Abstract
Analysts often use data-driven approaches to supplement their knowledge when selecting covariates for effect estimation. Multiple variable selection procedures for causal effect estimation have been devised in recent years, but additional developments are still required to adequately address the needs of analysts. We propose a generalized Bayesian causal effect estimation (GBCEE) algorithm to perform variable selection and produce double robust (DR) estimates of causal effects for binary or continuous exposures and outcomes. GBCEE employs a prior distribution that targets the selection of true confounders and predictors of the outcome for the unbiased estimation of causal effects with reduced standard errors. The Bayesian machinery allows GBCEE to directly produce inferences for its estimate. In simulations, GBCEE was observed to perform similarly or to outperform DR alternatives. Its ability to directly produce inferences is also an important advantage from a computational perspective. The method is finally illustrated for the estimation of the effect of meeting physical activity recommendations on the risk of hip or upper-leg fractures among older women in the study of osteoporotic fractures. The 95% confidence interval produced by GBCEE is 61% narrower than that of a DR estimator adjusting for all potential confounders in this illustration.
1 Introduction
Estimating causal effects using observational data requires important subject-matter knowledge. One notably needs to identify the confounding covariates that bias the association between the exposure and the outcome. Unfortunately, prior knowledge is often insufficient to undertake this task. For example, reviews of variable selection methods used in epidemiological journals indicate that data-driven approaches are frequently employed to select confounders [1,2].
Recent research has shown that many classical model selection approaches, including stepwise regression, Bayesian model averaging, lasso, and adaptive lasso, can have poor performances in a causal inference framework (e.g., [3,4, 5,6]). There are two important explanations of this phenomenon. First, many model selection approaches do not account for the model selection steps and therefore produce confidence intervals that include the true effect less often than they should, an issue known as the post-selection inference problem [7,8, 9,10]. Various approaches, including Bayesian methods and bootstraping, have been proposed to deal with this post-selection problem. However, note that not all Bayesian or bootstrap methods solve the post-selection problem, see, for example, [11]. Second, classical methods often fail to identify all important confounders [5,6,12,13]. This is because they generally focus on modeling the outcome as a function of the exposure and potential confounders. Confounders that are weakly associated with the outcome but strongly with the exposure may thus fail to be identified.
Multiple model selection methods for causal inference have been introduced in recent years, including the collaborative targeted maximum likelihood (C-TMLE [14]), Bayesian adjustment for confounding (BAC [6,15]), confounder selection via penalized credible regions (BP [16]), inference on treatment effects after selection among high-dimensional controls (HDM [13]), Bayesian causal effect estimation algorithm (BCEE [5]), model averaged double robust estimation (MADR [17]), outcome-adaptive lasso (OAL [4]), the group lasso and double robust estimation of causal effects (GLiDeR [18]), the outcome highly-adaptive lasso (OHAL [19]), and the high-dimensional confounding adjustment using continuous spike and slab priors (HD-SSL [20]). Despite these important developments, there is still a need for new methods or for the extension of existing methods. For instance, not all methods directly produce standard errors or confidence intervals for their causal effect estimator. Instead, they rely on bootstrap procedures for making inferences, which can be computationally intensive. Moreover, many methods were specifically elaborated for estimating the effect of a binary exposure on a continuous outcome. Only a few address additional situations. Finally, most methods model either the outcome as a function of the exposure and confounders, or the exposure as a function of confounders, and require that the postulated model is correct. In contrast, double robust (DR) methods combine both models and yield consistent estimators if either model, but not necessarily both, is correct. As such, DR methods require less stringent modeling assumptions than alternatives. Table 1 presents a summary of the characteristics of the causal model selection methods mentioned earlier.
Method | R package | Binary exposure | Continuous exposure | Binary outcome | Continuous outcome | Modeling | Inferences |
---|---|---|---|---|---|---|---|
C-TMLE | ctmle |
|
|
|
Double robust | Asymptotic | |
BAC | bacr, BACprior |
|
|
|
|
Outcome | Bayesian |
BP |
|
|
|
|
|
Outcome | Bayesian |
HDM | hdm |
|
|
Outcome | Asymptotic | ||
BCEE | BCEE |
|
|
|
Outcome | Bayesian | |
MADR | madr |
|
|
|
Double robust | Bootstrap | |
OAL | None
|
|
|
|
Exposure | Bootstrap | |
GLiDeR | None
|
|
|
|
Double robust | Bootstrap | |
OHAL | None
|
|
|
|
Double robust | Cross-validation | |
HD-SSL |
HDconfounding
|
|
|
|
|
Outcome | Bayesian |
In this article, we extend the BCEE algorithm in several directions. First, we generalize the algorithm to accommodate any combination of binary or continuous exposure and outcome. We also propose using a DR estimator of the causal effect within the algorithm to improve the robustness of the approach to modeling assumptions. Finally, we propose a revised implementation that substantially reduces computational time for running the algorithm. We call this extension the generalized Bayesian causal effect estimation algorithm (GBCEE). BCEE’s framework has various desirable features. First, its model selection algorithm is theoretically motivated using the graphical framework to causal inference (e.g., [21]). This algorithm favors models that include true confounders in addition to outcome risk factors, but that excludes instruments (covariates associated with the exposure, but not the outcome). Both simulation and theoretical results indicate that including outcome risk factors, in addition to true confounders, allows the unbiased estimation of causal effects with improved efficiency (reduced variance of the estimators) compared to models including all potential confounders or to models including only true confounders [4,18,22,23,24, 25,26]. Of particular interest to this work are the theoretical results of Koch et al. (2018) concerning DR estimators [18]. Moreover, BCEE takes advantage of the Bayesian model averaging framework to directly produce inferences that account for the model selection step, thus avoiding reliance on the bootstrap.
While Bayesian DR estimation of causal effects has previously been proposed, for example [27,28], GBCEE differs from previous proposals because it is specifically designed for confounder selection. The current work also shares similarities with BAC [6,15], which performs confounder selection and produces inferences that account for the model selection using a Bayesian framework. In fact, GBCEE’s algorithm is inspired by that of BAC. However, BAC targets the selection of all variables associated with either the exposure or the outcome, thus selecting instruments in addition to confounders and outcome risk factors, which may result in a loss of efficiency [24]. Moreover, GBCEE is a DR procedure, unlike BAC. Finally, GBCEE can also be seen as a special case of the general framework for MADR introduced by Cefalu et al. [17], where BCEE’s weights are used within this framework. However, there are important differentiating features between GBCEE and the specific MADR estimator they developed, including the use of the Bayesian framework to produce inferences, a different method for attributing weights to models in the model averaging procedure and a different choice of DR estimator. As will be seen in Section 4, these differences result in GBCEE being much more computationally efficient than MADR, in addition to having a noticeably lower standard error in some simulation scenarios.
The remainder of this article is structured as follows. In Section 2, we introduce some notation and further motivate the confounder selection problem. The GBCEE algorithm is presented in Section 3. Section 4 presents a simulation study to investigate and compare the finite sample properties of GBCEE. An illustration of GBCEE’s application for estimating the effect of physical activity on the risk of fractures using data from the Study of Osteoporotic Fractures is presented in Section 5. We conclude in Section 6 with a discussion of the results and perspectives for future research. All extensions presented in this article are implemented in the R package BCEE.
2 Notation and motivation
Let
For estimating the causal effect, a sample of independent and identically distributed observations
Confounder selection among the set
3 The GBCEE algorithm
We now describe the GBCEE algorithm. We first provide an overview of the algorithm in the next subsection. The construction of the prior distribution is presented in Section 3.2. This prior distribution is the same as in BCEE [5], but a brief presentation is repeated here for completeness. The DR estimators used within the algorithm are presented in Section 3.3. The implementation of the algorithm is described in Section 3.4.
3.1 Overview
GBCEE averages DR estimates of the causal effect
where
The weight attributed to each estimate
where
So far, GBCEE is thus similar to a classical Bayesian model averaging of
Before proceeding with the description of the prior distribution, we note that our procedure is only approximately Bayesian, notably because prior distributions are not specified for regression coefficients. While the core principle of the procedure, Bayesian model averaging, is Bayesian, we recognize that several aspects and approximations are borrowed from a frequentist framework. However, most of the frequentist approximations we propose making for the practical implementation of the approach are commonly used for implementing Bayesian model averaging, for example, in the R package BMA [31].
3.2 Prior distribution construction
The first step for constructing the prior distribution
where
Using the information from the first step, GBCEE’s prior distribution is defined as
Intuitively, if
We denote by
where
Classical Bayesian model averaging results ensure that the marginal likelihood of the outcome model
Now that the prior distribution of GBCEE has been presented, we turn our attention to the estimation of the causal effect for a given adjustment set
3.3 DR estimators
Targeted maximum likelihood estimation (TMLE) is a general framework for constructing DR estimators introduced by ref. [34], but based on earlier work [35]. TMLE is consistent if either the outcome or the exposure model is correctly specified and is semiparametric efficient when both are correctly specified. The augmented inverse probability weighting (AIPW) estimator that MADR and GLiDeR use share these asymptotic properties. However, it has been suggested that TMLE is more robust to near violations of the positivity assumption than AIPW [36,37], although this does not mean that TMLE is unaffected by positivity violations. For these reasons, we propose using TMLE in GBCEE to estimate
The estimators we propose using depend on the type of the outcome and of the exposure. For constructing these estimators, we use the generalized linear model (2) for the outcome with an identity link when
Briefly, TMLE first entails proposing an initial estimate for the causal contrast of interest based on the outcome model. The initial estimate is then “fluctuated” based on the output from the exposure model. The intuition is that the exposure model should contain no additional information for predicting the outcome if the initial outcome model is correctly specified. A residual association thus suggests that residual confounding is present. TMLE uses this residual association as an estimate of the residual bias of the initial estimator and modifies (fluctuates) the initial estimate in a clever way, that is, such that the estimating equations of the nonparametric efficient influence function of the causal contrast is solved. In general, these steps may need to be iterated to obtain a solution that solves the estimating equations of the efficient influence function. However, the TMLEs we use in GBCEE have been shown to converge in a single step. The influence function is a function of the observed data
We now describe the estimands we consider as well as their TMLE estimators. Because the GBCEE framework is fairly general, it could accommodate other estimands or estimators. The description below thus focuses on what is currently implemented in the R package BCEE. Remark that our proposal entails using frequentist estimates
3.3.1 Continuous or binary outcome, binary exposure
The estimands we consider are
where
The quantity
Note that
3.3.2 Continuous outcome, continuous exposure
The causal effect we consider is the slope associated with a unit increase in
where
3.3.3 Binary outcome, continuous exposure
Finally, when
where
The quantity
3.4 Implementation
Section 3.3 has provided the theoretical ingredients required for estimating the causal effect
Step 1. Determine the posterior distribution of the exposure model
Step 2. Because it is generally impossible to consider each possible outcome models individually (each
Step 3. For each
Step 4. Compute the posterior expectation
4 Simulation study
4.1 Scenarios
We consider seven main scenarios where the exposure is binary and the outcome is continuous. Some of these scenarios are then adapted to other cases. These scenarios were chosen for their diverse data generating equations and because they provide various challenges for the confounder selection methods. We focus on the binary exposure – continuous outcome case since it is the case for which the most comparators are available.
Scenario 1 is inspired by ref. [5]. The data-generating equations are
Scenario 2 is taken from ref. [4].
Scenario 3 is taken from ref. [3] and includes a large number of covariates. The data were generated as
Scenarios 4 and 5 are taken from ref. [18]. In both scenarios,
Scenario 6 features multiple weak confounders.
Scenario 7 includes an exposure–covariate interaction.
Scenarios 1, 2, 4, and 5 were adapted to the binary outcome case (Scenarios 1B, 2B, 4B, and 5B, respectively). Only the equations used for generating
In Scenarios 1–7, 2C, and 2E, both a sample size of
4.2 Analysis
For each scenario, 1,000 datasets were generated. The estimand of interest was the average treatment effect
As benchmarks, in all scenarios, we have additionally estimated
where
For all methods, only main terms of the covariates were considered, except in Scenario 7 where an
Most methods were implemented using the default options of their respective R function or package (see Table 1). For BAC, we used 500 burn-in iterations, followed by 5,000 iterations with a thinning of 5. For HD-SSL, we used 20,000 MCMC scans, 10,000 burn-in iterations, and a thinning of 10 (as in an example of its vignette). For GBCEE, we performed 2,000 iterations of the Markov chain Monte Carlo model composition algorithm. OAL and OAL-TMLE were implemented using a modified version of the code supplied in ref. [4] since the original code was not compatible with the latest version of R. All methods employed linear or logistic regression models for the outcome or exposure models. GBCEE, GLiDeR, OAL, OAL-TMLE, BAC, MADR, and HD-SSL further used parametric likelihoods for performing the model selection.
To evaluate the presence of structural positivity violations in each of Scenarios 1–7, we estimated
For each method, we computed the bias as the difference between the average of the estimates and the true effect. In Scenarios 1–7, 2C, and 2E, the true effect corresponds to the coefficient associated with
4.3 Results
According to Figure 1, the positivity assumption would be violated in Scenarios 1, 2, 3, and 5 when considering the true exposure model. Considering the target adjustment set reduces positivity issues in all scenarios: the positivity assumption becomes verified in Scenario 1, mild issues remain in Scenario 2 and 3, but a severe violation persists in Scenario 5.
The results of the simulation study for Scenarios 1–7 and
Scenario 1 | Scenario 2 | Scenario 3 | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Rel. | Rel. | Rel. | |||||||||||||
Method | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP |
Full-g | 0.00 | 0.09 | 0.10 | 1.41 | 0.96 | 0.00 | 0.09 | 0.09 | 1.06 | 0.95 | 0.01 | 0.18 | 0.18 | 1.14 | 0.95 |
Target-g | 0.00 | 0.06 | 0.07 | 1.00 | 0.96 | 0.00 | 0.08 | 0.08 | 1.00 | 0.95 | 0.00 | 0.16 | 0.16 | 1.00 | 0.95 |
Full-DR | 0.03 | 0.65 | 0.32 | 10.11 | 0.94 |
|
0.45 | 0.22 | 5.40 | 0.94 | 0.00 | 0.98 | 0.47 | 6.12 | 0.95 |
Target-DR | 0.00 | 0.06 | 0.07 | 1.00 | 0.95 | 0.00 | 0.08 | 0.06 | 1.00 | 0.86 | 0.00 | 0.16 | 0.13 | 1.00 | 0.89 |
GBCEE | 0.00 | 0.08 | 0.07 | 1.17 | 0.95 | 0.00 | 0.14 | 0.13 | 1.65 | 0.94 | 0.00 | 0.20 | 0.19 | 1.29 | 0.94 |
OAL | 0.02 | 0.08 | . | 1.34 | . | 0.24 | 0.28 | . | 4.40 | . |
|
0.30 | . | 2.84 | . |
OAL-TMLE | 0.00 | 0.08 | . | 1.20 | . | 0.00 | 0.14 | . | 1.71 | . |
|
0.25 | . | 2.68 | . |
C-TMLE | 0.01 | 0.10 | 0.06 | 1.50 | 0.87 | 0.02 | 0.11 | 0.07 | 1.39 | 0.77 |
|
0.28 | 0.14 | 1.78 | 0.65 |
GLiDeR | 0.01 | 0.08 | . | 1.18 | . | 0.00 | 0.12 | . | 1.38 | . |
|
0.21 | . | 1.45 | . |
BAC | 0.00 | 0.09 | 0.09 | 1.42 | 0.96 | 0.00 | 0.09 | 0.09 | 1.06 | 0.95 | 0.00 | 0.18 | 0.17 | 1.11 | 0.92 |
MADR | 0.01 | 0.07 | . | 1.05 | . | 0.00 | 0.15 | . | 1.75 | . | 0.00 | 0.22 | . | 1.41 | . |
BP | 0.00 | 0.09 | 0.10 | 1.42 | 0.95 | 0.00 | 0.09 | 0.09 | 1.06 | 0.95 | 0.00 | 0.18 | 0.17 | 1.13 | 0.94 |
HDM | 0.00 | 0.09 | 0.10 | 1.41 | 0.96 | 0.00 | 0.09 | 0.09 | 1.06 | 0.95 | 0.00 | 0.18 | 0.17 | 1.10 | 0.94 |
HD-SSL | 0.01 | 0.08 | 0.08 | 1.21 | 0.95 | 0.00 | 0.09 | 0.09 | 1.03 | 0.96 |
|
0.17 | 0.17 | 1.05 | 0.95 |
Scenario 4 | Scenario 5 | Scenario 6 | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Full-g |
|
1.67 | 1.81 | 1.96 | 0.26 | 0.00 | 0.10 | 0.10 | 1.02 | 0.95 |
|
0.07 | 0.07 | 1.00 | 0.95 |
Target-g |
|
1.67 | 1.81 | 1.96 | 0.26 | 0.00 | 0.09 | 0.10 | 1.00 | 0.95 |
|
0.07 | 0.07 | 1.00 | 0.95 |
Full-DR |
|
2.51 | 2.29 | 1.00 | 0.93 | 0.00 | 0.13 | 0.12 | 1.41 | 0.92 |
|
0.07 | 0.07 | 1.02 | 0.95 |
Target-DR |
|
2.51 | 2.29 | 1.00 | 0.93 | 0.00 | 0.09 | 0.06 | 1.00 | 0.76 |
|
0.07 | 0.07 | 1.00 | 0.96 |
GBCEE | 0.07 | 2.51 | 2.33 | 1.00 | 0.91 | 0.00 | 0.11 | 0.09 | 1.20 | 0.88 | 0.02 | 0.07 | 0.07 | 1.08 | 0.93 |
OAL |
|
5.97 | . | 2.38 | 0.89 | 0.12 | 0.23 | . | 2.76 | 0.94 | 0.04 | 0.07 | . | 1.29 | . |
OAL-TMLE |
|
2.51 | . | 1.00 | 0.90 | 0.02 | 0.12 | . | 1.25 | 0.96 | 0.04 | 0.07 | . | 1.29 | . |
C-TMLE | 2.45 | 1.10 | 1.13 | 1.07 | 0.39 | 0.02 | 0.17 | 0.08 | 1.78 | 0.72 | 0.01 | 0.07 | 0.06 | 1.03 | 0.92 |
GLiDeR | 0.01 | 3.34 | . | 1.33 | 0.89 | 0.02 | 0.13 | . | 1.41 | 0.95 | 0.02 | 0.07 | . | 1.05 | . |
BAC |
|
1.67 | 1.81 | 1.97 | 0.24 | 0.00 | 0.10 | 0.10 | 1.02 | 0.95 | 0.03 | 0.07 | 0.06 | 1.09 | 0.90 |
MADR | 0.02 | 3.22 | . | 1.28 | 0.90 | 0.03 | 0.26 | . | 2.82 | 0.95 | 0.03 | 0.07 | . | 1.15 | . |
BP |
|
1.67 | 1.81 | 1.96 | 0.26 | 0.00 | 0.10 | 0.10 | 1.02 | 0.95 |
|
0.07 | 0.07 | 1.00 | 0.95 |
HDM |
|
1.67 | 1.67 | 1.96 | 0.20 | 0.00 | 0.10 | 0.10 | 1.02 | 0.95 |
|
0.07 | 0.07 | 1.02 | 0.94 |
HD-SSL |
|
1.61 | 1.81 | 1.92 | 0.26 | 0.01 | 0.10 | 0.10 | 1.09 | 0.92 | 0.01 | 0.07 | 0.07 | 1.02 | 0.94 |
Scenario 7 | |||||
---|---|---|---|---|---|
Full-g | 0.00 | 0.09 | . | 1.02 | . |
Target-g | 0.00 | 0.08 | . | 1.00 | . |
Full-DR | 0.00 | 0.09 | 0.09 | 1.10 | 0.94 |
Target-DR | 0.00 | 0.08 | 0.08 | 1.02 | 0.94 |
GBCEE |
|
0.09 | 0.10 | 1.07 | 0.96 |
OAL | 0.03 | 0.12 | . | 1.45 | . |
OAL-TMLE |
|
0.09 | . | 1.10 | . |
C-TMLE |
|
0.09 | 0.09 | 1.06 | 0.96 |
GLiDeR |
|
0.09 | . | 1.10 | . |
BAC | 0.00 | 0.08 | 0.08 | 0.98 | 0.96 |
MADR |
|
0.09 | . | 1.11 | . |
BP |
|
0.09 | 0.08 | 3.36 | 0.12 |
HDM |
|
0.09 | 0.09 | 3.36 | 0.15 |
HD-SSL |
|
0.10 | 0.09 | 3.36 | 0.15 |
Conditional instruments are present in Scenario 1, covariates are strongly correlated in Scenario 2, a large number of covariates are present in Scenario 3, the true outcome model includes nonlinear terms in Scenario 4, the true exposure model includes nonlinear terms in Scenario 5, a large number of weak confounders are present in scenario 6, and an exposure–covariate interaction is present in Scenario 7.
First, we notice that the lowest bias, variance, and RMSE are generally achieved by both target-g and target-DR, except in Scenario 4 where the g-formula performs poorly because of the outcome model misspecification. However, recall that these estimators are only considered as benchmarks. Second, we remark that the variance and RMSE of full-g and full-DR are greater than their target counterparts. The increase in the variance and RMSE is particularly pronounced for the DR benchmark estimator. These results are important to keep in mind when considering those for the variable selection methods: the double robustness property, which protects against bias due to model misspecification, may come at the cost of increased variance when spurious variables are included. This is likely due, at least partly, to the sensitivity of DR methods to positivity violations.
Because BAC, BP, HDM, and HD-SSL are outcome modeling-based variable selection methods, it is not surprising that they generally perform better than the DR variable selection methods (GBCEE, OAL-TMLE, C-TMLE, GLiDeR, and MADR) in Scenarios 1–3, 5, and 6, where the correct outcome model is among the considered models. However, substantial bias is observed for BAC, BP, HDM, and HD-SSL in Scenario 4, and for BP, HDM, and HD-SSL in Scenario 7, that is, when the outcome model is not among the considered models. In addition, BAC, BP, and HDM offer no substantial RMSE reduction as compared to full-g when
When comparing together the DR methods, we first notice that GBCEE, GLiDeR, and MADR yield estimates with little or no bias in all scenarios when
The coverage probabilities of 95% confidence intervals for BP, HDM, and HD-SSL are close to their nominal level in all scenarios, except Scenario 4 and 7 where the correct outcome model is not among the considered models. Similarly, the coverage of BAC is good in all scenarios except Scenario 4. The coverage probability was much lower than 95% in all scenarios for C-TMLE. This is because the estimated standard error largely underestimates the true standard error, except in Scenario 4, where bias is responsible for the undercoverage. For GBCEE, the coverage was close to 95% in all scenarios, except in Scenario 5 where the coverage probability was 87%. This is likely due to the inconsistency of the empirical efficient influence function variance estimator when the exposure model is misspecified, as evidenced by the absence of notable bias and a slight, but notable, underestimation of the standard error. To verify this assumption, we ran again the simulation for GBCEE using a nonparametric bootstrap variance estimator with 200 replicates to estimate the within-model variance. A 94% coverage was then achieved. Similarly, in scenarios with
Plots of the inclusion probability of covariates are available in Appendix D. These figures reveal that GBCEE, OAL, and GLiDeR have relatively similar behaviors, having large probabilities of including both true confounders and outcome risk factors, and lower probabilities of including the other variables. MADR also has a similar behavior regarding the inclusion of variables in the outcome model, but lower overall probabilities of inclusion of variables in the exposure model. HD-SSL also had a somewhat similar behavior, but with much lower probabilities of inclusion of all variables. In general, the variables selected were closer to the target for
The results of two additional simulation scenarios are worth highlighting. In Scenario 2D, where the exposure is continuous and the outcome is binary, target-DR, full-DR, and GBCEE have a small but notable bias, and their mean estimated standard error is much larger than the standard deviation of the estimates. The latter seems to be due to extreme values in the estimated standard error since the median of the estimated standard error is smaller than the standard deviation of the estimates (result not shown). The coverage of their confidence intervals is also poor. We hypothesize that this is due to the strong relation between confounders and exposure that can cause the weighting component of the DR estimator to take extreme values. This scenario showcases the challenge of using DR estimators with continuous exposures. The other additional scenario whose results are particularly noteworthy is Scenario 2E, where the outcome error distribution is exponential. The results of this scenario were very similar to those of Scenario 2, even for methods that assumed a normal distribution for the outcome (BAC, MADR, GLiDeR, and GBCEE).
We evaluated the computational time of the variable selection methods in one replication of Scenario 1 with
5 Application for estimating the effect of physical activity on the risk of fractures
Osteoporosis is a disease where bones become fragile because of low bone mineral density or because of the deterioration of bone architecture. It is a common disease with a prevalence of 5% in men and 25% in women aged 65 years and older from the United States [44]. The prevalence of osteoporosis increases with age. Because of the bone fragility, individuals with osteoporosis are at higher risk of fractures. In addition to important pain, osteoporotic fractures may also induce a loss of mobility and quality of life, additional morbidities, and premature mortality. Preventing osteoporosis and the related fractures has thus become a public health priority. Physical activity may help prevent osteoporosis and osteoporotic fractures. Indeed, regular physical activity is associated with increased bone mineral density and lower risk of falls [45,46].
To illustrate the use of GBCEE, we estimated the effect of attaining the physical activity recommendations from the World Health Organization on the 5-year risk of fracture in older women using the publicly available limited data set from the study of osteoporotic fractures [47]. The study of osteoporotic fractures is a multicentric population based cohort study that recruited women aged 65 or older in four urban regions in the United states: Baltimore, Pittsburgh, Minneapolis, and Portland. An ethical exemption was obtained from the CHU de Québec – Université Laval Ethics Board (# 2020-4788).
In this application, we consider data on 9,671 white women enrolled in 1986. Subjects were considered as exposed if they spent 7.5 kilocalories per kilogram of body mass from moderate or high intensity physical activity per week, and unexposed otherwise. The outcome of interest was the occurrence of any hip or upper leg fracture in the 5 years following the baseline interview. A very rich set of 55 measured potential confounders were identified based on substantive knowledge and notably include data on age, ethnic origin, body mass index, smoking, alcohol use, education, various drug use, fall history, self and familial history of fracture, physical activity history when a teenager and during adulthood, fear of falling, and several health conditions (see Appendix E, Table A13).
Appendix E presents a comparison of the baseline characteristics of participants according to exposure to physical activity recommendations. Among others, subjects who were unexposed had a greater body mass index, were older, had less years of education, were more afraid of falling, practiced less physical activity in the past, drank less alcohol, had more difficulty walking, rated their overall health more poorly, and were less likely to still be married.
Only 3,458 participants had complete information for all variables. Most variables had only few missing data (
The fully adjusted AIPW yielded a 5-year risk difference of fracture of
6 Discussion
We have presented a generalization of the BCEE algorithm to estimate the causal effect of a binary or continuous exposure on a binary or continuous outcome using DR targeted maximum likelihood estimators. Like the original BCEE, GBCEE is theoretically motivated using the graphical framework to causal inference. It aims to select adjustment covariates such that the final estimator is unbiased and has reduced variance as compared to an estimator adjusting for all covariates. In addition, the Bayesian framework allows producing inferences that account for the model selection step in a principled way. We have also proposed an implementation of GBCEE that is more computationally efficient than that of BCEE.
We have compared GBCEE to alternative model selection methods in a simulation study. GBCEE produced estimates with little or no bias in all scenarios, whether the exposure and outcome were continuous or binary, and as long as either a correctly specified exposure or outcome model was available for selection. In addition, GBCEE’s estimates had improved precision as compared to a fully adjusted DR estimator in most scenarios. The only situation where no improvement in precision was observed was when unbiased estimation required adjusting for all potential confounders; hence, no variance reduction was possible. The good performance of GBCEE across simulation scenarios, even with a relatively small sample size of
The relative performance of the estimators was observed to depend on both the data-generating mechanism and sample size, but GBCEE often had the lowest or second lowest RMSE among DR methods. In large samples, GBCEE, OAL, and GLiDeR are all expected to select true confounders and outcome risk factors. MADR is also expected to select such variables in its outcome model, but only true confounders in the exposure model. This difference perhaps explains why MADR performed more poorly than GBCEE, OAL-TMLE, and GLiDeR in simulation Scenario 5. In a simulation study with scenarios inspired by large healthcare databases and with binary outcomes, GBCEE was one of the best performing approach in terms of bias and RMSE, together with BAC and Scalable C-TMLE, whereas GLiDeR performed very poorly in some scenarios [48]. It might be interesting for future studies to specifically investigate the factors that affect the efficiency of the different model selection methods and to provide insights regarding the situations where one method should be expected to outperform the others. Our simulation study results also indicated that outcome model-based algorithms outperformed DR methods in scenarios where the correct outcome model was within the considered model space. However, the former methods produced estimates with high bias when the correct outcome model was not in the model space, unlike the latter. Model misspecifications are likely to be common in practice, and using DR approaches may thus be preferred. Although we did not consider BCEE in our simulation study, since it is an outcome model-based algorithm, BCEE is sensitive to outcome model misspecifications. GBCEE, like most other methods we considered, employs parametric likelihoods for performing the variable selection. In a simulation scenario where the outcome error distribution was misspecified, we observed that results were practically unaffected for all methods. Moreover, some theoretical results indicate that Bayesian posteriors asymptotically concentrate on the model closest to the truth (according to the Kullback-Leibler divergence) when none of the considered model is correctly specified [49,50]. This suggests that even if none of the models considered is correctly specified, GBCEE’s model selection should converge to the best available approximation. Our simulation study also showcased the challenge of DR estimation with a continuous exposure. If DR estimation of the effect of a continuous exposure is attempted in practice, we warn analysts to be cautious of the potential influence of extreme weights on their results.
From a computational perspective, GBCEE produced both an estimate and inferences more quickly than the other DR algorithms when the empirical influence function based within-model variance estimator was used. This variance estimator produced valid inferences in most scenarios. Two exceptions were when the exposure model was misspecified and when the sample size was small. As such, it is advisable to use the bootstrap within-model variance estimator for inferences when the sample size is moderate or small. In such situations, using the bootstrap variance estimator is unlikely to be computationally prohibitive. The increased computational efficiency of GBCEE is particularly relevant when comparing with MADR, since GBCEE fits within the same general MADR framework that was developed by Cefalu et al. [17].
Regarding inferences for model selection procedures, it is unclear if employing the usual nonparametric bootstrap is appropriate. Indeed, model selection may yield estimators that lack the smoothness required for the bootstrap to be valid. This may be the reason why the coverage of 95% confidence intervals was around 90% for MADR and GLiDeR in Scenario 4 of our simulation study. A smoothed-bootstrap method has been proposed specifically for the variable selection problem [51] and was used by Shortreed and Erterfaie [4]. We have explored using this approach in our simulation study, but we have observed that a very large number of replications (5,000–10,000) was necessary to achieve appropriate variance estimation. Shortreed and Ertefaie [4] based their inferences on 10,000 replications. Overall, we believe that inference procedures for variable selection methods require further investigation. We note that GBCEE does not share this limitation when the bootstrap variance estimator is used, since the bootstrap is only used to estimate the within-model variance; the between-model variance remains estimated using the Bayesian machinery.
We have also illustrated the use of GBCEE for estimating the effect of reaching physical activity recommendations on the risk of osteoporotic fractures among older women. In this example, the estimate produced by GBCEE was similar to that of the fully adjusted DR estimator, but the confidence interval was much narrower. Since resources available for research are limited, it is important to make the most out of the available data. Confounder selection methods should be considered as a valuable tool for achieving this, especially when there are multiple potential confounders or when it is unknown whether some covariates are confounders, risk factors for the outcome, or instruments.
Model selection in causal inference is a rapidly growing field. It was thus impossible for us to compare our proposed method to all possible alternatives in the simulation study. For example, Antonelli et al. have introduced a general framework for Bayesian DR estimation of causal effects with a focus on high-dimensional settings [28]. We considered including their method in our simulation study, but we ran into multiple errors when trying to use their R package. However, we hypothesize that GBCEE is more efficient since it connects the exposure and outcome modeling steps to improve variable selection, contrary to Antonelli et al.’s proposal. We could not include the OHAL either in our simulation study because of the excessive computing time required to obtain even only point estimates for this method in the scenarios we considered.
There are multiple directions in which the current work could be extended. For instance, to the best of our knowledge, no confounder selection method is applicable to the censored time-to-event outcome case yet. To reduce the risk of model misspecifications, it would also be possible to employ machine learning within GBCEE. In fact, TMLE is often paired with Super Learner, an ensemble method that combines the predictions from multiple methods using a cross-validation procedure [37]. A simple possibility for doing so would entail computing the prior and posterior probability of a given adjustment set
-
Funding information: This work was supported by grants from the Natural Sciences and Engineering Research Council of Canada and the Fonds de recherche du Québec - Santé. D Talbot is a Fonds de recherche du Québec - Santé Chercheur-Boursier. The Study of Osteoporotic Fractures (SOF) was supported by National Institutes of Health funding. The National Institute on Aging (NIA) provided support under the following grant numbers: R01 AG005407, R01 AR35582, R01 AR35583, R01 AR35584, R01 AG005394, R01 AG027574, and R01 AG027576.
-
Conflict of interest: The authors state no conflict of interest.
-
Data availability statement: The R code required to reproduce the simulation study is available at https://github.com/detal9/GBCEE. The public access data from the Study of Osteoporotic Fractures is available at https://sofonline.ucsf.edu/.
Appendix A Asymptotic properties of the maximum likelihood estimation of the prior’s parameters
We provide greater details on the properties of the estimated prior distribution. Denote by
B Detailed presentation of GBCEE’s implementation
In this Appendix, we provide further details concerning Steps 2 and 3 of our proposed implementation of GBCEE.
Step 2. Use a Markov chain Monte Carlo model composition algorithm to explore the outcome model space
Recalling that
Equation (A1) can be expensive to compute. Indeed, evaluating the value of each term of the product
Under the assumption that the exposure model is correctly specified, remark that the approximation (A2) has the following form asymptotically:
because the posterior mass of
Since
Step 3. For each
The posterior probability of each outcome model
because
B.1 Simulation study comparing oracle GBCEE to MLE GBCEE
Implementing the oracle GBCEE is challenging because it requires calculating the true regression parameters of multiple regression models. As a consequence, we only examined three simple scenarios.
In Scenario 1S,
For each scenario, we generated 1,000 samples of sizes
Method |
|
|
|
|
|
---|---|---|---|---|---|
Bias | |||||
Oracle | 0.00 | 0.03 | 0.01 | 0.00 | 0.00 |
MLE | 0.00 | 0.03 | 0.02 | 0.01 | 0.00 |
Standard deviation | |||||
Oracle | 0.49 | 0.32 | 0.22 | 0.16 | 0.07 |
MLE | 0.49 | 0.32 | 0.22 | 0.16 | 0.07 |
Coverage | |||||
Oracle | 0.95 | 0.94 | 0.94 | 0.94 | 0.95 |
MLE | 0.95 | 0.94 | 0.93 | 0.94 | 0.94 |
Method |
|
|
|
|
|
---|---|---|---|---|---|
Bias | |||||
Oracle | 0.03 | 0.04 | 0.06 | 0.05 | 0.03 |
MLE | 0.03 | 0.05 | 0.07 | 0.06 | 0.03 |
Standard deviation | |||||
Oracle | 0.54 | 0.33 | 0.23 | 0.16 | 0.08 |
MLE | 0.54 | 0.33 | 0.23 | 0.16 | 0.08 |
Coverage | |||||
Oracle | 0.94 | 0.94 | 0.94 | 0.94 | 0.92 |
MLE | 0.93 | 0.94 | 0.94 | 0.92 | 0.90 |
Method |
|
|
|
|
|
---|---|---|---|---|---|
Bias | |||||
Oracle |
|
0.00 | 0.00 |
|
0.00 |
MLE |
|
0.01 | 0.01 | 0.00 | 0.00 |
Standard deviation | |||||
Oracle | 0.52 | 0.31 | 0.22 | 0.15 | 0.07 |
MLE | 0.53 | 0.32 | 0.22 | 0.16 | 0.07 |
Coverage | |||||
Oracle | 0.94 | 0.95 | 0.95 | 0.95 | 0.94 |
MLE | 0.94 | 0.94 | 0.95 | 0.94 | 0.94 |
C Additional simulation results
Scenario 1 | Scenario 2 | Scenario 3 | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Rel. | Rel. | Rel. | |||||||||||||
Method | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP |
Full-g |
|
0.24 | 0.24 | 1.49 | 0.95 |
|
0.21 | 0.21 | 1.10 | 0.95 | 0.00 | 0.54 | 0.53 | 1.52 | 0.95 |
Target-g | 0.00 | 0.16 | 0.16 | 1.00 | 0.95 |
|
0.19 | 0.19 | 1.00 | 0.95 |
|
0.35 | 0.35 | 1.00 | 0.94 |
Full-DR | 1.13 | 31.36 | 4.60 | 198.12 | 0.82 |
|
2.27 | 0.55 | 12.01 | 0.88 | . | . | . | . | . |
Target-DR | 0.00 | 0.16 | 0.15 | 1.00 | 0.93 |
|
0.19 | 0.14 | 1.00 | 0.85 |
|
0.35 | 0.29 | 1.00 | 0.88 |
GBCEE | 0.02 | 0.20 | 0.19 | 1.29 | 0.93 |
|
0.26 | 0.24 | 1.39 | 0.91 |
|
0.48 | 0.34 | 1.41 | 0.79 |
OAL | 0.05 | 0.20 | . | 1.29 | . | 0.43 | 0.43 | . | 3.20 | . |
|
0.60 | . | 1.84 | . |
OAL-TMLE | 0.00 | 0.19 | . | 1.22 | . | 0.01 | 0.30 | . | 1.56 | . |
|
0.46 | . | 1.55 | . |
C-TMLE | 0.06 | 0.28 | 0.13 | 1.82 | 0.74 | 0.05 | 0.28 | 0.14 | 1.48 | 0.70 | 0.02 | 1.08 | 0.26 | 3.05 | 0.56 |
GLiDeR | 0.03 | 0.18 | . | 1.13 | . | 0.01 | 0.24 | . | 1.25 | . |
|
0.38 | . | 1.18 | . |
BAC | 0.01 | 0.22 | 0.21 | 1.38 | 0.94 |
|
0.20 | 0.19 | 1.09 | 0.94 |
|
0.47 | 0.45 | 1.32 | 0.94 |
MADR | 0.03 | 0.17 | . | 1.12 | . |
|
0.40 | . | 2.13 | . |
|
0.48 | . | 1.41 | . |
BP | 0.02 | 0.22 | 0.20 | 1.42 | 0.92 |
|
0.21 | 0.20 | 1.10 | 0.94 | 0.01 | 0.45 | 0.37 | 1.28 | 0.93 |
HDM |
|
0.18 | 0.18 | 1.14 | 0.95 |
|
0.20 | 0.20 | 1.07 | 0.95 |
|
0.36 | 0.36 | 1.01 | 0.95 |
HD-SSL | 0.01 | 0.17 | 0.18 | 1.07 | 0.96 | 0.01 | 0.22 | 0.20 | 1.14 | 0.93 |
|
0.39 | 0.39 | 1.10 | 0.94 |
Scenario 4 | Scenario 5 | Scenario 6 | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Full-g |
|
3.82 | 4.07 | 1.15 | 0.83 |
|
0.22 | 0.22 | 1.03 | 0.95 |
|
0.16 | 0.16 | 1.03 | 0.95 |
Target-g |
|
3.82 | 4.07 | 1.15 | 0.83 |
|
0.21 | 0.22 | 1.00 | 0.95 |
|
0.16 | 0.16 | 1.00 | 0.95 |
Full-DR | 0.23 | 5.30 | 4.53 | 1.00 | 0.94 |
|
1.53 | 0.39 | 7.14 | 0.86 | 0.00 | 0.18 | 0.16 | 1.11 | 0.92 |
Target-DR | 0.23 | 5.30 | 4.53 | 1.00 | 0.94 |
|
0.21 | 0.12 | 1.00 | 0.73 |
|
0.16 | 0.14 | 1.00 | 0.91 |
GBCEE | 0.47 | 5.36 | 4.47 | 1.01 | 0.92 | 0.01 | 0.26 | 0.22 | 1.23 | 0.85 | 0.03 | 0.16 | 0.15 | 1.04 | 0.92 |
OAL | 1.09 | 12.00 | . | 2.27 | 0.84 | 0.25 | 0.39 | . | 2.17 | 0.92 | 0.06 | 0.16 | . | 1.10 | . |
OAL-TMLE | 0.22 | 5.29 | . | 1.00 | 0.87 | 0.02 | 0.27 | . | 1.24 | 0.97 | 0.05 | 0.16 | . | 1.08 | . |
C-TMLE | 2.50 | 2.79 | 2.49 | 0.71 | 0.78 | 0.05 | 0.40 | 0.16 | 1.87 | 0.66 | 0.03 | 0.16 | 0.13 | 1.02 | 0.89 |
GLiDeR | 0.21 | 5.43 | . | 1.02 | 0.90 | 0.04 | 0.23 | . | 1.10 | 0.96 | 0.05 | 0.17 | . | 1.08 | . |
BAC |
|
3.82 | 4.09 | 1.13 | 0.84 | 0.01 | 0.22 | 0.22 | 1.02 | 0.95 |
|
0.16 | 0.13 | 0.99 | 0.87 |
MADR | 0.28 | 5.13 | . | 0.97 | 0.90 | 0.01 | 0.65 | . | 3.04 | 0.97 | 0.05 | 0.16 | . | 1.06 | . |
BP |
|
3.81 | 4.07 | 1.15 | 0.83 |
|
0.22 | 0.22 | 1.03 | 0.95 |
|
0.16 | 0.15 | 1.02 | 0.94 |
HDM |
|
3.82 | 3.74 | 1.15 | 0.76 |
|
0.22 | 0.22 | 1.03 | 0.95 |
|
0.16 | 0.15 | 1.04 | 0.92 |
HD-SSL |
|
3.79 | 4.04 | 1.09 | 0.84 | 0.03 | 0.22 | 0.22 | 1.02 | 0.94 | 0.03 | 0.15 | 0.15 | 0.98 | 0.94 |
Scenario 7 | |||||
---|---|---|---|---|---|
Full-g |
|
0.19 | . | 1.02 | . |
Target-g |
|
0.18 | . | 1.00 | . |
Full-DR |
|
0.21 | 0.20 | 1.14 | 0.94 |
Target-DR |
|
0.19 | 0.19 | 1.05 | 0.94 |
GBCEE |
|
0.22 | 0.22 | 1.20 | 0.95 |
OAL | 0.08 | 0.25 | . | 1.46 | . |
OAL-TMLE |
|
0.22 | . | 1.21 | . |
C-TMLE |
|
0.22 | 0.18 | 1.23 | 0.89 |
GLiDeR |
|
0.22 | . | 1.20 | . |
BAC | 0.01 | 0.19 | 0.19 | 1.07 | 0.96 |
MADR |
|
0.22 | . | 1.22 | . |
BP |
|
0.20 | 0.19 | 1.82 | 0.69 |
HDM |
|
0.20 | 0.20 | 1.83 | 0.73 |
HD-SSL |
|
0.20 | 0.19 | 1.81 | 0.71 |
Conditional instruments are present in Scenario 1, covariates are strongly correlated in Scenario 2, a large number of covariates are present in Scenario 3, the true outcome model includes nonlinear terms in Scenario 4, the true exposure model includes nonlinear terms in Scenario 5, a large number of weak confounders are present in scenario 6, and an exposure–covariate interaction is present in Scenario 7.
Scenario 1B | Scenario 2B | |||||||||
---|---|---|---|---|---|---|---|---|---|---|
Rel. | Rel. | |||||||||
Method | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP |
Full-g | 0.00 | 0.04 | . | 1.46 | . | 0.00 | 0.04 | . | 1.11 | . |
Target-g | 0.00 | 0.03 | . | 1.00 | . | 0.00 | 0.04 | . | 1.00 | . |
Full-DR | 0.00 | 0.58 | 0.18 | 19.36 | 0.93 |
|
0.09 | 0.07 | 2.67 | 0.93 |
Target-DR | 0.00 | 0.03 | 0.03 | 1.01 | 0.95 | 0.00 | 0.05 | 0.04 | 1.33 | 0.93 |
GBCEE | 0.00 | 0.04 | 0.04 | 1.27 | 0.94 | 0.00 | 0.05 | 0.06 | 1.54 | 0.93 |
OAL | 0.01 | 0.04 | . | 1.49 | . | 0.04 | 0.07 | . | 2.18 | . |
OAL-TMLE | 0.00 | 0.05 | . | 1.52 | . | 0.00 | 0.06 | . | 1.66 | . |
C-TMLE | 0.01 | 0.05 | 0.03 | 1.57 | 0.80 | 0.04 | 0.08 | 0.08 | 2.39 | 0.53 |
BAC | 0.00 | 0.04 | 0.04 | 1.45 | 0.91 |
|
0.04 | 0.04 | 1.10 | 0.91 |
Scenario 4B | Scenario 5B | |||||||||
---|---|---|---|---|---|---|---|---|---|---|
Full-g | 0.00 | 0.02 | . | 1.02 | . | 0.00 | 0.04 | . | 1.04 | . |
Target-g | 0.00 | 0.02 | . | 1.02 | . | 0.00 | 0.04 | . | 1.00 | . |
Full-DR |
|
0.02 | 0.02 | 1.00 | 0.94 | 0.00 | 0.05 | 0.06 | 1.43 | 0.96 |
Target-DR |
|
0.02 | 0.02 | 1.00 | 0.94 | 0.00 | 0.04 | 0.04 | 1.14 | 0.93 |
GBCEE |
|
0.02 | 0.02 | 1.00 | 0.94 | 0.00 | 0.04 | 0.05 | 1.10 | 0.96 |
OAL | 0.00 | 0.02 | . | 1.06 | . | 0.01 | 0.08 | . | 2.13 | . |
OAL-TMLE | 0.00 | 0.02 | . | 1.02 | . | 0.00 | 0.04 | . | 1.14 | . |
C-TMLE | 0.00 | 0.02 | 0.02 | 1.00 | 0.96 | 0.00 | 0.05 | 0.04 | 1.04 | 0.94 |
BAC | 0.00 | 0.02 | 0.02 | 1.02 | 0.95 | 0.00 | 0.04 | 0.04 | 1.04 | 0.94 |
Conditional instruments are present in Scenario 1B, covariates are strongly correlated in Scenario 2B, the true outcome model includes nonlinear terms in Scenario 4B, and the true exposure model includes nonlinear terms in Scenario 5B.
Rel. | |||||
---|---|---|---|---|---|
Method | Bias | SD | ESE | RMSE | CP |
Full-g |
|
0.08 | 0.07 | 1.63 | 0.95 |
Target-g |
|
0.05 | 0.05 | 1.00 | 0.96 |
Full-DR |
|
0.08 | 0.07 | 1.63 | 0.91 |
Target-DR |
|
0.05 | 0.05 | 1.00 | 0.96 |
GBCEE |
|
0.05 | 0.05 | 1.18 | 0.94 |
BAC |
|
0.07 | 0.07 | 1.62 | 0.94 |
BP |
|
0.07 | 0.07 | 1.62 | 0.93 |
HD-SSL | 0.02 | 0.07 | 0.07 | 1.49 | 0.94 |
Rel. | |||||
---|---|---|---|---|---|
Method | Bias | SD | ESE | RMSE | CP |
Full-g |
|
0.03 | 0.03 | 1.61 | 0.95 |
Target-g | 0.00 | 0.02 | 0.02 | 1.00 | 0.95 |
Full-DR |
|
0.03 | 0.03 | 1.61 | 0.95 |
Target-DR | 0.00 | 0.02 | 0.02 | 1.00 | 0.95 |
GBCEE |
|
0.02 | 0.02 | 1.09 | 0.94 |
BAC |
|
0.03 | 0.03 | 1.61 | 0.94 |
BP |
|
0.03 | 0.03 | 1.61 | 0.95 |
HD-SSL | 0.00 | 0.03 | 0.03 | 1.28 | 0.96 |
Rel. | |||||
---|---|---|---|---|---|
Method | Bias | SD | ESE | RMSE | CP |
Full-g | 0.01 | 0.19 | . | 1.23 | . |
Target-g | 0.01 | 0.16 | . | 1.00 | . |
Full-DR | 0.03 | 0.23 | 0.46 | 1.48 | 0.86 |
Target-DR | 0.02 | 0.19 | 0.34 | 1.24 | 0.92 |
GBCEE | 0.03 | 0.22 | 0.33 | 1.43 | 0.81 |
Rel. | |||||
---|---|---|---|---|---|
Method | Bias | SD | ESE | RMSE | CP |
Full-g | 0.01 | 0.20 | 0.21 | 1.11 | 0.96 |
Target-g | 0.00 | 0.18 | 0.19 | 1.00 | 0.95 |
Full-DR | 0.14 | 3.97 | 0.53 | 21.75 | 0.87 |
Target-DR | 0.00 | 0.18 | 0.14 | 1.00 | 0.87 |
GBCEE |
|
0.27 | 0.22 | 1.45 | 0.91 |
OAL | 0.44 | 0.44 | . | 1.45 | . |
OAL-TMLE | 0.00 | 0.30 | . | 1.44 | . |
C-TMLE | 0.07 | 0.25 | 0.13 | 1.44 | 0.72 |
GLiDeR | 0.01 | 0.22 | . | 1.19 | . |
BAC |
|
0.19 | 0.19 | 1.07 | 0.95 |
MADR |
|
0.29 | . | 1.56 | . |
BP | 0.00 | 0.20 | 0.20 | 1.10 | 0.96 |
HDM | 0.01 | 0.20 | 0.20 | 1.07 | 0.96 |
HD-SSL | 0.01 | 0.20 | 0.20 | 1.09 | 0.94 |
Rel. | |||||
---|---|---|---|---|---|
Method | Bias | SD | ESE | RMSE | CP |
Full-g |
|
0.09 | 0.09 | 1.09 | 0.95 |
Target-g |
|
0.08 | 0.08 | 1.00 | 0.94 |
Full-DR |
|
0.29 | 0.18 | 3.48 | 0.94 |
Target-DR |
|
0.08 | 0.06 | 1.00 | 0.86 |
GBCEE |
|
0.13 | 0.12 | 1.56 | 0.95 |
OAL | 0.27 | 0.27 | . | 4.65 | . |
OAL-TMLE | 0.00 | 0.14 | . | 1.66 | . |
C-TMLE | 0.02 | 0.11 | 0.07 | 1.34 | 0.79 |
GLiDeR | 0.00 | 0.11 | . | 1.33 | . |
BAC |
|
0.09 | 0.09 | 1.09 | 0.95 |
MADR |
|
0.14 | . | 1.71 | . |
BP |
|
0.09 | 0.09 | 1.09 | 0.95 |
HDM |
|
0.09 | 0.09 | 1.09 | 0.94 |
HD-SSL | 0.00 | 0.08 | 0.09 | 1.02 | 0.95 |
Scenario 1 | Scenario 2 | Scenario 3 | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Rel. | Rel. | Rel. | |||||||||||||
Method | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP |
|
0.01 | 0.19 | 0.17 | 1.22 | 0.92 |
|
0.28 | 0.23 | 1.48 | 0.89 |
|
0.50 | 0.34 | 1.45 | 0.78 |
|
0.02 | 0.20 | 0.19 | 1.29 | 0.93 |
|
0.29 | 0.24 | 1.53 | 0.91 |
|
0.48 | 0.34 | 1.41 | 0.79 |
|
0.01 | 0.22 | 0.20 | 1.39 | 0.93 |
|
0.29 | 0.24 | 1.53 | 0.90 |
|
0.50 | 0.35 | 1.45 | 0.80 |
w/o model averaging | 0.01 | 0.22 | 0.17 | 1.39 | 0.86 |
|
0.29 | 0.22 | 1.53 | 0.86 |
|
0.50 | 0.34 | 1.45 | 0.77 |
Scenario 4 | Scenario 5 | Scenario 6 | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
0.30 | 5.69 | 4.47 | 1.07 | 0.91 | 0.02 | 0.25 | 0.20 | 1.18 | 0.85 | 0.04 | 0.16 | 0.15 | 1.05 | 0.92 |
|
0.47 | 5.36 | 4.47 | 1.01 | 0.92 | 0.01 | 0.26 | 0.22 | 1.23 | 0.85 | 0.03 | 0.16 | 0.15 | 1.04 | 0.92 |
|
0.29 | 5.69 | 4.47 | 1.08 | 0.91 | 0.02 | 0.26 | 0.22 | 1.23 | 0.86 | 0.03 | 0.16 | 0.15 | 1.07 | 0.92 |
w/o model averaging | 0.28 | 5.70 | 4.44 | 1.08 | 0.90 | 0.02 | 0.26 | 0.18 | 1.23 | 0.80 | 0.03 | 0.17 | 0.15 | 1.07 | 0.90 |
Scenario 7 | |||||
---|---|---|---|---|---|
|
|
0.22 | 0.22 | 1.19 | 0.95 |
|
|
0.22 | 0.22 | 1.20 | 0.95 |
|
|
0.22 | 0.22 | 1.21 | 0.95 |
w/o model averaging |
|
0.22 | 0.21 | 1.21 | 0.94 |
Conditional instruments are present in Scenario 1, covariates are strongly correlated in Scenario 2, a large number of covariates are present in Scenario 3, the true outcome model includes nonlinear terms in Scenario 4, true exposure model includes nonlinear terms in Scenario 5, and an exposure–covariate interaction is present in Scenario 7.
Scenario 1 | Scenario 2 | Scenario 3 | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Rel. | Rel. | Rel. | |||||||||||||
Method | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP | Bias | SD | ESE | RMSE | CP |
|
0.01 | 0.08 | 0.07 | 1.17 | 0.94 |
|
0.13 | 0.12 | 1.56 | 0.93 |
|
0.20 | 0.19 | 1.29 | 0.92 |
|
0.00 | 0.08 | 0.07 | 1.17 | 0.95 | 0.00 | 0.14 | 0.13 | 1.65 | 0.94 | 0.00 | 0.20 | 0.19 | 1.29 | 0.94 |
|
0.01 | 0.08 | 0.08 | 1.26 | 0.94 |
|
0.13 | 0.13 | 1.58 | 0.94 |
|
0.21 | 0.19 | 1.29 | 0.93 |
w/o model averaging | 0.01 | 0.08 | 0.07 | 1.26 | 0.92 |
|
0.13 | 0.11 | 1.58 | 0.93 |
|
0.20 | 0.19 | 1.29 | 0.92 |
Scenario 4 | Scenario 5 | Scenario 6 | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
0.06 | 2.47 | 2.32 | 0.99 | 0.92 | 0.01 | 0.11 | 0.09 | 1.21 | 0.86 | 0.03 | 0.07 | 0.07 | 1.10 | 0.93 |
|
0.07 | 2.51 | 2.33 | 1.00 | 0.91 | 0.00 | 0.11 | 0.09 | 1.20 | 0.88 | 0.02 | 0.07 | 0.07 | 1.08 | 0.93 |
|
0.06 | 2.47 | 2.32 | 0.99 | 0.92 | 0.00 | 0.11 | 0.09 | 1.24 | 0.89 | 0.02 | 0.07 | 0.07 | 1.09 | 0.93 |
w/o model averaging | 0.06 | 2.47 | 2.32 | 0.99 | 0.92 | 0.00 | 0.12 | 0.08 | 1.24 | 0.83 | 0.02 | 0.07 | 0.07 | 1.09 | 0.92 |
Scenario 7 | |||||
---|---|---|---|---|---|
|
|
0.09 | 0.10 | 1.07 | 0.96 |
|
|
0.09 | 0.10 | 1.07 | 0.96 |
|
|
0.09 | 0.10 | 1.07 | 0.96 |
w/o model averaging |
|
0.09 | 0.10 | 1.07 | 0.96 |
Conditional instruments are present in Scenario 1, covariates are strongly correlated in Scenario 2, a large number of covariates are present in Scenario 3, the true outcome model includes nonlinear terms in Scenario 4, true exposure model includes nonlinear terms in Scenario 5, and an exposure–covariate interaction is present in Scenario 7.
D Covariates inclusion probabilities
In the following images, each line represents the marginal probability of inclusion of covariates for a given method. The gray line represents the ideal situation where confounders and outcome risk factors are included, whereas instruments and spurious variables are excluded. MADRx and MADRy represent the covariates included to model the exposure and the outcome, respectively. Unlike GBCEE, MADR does not necessarily include the same covariates in both models.
Note: Lines have been slightly adjusted vertically to avoid superimposition, but GBCEE, GLiDeR, BAC, MADRy, and BP all have probability of inclusions of 1 for all variables.
E Additional information on the application
Unexposed | Exposed | Standardized mean difference | |
---|---|---|---|
|
2,984 | 474 | |
Body mass index (in kg/m
|
26.45 (4.37) | 25.15 (4.04) | 0.309 |
Waist-to-hip ratio | 0.81 (0.06) | 0.80 (0.06) | 0.183 |
Short Mini Mental Status Exam | 24.68 (1.61) | 24.95 (1.33) | 0.184 |
Clinic | 0.267 | ||
|
751 (25.2) | 153 (32.3) | |
|
595 (19.9) | 122 (25.7) | |
|
780 (26.1) | 103 (21.7) | |
|
858 (28.8) | 96 (20.3) | |
Age | 71.67 (5.18) | 70.07 (4.31) | 0.335 |
Highest grade of school completed | 12.56 (2.70) | 13.62 (2.76) | 0.388 |
Northern European,
|
1,845 (61.8) | 298 (62.9) | 0.021 |
Central European,
|
1,512 (50.7) | 235 (49.6) | 0.022 |
Southern European,
|
198 (6.6) | 29 (6.1) | 0.021 |
Jewish,
|
56 (1.9) | 17 (3.6) | 0.105 |
Native American,
|
21 (0.7) | 1 (0.2) | 0.073 |
Russian,
|
83 (2.8) | 11 (2.3) | 0.029 |
Other origin,
|
121 (4.1) | 16 (3.4) | 0.036 |
Hysterectomy,
|
759 (25.4) | 117 (24.7) | 0.017 |
Ovary removed,
|
756 (25.3) | 123 (25.9) | 0.014 |
Fracture before 50,
|
997 (33.4) | 145 (30.6) | 0.060 |
Fell in past year,
|
798 (26.7) | 128 (27.0) | 0.006 |
Injury from fall in past year,
|
562 (18.8) | 80 (16.9) | 0.051 |
Fear of falling,
|
1,323 (44.3) | 131 (27.6) | 0.353 |
Time/year of moderate physical activity at 50 | 16.58 (45.32) | 50.24 (77.91) | 0.528 |
Time/year of high physical activity at 50 | 1.93 (16.05) | 7.89 (29.92) | 0.248 |
Time/year of moderate physical activity at 30 | 16.81 (46.74) | 38.70 (74.39) | 0.352 |
Time/year of high physical activity at 30 | 3.28 (19.89) | 7.86 (30.47) | 0.178 |
Time/year of moderate physical activity when teenager | 46.61 (80.66) | 69.87 (94.15) | 0.265 |
Time/year of high physical activity when teenager | 16.20 (48.09) | 20.68 (51.64) | 0.090 |
Bed ridden for 7 days in past year,
|
126 (4.2) | 11 (2.3) | 0.107 |
Smoking,
|
0.080 | ||
|
1,887 (63.2) | 291 (61.4) | |
|
803 (26.9) | 143 (30.2) | |
|
294 (9.9) | 40 (8.4) | |
Caffeine intake (mg/day) | 134.06 (137.32) | 138.69 (136.24) | 0.034 |
Any alcoholic beverage in past year,
|
2,046 (68.6) | 376 (79.3) | 0.247 |
At least 1 alcoholic beverage in past 30 days,
|
1,576 (52.8) | 314 (66.2) | 0.276 |
How often
|
0.18 (0.76) | 0.22 (0.73) | 0.050 |
Osteoporosis,
|
426 (14.3) | 39 (8.2) | 0.192 |
Diabetes,
|
174 (5.8) | 19 (4.0) | 0.084 |
Ever had a stroke,
|
84 (2.8) | 4 (0.8) | 0.147 |
Ever had hypertension,
|
1,114 (37.3) | 142 (30.0) | 0.157 |
Parkinson’s disease,
|
12 (0.4) | 1 (0.2) | 0.035 |
Arthritis,
|
1,772 (59.4) | 253 (53.4) | 0.121 |
Stayed in hospital overnight in past year,
|
311 (10.4) | 39 (8.2) | 0.076 |
Pain around hip for most days in a month in past year,
|
1,015 (34.0) | 146 (30.8) | 0.069 |
Thiazide use,
|
0.116 | ||
|
1,971 (66.1) | 335 (70.7) | |
|
244 (8.2) | 40 (8.4) | |
|
769 (25.8) | 99 (20.9) | |
Non-thiazide diuretice use,
|
0.167 | ||
|
2,750 (92.2) | 455 (96.0) | |
|
79 (2.6) | 8 (1.7) | |
|
155 (5.2) | 11 (2.3) | |
Benzodiazapine use in past year,
|
414 (13.9) | 63 (13.3) | 0.017 |
Sedative hypnotic use in past year,
|
37 (1.2) | 13 (2.7) | 0.108 |
Antidepressants use in past year,
|
101 (3.4) | 19 (4.0) | 0.033 |
Oral estrogen use,
|
0.286 | ||
|
1,913 (64.1) | 238 (50.2) | |
|
743 (24.9) | 158 (33.3) | |
|
328 (11.0) | 78 (16.5) | |
Progestin use,
|
0.171 | ||
|
2,819 (94.5) | 429 (90.5) | |
|
76 (2.5) | 14 (3.0) | |
|
89 (3.0) | 31 (6.5) | |
Difficulty walking 2 or 3 blocks,
|
356 (11.9) | 18 (3.8) | 0.306 |
Back pain in past year,
|
1,922 (64.4) | 274 (57.8) | 0.136 |
Use walking aids,
|
114 (3.8) | 11 (2.3) | 0.087 |
Problems that prevent getting up or walking up stairs –
|
283 (9.5) | 27 (5.7) | 0.143 |
Comparative self-ratted health (1 = Excellent, 5 = Very poor) | 1.84 (0.70) | 1.53 (0.63) | 0.461 |
Not married,
|
1,500 (50.3) | 182 (38.4) | 0.241 |
Years since menopause | 23.76 (8.03) | 21.04 (7.29) | 0.354 |
Mother ever had a fracture,
|
974 (32.6) | 158 (33.3) | 0.015 |
Father ever had a fracture,
|
626 (21.0) | 106 (22.4) | 0.034 |
Numbers are mean (standard deviation) unless otherwise stated.
References
[1] Walter S, Tiemeier H. Variable selection: current practice in epidemiological studies. European J Epidemiol. 2009;24(12):733–6. 10.1007/s10654-009-9411-2Search in Google Scholar PubMed PubMed Central
[2] Talbot D, Massamba VK. A descriptive review of variable selection methods in four epidemiologic journals: there is still room for improvement. European J Epidemiol. 2019;34(8):725–30. 10.1007/s10654-019-00529-ySearch in Google Scholar PubMed
[3] Ertefaie A, Asgharian M, Stephens DA. Variable selection in causal inference using a simultaneous penalization method. J Causal Inference. 2018;6(1):1–16. 10.1515/jci-2017-0010Search in Google Scholar
[4] Shortreed SM, Ertefaie A. Outcome-adaptive lasso: variable selection for causal inference. Biometrics. 2017;73(4):1111–22. 10.1111/biom.12679Search in Google Scholar PubMed PubMed Central
[5] Talbot D, Lefebvre G, Atherton J. The Bayesian causal effect estimation algorithm. J Causal Inference. 2015;30(2):207–36. 10.1515/jci-2014-0035Search in Google Scholar
[6] Wang C, Parmigiani G, Dominici F. Bayesian effect estimation accounting for adjustment uncertainty. Biometrics. 2012;68(3):661–71. 10.1111/j.1541-0420.2011.01731.xSearch in Google Scholar PubMed
[7] Fithian W, Sun D, Taylor J. Optimal inference after model selection. 2014. Available from: https://arxiv.org/abs/1410.2597. Search in Google Scholar
[8] Berk R, Brown L, Buja A, Zhang K, Zhao L. Valid post-selection inference. Ann Statist. 2013;41(2):802–37. 10.1214/12-AOS1077Search in Google Scholar
[9] Leeb H, Pötscher BM. Model selection and inference: facts and fiction. Econometric Theory. 2005;21(1):21–59. 10.1017/S0266466605050036Search in Google Scholar
[10] Leeb H, Pötscher BM, Can one estimate the unconditional distribution of post-model-selection estimators?. Econometric Theory. 2008;24(2):338–76. 10.1017/S0266466608080158Search in Google Scholar
[11] Panigrahi S, Taylor J, Asaf W. Bayesian post-selection inference in the linear model. 2016. Available from: https://arxiv.org/abs/1605.08824. Search in Google Scholar
[12] Crainiceanu CM, Dominici F, Parmigiani G. Adjustment uncertainty in effect estimation. Biometrika. 2008;73(2):635–51. 10.1093/biomet/asn015Search in Google Scholar
[13] Belloni A, Chernozhukov V, Hansen C. Inference on treatment effects after selection among high-dimensional controls. Rev Econ Stud. 2014;81(2):608–50. 10.1093/restud/rdt044Search in Google Scholar
[14] van der Laan MJ, Gruber S. Collaborative double robust targeted maximum likelihood estimation. Int J Biostatist. 2010;6(17):1–61. 10.2202/1557-4679.1181Search in Google Scholar PubMed PubMed Central
[15] Wang C, Dominici F, Parmigiani G, Zigler CM. Accounting for uncertainty in confounder and effect modifier selection when estimating average causal effects in generalized linear models. Biometrics. 2015;71(3):654–65. 10.1111/biom.12315Search in Google Scholar PubMed PubMed Central
[16] Wilson A, Reich BJ. Confounder selection via penalized credible regions. Biometrics. 2014;70(4):852–61. 10.1111/biom.12203Search in Google Scholar PubMed
[17] Cefalu M, Dominici F, Arvold N, Parmigiani G. Model averaged double robust estimation. Biometrics. 2017;73(2):410–21. 10.1111/biom.12622Search in Google Scholar PubMed PubMed Central
[18] Koch B, Vock DM, Wolfson J. Covariate selection with group lasso and double robust estimation of causal effects. Biometrics. 2018;74(1):8–17. 10.1111/biom.12736Search in Google Scholar PubMed PubMed Central
[19] Ju C, Benkeser D, van der Laan MJ. Robust inference on the average treatment effect using the outcome highly adaptive lasso. Biometrics. 2019;76(1):109–18. 10.1111/biom.13121Search in Google Scholar PubMed
[20] Antonelli J, Parmigiani G, Dominici F. High-dimensional confounding adjustment using continuous spike and slab priors. Bayesian Anal. 2019;14(3):805–28. 10.1214/18-BA1131Search in Google Scholar PubMed PubMed Central
[21] Pearl J. Causality: Models, reasoning, and inference. 2nd ed. New York: Cambridge University Press; 2009. 10.1017/CBO9780511803161Search in Google Scholar
[22] Brookhart MA, Schneeweiss S, Rothman KJ, Glynn RJ, Avorn J, Stürmer T. Variable selection for propensity score models. Am J Epidemiol. 2006;163(12):1149–56. 10.1093/aje/kwj149Search in Google Scholar PubMed PubMed Central
[23] de Luna X, Waernbaum I, Richardson TS. Covariate selection for the nonparametric estimation of an average treatment effect. Biometrika. 2011;98(4):861–75. 10.1093/biomet/asr041Search in Google Scholar
[24] Lefebvre G, Atherton J, Talbot D. The effect of the prior distribution in the Bayesian adjustment for confounding algorithm. Comput Statist Data Anal. 2014;70:227–40. 10.1016/j.csda.2013.09.011Search in Google Scholar
[25] Rotnizky A, Smucler E. Efficient adjustment sets for population average causal treatment effect estimation in graphical models. J Mach Learn Res. 2020;21(188):1–86. Search in Google Scholar
[26] Tang D, Kong D, Pan W, Wang L. Ultra-high dimensional variable selection for doubly robust causal inference. 2020. Available from: https://arxiv.org/abs/2007.14190. Search in Google Scholar
[27] Saarela O, Belzile LR, Stephens DA. A Bayesian view of doubly robust causal inference. Biometrika. 2016;103(3):667–81. 10.1093/biomet/asw025Search in Google Scholar
[28] Antonelli J, Papadogeorgou G, Dominici F. Causal inference in high dimensions: a marriage between Bayesian modeling and good frequentist properties. Biometrics. 2022;78(1):100–14. 10.1111/biom.13417Search in Google Scholar PubMed PubMed Central
[29] Pearl J. Invited commentary: understanding bias amplification. Am J Epidemiol. 2011;174(11):1223–7. 10.1093/aje/kwr352Search in Google Scholar PubMed PubMed Central
[30] Raftery A. Bayesian model selection in structural equation models. In: Bollen K, Long J, editor, Testing structural equation models. Newbury Park, CA: Sage; 1993. p. 163–80. Search in Google Scholar
[31] Raftery AE, Hoeting J, Volinsky C, Painter I, Yeung KY. BMA: Bayesian model averaging. 2018. R package version 3.18.9. Available from: https://CRAN.R-project.org/package=BMA. Search in Google Scholar
[32] Haughton DMA. On the choice of a model to fit data from an exponential family. Ann Statist. 1988;14(1):342–65. 10.1214/aos/1176350709Search in Google Scholar
[33] Wasserman L. Bayesian model selection and model averaging. J Math Psych. 2000;44(1):92–10. 10.1006/jmps.1999.1278Search in Google Scholar PubMed
[34] van der Laan MJ, Rubin D. Targeted maximum likelihood learning. Int J Biostatist. 2006;2(1):1–38.10.2202/1557-4679.1043Search in Google Scholar
[35] Scharfstein DO, Rotnitzky A, Robins JM. Adjusting for non-ignorable drop-out using semi-parametric nonresponse models (with discussion and rejoinder). J Amer Statist Assoc. 1999;94(448):1096–120. 10.1080/01621459.1999.10473862Search in Google Scholar
[36] Porter KE, Gruber S, van der Laan MJ, Sekhon JS. The relative performance of targeted maximum likelihood estimators. Int J Biostatist. 2011;7(1):1–34. 10.2202/1557-4679.1308Search in Google Scholar PubMed PubMed Central
[37] van der Laan MJ, Rose S. Targeted learning: causal inference for observational and experimental data. New York: Springer Series in Statistics; 2011. 10.1007/978-1-4419-9782-1Search in Google Scholar
[38] Tsiatis A. Semiparametric theory and missing data. New York: Springer Science & Business Media; 2007. Search in Google Scholar
[39] Luque-Fernandez MA, Schomaker M, Rachet B, Schnitzer ME. Targeted maximum likelihood estimation for a binary treatment: a tutorial. Stat Med. 2018;37(16):2530–46. 10.1002/sim.7628Search in Google Scholar PubMed PubMed Central
[40] Siddique AA, Schnitzer ME, Bahamyirou A, Wang G, Holtz TH, Migliori GB, et al. Causal inference with multiple concurrent medications: a comparison of methods and an application in multidrug-resistant tuberculosis. Stat Methods Med Res. 2019;28(12):3534–49. 10.1177/0962280218808817Search in Google Scholar PubMed PubMed Central
[41] Rosenblum M, van der Laan MJ. Targeted maximum likelihood estimation of the parameter of a marginal structural model. Biostatistics. 2010;6(2):1–27. 10.2202/1557-4679.1238Search in Google Scholar PubMed PubMed Central
[42] Neugebauer R, van der Laan MJ. Why prefer double robust estimators in causal inference? J Statist Plann Inference. 2005;129:405–26. 10.1016/j.jspi.2004.06.060Search in Google Scholar
[43] Madigan D, York J, Allard D. Bayesian graphical models for discrete data. Int Stat Rev. 1995;63(2):215–32. 10.2307/1403615Search in Google Scholar
[44] Looker AC, Frenk SM. Percentage of adults aged 65 and over with osteoporosis or low bone mass at the femur neck or lumbar spine: United States, 2005-2010. National Center for Health Statistics. 2020. Available from: https://www.cdc.gov/nchs/data/hestat/osteoporsis/osteoporosis2005_2010.htm. Search in Google Scholar
[45] Bonaiuti D, Shea B, Iovine R, Negrini S, Welch V, Kemper HH, et al. Exercise for preventing and treating osteoporosis in postmenopausal women. Cochrane Database of Systematic Reviews. 2002;(2):1–37. 10.1002/14651858.CD000333Search in Google Scholar PubMed
[46] Thibaud M, Bloch F, Tournoux-Facon C, Brèque C, Rigaud S, Dugué B, et al. Impact of physical activity and sedentary behaviour on fall risks in older people: a systematic review and meta-analysis of observational studies. European Rev Aging Phys Activity. 2011;9:5–15. 10.1007/s11556-011-0081-1Search in Google Scholar
[47] World Health Organization. Global recommandations on physical activity for health. 2010. Available from: https://www.who.int/dietphysicalactivity/publications/9789241599979/en/. Search in Google Scholar
[48] Benasseur I, Talbot D, Durand M, Holbrook A, Matteau A, Potter BJ, et al. A comparison of confounder selection and adjustment methods for estimating causal effects using large healthcare databases. Pharmacoepidemiol Drug Safety. 2022;31(4):424–33. 10.1002/pds.5403Search in Google Scholar PubMed PubMed Central
[49] Ramamoorthi RV, Siriam K, Martin R. On the posterior concentration in misspecified models. Bayesian Anal. 2015;10(4):759–89. 10.1214/15-BA941Search in Google Scholar
[50] Lv J, Liu JS. Model selection principles in misspecified models. J R Stat Soc Ser B Stat Methodol. 2014;76(1):141–67. 10.1111/rssb.12023Search in Google Scholar
[51] Efron B. Estimation and accuracy after model selection. J Amer Statist Assoc. 2013;109(507):991–1007. 10.1080/01621459.2013.823775Search in Google Scholar PubMed PubMed Central
[52] Walker AM. On the asymptotic behavior of posterior distributions. J R Stat Soc Ser B Stat Methodol. 1969;31(1):80–8. Search in Google Scholar
[53] Dawid AP. On the limiting normality of posterior distributions. Math Proc Cambridge Philos Soc. 1970;67(6):25–33. 10.1017/S0305004100045953Search in Google Scholar
© 2022 Denis Talbot and Claudia Beaudoin, published by De Gruyter
This work is licensed under the Creative Commons Attribution 4.0 International License.