7.7 Assessing causality

The potential is high for drawing incorrect conclusions about how environmental exposure affects disease from GIS studies. This is both because of data availability and methodological problems in such studies. It is important to understand what potential problems may occur when assessing causality in spatial epidemiological studies of exposure, so that findings can be interpreted in an appropriately cautious manner.

There are potentially two mistaken conclusions that can be drawn about the relationship between an environmental exposure and a disease. Firstly, a GIS-based study can suggest that the environmental exposure affects the disease, when in fact it does not (a type I error). Alternatively, the GIS-based study can suggest that the exposure does not affect the disease, when in fact it does (a type II error).

The most common cause of a type I error (i.e. mistakenly finding a relationship between exposure and disease where none exists) is confounding. This occurs where an unmeasured factor affects disease risk and is also related to the exposure being studied. For example, a study of poor air quality might show a relationship between asthma and residence in inner-city areas with high traffic densities. However, if lack of exercise is a risk factor for asthma and fewer people in inner city areas exercise regularly, then this apparent relationship may be due entirely to spatial variation in physical exercise, not pollutants from traffic. In this instance, the group exposed to traffic air pollution is not comparable with the group who are not exposed to pollution because they differ in terms of their levels of physical exercise. The health effect is related to a different causal pathway altogether – physical activity, not inhaled air pollutants. Cross-sectional studies are particularly vulnerable to this problem and particular attention should be paid to the potential effects of confounding when interpreting cross-sectional study findings.

Another problem inherent in GIS studies of exposure is bias. There are two types of bias that affect GIS studies. Selection bias occurs where the subjects being selected for study are not drawn randomly from the population at risk. For example, if controls in a case-control study are selected randomly from healthy patients attending at a doctor’s surgery, this sampling strategy may give a biased picture of the population at risk. Young men in particular very seldom attend surgeries and so such a sample of controls could be biased away from this age cohort. Selection bias may well affect the findings of a study, but it is difficult to assess from the outset whether such bias will result in a type I or type II error.

Accurate exposure measures are essential if the complex relationships between environment and health are to be understood.

(Briggs, 2000: p. 355)

 
The second type of bias is known as information bias. This occurs when the individuals being studied are mis-classified either in terms of their level of exposure or in terms of their health status. For example, a GIS study may measure exposure based on buffer zones around an incinerator chimney. If there are persistent winds in one direction, this buffer zone will under-estimate exposure downwind of the chimney and over-estimate exposure upwind of the chimney, so exposure mis-classification will occur. Health status can also be mis-classified. For example, many cancers have a latent period where they may remain undiagnosed. A case-control study may mistakenly include individuals with latent cancers in the healthy control group, thereby producing mis-classification of health status. In most situations, information bias will increase the chances of a type II error (i.e. not identifying a relationship between exposure and disease when in fact there is one), although there are specific circumstances where this is not the case.

The possibility that chance accounts for the study findings also needs to be considered. If the relationship between 10 different exposure measures and a health outcome is assessed statistically, it is quite likely that at least one of the exposure measures will appear significant at the 95% level. The possibility that an exposure-health outcome relationship is spurious should therefore also be borne in mind.

There are criteria for evaluating how far a GIS study of environmental exposure may have been affected by these problems, originally proposed by Bradford-Hill in 1965. Firstly, in interpreting the study findings, it is useful to consider whether there is a biologically plausible mechanism why the exposure might have led to the disease. For example, it is difficult to conceive of a mechanism for a significant spatial cluster of fall-related injuries among the elderly occurring near nuclear power plants. Secondly, it is worth considering whether several studies have drawn the same conclusions or whether the findings of the GIS study are completely at odds with previous research. Finally, the strength of the relationship in the GIS study between the environmental exposure and the disease should be considered. If the relationship is very pronounced, it deserves more attention. The latter two criteria form the basis of meta-analysis, a statistical technique for analyzing the findings of many different studies with similar designs.

Note: The slides below are missing the normal navigation bar for technical reasons linked to download time. Click on the presentation to move through the slides

 

---

---