The C-Word: The More We Discuss It, the Less Dirty It Sounds

I thank Chiolero (p. 622), Ahern (p. 621), Glymour and Hamad (p. 623), Jones and Schooling (p. 624), and Begg and March (p. 620) for sharing their reactions to my commentary (p. 625). My impression is that there are few substantial disagreements among us, just differences in emphasis or, in one case, a misunderstanding.

Chiolero and Ahern zero in on a key issue: the need to distinguish the causal question from the procedure used to answer it. As Chiolero puts it, “How to formulate adequate causal questions had [not] been formalized” until recently in health research, and much of the teaching is devoted to “data management and analysis, leaving no room for causal thinking or for the formulation (before running the analyses) of research questions.” Ahern stresses the importance of a structured process, or a roadmap, to ask and answer causal questions using observational data.

SPECIFY THE TARGET TRIAL

The first step of that process is, in Ahern’s words, “to articulate the scientific question, including definition of the causal parameter of interest.” Glymour and Hamad also highlight the importance of this first step when they state, “We must first start by articulating clear causal questions,” which is especially true in social epidemiology when the goal is translating causal inferences into action.

One way of performing this step precisely is to specify the protocol of the hypothetical randomized trial that would allow us to estimate the causal parameter of interest. We refer to that hypothetical trial as the “target trial.”¹ Some of us have argued that causal questions that cannot be translated into a hypothetical experiment are ill defined.² As a consequence of ill-defined questions, data analyses yield numerical estimates that are not easily interpretable as estimates of causal effect.

EMULATE THE TARGET TRIAL

The second step of the process is to emulate the target trial using a combination of data, empirically unverifiable assumptions, and statistical methods. Jones and Schooling are concerned that trying to emulate a target trial may drive too much attention to sophisticated statistical techniques (e.g., inverse probability weighting) for confounding adjustment at the expense of a thoughtful consideration of design issues and of expert knowledge summarized in causal theories. I suppose that the process of specifying and emulating a target trial can be misused but, if that happens, Jones and Schooling will find me by their side fighting for sound design and appropriate incorporation of expert knowledge in the process. Indeed, incorrect causal inferences from observational data are often the result of a flawed emulation of the basic design of the target trial (e.g., choice of time zero and classification of treatment groups) rather than of emulation of its randomized assignment (i.e., insufficient confounding adjustment).

Of course, specifying and emulating the target trial do not imply that our observational study “has fulfilled its purpose and correctly identified a causal effect,” as Jones and Schooling warn us. It just means that (1) we can provide a scientific description of the causal effect that we are estimating, and (2) we have provided our best estimate of that causal effect. But, as Begg and March remind us, even our best estimate may be affected by systematic bias attributable to selection, confounding, or mismeasurement (reverse causation, also cited by Begg and March, can often be viewed as a form of confounding in which an undetected outcome or its precursors confound the effect of treatment on the detected outcome). Because these biases induce associations that do not have a causal interpretation, the association estimated from any data analysis is always causally suspect. Again, the process of specifying and emulating a target trial helps by providing a systematic way to explore each type of bias and its potential influence on the effect estimate. The Cochrane tool has adopted this target trial–based approach to assess the risk of bias of nonrandomized studies.³

TRIANGULATE

Ultimately, no single study can produce uncontroversial estimates of causal effect. As Glymour and Hamad point out, some form of “triangulation” of studies will be needed. To quantify a causal effect, triangulation consists in explicitly emulating the target trial of interest using different methods and data sources. When some of those emulations are expected to be differentially affected by bias, investigators can use the imperfect estimates from each emulation to try to pinpoint or bound the magnitude of the true causal effect. The idea is analogous to the process by which travelers obtaining readings of radio waves at different positions can triangulate the position of the radio transmitter.

But the success of triangulation efforts to estimate causal effects requires that “causal” stop being considered the C-word that investigators and editors avoid. Only by precisely defining the causal effect of interest will we have a chance of estimating it accurately. In the absence of a precise definition of the causal effect of interest in each study, researchers will end up trying to triangulate study estimates that cannot be triangulated, just as travelers who obtain mixed readings from multiple transmitters cannot locate the position of any of them. Who could blame them for being confused?