ouR data generation

I am putting together a brief lecture introducing causal inference for graduate students studying biostatistics. As part of this lecture, I thought it would be helpful to spend a little time describing directed acyclic graphs (DAGs), since they are an extremely helpful tool for communicating assumptions about the causal relationships underlying a researcher’s data.

The strength of DAGs is that they help us think how these underlying relationships in the data might lead to biases in causal effect estimation, and suggest ways to estimate causal effects that eliminate these biases. (For a real introduction to DAGs, you could take a look at this paper by Greenland, Pearl, and Robins or better yet take a look at Part I of this book on causal inference by Hernán and Robins.)

I’ve never taught an intro probability/statistics course. If I ever did, I would certainly want to bring the underlying wonder of the subject to life. I’ve always found it almost magical the way mathematical formulation can be mirrored by computer simulation, the way proof can be guided by observed data generation processes, and the way DGPs can confirm analytic solutions.

I would like to begin such a course with a somewhat unusual but accessible problem that would evoke these themes from the start. The concepts would not necessarily be immediately comprehensible, but rather would pique the interest of the students.

I was recently contacted to see if simstudy can create a data set of correlated outcomes that are measured over time, but at different intervals for each individual. The quick answer is there is no specific function to do this. However, if you are willing to assume an “exchangeable” correlation structure, where measurements far apart in time are just as correlated as measurements taken close together, then you could just generate individual-level random effects (intercepts and/or slopes) and pretty much call it a day. Unfortunately, the researcher had something more challenging in mind: he wanted to generate auto-regressive correlation, so that proximal measurements are more strongly correlated than distal measurements.

In part 1 of this 2-part series, I introduced the notion of sensitivity to unmeasured confounding in the context of an observational data analysis. I argued that an estimate of an association between an observed exposure \(D\) and outcome \(Y\) is sensitive to unmeasured confounding if we can conceive of a reasonable alternative data generating process (DGP) that includes some unmeasured confounder that will generate the same observed distribution the observed data. I further argued that reasonableness can be quantified or parameterized by the two correlation coefficients \(\rho_{UD}\) and \(\rho_{UY}\), which measure the strength of the relationship of the unmeasured confounder \(U\) with each of the observed measures. Alternative DGPs that are characterized by high correlation coefficients can be viewed as less realistic, and the observed data could be considered less sensitive to unmeasured confounding. On the other hand, DGPs characterized by lower correlation coefficients would be considered more sensitive.

Principled causal inference methods can be used to compare the effects of different exposures or treatments we have observed in non-experimental settings. These methods, which include matching (with or without propensity scores), inverse probability weighting, and various g-methods, help us create comparable groups to simulate a randomized experiment. All of these approaches rely on a key assumption of no unmeasured confounding. The problem is, short of subject matter knowledge, there is no way to test this assumption empirically.

It’s always nice to be able to speed things up a bit. My first blog post ever described an approach using Rcpp to make huge improvements in a particularly intensive computational process. Here, I want to show how simple it is to speed things up by using the R package parallel and its function mclapply. I’ve been using this function more and more, so I want to explicitly demonstrate it in case any one is wondering.

In my previous post, I described a (relatively) simple way to simulate observational data in order to compare different methods to estimate the causal effect of some exposure or treatment on an outcome. The underlying data generating process (DGP) included a possibly unmeasured confounder and an instrumental variable. (If you haven’t already, you should probably take a quick look.)

A key point in considering causal effect estimation is that the average causal effect depends on the individuals included in the average. If we are talking about the causal effect for the population - that is, comparing the average outcome if everyone in the population received treatment against the average outcome if no one in the population received treatment - then we are interested in the average causal effect (ACE).

I’ve been inspired by two recent talks describing the challenges of using instrumental variable (IV) methods. IV methods are used to estimate the causal effects of an exposure or intervention when there is unmeasured confounding. This estimated causal effect is very specific: the complier average causal effect (CACE). But, the CACE is just one of several possible causal estimands that we might be interested in. For example, there’s the average causal effect (ACE) that represents a population average (not just based the subset of compliers). Or there’s the average causal effect for the exposed or treated (ACT) that allows for the fact that the exposed could be different from the unexposed.

I put together a series of demos for a group of epidemiology students who are studying causal mediation analysis. Since mediation analysis is not always so clear or intuitive, I thought, of course, that going through some examples of simulating data for this process could clarify things a bit.

Quite often we are interested in understanding the relationship between an exposure or intervention on an outcome. Does exposure \(A\) (could be randomized or not) have an effect on outcome \(Y\)?

Every new study presents its own challenges. (I would have to say that one of the great things about being a biostatistician is the immense variety of research questions that I get to wrestle with.) Recently, I was approached by a group of researchers who wanted to evaluate an intervention. Actually, they had two, but the second one was a minor tweak added to the first. They were trying to figure out how to design the study to answer two questions: (1) is intervention \(A\) better than doing nothing and (2) is \(A^+\), the slightly augmented version of \(A\), better than just \(A\)?