![MathJax Logo](/templates/jsp/_style2/_tandf/pb2/images/math-jax.gif)
Abstract
Cancer diagnosis is part of a complex stochastic process, in which patients' personal and social characteristics influence the choice of diagnosing methods, diagnosing methods in turn influence the initial assessment of cancer stage, cancer stage in turn influences the choice of treating methods, and treating methods in turn influence cancer outcomes such as cancer survival. To evaluate the performance of diagnoses, one needs to estimate and test the sequential causal effect (SCE) under a specified regime of diagnoses and treatments in such a complex observational study, where the data-generating mechanism is unknown and modeling is needed for statistical inference. In this article, we introduce a method of statistical modeling to estimate and test SCEs under regimes of treatments (diagnoses and treatments in cancer diagnosis) in complex observational studies. By applying the alternative G-formula, we express the SCE in terms of the point effects of treatments in the sequence, so that the modeling can be conducted via the point effects in the framework of single-point causal inference. We illustrate our method by a medical example of cancer diagnosis with data from a Swedish prognosis study of cardia cancer.
1. Introduction
Cancer diagnosis is part of a complex stochastic process, in which patients' personal and social characteristics influence the choice of diagnosing methods, diagnosing methods in turn influence the initial assessment of cancer stage, cancer stage in turn influences the choice of treating methods, and treating methods in turn influence cancer outcomes such as cancer survival. To evaluate the performance of cancer diagnoses, one needs to estimate and test the sequential causal effect (SCE), which is the causal effect under a specified regime of diagnoses and treatments on a certain outcome of interest. An important example of SCEs is the one under optimal regime.
In sequential causal inference, Robins (Citation1986, Citation1997, Citation2009) derived the well-known G-formula, which identifies the SCE under a specified regime of treatments (diagnoses and treatments in cancer diagnosis) via standard parameters. Based on the G-formula, a parametric method has been developed for estimating the SCE via standard parameters (Hernan and Robins Citation2020; Robins Citation1997, Citation2009; Taubman et al. Citation2009). However, this method may suffer from the curse of dimensionality and the null paradox. See also the improved versions of the parametric method. Still, these versions have the difficulty of specifying the influence of the time-dependent covariates due to the curse of dimensionality and the null paradox (Almirall, Have, and Murphy Citation2010; Henderson, Ansell, and Alshibani Citation2010; Chaffe and van der Laan Citation2012).
To avoid these problems, the literature focuses on semi-parametric methods. One class of methods is the marginal structural model based on the inverse probability of treatment weighting (Hernan and Robins Citation2020; Hernan, Brumback, and Robins Citation2000; Robins, Brumback, and Hernan Citation2000; Robins Citation2009) or the augmented the inverse probability of treatment weighting (Zhang et al. Citation2013). Another class of methods is the G-estimation based on the structural nested mean model (SNMM) or optimal SNMM (Hernan and Robins Citation2020; Robins Citation1997, Citation2004, Citation2009); the G-estimation incorporates the A-learning (for instance, Murphy Citation2003) and Q-learning (for instance, Sutton and Barto Citation1998). See also a summary of the development of these methods in the area of optimal regimes (Kosorok et al. Citation2021). These methods achieve consistent estimates of SCEs. In a recent review, Kosorok and Laber (Citation2019) highlighted the need for methods of evaluating the uncertainty in estimating SCEs in the context of optimal regimes (Sec. 6, statistical inference).
Recently, Wang and Yin (Citation2015, Citation2020) derived an alternative G-formula, which expresses the SCE in terms of the point effects of treatments in the sequence. The point effect is simply the effect of single-point treatment in single-point causal inference (Rosenbaum and Rubin Citation1983). In designed experiments, where data-generating mechanisms are known, they estimated the SCE by estimating point effects of treatments without the need for modeling. In observational studies such as the influence of early cancer diagnosis, where data-generating mechanism is unknown, one needs to conduct modeling to estimate and test the SCE.
In this article, we will introduce a method of statistical modeling to estimate and test the SCE in observational studies and illustrate our method by a medical example using data from a Swedish prognosis study of cardia cancer. In Sec. 2, we describe our motivating example. In Sec. 3, we introduce the general framework of our method. We apply the alternative G-formula to express the SCE in terms of point effects of treatments in the sequence and describe how to conduct modeling to estimate and test the point effects and then the SCE in the framework of single-point causal inference. In Sec. 4, we apply our method to estimate and test the SCE under any regime of diagnosing and treating hospitals on one-year survival in our medical example. In Sec. 5, we compare our method with the three available methods mentioned earlier by applying all these four methods to the data. Finally, in Sec. 6, we conclude the article with discussion.
2. Data and the medical background
In Sweden, patients usually seek medical help at hospitals near their residential areas, namely, catchment areas. When cancer is diagnosed, they may stay at the diagnosing hospital or transfer to another hospital for treatment. The hospital diagnosing cancer is called the diagnosing hospital, while the one treating cancer is called the treating hospital. To evaluate the performance of diagnosing and treating hospitals, we may study cancer outcomes such as survival under various regimes of diagnosing and treating hospitals among cancer patients after adjusting for patients' differences.
The data used in this study is from a prognosis study conducted during a period between 1988 and 1995 in hospitals in central and northern Sweden (Hansson et al. Citation2000). It contained information of 157 patients of cardia cancer. Cardia cancer is highly malignant with bad prognosis, and one-year survival is a good measure of the performance of both diagnosing and treating hospitals. A question of relevance to the public health policy is which types of the diagnosing and treating hospitals, large versus small, perform better on cancer outcomes, where the large type refers to the regional or county hospitals and the small type to local hospitals.
In Sec. 4 below, we will estimate and test the SCE on one-year survival under any regime of diagnosing and treating hospitals and then use the SCEs to evaluate the performance of diagnosing and treating hospitals.
3. General framework of estimating and testing SCEs based on the alternative G-formula
In this section, we will describe the alternative G-formula, which identifies the SCE by the point effects of treatments, and the general framework of our method for estimating and testing SCEs in observational studies.
3.1. Potential variables and the causal effect from a sequence of treatments
Let at time
be the treatment variable, which potentially and deterministically assigned treatments
to each unit of the population. A regime is a sequence of such treatment variables,
Prior to
there exists a set of stationary covariates
Under
each unit could have a set of potential time-dependent covariates
between
and
(
) and a potential outcome
of interest after the last treatment
Noticeably,
can be dynamic, namely, the assignment of the treatment depends on the earlier treatments and covariates of each unit.
Let be the set of all stationary covariates and potential covariates, and {
be the set of all stationary and potential covariates, treatment variables, and the potential outcome. The realization of these variables is
In stratum
we document the regime as
its potential covariates as
and potential outcome as
Let be the control regime given
The blip effect
of treatment
in stratum
is an increase of the mean of potential outcome
under regimes
versus
in the stratum,
that is
The blip effect is called the net effect (Wang and Yin Citation2015). Clearly,
The SCE under regime
in stratum
is an increase of the mean of potential outcome
under
versus
under
in the stratum, namely,
Here we see that if
then
Therefore the blip effect is a special type of SCEs. As it will become clear in Sec. 3.4, the blip effects determine the SCE under any regime of treatments.
The SCE under regime in the population,
is an increase of the mean of potential outcome
under
versus
under
in the population, namely,
3.2. Observable variables and the point effects of treatments in the sequence
Now we consider observable variables. Let be the treatment variable at
be the set of stationary covariates,
be the set of time-dependent covariates between
and
and
be the outcome of interest. Let
be the sequence of treatment variables;
be the set of all stationary and time-dependent covariates;
be the set of all treatments, covariates and outcome.
Suppose that the observable variables have the same support as the potential ones
The observed values of these variables are
In the following, we will use
to denote the probability distribution of a discrete variable or the density distribution of a continuous variable. The joint distribution of
is given by
(1)
(1) The standard parameters for these conditional distributions are simply their conditional means, for instance, the standard parameter for
is the conditional mean
Now we describe the point effects of treatments in the sequence, which are also parameters for the outcome (Wang and Yin Citation2015). Let be the conditional mean of observable outcome
given
Without loss of generality, we take
as the control treatment. Then the point effect of treatment
in the stratum
is the following difference between the means
(2)
(2) Clearly,
The point effect is simply the point effect of treatment in single-point causal inference and its estimation and hypothesis testing is well studied (for instance, Rosenbaum and Rubin Citation1983).
To identify the SCEs by the observable variables Robins (Citation1986, Citation1997, Citation2009) introduced the identifying condition, which consists of the consistency assumption, the positivity assumption for treatments and the assumption of no unmeasured confounders. The identifying condition is satisfied in sequential randomized experiments where treatment
is randomly assigned according to a history
) of the earlier covariates and treatments. It is approximately satisfied in observational studies with a sufficient number of covariates. Throughout, we assume the identifying condition.
Under the identifying condition, Robins (Citation1986, Citation1997, Citation2009) derived the well-known G-formula, which expresses the SCE in terms of the standard parameters If
(
) has influence from earlier treatments
and influence on subsequent treatments
then
are essentially all different for different
even if each
has null effect on
So any model imposing equalities between
for different
is misspecified (null paradox). Obviously, the number of
is huge in case of large
or even uncountable (curse of dimensionality).
Under the same identifying condition, Wang and Yin (Citation2020) derived an alternative G-formula, which expresses the blip effect and SCE in terms of point effects instead of the standard parameters. In the rest of this section, we will first describe the G-formula, then describe our method of conducting modeling to estimate and test blip effects and SCEs based on the alternative G-formula.
3.3. Alternative G-formula to estimate and test the blip effect in observational studies
Intuitively, the point effect is the result of treatment
and the subsequent treatments
and it may decompose into a sum of the blip effects of these treatments. Therefore, we should have the following decomposing formula
(3)
(3) where the expectations are with respect to the conditional distributions of observable variables
and
Wang and Yin (Citation2020) theoretically proved (3) under the identifying condition. Formula (3) is the alternative G-formula and expresses the point effects in terms of blip effects.
Now we describe our method of conducting modeling to estimate and test blip effects. First, the point effect is simply the point effect of treatment in single-point causal inference, so we can estimate it by modeling the mean in the framework of single-point causal inference. As an illustration, if
influences the outcome Y via
then we have
denoted by
i.e.,
Thus, we may rewrite (2) as
Second, Robins (Citation1997, Citation2009) and Hernan and Robins (Citation2020) pointed out that the blip effects follow a certain pattern described by structural nested mean model (SNMM), as an illustration,
where
is a deterministic function of
and is indexed by a parameter vector
of small dimension. With
and SNMM, the alternative G-formula becomes the following decomposing formula
where the expectations are with respect to the conditional distributions of observable variables
and
To estimate and test we treat the above decomposing formula as a regression model, where the response variables are the estimates
and the explanatory variables are the observed proportions
which are obtained from the data without modeling. In the regression, we need the conditional variance of the estimated point effect given all covariates and treatments
which is obtained by specifying a constant dispersion parameter for the distribution
. Noticeably, the conditional covariance between the estimated point effects at different times is negligible, for instance, it is equal to zero for the normal outcome (Wang and Yin Citation2015, Citation2020). The bootstrap method is used to obtain the covariance matrix
incorporating the variability of
If the estimates for the point effects are consistent and asymptotically normal as typically required in single-point causal inference and if is of finite dimension, then the estimate
is consistent and asymptotically normal. In this case, with
and
we conduct the Wald test on
3.4. Alternative G-formula to estimate and test SCEs in observational studies
For the sake of explication, we first describe the alternative G-formula for SCEs in the population. Intuitively, it may be decomposed into a sum of the blip effects of treatments in the regime, as follows. Prior to the covariate distribution in the population is
so when
potentially and deterministically assigned
to the population, the contribution of
to SCE would be the expectation
with respect to
as well-known in single-point causal inference (for instance, Rosenbaum and Rubin Citation1983). Between
and
the covariate distribution in the population under
would be
so the contribution of
to SCE would be
with respect to
and so forth. Therefore, we have the following decomposing formula,
(4)
(4) where the expectation is with respect to the distribution of observable covariates
Similarly, for the SCE in stratum
we have the decomposing formula
(5)
(5) where the expectation is with respect to the distribution of observable covariates
Theoretically, Wang and Yin (Citation2020) proved (4) and (5). The formula (4) or (5) is the alternative G-formula for the SCE and expresses the SCE in terms of the blip effects.
Now we describe our method of conducting modeling to estimate and test SCEs in observational studies. The blip effects are already estimated from the previous subsection. Then we only need to model the probabilities
by regression in the framework of single-point causal inference. As an illustration of our method, suppose that we obtain SNMM
and
By decomposing the SCE into the blip effects or applying (4) and (5), we obtain
with respect to
and
with respect to
With the estimates and
we obtain the estimate
under regime
in stratum
and
under regime
in the population. The covariance matrix between the estimated SCEs under different regimes is estimated by the bootstrap method. If
in addition to
are consistent and asymptotically normal as typically required in single-point causal inference, so are
and
With the estimates and the covariance matrix, we conduct the Wald test on the SCE under any regime. We may also compare different regimes by testing the differences between their SCEs.
Several advantages of our method of modeling to estimate and test blip effects and SCEs are summarized as follows.
When estimating the blip effect and SNMM, there is little risk for the model misspecification due to the following reasons. First, the regression model for the blip effect is the alternative G-formula (the decomposing formula) for blip effects, which is subject to no model misspecification. Second, the point effects as the response variables are estimated by regression in the framework of single-point causal inference. Third, the probability of treatments and covariates as the explanatory variables are estimated by the observed proportion in the data without modeling. Fourth, there is no restriction on the functional form of SNMM, and SNMM is testable in the framework of regression. SNMM improves the efficiency of estimating and testing blip effects. Notably, the modeling does not suffer from the curse of dimensionality and the null paradox as compared to the modeling via standard parameters (Hernan and Robins Citation2020; Robins Citation1997, Citation2009; Taubman et al. Citation2009).
The covariate probabilities in the alternative G-formula (the decomposing formula) for SCEs are estimated by regression in the framework of single-point causal inference. With the estimated blip effects/SNMM and the estimated covariates probabilities, we use the alternative G-formula to calculate the estimate of SCE under any regime of treatments. We do not need to specify models for the baselines such as
which are subject to high risks for model misspecification (Hernan and Robins Citation2020; Robins Citation1997, Citation2004, Citation2009).
4. Estimating and testing SCEs of diagnosing and treating hospitals on cardia cancer survival
4.1. Setting
In the medical example of Sec. 2, the diagnosing hospital is the treatment variable which takes the value
for small type and
for large type. The treating hospital is the treatment variable
which takes
for small type and
for large type. The following stationary covariates before
are measured: gender
geographic area
and age
Gender is
for female and
for male. Geographic area is categorized into
for rural and
for urban. Age takes continuous values
The time-dependent covariate between
and
is cancer stage
taking the values
The outcome of interest is
which takes
for death and
for survival within one year after diagnosis. In Supporting Material, the data is available together with the code for analyses in this article. The descriptive statistics are given in .
Table 1. Frequencies or means (standard deviations) of covariates and outcome across the diagnosing and treating hospitals for 157 cardia cancer patients.
Due to the long-term social welfare system and relatively uniform culture in Sweden, most of the stationary covariates, such as education and socioeconomic status, have similar distributions across different hospitals and thus do not confound the causal effect. As a common practice in many epidemiologic researches in Sweden, we assume no unmeasured confounders for diagnosing hospitals at least after conditional on gender
residential area
and age
Similarly, we assume no unmeasured confounders for treating hospitals
conditional on
and cancer stage
besides
and
Together with the positivity and consistency assumptions, the assumption of no hidden confounders is the identifying condition (Robins Citation1986, Citation1997, Citation2009). Under the identifying condition, we identify the SCE under any regime of diagnosing and treating hospitals including the optimal regime by our data.
In the rest of this section, we will apply our method described in Sec. 3 to estimate and test the SCE under any regime of diagnosing and treating hospitals and evaluate the diagnosing hospitals.
4.2. Estimating and testing point effects of diagnosing and treating hospitals on one-year survival
In the area of cancer diagnosis, one usually studies one-year, three-year and five-year survivals. One-year survival is often the focus of highly malignant cancers such as cardia cancer.
To estimate the point effect of diagnosing hospital we model the mean
which is the probability of one-year survival
in stratum
where
is the variable of interest while all others are possible confounders of
Notably, the model does not contain any posttreatment variable after
so the null paradox does not necessarily occur, and the model can be unsaturated without biasing the point effect. At the significance level of 0.10, we exclude the geographic area
and age
and obtain
(6)
(6)
From this model, we obtain the estimate for
and then the point effect of large diagnosing hospital
Here, the point effects for both
and
are approximately equal, i.e.,
with the p-value
0.6.
The estimate is presented in . Please note that the point effect is a mixed effect of both
and
and thus of little medical interest.
Table 2. Point effects, blip effects and optimal SCEs of diagnosing and treating hospitals on one-year survival of cardia cancer: estimate, p-value and 95 % CI.
To estimate the point effect of treating hospital we model the mean
which is the probability of one-year survival
in stratum
where
is the variable of interest while all others are possible confounders of
At the significance level of 0.10, we exclude
and
and obtain
(7)
(7)
From this model, we obtain the estimate for the point effect of large treating hospital
Here, the point effects in all strata
are approximately equal, i.e.,
with the p-value
The estimate
is presented in together with
As it will be clear in the next subsection, we will need the conditional variances of the estimated point effects given all the covariates, the diagnosing and treating hospitals. The conditional variances are obtained by specifying a constant dispersion parameter for the binomial distribution and applying the obtained models (6) and (7). Noticeably, the conditional covariance between the estimated point effects of diagnosing and treating hospitals is approximately zero (Wang and Yin Citation2015, Citation2020).
4.3. Estimating and testing blip effect of diagnosing and treating hospitals on one-year survival
Because neither (6) nor (7) contains geographic area and age
we do not include them in the blip effects and SCEs below. Because the blip effects are a special type of SCEs and determine all other SCEs (see the next subsection), we will estimate the blip effects here and SCEs in the next subsection. The blip effect
of diagnosing hospital
in the stratum
is an increase of the mean of potential survival
under regimes
versus to
in the stratum, namely
Clearly, we have
The blip effect
of treating hospital
in the stratum
is an increase of the mean of potential survival
under regimes
versus
in the stratum, namely,
clearly,
Because the treating hospital
is the last treatment variable, the blip effect
is equal to the point effect
obtained in the previous subsection. Furthermore, a question of medical interest is whether the blip effect of
depends on gender
Motivating by all these together, we suppose that the blip effects satisfy SNMM of the form
(8)
(8)
Let be the vector of all parameters in (9). The parameter
is the blip effect of large diagnosing hospital among women (
and
is blip effect of large diagnosing hospital among men (
They describe modification of the blip effect of diagnosing hospital by gender. The average of the blip effect
with respect to
is blip effect of large diagnosing hospital in the population and equal to
Intuitively, the point effect results from the blip effects of
and
For
the contribution is
if
and
if
For
the contribution comes from the two strata
and
and it is equal to
where
Thus, we have
and
The
is estimated by the observed proportion in the data without modeling. For the last treatment variable
the point effect of
is equal to the blip effect of
namely,
Taking these together, we have the following decomposing formula under SNMM (8)
(9)
(9)
The formula (9) is the alternative G-formula for the blip effect under SNMM (8) and it can also be derived by applying the formula (3) and SNMM (8) as follows. By applying (3) to we obtain
Applying (8), we obtain for
Thus, we have
Similarly we have
By applying (3) to
we obtain
That is
Because for both
as seen from model (6), we have
but the variance
is obtained by adjusting
to the size of stratum
All probabilities in
are estimated by the corresponding proportions in the data without modeling. The
is estimated from model (7).
Now, conditional on all covariates and diagnosing and treating hospitals, we use (9) as a regression model to estimate
), where the response variables are
and
while the explanatory variables are the estimates
and the ones. The bootstrap method is used to obtain the covariance matrix
incorporating the variability of all covariates and diagnosing and treating hospitals. With
and
we conduct the Wald test on
The result is presented in .
4.4. Estimate and testing SCE under any regime of diagnosing and treating hospitals on one-year survival
First, we consider It is an increase of the mean of potential survival under regimes
versus
in stratum
namely
where
potentially and deterministically assigned treating hospital
to each patient. Noticeably,
can be dynamic, for instance,
for patients younger than 55 years and
otherwise. By applying (5) to
we see that
Thus, by applying
in SNMM (8), we obtain following decomposing formula for the SCE
(10)
(10)
Second, we consider which is an increase of the mean of potential survival under regime
versus
in stratum
namely
where
potentially and deterministically assigned
of diagnosing and treating hospitals and it can be dynamic. If regime
then
which implies that the blip effect is a special type of SCEs. Intuitively the SCE should be a result of the blip effects of
and
in the regime. The contribution of
is
if
and
if
The contribution of
is
Therefore, we have the following decomposing formula
(11)
(11)
The formulas (10) and (11) implies that all SCEs are determined by blip effects. They are the alternative G-formula for the SCE under SNMM (8) and can also be derived by applying the formula (5) and the SNMM (8) as follows. By applying (5) to we obtain
Now by applying (8), we obtain(11).
Taking the average of (11) with respect to we obtain the alternative G-formula for
which is an increase of the mean of the potential survival under regimes
versus
in the population.
In (10) replacing
) by estimates
obtained from (9), we obtain the estimate for
under any treating hospital
In (11) replacing
) by estimates
obtained from (9), we obtain the estimate for
under any regime
Averaging
with respect to proportion
we obtain the estimate for
The variance of the estimated SCE is obtained by the bootstrap method. With the estimate and its variance, we obtain the confidence interval of the SCE. The result is presented in .
4.5. Estimating and testing optimal regime of diagnosing and treating hospitals on one-year survival
By the dynamic programming procedure (Bellman, Citation1957), we estimate and test optimal regime of diagnosing and treating hospitals by estimating and testing SCEs, as follows. We first estimate the optimal treating hospital
by
which is the treating hospital such that
achieves the maximum value among all possible values (0 or 1) that
assigns for given
Because
(see ), we have
(the large treating hospital) and
for all given
We test
by testing
and obtain the 95% CI and the p-value. Now we estimate the optimal diagnosing hospital
by
Because and
we have
We see that take maximum value when
for both
and
Therefore
(small diagnosing hospital) and
for any given gender
Thus, the estimates for
are all equal to
We test
by testing
and obtain the 95% CI and the p-value for
The result is presented in .
4.6. Causal analysis of diagnosing and treating hospitals based on
For the treating hospital, the point effect or the blip effect
of large treating hospital is estimated at 0.194 with the p-value equal to 0.009, implying that the large treating hospital is superior to one-year survival. There is no effect modification by the cancer stage, gender and age. This observation is consistent with the medical knowledge. Cardia cancer is highly malicious regardless of age and gender. It progressed quickly regardless of cancer stage. Patients typically have comorbidities and large treating hospitals are probably better in dealing with comorbidities.
For the diagnosing hospital, the point effect is estimated at
0.020 with the p-value 0.799, suggesting large and small diagnosing hospitals perform equally well. But it is misleading, because the point effect is a mixed effect of diagnosing and treating hospitals. The blip effect
is estimated at
0.121 with the p-value 0.142, implying that the small diagnosing hospital performs better for one-year survival. This observation indicates that the early diagnosis of such a malicious cancer as cardia cancer depends on the short waiting queue and attention typically at small diagnosing hospitals. The blip effect
for women is estimated at
with the p-value equal to 0.090, and the blip effect
for men at
with the p-value equal to 0.179, suggesting some modification of the blip effect of diagnosing hospital by gender.
The optimal regime is estimated as the small diagnosing hospital and large treating hospital. The optimal SCE is estimated at 0.194 with the p-value 0.009. This reveals the potential improvement for one-year survival by the optimal regime.
Here, all effects are measured by the difference in mean survival, but they can also be measured by the difference in a function of mean survival. We may apply the alternative G-formula (namely, the decomposing formula) to estimate the causal effect measured by odds ratio, rate ratio and hazard ratio. Though the estimates are consistent, the problem of non-collapsibility of a non-linear measure becomes far worse in the context of sequential treatments, leading to a biased estimate for finite sample. Therefore, it is recommended that one uses the linear measure for the SCE.
5. Comparison of our method with available methods
Method (i) is our method described in Sec. 4. Method (ii) is based on the well-known G-formula (Hernan and Robins Citation2020; Robins Citation1997, Citation2009; Taubman et al. Citation2009). Method (iii) is the marginal structural model based on inverse probability of treatment weighting (Hernan and Robins Citation2020; Hernan, Brumback, and Robins Citation2000; Robins, Brumback, and Hernan Citation2000; Robins Citation2009). Method (iv) is the G-estimation based on SNMM or optimal SNMM (Hernan and Robins Citation2020; Robins Citation1997, Citation2004, Citation2009). Method (i) and (ii) are parametric whereas methods (iii) and (iv) are semi-parametric. Methods (ii)-(iv) are also described in the context of this article in Supplemental Material.
Because methods (i)-(iv) may lead to the same inference of causal effect of treating hospital, we only present the causal effect of diagnosing hospital. In Sec. 4, we have applied method (i) to estimate the blip effect of large diagnosing hospital among women;
of large diagnosing hospital among men;
of large diagnosing hospital in the population; optimal SCE
under optimal regime of diagnosing and treating hospitals versus small ones in the population. Here, we additionally estimate the optimal blip effect of large diagnosing hospital:
because method (iv) does not estimate
The result from these methods is presented in . From this table, we have the following observations.
Table 3. Comparison of our method with available methods in Sec. 5: estimate, p-value and 95 % confidence interval (95% CI) for causal effects of diagnosing and treating hospitals on one-year survival of cardia cancer.
Method (i) yields the estimates for all five causal effects:
Here we only need to estimate a total of two parameters:
for the point effect of
and
for the point effect of
to estimate the blip effect and the SCE under any regime, as described in Sec. 4. The estimates are in consistency with the medical knowledge. Because SCEs are all estimated from the same estimated SNMM, we can compare different regimes by testing their SCEs, for instance, we compare
with
by testing
Method (ii) only yields the estimates for three causal effects:
because of imbalanced and sparse data between exposure sequence
and gender
Here one needs to estimate a total of
standard parameters
to estimates these causal effects (curse of dimensionality), which leads to wide confidence intervals. The estimate
is biased, because the optimal SCE should be larger than or equal to zero. The bias may be due to the null paradox (Hernan and Robins Citation2020; Robins Citation1997, Citation2009; Taubman et al. Citation2009). For a long sequence of cancer diagnoses and treatments, the problems with the null paradox and the curse of dimensionality in method (ii) become far worse. In this case, method (i) may provide a useful tool for estimating and testing the causal effect of cancer diagnosis.
Method (iii) also yields the estimates for only three causal effects:
The estimate is not medically sensible. Also notice the wide confidence intervals. The reason is due to imbalanced and sparse data between exposure sequence
and gender
(Hernan and Robins Citation2020; Hernan, Brumback, and Robins Citation2000; Robins, Brumback, and Hernan Citation2000; Robins Citation2009).
Method (iv) yields estimates for four causal estimates:
The estimates
and
are rather close to those obtained from method (i), albeit with lower significance. However, it does not estimate
Generally, method (iv) does not facilitate comparison between different regimes.
6. Conclusion
In many practices, a single-point treatment can hardly achieve the desirable result. More often, a sequence of treatments are assigned or implemented, where a new situation often arises from the early treatments and also influences the assignment of subsequent treatments. A typical example is the stochastic process of cancer diagnosis. In such circumstances, it is highly difficult to design and conduct sequential randomized trials in comparison to randomized trials for single-point treatments. Consequently, data arising from a sequence of treatments are often observational, and this requires statistical modeling when estimating and testing SCEs under specified regimes of treatments.
In this article, we introduce a parametric method of estimating and testing SCEs in the observational study settings by modeling based on the alternative G-formula. Our method is implemented in three steps: first to estimate and test the point effect, second to use the estimated point effect to estimate and test the blip effect, and finally to use the estimated blip effects to estimate and test SCEs. Each of these steps is intuitive and can be carried out using regressions familiar to epidemiologists in single-point causal inference.
With its simple setting, we can readily extend the example in this article to other observational studies. As to cancer diagnosis, we can similarly study various cancer types, different cancer outcomes such as cancer progression, quality of life and others, different diagnosing techniques such as the biomarkers or a sequence of screening steps, and different modification factors such as social-economic status, comorbidity, and family history.
Ethics approval
The proposed research is covered by the ethical committee approval (DNR880113/13, §121) from the ethical review board of Uppsala University.
This work was supported by Vetenskapsrådet, Sweden (2019-02913).
Supplemental Material
Download MS Word (38.5 KB)Supplemental material
Data and code available in [Zenodo], at https://doi.org/10.5281/zenodo.6367502
A description of available methods (ii), (iii) and (iv) in the context of the medical example.
Additional information
Funding
References
- Almirall, D., T. T. Have, and S. A. Murphy. 2010. Structural nested mean models for assessing time-varying effect moderation. Biometrics 66 (1):131–9. doi:10.1111/j.1541-0420.2009.01238.x.
- Bellman, R. 1957. Dynamic programming. Princeton, NJ: Princeton University Press.
- Chaffe, P. H, and M. J. van der Laan. 2012. Targeted maximum likelihood estimation for dynamic treatment regimes in sequentially randomized controlled trials. International Journal of Biostatistics 8:14.
- Hansson, L. E., A. M. Ekström, R. Bergström, and O. Nyrén. 2000. Surgery for stomach cancer in a defined Swedish population: Current practices and operative results. The European Journal of Surgery = Acta Chirurgica 166 (10):787–975. doi:10.1080/110241500447425.
- Henderson, R., P. Ansell, and D. Alshibani. 2010. Regret-regression for optimal dynamic treatment regimes. Biometrics 66 (4):1192–201. doi:10.1111/j.1541-0420.2009.01368.x.
- Hernan, M. A., B. Brumback, and J. M. Robins. 2000. Marginal structural models to estimate the causal effect of zidovudine on the survival of HIV-positive men. Epidemiology 11:561–70.
- Hernan, M. A, and J. M. Robins. 2020. Causal inference: What if. Boca Raton, FL: Chapman & Hall/CRC.
- Kosorok, M. R, and E. B. Laber. 2019. Precision medicine. Annual Review of Statistics and Its Application 6:263–86. doi:10.1146/annurev-statistics-030718-105251.
- Kosorok, M. R., Laber, E. B. Small, D. S. Donglin, and Zeng, D. 2021. Introduction to the theory and methods special issue on precision medicine and individualized policy discovery. Journal of the American Statistical Association 116 (533):159–61. doi:10.1080/01621459.2020.1863224.
- Murphy, S. 2003. Optimal dynamic treatment regimes. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 65 (2):331–66. doi:10.1111/1467-9868.00389.
- Robins, J. M. 1986. A new approach to causal inference in mortality studies with sustained exposure periods – Application to control of the healthy worker survival effect. Mathematical Modelling 7 (9–12):1393–512. doi:10.1016/0270-0255(86)90088-6.
- Robins, J. M. 1997. Causal inference from complex longitudinal data. In Latent variable modeling and applications to causality, Lecture Notes in Statistics 120, ed. M. Berkane, 69–117. New York: Springer-Verlag.
- Robins, J. M. 2004. Optimal structural nested models for optimal sequential decisions. In Proceedings of the second seattle symposium in biostatistics, analysis of correlated data, eds, D. Y. Lin, and P. J. Heagerty, 189–326. New York: Springer-Verlag.
- Robins, J. M. 2009. Longitudinal data analysis. In Handbooks of modern statistical methods, ed. G. Fitzmaurice, 553–99. Boca Raton, FL: Chapman & Hall/CRC.
- Robins, J. M., B. Brumback, and M. A. Hernan. 2000. Marginal structural models and causal inference in epidemiology. Epidemiology (Cambridge, Mass.) 11 (5):550–60. doi:10.1097/00001648-200009000-00011.
- Rosenbaum, P. R, and D. B. Rubin. 1983. The central role of the propensity score in observational studies for causal effects. Biometrika 70 (1):41–55. doi:10.1093/biomet/70.1.41.
- Sutton, R, and A. Barto. 1998. Reinforcement learning: An introduction. MIT Press, Cambridge.
- Taubman, S. L., J. M. Robins, M. A. Mittleman, and M. A. Hernán. 2009. Intervening on risk factors for coronary heart disease: An application of the parametric G-formula. International Journal of Epidemiology 38 (6):1599–611. doi:10.1093/ije/dyp192.
- Wang, X, and L. Yin. 2015. Identifying and estimating net effects of treatments in sequential causal inference. Electronic Journal of Statistics 9 (1):1608–43. doi:10.1214/15-EJS1046.
- Wang, X, and L. Yin. 2020. New G-formula for the sequential causal effect and blip effect of treatment in sequential causal inference. Annals of Statistics 48:138–60.
- Zhang, B., A. A. Tsiatis, E. B. Laber, and M. Davidian. 2013. Robust estimation of optimal dynamic treatment regimes for sequential treatment decisions. Biometrika 100 (3):681–94. doi:10.1093/biomet/ast014.