Our findings suggest that in hypertensive patients with CAD, aggressive antihypertensive therapy is associated with reduced risk for adverse cardiovascular outcomes compared with more conventional treatment. This effect was only apparent when using a MSCM that adjusted for time-dependent confounding by systolic blood pressure and differential loss to follow-up, not when a standard Cox model was used. This suggests that time-dependent confounding by systolic blood pressure and/or differential loss to follow-up introduces bias into the estimates of treatment effects for antihypertensive combination therapies resulting in an under-estimation of the benefits of more aggressive treatment regimens. Similar findings were observed across the two additional exposure specifications used in sensitivity analysis. However, the observed differences between the MSCMs and standard Cox models did not rise to statistical significance in two out of the three exposure specifications. The magnitude of the between model differences increased with the number of concomitant drugs required for the aggressive treatment definition, suggesting that time-dependent confounding by blood pressure may be of increasing concern the more aggressive the examined treatment strategy becomes.
Assuming a correctly specified model and no violation of the underlying model assumptions, the MSCM estimate for the effect of aggressive versus conventional antihypertensive therapy is interpretable as the effect that would have been observed in a RCT that compared the effects of aggressive to conventional treatment without making any adjustments in therapy to assure blood pressure control. In the absence of direct experimental evidence on treatment intensity, our results thus suggest that more aggressive treatment strategies are beneficial in adults older than 50
years with hypertension and prevalent CAD. Importantly, this finding is limited in its clinical utility as antihypertensive therapy in practice, as well as in typical RCTs, is highly dynamic (i.e., routinely adjusted to achieve and maintain target blood pressure goals) while our analysis makes inferences about a static treatment comparison. Our findings also need to be applied in the context of the known fact that more complex treatment regimens generally result in lower adherence [21
] and increased risk of adverse effects [22
Our results are consistent with other studies that have considered blood pressure as a mediating, time-dependent confounder [11
]. Specifically, our results extend the findings by Sugihara
from blood pressure response to clinical outcomes, and suggest that aggressive management of hypertension may not only result in higher achievement rates of target blood pressure [12
] but also in clinically important reduction in cardiovascular outcomes. In these previous studies, the use of IPTW to handle time dependent confounding by blood pressure enabled observational analyses of key associations that were consistent with the associations found from RCTs. In contrast, conventional approaches led to significant amounts of bias in the estimation of the association between antihypertensive medications and cardiovascular outcomes [11
]. The previous work on this subject as well as the known theoretical results give us confidence that a MSM approach is providing the best estimate of the effectiveness of aggressive antihypertensive therapy and is less susceptible to bias when compared with standard analytical approaches.
Of course, there are other possible explanations for the observed differences between the two approaches. The estimates from a MSCM are marginal estimates and thus estimate the contrast between treating everyone and treating no-one. In contrast, the conventional estimates are conditional estimates and thus estimate the effect of treatment holding all other variables constant. In the presence of non-linear covariates or effect measure modification, these estimators will be estimating different parameters which could explain some of the observed differences. The MSCM approach also accounted for loss to follow-up (via IPCW) whereas the conventional approach assumed random censoring.
Importantly, our results do not speak to the mechanisms by which aggressive antihypertensive therapy reduces the risk for cardiovascular outcomes, which may include both non-blood pressure mediated effects of individual agents, e.g., heart rate lowering effects of beta blockers in patients with a history of myocardial infarction [23
], as well as tighter blood pressure control. That tight blood pressure control might be of benefit to patients with hypertension and CAD is not unexpected, as strong evidence suggests that the cardio-protective effects of blood pressure lowering medications are largely mediated by blood pressure reduction [24
This study illustrates how properly constructed and appropriately analyzed observational studies can provide some clinical guidance until direct RCT evidence is available. Further extensions of this work could aim to increase clinical utility by using dynamic treatment regimens that consider specific blood pressure thresholds for aggressive versus conventional treatment to allow identification of high-risk sub-groups [28
], as there has been some evidence that aggressive treatment may be less effective in some sub-populations [18
], and examine specific drug and drug/dose combinations.
The present study has several limitations. First, the generalizability of our study is limited to hypertensive patients with CAD. However, this population represents a large and rapidly growing important high risk population that has not been at the center of antihypertensive research efforts and therefore warrants comprehensive investigation [30
Second, measurement error is inherently problematic in studies of hypertension [31
]. While the INVEST protocol tried to minimize measurement error by providing to all site investigators standardized blood pressure measurement instructions following JNC VI guidelines [34
], some measurement error is unavoidable. However, it is unlikely that this measurement error results in important levels of bias and more likely reduces the precision of the estimates due to additional random error in the estimates.
Third, we imputed data for missing visits during follow-up which may result in limited misclassification of both the exposure and the time-dependent confounder. Such misclassification would likely bias results (effect estimates of aggressive vs. conventional therapy as well as differences between the two analytic approaches) towards the null.
Fourth, to simplify the analytic approach, we focused the adjustment for time-dependent confounding on systolic blood pressure, and did not include additional time-dependent variables such as diastolic blood pressure or heart rate. More complex modeling of multiple time-dependent confounding factors is an active area for future research in this field.
The use of a MSCM to adjust for time-dependent confounding through inverse probability of treatment and censoring weighting introduces additional limitations to the study. MSMs, as implemented in this study, require a number of assumptions and a rather simple data structure. First, the independent variable of interest has to be binary. Our study categorized antihypertensive drug use into aggressive vs. conventional antihypertensive treatment. This definition does not distinguish between specific drugs and/or drug classes nor does it take dosing of each specific drug into account. If treatment effects vary between specific drugs and doses, our model will not show these differences and rather produce an estimate reflective of the average treatment combinations used in the aggressive and conventional treatment groups. Thus, our results do not address the question of the optimal selection or dosing of specific antihypertensive drugs for the aggressive therapy regimen.
Second the method requires the assumption, that once initiated, exposure is not discontinued until the end of follow-up or censoring (an intention to treat approach to exposure). This assumption is not fully met in INVEST using actual medication use. We therefore focused on the decision to begin aggressive (as opposed to conventional) treatment as our object of inference rather than an “as treated” analysis, which likely makes the analysis more conservative (i.e., introduces a bias towards the null).
Third, the pooled logistic regression model that was used to estimate the MSCM assumes that observations are equally spaced, which was not the case in the INVEST where the initial five visits occurred in six week intervals while the remaining visits occurred in six month intervals.
Lastly, as discussed in more detail in the introduction, in the presence of a time-dependent confounder standard methods allow two options, both expected to result in biased estimates — to disregard the time-dependent confounder (and accept the resulting bias introduced by the time-dependent confounder) or to adjust for the time-dependent confounder (and introduce bias by adjusting for a variable on the causal pathway) [7
]. Our study focused on the comparison between the MSCM to the former standard option, i.e., compared the MSCM to a standard Cox model that allowed exposure to vary over time but held all covariates, including systolic blood pressure, fixed at their baseline value. We chose this comparison because the standard Cox model with fixed covariates is more widely used and because its comparison to the MSCM provides a direct estimate of the bias introduced by the time-dependent confounder which is more readily interpretable than the bias introduced by inappropriate adjustment (adjustment for a time-dependent intermediate variable that is affected by prior treatment). However, the alternative approach (a comparison to a standard Cox model that allows both exposure and covariates to change over time) [35
] could be considered fairer as the both approaches ‘adjust’ for the time-dependent confounder and may, in our example, have shown better performance in terms of bias. One should therefore be cautious in generalizing the estimates of bias provided in this paper to scenarios involving Cox models with time-dependent covariates.
Summary and conclusions
It is conceptually clear that differential loss to follow-up and time-dependent confounding by a surrogate are problematic for the estimation of treatment effects in observational studies. Our findings empirically illustrate this issue by showing that standard Cox models systematically underestimate the effectiveness of aggressive antihypertensive combination treatment in a post-hoc analysis of a large hypertension treatment RCT and thus demonstrate the importance of using appropriate methods to address time-dependent confounding and differential loss to follow-up in observational antihypertensive drug studies. Although they do not support inferences regarding specific mechanisms of action, the results from a MSCM clearly suggest that more aggressive antihypertensive regimens are associated with a reduced rate of serious cardiovascular events in patients with hypertension and prevalent CAD.