational epidemiological studies

Although the randomized controlled trial (RCT) is the gold standard to measure causality in its unique advantage of random assignment, we must consider alternative research designs that permit relatively strong causal inferences. Therefore, when conducting observational studies researchers need to ensure that the internal validity of the study is not compromised by bias and that the results found are close to the “truth” [14, 27]. Biases common to all observational studies include selection bias and information bias (Table 2, Adapted from: Bonita et al. [28], Examples of low and high risk of bias in observational studies can be found in [26, 29]). Studies that attempt to show a causal association between an exposure and outcome can also be affected by confounding–these are the observational analytical designs, usually cohort and case-control studies.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Common biases in observational epidemiological studies, along with the extent of their occurrence.

https://doi.org/10.1371/journal.pone.0180220.t002

While all designs have their advantages and disadvantages, researchers need to be aware that depending on the research question [28, 30]:

• The probability of selection bias due to inclusion bias is highest in case-control studies.
• The probability of selection bias due to loss to follow-up is highest in cohort studies.
• The probability of measurement bias due to recall bias is highest in cross-sectional and case-control studies.
• Information bias due to instrument error and observer bias is common to all study designs.
• Confounding is an issue that is common to all observational analytical study designs.

Measurement of the exposure of interest–challenges in accurately identifying ZIKV infection

The symptomatology of ZIKV infection includes fever, rash, and conjunctivitis, as well as other minor symptoms commonly associated with influenza, such as body aches and chills [31]. However, an individual may be infected with ZIKV but either have none or some of the aforementioned symptoms [1]. The range of symptoms of the ZIKV infection is so vast that a myriad of definitions have been generated as to what constitutes an infected individual (i.e. number and severity of symptoms present to qualify as a case of ZIKV), thus limiting the external validity of studies if they use variable criteria.

Perhaps the main critical issue for all observational studies conducted since the ZIKV outbreak of 2015 is the lack of a standardized definition of ZIKV infection due to challenges and limitations of available molecular and serologic diagnostic tests for ZIKV [32]. This can lead to risks of both selection bias and information bias due to misclassification of infection. In S1 Fig, this issue is depicted at the center of the diagram (E) at the intersection of the four analytical designs [17–20] described in Table 1, and also applies to the two descriptive observational protocols [21, 22]. In studies where ZIKV infection determines participation, there is a risk of selection bias if individuals do not meet the inclusion criterion for infection and may therefore be systematically excluded from participation. In studies where ZIKV infection does not determine participation, but is still considered an exposure of interest, there is a risk of information bias due to misclassification of exposure.

Serologic assays use certain biomarkers present in blood and other bodily fluids to test for the presence of antibodies suggestive of previous or ongoing ZIKV infection. However, immunology assays (immunofluorescence assay (IMF) and enzyme-linked immunosorbent assay (ELISA)) for ZIKV may result in misclassification errors due to cross-reactivity between ZIKV and dengue virus (DENV) or other flaviviruses co-circulating in the proximity of ZIKV patients and in addition the cross-reactivity but also the actual presence of anti-flavivirus antibodies for two different viruses in the same patient [5, 33]. At the time of writing, the United States Centers for Disease Control (CDC) stated that, “Due to serological cross-reactivity between flaviviruses, current IgM antibody assays cannot reliably distinguish between Zika and dengue virus infections” [5]. The lack of reliable distinction between ZIKV and DENV is a significant source of information bias for all studies of ZIKV where DENV is circulating, or has circulated, and could contribute to confounding of the association between exposure to the ZIKV and disease outcomes (see section 4 of this paper). However, a recent study using a novel anti-ZIKV ELISA based on recombinant ZIKV non-structural protein1 found that cross-reactivity with high-level dengue virus antibodies was not detected [34].

Interim case definitions for suspected, probable, and confirmed cases of zika have been defined by the PAHO/WHO and their use should enable standardization and help limit selection bias and information bias due to misclassification [35]. In the context of research confirmed cases are required. The standardized protocols reference the PAHO/WHO case definitions and include clear a description of which laboratory tests should be used during the course of infection and when, the procedures to be followed, and definitions of cases [36]. These include lab algorithms to rule out other flaviviruses, such as DENV and the collection of samples for all study participants, regardless of symptom status or case status. They also include recommendations for reporting of results stratified by symptom status. Further, all of the protocols include a standardized questionnaire [36] to ensure standardization of collection of information on symptoms and other relevant information. Questionnaires were specifically developed for each of the six zika standardized research protocols based on existing case report forms [16, 36].

While the test with highest specificity and sensitivity to detect ZIKV-infected patients is reverse-transcriptase polymerase chain reaction (RT-PCR), which detects the presence of ZIKV RNA in human body fluid samples [37, 38], the use of this method is limited as ZIKV can only be reliably detected in up to the first 7 days after infection [5, 32]. RT-PCR is expensive and samples must be sent to specialized laboratories for analysis, which are not yet readily accessible in all countries affected by ZIKV. Thus, for epidemiological studies, seropositivity is mostly defined as positive anti-ZIKV IgM antibodies AND plaque reduction neutralization (PRNT90) for ZIKV titers ≥20 and four or more times higher than for other flavivuruses; AND exclusion of other flavivirus [22]. Where the sensitivity and specificity of the serologic test is known, these can be adjusted for to correct for misclassification. Alternatively, sensitivity analysis can be undertaken using a range of plausible estimates for the misclassification. The testing strategy for patients presenting ≥ 7days after onset of symptoms focuses on IgM serology due to the availability of reagents. The WHO Interim Guidance on laboratory testing for Zika virus infection indicates that IgM detection should be performed for pregnant women in areas of endemic transmission or pregnant women who could have had contact with vector-borne or sexually transmitted Zika virus [39].

However, some issues remain and need to be addressed by the research community. These include development of better diagnostic tools; determination of the sensitivity, specificity and predictive value of serum ZIKV IgM; and assurance of reliable, accurate, and standardized testing [12, 40]. Other issues which are generally found with biomarkers are exacerbated by problems specific to the countries and regions where the zika epidemic is currently found. For example, there are limited labs with the capability of testing biological samples for ZIKV infection, which may lead to difficulties shipping samples to the labs in a timely manner and may put the integrity of the samples at risk due to the time of collection, maintaining of the cold chain, and material transfer issues [41]. Until more laboratories have the ability to test for ZIKV infection in biological samples, the possibility of reduced integrity of samples should be considered a potential source of information bias as it can lead to misclassification of exposure status [42].

Measurement of the outcomes of interest—Challenges in defining the spectrum of diseases caused by ZIKV infection: Microcephaly and other congenital birth defects and Guillain-Barré Syndrome (GBS)

Measurement of the outcomes of interest is a risk of bias common to the observational analytical studies on ZIKV infection [17–20]. Research conducted to date has used a variety of case definitions for microcephaly, congenital ZIKV syndrome and GBS [12], which makes the comparison and pooling of results difficult.

For microcephaly–the International Fetal and Newborn Growth (INTERGROWTH) standard is considered the best method to classify a case of microcephaly in a fetus or newborn [43]. The WHO recognizes that there is a large variability in the levels of microcephaly present in different geographic populations, but it is likely that some of the variability is due to differences in measurement and subsequent classification, which would be a potential source of information bias [43]. Several methods can be used to determine head size, such as the ultrasound (during pregnancy), Magnetic Resonance Imaging (MRI) scan, and Computed Tomography (CT) scan, where available (see S1D Fig) [44].

For Guillain Barré Syndrome–WHO has published an interim guidance on the Identification and management of GBS syndrome in the context of ZIKV, which provides recommendations for clinical assessment based on the Brighton Collaboration criteria case definitions for GBS [45, 46]. In the standardized protocol relevant to GBS, the definition of a case is: GBS meeting levels 1–3 of diagnostic certainty for the Brighton Collaboration criteria case definitions for GBS [18, 45, 47].

Standardized and validated methods for measuring outcomes will help to prevent misclassification of outcomes leading to information bias. For the case-control study designs [17, 18], misclassification of outcome can also lead to selection bias, with those not meeting the inclusion criteria for a “case” and thus being systematically excluded from participation. To help overcome this issue clear definitions of cases are given in the ZIKV standardized protocols, though both protocols include a note that the studies will be used to refine and update recommendations for surveillance and case definitions [17, 18].

Confounding and effect modification

In addition to selection and information bias due to challenges with currently available ZIKV diagnostic tests and clear definitions of key outcomes, for the analytical designs researchers need to account for potential confounders [17–20]. In order to prove causation between an exposure (in this case ZIKV infection) and an outcome (e.g., microcephaly in the newborn or GBS in adults) confounding needs to be minimized or, if possible, ruled out.

Confounding can occur when another exposure exists in the study population and is associated with both the disease and the exposure under study. Potential confounders include independent risk factors for the outcomes of interest, though to be confirmed as confounders they also need to be associated with the exposure of interest in the particular study and not be on the causal pathway between the exposure and outcome [28].

Potential confounders need to be considered at both the design and analysis stages of observational studies [28]. At the design stage, approaches that can be taken to prevent confounding include matching and restriction. Randomization is the best approach for preventing confounding but its use is limited to experimental designs. At the analysis stage confounding can be controlled by stratification or multivariate modeling. However, all of these approaches require that all potential confounders are identified and measured. Potential confounders are included as part of the standardized questionnaires for each of the standardized protocols and/or in the laboratory testing regime. In addition, for the case-control designs controls are matched to cases on some key variables.

In the case of microcephaly and other congenital birth defects of the newborn independent risk factors include certain infections during pregnancy, such as rubella, toxoplasmosis, syphilis, varicella-zoster, rubella, cytomegalovirus, and herpes infections (TORCH) in utero (see S1D Fig) [48]. Testing of serum for these potential confounders is recommended in the standardized protocols relevant for microcephaly. Other independent risk factors for microcephaly may include: severe malnutrition; maternal alcohol and tobacco use, maternal sociodemographic characteristics including age, race, education, marital status, body mass index and interruption of the blood supply to the baby’s brain during development and these are included in the standardized questionnaires [49, 50].

Given the possibility of misclassification of DENV and ZIKV, DENV should also be considered as a potential confounder in all cohort and case-control studies of ZIKV [12]. Questions to address exposures to these potential risk factors and others are included in the standardized questionnaires and/or in the laboratory testing regimes for the cohort and case-control studies for microcephaly [17, 19, 20].

For GBS, although the exact cause is unknown, most cases of GBS occur after a i) virus infection–such as the cytomegalovirus (a member of the herpes group), the Epstein-Barr virus or HIV; or ii) bacterial infection–such as infection from Campylobacter bacteria, a common cause of food poisoning [51]. Questions to address exposures to these potential risk factors and others are included in the standardized case-control study for GBS [18].

Despite the attempts to account for all potential confounders in the ZIKV standardized protocols, researchers using these protocols need to ensure that appropriate statistical analyses are undertaken to test and control for these. Some variables may also be modifiers of the effect and this need to be considered at the analysis stage of the study. Both users of the protocol and of the research should also be aware of unknown confounders.

Risks of bias common to cohort designs of ZIKV infection

The risk of selection bias due to loss to follow-up is higher in cohort studies than in other observational designs (Table 2). This is particularly important if there is differential loss to follow-up due to the exposure. In the case of ZIKV infection, cohort studies are ideal for investigating multiple outcomes from a single exposure [52], which is applicable when the primary exposure of interest is ZIKV infection and there is a wide range of outcomes of interest [19, 20]. If performed in a conscientious and rigorous manner, cohort studies can be used to determine causality, given that exposure and outcome are investigated in a temporal framework [53], i.e. the exposure (ZIKV) clearly comes before the outcome (e.g. microcephaly or GBS).

The two main sources of selection bias in cohort studies arise during the recruitment and the follow-up phases (see S1G Fig). In the recruitment phase, the considerations regarding classification of ZIKV infection are similar to those for all observational designs as described in section 3. In addition, care is needed to ensure that the exposed and unexposed individuals are recruited from the same source population so that they are similar in all regards except for the particular exposure of interest [14, 52, 54]. If the exposed and unexposed individuals differ in extraneous factors (potential confounders), then internal validity is at risk and any differences observed in the outcome of interest may not be attributable to a causal relationship [14, 55], unless these factors can be controlled for in the analysis.

In the follow-up phase, time is crucial to minimize losses to follow-up. In the case of ZIKV infection, follow-up is important for analyzing causal considerations such as temporality, dose-response, and consistency [55]. The follow-up period must be long enough for a sufficient number of outcome events to occur or be observed, increasing the statistical power to detect differences between the exposed and unexposed individuals [56]. Depending on the type of outcomes sought, minimum length to follow-up may vary. However, even if intended follow-up times are appropriate for the outcomes of interest, loss to follow-up still represents a significant risk of selection bias and a threat to validity if there is differential loss between the exposed and unexposed cohorts. That is, if loss is classified as “missing at random”, then theoretically, there is no risk for selection bias, but if the loss is classified as “missing not at random”, then one must make the conservative assumption that the loss is related to exposure status, which then introduces selection bias [57].

Risks of bias specific to case-control designs

In addition to challenges in accurate diagnosis of ZIKV infection (section 2), the main source of bias common to all case-control studies of ZIKV infection occurs at the point of selection of cases and controls to be included in the study. Both cases and controls should represent the same base set or same source population [62]. That is, cases from one population should not be compared with controls arising from a different population. However, the cases and controls do not necessarily need to be representative of the total underlying source population, but to have a similar baseline level of risk [62].

The selection of controls is a particular challenge, and there are a few different ways to recruit them. Controls can be recruited from the general population, from amongst friends and family of the cases, or from the same hospital where cases were recruited [63]. However, care must be exercised when selecting controls from hospitals since they may have comorbid conditions or exposures that make them systematically different from the cases, increasing the risk of confounding [63–65].

The selection of controls provides an important opportunity to reduce the risk of confounding in the study design. Because the important outcomes of ZIKV infection are relatively rare, matching should be strongly considered. Matching is a method used to ensure comparability between cases and controls with regard to potential confounding factors. In an individual matching scheme, each case is matched to a control based on various potential confounding factors, such as age, sex, or other variables that are known to be associated with the outcome of interest [62, 65, 66]. However, over-matching can occur when cases and controls are matched by a non-confounding variable associated to the exposure but not to the disease, and may result in bias. Over-matching can underestimate associations and cannot be corrected in the analysis. Also, it is difficult to match for every possible confounder and unmatched confounders must still be considered in the analysis phases of the study.

Given the retrospective measurement of exposure status and potential confounders in case-control studies the risk of bias due to measurement error is also greater than in the prospective cohort studies. The advantage, however, of the case-control study is that smaller sample sizes are needed than for cohort studies and the follow-up time much shorter–which reduces the risk of selection bias due to loss to follow-up. The risks of bias due to measurement of the exposure, outcomes and confounders (as detailed above) also apply to these two protocols (17,18).

Effect modifiers (interaction variable) can over- or underestimate the risk of bias by modifying the observed effect of a risk factor on disease status. In the case of Zika, this includes the variable gestational age, and it is not recommendable to match for this type of variable as it is necessary to show the directionality and the magnitude of the modification of the effect granted.

Implications for policy and research

• Researchers undertaking ZIKV infection studies should use the standardized protocols and contribute to efforts to share and pool data.
• Funders of ZIKV research should prioritize research that uses these protocols and/or fills gaps in knowledge related to the protocol.
• Policy makers should use the information from research undertaken using the standardized protocols to guide decision making regarding public health advice and resource allocation.
• Researchers, policy makers and funders can help to improve the protocols by contributing to the generation of better information to calculate sample size, to determine the epidemiological spread of ZIKV and timing for research, expertise needed to run some tests, and legislation / regulations that enable the sharing of samples.