Talk:Disease burden of air pollution

Value judgements by Lelieveld2015

This is a hierachical representation of value judgements of the Lelieveld et al 2015 assessment. The identifiers starting with Q and P refer to items and properties in Wikidata.

Case-specific values

Disease burden (Q5282120) (Method for estimating disease burden of risk factors.)

⇐ instance of (P31)

The contribution of outdoor air pollution sources to premature mortality on a global scale (Lelieveld2015) (Q23670156, in Opasnet)

⇒ interested in (P2650) number of cases (Q23696805) qualifier: of (P642) death (Q4)

⇐ Interest is derived from value judgements:

• ⇤--#: . No life expectancy (Q188419) because we want to reflect the size of population in the burden of disease estimate. --Jouni (talk) 08:58, 18 March 2016 (UTC) (type: truth; paradigms: science: attack)
• ⇤--#: . No disability-adjusted life year (Q55627) or quality-adjusted life year (Q614165), because we want to have an easily understandable and comparable metric. --Jouni (talk) 08:58, 18 March 2016 (UTC) (type: truth; paradigms: science: attack)
• ⇤--#: . No resolution 2 of discussion #2, because we don't want premature mortality or probability of causation (Op_en6211). Instead, we want a counterfactual difference in disease burden estimates. --Jouni (talk) 08:58, 18 March 2016 (UTC) (type: truth; paradigms: science: attack)

⇒ used by (P1535)

Disease burden of air pollution (Q23680551)

Generic health indiator advice

The answer from page Health indicator concludes:

Disability-adjusted life year

Use when you want to combine death and disease, or impacts of several different diseases (especially when some are mild and some severe).

Quality-adjusted life year

Use when you want to combine death and suffering or lack of functionality, especially when the health outcomes are such that are not easily found from health statistics such as disease diagnoses.

Number of cases of death or disease

Use when the health impact is predominantly caused by a single outcome or when there is no need to aggregate different outcomes into a single metric. This is an easily understandable concept by lay people.

Life expectancy

Use when you want to describe public health impacts to a whole population and possibly its implications to the public health system. This is also a useful indicator if you want to avoid discussions about "what is premature" or "everybody dies anyway".

Welfare indicators

Use when you want to describe impacts on welfare rather than disease or health. There are a number of welfare indicators, but none of them has become the default choice. Consideration about the case-specific purpose is needed.

Case-specific advice

Therefore, based on the values of the Lelieveld2015 assessment,

• Number of attributable deaths should be used.

Bias in attributable fraction

Darrow and Steenland studied the direction and magnitude of bias in attributable fraction with different confounding situations. For details, see Attributable risk#Impact of confounders.

Darrow and Steenland studied the direction and magnitude of bias in attributable fraction with different confounding situations. For details, see Attributable risk#Impact of confounders. In brief, if there is a confounding factor that would make the apparent (crude) risk ratio larger than the true (adjusted) risk ratio, the apparent attributable fraction would be smaller than the true one and vice versa. This bias is more important when the fraction of exposed people in the population is small and the impact of confounding large.

So, we need to ask: a) what is the fraction of exposed population in Lelieveld2015, b) what is the impact of potential confounding, and c) taken these together, what is the likely direction and magnitude of bias in the attributable fraction estimates?

Exposed population

With fine particles (the most important air pollutant in Lelieveld2015), practically everyone is exposed. The exposure assessment was based on global atmospheric modelling with resolution of tens of kilometres. This reflects the background levels and misses the high peak-levels that occur when people are close to an emission source. In contrast, it is a good estimate of the lower end of exposure distribution in any given grid cell, because fine particles penetrate well into indoor environments. Only effective particle filters can remove the majority of fine particles indoors and thus reduce the exposure significantly. But such equipment are available to a tiny fraction of the population in the world. It is therefore reasonable to assume that the exposures modelled are fair estimates of the median or mean exposures, although they understimate the very highest exposures.

In conclusion, everyone is exposed to levels estimated by Lelieveld2015. This would lead to low bias in attributable fraction.

Confounding in RR

Risk ratios from the scientific literature were used. These can be biased in all kinds of unknown ways, but they are the best estimates we have and there is no point in questioning their practical usability. Instead, we should examine what local confounders there may be that would lead us to identify (possibly quantifiable) biases in Lelieveld2015.

The most obvious potential confounder is age, and age distribution varies greatly in different parts of the world. The age structures come from national statistics ----#: . Is this true? --Jouni (talk) 13:52, 9 April 2016 (UTC) (type: truth; paradigms: science: comment), so they may vary locally. A key question is: is age correlated with both exposure (fine particle concentration maps) and disease (cardiovascular and other mortality)? It is strongly positively correlated with disease, but exposure is not obvious. But because of young people moving from urban areas to cities it is reasonable to assume that age is negatively correlated with exposure within the fine particle concentration map grid cells. The correlation may even be moderate but not high because a grid cell mostly contains either rural or urban area and therefore such correlation mostly happens between grid cells, not within.

If age has positive correlation with disease and negative with exposure, it means that the risk ratio is biased downward and the attributable risk upward, i.e. the true risk is smaller than the assessment predicts. It is difficult to estimate the possible confounding, but because it arises from correlations within grid cells, it is hard to imagine that it would be more and twofold.

Overall bias

Darrow and Steenland offer quantitative graphs for estimating bias. If we assume that practically everyone is exposed and age confounding (the largest potential confounding factor) decreases RR estimates by half, we can conclude that the overall bias in AF is still very small.