Diagram-based Analysis of Causal Systems (DACS): elucidating inter-relationships between determinants of acute lower respiratory infections among children in sub-Saharan Africa

Acute lower respiratory infections (ALRI) represent a permanent global emergency [3]. They were responsible for 6.8% of 2.8 million neonatal deaths, 20.1% of 2.0 million infant deaths and 12.4% of 2.0 million deaths among children aged 1 to 4 years respectively in 2010 [4], with sub-Saharan Africa being the most affected world region [5]. Childhood ALRI is one of several enormous public health problems in developing countries that fail to attract the necessary research to elucidate more of their epidemiology, and how this relates to the environmental and social context.

One of the reasons for this lack of attention is the difficulty of obtaining high-quality health information. In most African countries, vital registration systems are either non-existent or do not reach the whole population, and health surveillance is limited [6, 7]. Instead, national and international decision-making heavily relies on nationally representative household surveys that generate cross-sectional datasets that are largely comparable across countries.

Childhood ALRI is the outcome of a web of interacting socio-economic, environmental, behavioral and healthcare factors. Interventions can thus be directed at curbing mortality (e.g. access to healthcare, timely treatment with antibiotics) [8, 9], at reducing the risk of infection (e.g. improved nutrition, promotion of breastfeeding, reduced exposure to indoor air pollution) [10‐13], or at instigating longer-term socio-economic changes to create healthier societies [14].

Context is crucial. Fewer than 20% of children with ALRI receive appropriate treatment [15]. Even with successful cure, children return to a home with high risk of re-infection. Interestingly, childhood ALRI deaths in the United States fell by two thirds during the first three decades of the 20th century, prior to the introduction of antibiotics or vaccines [3]. This implies that interventions to improve living conditions play a critical role in reducing ALRI morbidity and mortality [16].

Developing effective interventions requires evidence not only on how individual causal mechanisms influence ALRI risk but also on how these jointly determine disease. This concept could be called systems epidemiology [17]: analysis of the whole system of humans in their total environment within which disease occurs. While systems thinking is increasingly being discussed as important [18, 19], a whole-system approach to analysis is still rare in epidemiology.

Causal diagrams provide an important basis for such an approach. Traditional epidemiology is often confined to the analysis of single links between a proximal determinant and disease, and causal diagrams were introduced in this context, mainly to aid causal inference [20, 21]. In addition, systems approaches using causal diagrams have been introduced in specific contexts, notably infectious disease modeling using flow diagrams and differential equations [22, 23]; causal mediation analysis [24]; and the emerging field of dynamic systems modeling [25, 26].

In principle, a systems perspective can also be applied to large and complex causal networks involving environmental or social factors [27] and/or bio-molecular or genetic pathways [17, 28, 29]. A causal diagram then constitutes a means of organizing the current state of knowledge about that specific area of research, and a framework for integrating statistical analyses [17, 28, 29]. It consists of all the constituent hypotheses within the system, each represented by a link, plus the additional pathways that are necessary for causal inference (e.g. the potential confounders).

In systems epidemiology, diagram construction involves a partly a priori approach: the causal diagram does not emerge from the data but is developed by combining prior knowledge with empirical testing. The former is obtained from pre-existing evidence of various types (e.g. mechanistic or statistical). Thus, constructing such a diagram requires that its structure and content be specified upfront from existing knowledge about the system, supplemented by plausible assumptions where evidence is lacking. It is vital that this includes a complete or near-complete set of potential confounding and selection effects for all the hypotheses embodied in the diagram. Several studies have combined a priori conceptualization of health determinants with modeling, using, for example, quantile regression [30], graphical chain models [31‐33] or hierarchical models [34]. Other possible statistical tools for estimating causal diagrams include structural equation modeling [24, 35, 36], propensity score matching [37, 38] and instrumental variables [28, 39, 40].

To our knowledge, all of the studies published to date were based on a single dataset. In systems epidemiology, the causal diagram represents the overall state of knowledge about the system. This means that empirical testing can, and often has to, involve multiple datasets. One advantage is repeated examination, where the same link is explored using more than one dataset, allowing the robustness of the evidence to be assessed. A second is that, for complex systems which cannot be covered by an individual dataset, different parts of the system can be examined using different datasets. This may entail decomposing the diagram into corresponding sub-diagrams. Ideally, this decomposition can be done using conditional independence, so that each relationship or group of relationships becomes testable as an individual hypothesis, without introducing further confounding or other distortions [41], but this is not always possible. Where appropriate, selection effects (colliders) can also readily be introduced into a causal diagram [42].

In this paper, we seek to contribute to the concept of systems epidemiology [17] by demonstrating how Diagram-based Analysis of Causal Systems (DACS) can be used to investigate complex questions of disease causation, including both proximal health risks and “causes of causes” [43]. This approach to analysis combines the use of causal diagrams with multiple routinely available data sources, employing a variety of statistical techniques. In the following, we demonstrate the step-by-step progression from a merely plausible a priori causal diagram towards one supported by empirical analysis, aimed at deciphering the contributions of proximal and distal risk factors to childhood ALRI.

The approach follows the notational conventions of Earp and Ennett [44], Robins [21] and Best and Green [45] (see Figure 1). It relies on the initial assumptions that causal relationships apply independently of the data source used to assess them; and that they are stable between populations and over time. These assumptions can be tested empirically and rejected if appropriate. The timeframe must be clearly specified, as must the unit of analysis (i.e. individual vs. aggregate) and level of aggregation or spatial scale, as they influence the identification and definition of variables included. We focus on relatively short-term processes affecting health outcomes among children under five years of age.

Step 1: a conceptual causal diagram was constructed a priori based on state-of-the-art knowledge, independent of data availability (Table 1). This was organized into hierarchical layers, which represent the causal and temporal ordering of the different groups of variables (we prefer the term “layer” to “level”, as the latter could be confused with its quite different usage in the context of multilevel modeling). A four-layer diagram resulted, with layers representing context, household socio-economic status, proximal health risks and health outcomes; as a convention, health outcomes are depicted as being harmful [27]. Specificity was introduced by attempting to identify all distinct variables in each layer and all potential causal pathways that connect variables within and across layers, including potential confounders.

This diagram resembles a graphical chain model [46, 47] in the sense that its variables are partitioned into hierarchically arranged subsets (i.e. conceptual groupings), with arrows representing causal associations linking variables in adjacent subsets. Within these conceptual groupings associations between variables are assumed to be non-causal. We represent this assumption by the absence of arrows (e.g. no arrow connecting the different variables within material situation), whereas in graphical chain models, non-causal associations (correlations) between variables within a subset are empirically tested and represented by undirected links.

Step 2A: data availability informed which parts of this diagram could be examined using a specific dataset. Conceptual variables were replaced with measured variables, resulting in an operational single-dataset causal diagram (see Table 1). As omission of relevant variables and relationships, where these cannot be measured, is likely to result in residual confounding, conceptual variables and the paths connecting them were retained in the diagram. This serves to ensure that all relevant common causes and colliders that could affect a particular path between exposure and outcome are included. Where this is the case, each arrow in the operational single-dataset causal diagram represents a separate hypothesis. These were tested – in our case using a Demographic and Health Survey (DHS) dataset for Benin – to assess overall support for the hypothesis (see Data and statistical analysis and Criteria for testing causal diagrams).

Step 2B: testing across multiple similar datasets generated an operational multiple-dataset causal diagram (see Table 1) in our case using DHS datasets for Kenya, Ethiopia and Namibia. This step has a confirmatory function, the aim being to assess the degree of consistency/robustness of the step 2A diagram across populations. In making these comparisons, attention must be paid to selection factors or colliders which may differ between populations; these could be examined explicitly [42] and, where appropriate, included in country-specific operational causal diagrams. Hypothesis testing was repeated and conclusions were drawn about the sign of the relationships in each hypothesis (i.e. a decreasing or increasing function), and about its strength (i.e. strong or limited). Where a given path between measured variables did not show any effect across settings, the path was removed from the diagram (see Data and statistical analysis and Criteria for testing causal diagrams).

Step 3: as a single type of dataset rarely includes all the information required for testing a whole conceptual diagram, other datasets with a complementary function can be brought in. We used a pooled World Health Survey (WHS) dataset to examine the pathway linking socio-economic status, solid fuel use and ALRI mortality. The results obtained from WHS testing and DHS testing were combined in an integrated causal diagram (see Table 1). This represents our current best understanding of the relationships specified in the original conceptual diagram, but it is not in any sense a final version; such diagrams continue to evolve as more evidence becomes available to confirm, extend and refine the operational version. This involves repeating steps 2 and/or 3, using the most appropriate statistical approaches in view of data availability and the strength and consistency of evidence already represented in the diagram. Suitable methods include regression techniques, structural equation modeling and instrumental variable approaches, depending on the nature of the relationships to be examined and the data available; Bayesian approaches and/or various types of meta-analysis [48] can also be used to formally combine new evidence with the existing evidence represented in the diagram. Also, periodic reassessment of the adequacy of the original conceptual diagram may be necessary (step 1).

We used DHS [49] and WHS data [50], nationally representative household surveys with response rates well above 90%, that provide high-quality and comparable information on health and risk factors for a large number of developing countries. DHS datasets for Benin, Kenya, Ethiopia and Namibia were selected because they included information on the main cooking fuel used; they were analyzed separately to allow for a distinction between those causal links that are likely to be consistent between populations and those that may operate in a population-specific fashion (Table 2). WHS data for sixteen countries of sub-Saharan Africa were pooled (Table 3; [51]).

While the DHS provides detailed information on the recent occurrence of childhood illnesses as well as child mortality rates, the WHS assesses child mortality and symptoms prior to death. Variables representing ALRI determinants also vary between the data sources (Table 4).

Given the characteristics of the hypotheses, we used logistic regression, ordered logistic regression and survival analysis to test individual hypotheses of different outcomes [52, 53]; survival analysis was also used to examine the impact of cooking-related parameters on ALRI mortality [51]. In this way, empirical evidence (or lack thereof) was assessed for every arrow connecting two measured variables in the operational causal diagrams. In view of non-independence due to children born to the same mother, all analyses were adjusted for clustering to produce robust standard errors, resulting in wider confidence intervals. We did not adjust for stratification and cluster samples and omitted sample weights, a strategy considered conservative for multivariable analysis of household survey data [54]. Models were run in Stata Special Edition 9 for the whole dataset, and separately for urban and rural settings; exploratory analyses, confirmed by cluster analysis, had revealed striking urban–rural differences in all countries [55].

For each hypothesis, the full model including all measured explanatory variables and all possible nested models were run. Initially, provided the univariate model explained the data significantly better than the null model with p < 0.05, all variables were retained. Subsequently, nested models were compared against the full model using the Akaike Information Criterion (AIC) to explore whether different explanatory variables contributed independently to the outcome of interest. Only differences of 3 or more points are considered meaningful in terms of indicating better model fit [56]; for differences of less than 3 points, the more complete model was reported in line with a conservative approach. Where the diagram showed unmeasured explanatory variables, their likely influence in the form of residual confounding was carefully considered in the interpretation of findings.

As we relied on different data sources to inform different parts of the diagram, we decomposed it into a series of separate hypothesis tests and, similar to Weng and colleagues [57], employed conventional covariate selection techniques for testing. The criteria described here provide a basis for replicating the analysis, conducting sensitivity analyses and incorporating new datasets and data sources as they become available. They were developed in an attempt to minimize the role of researcher judgment, and where this is unavoidable make its use explicit and transparent; they represent but one way of doing so.

As inclusion is more conservative than omission, it is important to ensure that no hypothesis is rejected prematurely because of a country-specific situation or data quality issues. Therefore, in step 2A, if any of the univariate models explained the data significantly better than the null model in at least one of the three settings, this was considered to provide evidence towards the hypothesis. This initial testing stage only assessed overall support for a given hypothesis involving several variables, not the role of individual variables.

Figure 2 depicts the decision-making process to assess the sign of a relationship in individual settings. For any hypothesis tested in step 2B, each explanatory variable was considered in turn and separately for urban and rural settings; only variables included in the model selected based on the AIC were judged to have a direct and at least partly independent effect on the outcome variable. Absolute sample sizes, relative sample sizes (i.e. small sub-groups with respect to variables of interest) and measurement error can all affect whether a true relationship is measured as statistically significant or not. We therefore contend that consistency (i.e. central estimates of odds ratios are always above or always below 1) across different levels of a given variable (e.g. low, intermediate, high electric goods index) is potentially a more reliable guide for assessing the sign of a relationship than statistical significance. As detailed in Figure 2, this consistency criterion was applied to distinguish between strong versus limited support for an increasing or decreasing effect on the outcome variable, or to conclude that the relationship is non-ordered.

Once the assessment on a country-by-country and setting-by-setting basis was completed, a judgment needed to be made about homogeneity/heterogeneity. Three situations put consistency at stake:

➢ A relationship is classified as increasing across all levels of a variable in some settings but absent in the remaining settings. Such findings contradict consistency but may be explained by small differences in causal mechanisms or by random error.

➢ A relationship is classified as increasing across all levels of a variable in some settings but as decreasing in the remaining settings. Such findings strongly contradict consistency as they point towards distinct setting-specific causal mechanisms unlikely to be explained by chance.

➢ A relationship is classified as non-ordered (i.e. showing an effect, with the sign of the relationships varying at different levels of a variable) in some but as increasing in other settings. Consistency can easily accommodate findings that go further in a sub-set of settings relative to the rest by “downgrading” the classification of the relationship to non-ordered.

In deriving an operational multiple-dataset causal diagram in step 2B, we distinguish between an increasing/decreasing function with strong or limited support, a non-ordered relationship, a setting-specific increasing or decreasing function and no effect, taking into account the possibility of “false negatives” due to measurement error, other data quality issues or random error. As illustrated in Figure 3, we made the somewhat arbitrary decision to allow for one exception regarding the presence or absence of an effect (n_total-1) and the sign of a relationship (n_effect-1). We only considered there to be sufficient evidence of a null effect, when an effect was tested across multiple settings and no relationship was observed in any of these; only under these circumstances was an arrow removed from the diagram.

The basic idea of DACS is the use of causal diagrams as a formal tool for representing state-of-the-art knowledge of a specific research area. This whole-system approach makes the analyst’s view of reality and assumptions explicit; generates testable hypotheses; provides a framework for model building and statistical analysis, potentially using a broad range of statistical methods; accommodates multiple sources of evidence; evolves from conceptual to empirically supported; identifies research gaps; highlights entry-points for public health interventions; and facilitates communication between stakeholders.

Constructing a causal diagram involves a dialog between an a priori statement of causal structure and an inductive a posteriori method of testing. Each informs the other: the structure facilitates design of studies to test specific hypotheses, and accumulating evidence stimulates revision of the causal structure. Each diagram explicitly represents not only pathways of substantive interest, but also those that could bring about non-random statistical associations in the absence of direct causation: i.e. reverse causation, confounding (common ancestor(s)) or Berksonian bias (common descendant(s)). By representing all relevant potential causal pathways, this resembles established epidemiological methodology of directed acyclic graphs for addressing biases and selection effects relating to a single causal pathway [20, 21, 44, 76, 77].

While individual causal pathways in our approach resemble directed acyclic graphs, the overall DACS approach is more closely related to graphical chain models, where complex association structures are explicitly recognized, sets of variables are placed along a dependence chain and then modeled using a series of regression analyses [46, 47]. DACS goes beyond conventional graphical chain models, in particular by: (i) highlighting the importance of evidence in developing the initial conceptual diagram, which leads to a more finely-structured diagram comprising conceptual groupings within layers (=blocks) rather than only blocks and which explicitly promotes a hypothesis-driven analysis strategy; (ii) emphasizing model building through a structured step-by-step method, and embracing a forward-looking approach, where the causal diagram is expected to continue to evolve over time; and (iii) allowing for the possibility of drawing on multiple datasets, both for the original analysis and for the subsequent addition of information.

The last point in particular is a key strength of DACS and carries two advantages. First, repeated assessment allows the robustness of a single causal pathway across data sources and settings to be examined. Secondly, following decomposition of the system, relevant sub-structures can be assessed using different data sources, although this does entail specifying potential confounders that may be introduced where conditional independence cannot be used.

The requirement to specify all possible causal pathways, including those that cannot be observed, is formidable both with a causal diagram-based and the traditional approach. Researcher judgment plays a role in both, but is more explicit in the combined a priori/empirical method because it is exercised prior to the analysis, and diagramming makes it visible and transparent. With a purely inductive approach, findings that are not based on previously specified hypotheses may be easier to ignore or dismiss when inconvenient [77], and such findings are suspect when novel – the “fishing expedition” problem. With a partly a priori approach, judgments and assumptions are uninfluenced by the data and have to be made transparent; any changes following the statistical analysis must be explicitly stated and justified. Indeed, incorrect researcher judgment is also potentially reversible: the conceptual diagram is periodically revised as further evidence becomes available. Initial conceptual diagrams are, however, not necessarily unique, and it is good practice for researchers to use their diagrams to explore underlying differences in their assumptions [27]. Importantly, it is impossible to specify a correct analytical strategy without causal knowledge or assumptions [57, 76, 78].

The application of this method to the problem of ALRI in developing countries here shows several specific weaknesses. In most complex systems there are likely to be some unspecified or unmeasured common causes, and this almost certainly holds true for the present analysis, which was limited by data availability and quality [55]. Neither the DHS nor the WHS include all relevant variables and, as a result, we were unable to control adequately for confounding in several of the observed relationships. Moreover, several variables are subject to significant measurement error (e.g. a household’s main cooking fuel as proxy for a child’s exposure to indoor air pollution in DHS and WHS; a combination of reported symptoms prior to child death as proxy for clinically-confirmed cause of death in WHS).

In addition, the cross-sectional nature of the datasets precluded analysis of the temporal sequence of events. We carefully assessed the likely impact of this limitation, and consider that reverse causation and feedback loops were unlikely to be important given the short timeframe of interest specified.

Another limitation is that our models assume simple (e.g. linear) functional form and an absence of effect modification or statistical interaction. With the future availability of more and richer datasets on ALRI determinants, it will become possible to test, and if necessary modify, this assumption. Finally, the countries investigated here represent a convenience sample – they were not chosen for their representativeness of the African continent. Nevertheless, based on their geography, climate and distinct colonial histories, Benin, Ethiopia, Kenya and Namibia can be considered representative of a variety of living conditions in sub-Saharan Africa, and so can the pooled WHS dataset of sixteen countries. In summary, the results presented here have only interim status, reflecting our understanding of the system based on currently available evidence.

Our research has only been able to shed light on one of the many pathways that link distal and proximal determinants of childhood ALRI, the inter-linkages between socio-economic status, solid fuel use and ALRI mortality.

The risk of dying from ALRI during childhood varies between population groups but socio-economic differences may be less marked than those commonly observed with respect to all-cause child and infant mortality [79‐87]; indeed most of the observed socio-economic gradients in ALRI mortality do not reach significance. This is consistent with previous reports of minor socio-economic gradients in childhood ALRI mortality [88, 89] and morbidity [58, 90‐93].

Wealth, maternal education and, to a lesser extent, paternal education are highly protective against solid fuel use. This suggests that these play at least partly independent roles in structuring cooking fuel use in sub-Saharan Africa. Previous reports [58, 59] that cooking with liquefied petroleum gas is more prevalent among sub-populations with higher income and higher educational attainment accord with these findings.

Our research confirms solid fuel use as a major risk factor for ALRI among African children, suggesting that exposure more than doubles the risk of ALRI mortality. These results are comparable to findings in two case–control studies of ALRI mortality conducted in Tanzania [89] and the Gambia [88], and a case–control study of all-cause child mortality in India [94]. Reducing solid fuel use could therefore be an important means of reducing ALRI mortality.

We found that socioeconomic variables affected solid fuel use, and that solid fuel use affected ALRI mortality. For solid fuel use to qualify as a mediator of the impact of socio-economic status on ALRI mortality, the relationship between maternal education/wealth and ALRI mortality should be attenuated once solid fuel use is included in the model [95]. Surprisingly, this was not the case. Common concerns in mediation analysis are unmeasured confounding and measurement error, but in this instance a third issue proved to be the explanation: heterogeneity in the population. Detailed investigations revealed the presence of a threshold effect in the wealth-fuel use relationship: i.e. differences in fuel use are limited among the lower wealth quintiles but stark between the bottom four quintiles and the richest quintile. Within the richest quintile solid fuel use acts as a mediator of the wealth-ALRI mortality relationship: when the variable is included in the model the hazard ratio is attenuated from 0.66 (0.41; 1.07) to 0.87 (0.52; 1.44). Causal diagrams therefore play an important role in making complex issues of mediation and non-transmission apparent.

The integrated causal diagram in Figure 7 illustrates two potential entry-points for interventions. Increasing household wealth and, to a lesser extent, improving education and knowledge represent entry-points for social intervention, which reduce solid fuel use indirectly over long periods of time. A switch to cleaner fuels represents an entry-point for environmental intervention, which directly and immediately addresses hazardous household energy practices. An alternative would be promotion of cleaner-burning, more fuel-efficient stoves to reduce indoor air pollution, as demonstrated by a randomized controlled trial in Guatemala [96] and a systematic review of the effectiveness of different improved stove models in reducing household air pollution (Bruce et al., manuscript in preparation). As many projects and programs to improve access to modern household energy have either failed to reach the lowest-income group (e.g. Indonesian national LPG program [97]) or failed to bring about sustainable benefits (e.g. Indian National Improved Cookstove program [98]) we believe that technical interventions to reduce solid fuel use must be embedded in an integrated approach that considers and, to some extent, addresses socio-economic determinants if they are to benefit those who need them most. This conclusion accords with the findings of a recent systematic review of enablers and barriers to uptake of improved stoves [61].

Importantly, Figure 7 exposes neglected areas of research and can thus be useful in setting the research agenda. For example, it is noteworthy how few of the proximal risk factors are actually measured in DHS/WHS and that, based on routinely available data, no analysis of the multiple determinants of ALRI morbidity is feasible. Apart from further research along the lines reported here using additional datasets, efficacy and effectiveness trials of specific technical interventions, and natural experiments relating to policy changes, could illuminate the interrelationship of socio-economic conditions, solid fuel use and ALRI. The design of research could also benefit, e.g. by ensuring that omitted variables (residual confounding) are kept to a minimum, and, where possible, that datasets refer to a section of the diagram that can be disaggregated using conditional independence.