Applying Evidence Synthesis for Constructing Directed Acyclic Graphs to Identify Causal Pathways Affecting U.S. Early-Stage Non-Small Cell Lung Cancer Treatment Receipt and Overall Survival

Abstract

Background/Objectives: Directed acyclic graphs (DAGs) inform the epidemiologic statistical modeling confounders to determine close to true causal relationships in a study context. They inform the inclusion of the predictive model variables that affect the causal relationship. Non-small cell lung cancer (NSCLC) is frequently diagnosed, aggressive, and the second leading cause of cancer deaths in the United States. Determining factors affecting both the guideline-concordant treatment receipt and survival outcomes for early-stage lung cancer will help inform future statistical models aiming to achieve a close to true causal relationship. Methods: Peer-reviewed original research published during 2002–2023 was identified through PubMed, Embase, Web of Sciences, Clinical trials registry, and the gray literature. DAGitty version 3.1, an online software program, developed implied DAGs and integrated DAG graphics. The evidence synthesis for constructing directed acyclic graphs (ESC-DAGs) protocol was utilized to guide DAG development. The conceptual models utilized were Andersen and Aday for factors affecting treatment receipt and Shi and Steven for survival outcome factors. Results: A total of 36 studies were included in the DAG synthesis out of 9421 retrieved across databases. Eight studies served in the synthesis of treatment receipt DAG, while 28 studies were used for the survival outcomes DAG. There were 10 causal paths and 13 covariates for treatment receipt and 2 causal pathways and 32 covariates for survival outcomes. Conclusions: There are very few studies reporting on factors affecting early-stage NSCLC guideline-concordant care receipt compared to factors affecting its survival outcomes in the past two decades of original research. Future investigations can utilize data extracted in the current study to develop a meta-analysis informing effect size.

1. Introduction

A long-standing association exists between lung cancer survival and socioeconomic factors in epidemiologic studies on early-stage non-small cell lung cancer (NSCLC) [1]. Several factors contribute to the geographic differences among early-stage NSCLC regarding comorbidity status and carcinogen exposure [2]. Geographic area is a critical factor in treatment utilization and survival for early-stage NSCLC [2]. Treatment modalities for early NSCLC include surgery for medically fit candidates or radiation therapy for medically unfit candidates [3]. Differentiating treatment modalities are associated with survival outcome differences [2,4]. Hence, it is important to establish factors that cause differences in treatment receipt and survival outcomes for this frequently diagnosed, differentially treated, and aggressive cancer to achieve a close to true causal relationship through predictive modeling in cancer epidemiology.

DAG is a simple graphical representation of causal relation assumptions in the study context and multiple factors that must be accounted for to obtain the unconfounded relationship between the exposure and the outcome variable [5]. Conventional statistical models contain several parametric assumptions that may or may not be correct [5]. This is a drawback when identifying assumptions in a study context and model violation [5]. Causal diagrams depicted through DAG represent those study assumptions that can complement statistical models [5]. Causal diagrams illustrate causal relationships without any parametric assumptions, as in the case of conventional statistical models [5]. However, causal diagrams capture the series of causation in the current study context, which a conventional model might not be equipped to perform. Some causal relation assumptions might be untested or unknown, but a causal diagram can capture all possible causal pathways [5]. However, no scientific literature exists that develops directed acyclic graphs (DAGs) through a historic (past 21 years) systematic review regarding factors affecting treatment receipt and survival outcomes among early-stage NSCLC according to a search carried out on PubMed, Embase, Web of Science, and Google.

This study aimed to identify factors associated with non-treatment among TN0M0 NSCLC with the first primary tumor and determine risk factors associated with lung cancer-specific survival after surgery and radiation therapy through a systematic review. What factors affect overall survival (OS) in patients with early-stage primary NSCLC in the United States (U.S.)? What determines the treatment choice? These are the key questions that we aimed to seek through a comprehensive DAG-guided systematic literature review of the topics. We reviewed the existing literature that has sought to measure these factors, especially from the perspective of treatment selection and lung cancer-specific survival among national cancer registry data or clinical trials. To the best of our knowledge, this is the first study to build DAGs informed by the health services research theory conceptual model and to utilize the longest research published year to inculcate evolving medical advances focused on a specific stage of lung cancer. This study will guide future statistical model decision-making by determining the pathways that need to be accounted for to achieve close to true causal relationships. This study will also serve as a foundation for future meta-analysis investigations for early-stage NSCLC in the US.

2. Materials and Methods

A literature search strategy was developed with the assistance of a librarian expert, oncologist, and health economist to ensure that exhaustive literature was included. Three key databases were identified: Embase, PubMed, and Web of Science. For the gray literature, the Connected Papers database [6], manual searching by bibliographically browsing key research articles relevant to the study topic, and the Clinicaltrials.gov [7] database were used. An approach was strategized during the screening phase of the study in an attempt to be consistent across study periods dealing with three different AJCC stages. The hierarchical strategy was informed by the American College of Surgeons (ACS) and the American Cancer Society, which emphasize TNM staging serving as a foundation for defining the overall AJCC staging system. The definition of early-stage from AJCC 6th to 8th moves from broader categorization of T staging to more granular categorization, and for the same reason, the study characteristics Appendix A Table A1 informs about the particular tumor staging each finalized study included, since the definition of stage 1 was relative across AJCC 6th to 8th. Clinical staging informs definitive treatment decision-making and affects survival outcomes, so it is very important to consider studies published after 2001, as the AJCC 6th edition was implemented after that year. No publication year filter was applied for studies found through gray literature searching to capture studies that might be important and relevant to the current topic context. Limited empirical evidence is available to develop a systematic review bibliographic search strategy for healthcare. However, an experiment determined that significant relevant studies were found on Embase, ranking it the second highest of all the pertinent databases in terms of search results [8]. PubMed comprises citations from Medline, another relevant medical literature database, while Web of Science provides only peer-reviewed studies. The decision to use these three databases was made after consultation with a librarian expert and an oncologist. The search strategy across each database is described in the Appendix A. DAGitty version 3.1, an online software program, developed implied DAGs and integrated DAG graphics [9]. Abstrackr, an online free literature screening software, was utilized to maintain objectivity [10]. All the citations from identified databases were imported to Mendley, a reference manager citation software, in which duplicates were removed. The citations were then moved to an Excel sheet in a format in which three researchers (NP, SK, and ME) independently reviewed the first 100 records for title and abstract screening, after which reconciliation was reached through an in-person discussion. Then, independently, two researchers (NP and SK) screened 4646 records for title, abstract, and full-text screening, and in cases in which a consensus was not reached, the third researcher (ME) was reached to resolve the disagreement. The total agreement rate was 89.4%, while the disagreement rate was 10.6%. The data was extracted manually from the final sample by two researchers, NP and SK, independently and later discussed for reconciliation if necessary. The current review complied with PRISMA guidelines, and the PRISMA checklist is described in Appendix B. The current review is not registered.

2.1. Study Inclusion and Exclusion Criteria

Only studies focused on the U.S. were moved toward the final sample from the body screening stage, as the clinical staging and treatment guidelines differ internationally. Additionally, the current study aims to develop causal diagrams to supplement the future statistical model variables utilizing U.S. data. The homogeneity of the included sample in terms of the country was emphasized better to understand the causal relationship within the location context; for the literature database, the publication year was set to 2002–2023 to only capture studies relevant to recent medical advances in this field, as well as clinical staging AJCC 6th and higher. The included studies ranged from AJCC 6th to 8th edition; hence, a strategy of following TN0M0 staging convention first in hierarchy decision was developed to be consistent in the definition of non-metastatic tumors, as mentioned in the protocol section of this paper, in an attempt to avoid the exclusion of studies that do not refer to specific AJCC staging information yet focus on overall stage 1. The title, abstract, and body screening questions were as follows: (1) Is it about stage 1 first primary NSCLC TN0M0? (2) Is the document an article? (3) Is it quantitative research? (4) Is it about assessing the factors affecting treatment receipt? OR is it about assessing the factors affecting survival outcomes?

2.2. Protocol

The protocol for developing the final integrated DAGs (iDAGs) from the shortlisted literature was informed by “Evidence Synthesis for Constructing Directed Acyclic Graphs” (ESC-DAGs) [11]. The protocol comprises three main stages, i.e., mapping, translation, and integration (synthesis and recombination). The current review is divided into two components: (a) factors affecting treatment receipt and (b) factors affecting survival outcomes. The conceptual model utilized in the translation stage of ESC-DAGs for factors affecting treatment receipt is Andersen and Aday’s [12] health services research behavioral model. As for factors affecting survival outcomes, the conceptual model by Shi and Steven [13] for vulnerable populations is utilized. Therefore, the translation stage decisions were guided solely by these conceptual frameworks for each component regarding temporality and construct validity.

3. Results

A total of 36 of 9421 studies qualified for final data extraction, as reflected in the PRISMA flowchart in Figure 1. The detailed PRISMA checklist is provided in Appendix B. The baseline characteristics of each study are described in Appendix A, Table A1. It describes the study setting, study period, data registry utilized for the study, age range of the population included, sample size, type of intervention (exposure variable), outcome, AJCC staging version used for study inclusion, and component under which the study falls. The qualified literature study period ranged from 1988 to 2021, and the study designs were both observational and clinical trials. The observational studies utilized the SEER, National Cancer Database (NCDB), SEER-Medicare linked, California Cancer Registry (CCR part of the SEER registry), and primary data collection in clinical trials.

Figure 1. PRISMA flowchart.

3.1. Mapping Stage

In this stage, each qualified study was used to extract data to develop implied DAGs using DAGitty software separately for each component of the review (Table 1 and Table 2). The studies were read in detail to determine the significant confounders, unobserved/unadjusted confounders, mediators, and colliders. In the implied DAG, the gray bubbles are the identified confounders in each study that were not adjusted in their statistical analysis model. The green bubbles are the study’s exposure variables, and the blue bubble is the study’s outcome variable. Green arrows are the front door causal pathways, while purple indicates the back door pathways that must be closed to achieve a true causal relationship within the study relationship context. In this stage, implied DAGs were developed as determined by the studies, and arrow edges were directed as identified in the study results or conclusions.

Table 1. Factors affecting treatment receipt mapping stage of ESC-DAG.

	Exposure Variable Group	Outcome Variable	Odds Ratio (95% CI)	Confounders Based on Study Conclusion	Identified Mediators	Identified Colliders	Statistical Analysis Approach
1 [14]	Race	Treatment receipt	Black men: 0.13 (0.04–0.47) Black women: 0.87 (0.13–3.69)	Age at diagnosis Smoking status Coronary artery disease COPD Tumor histology Diagnosis after screening	None	None	Multinomial logistic regression
2 [15]	Race	Treatment receipt	Black: 1.43(1.18–1.72)	Age Race Gender Marital status Insurance type Neighborhood Socioeconomic Status NCI designation of hospital Tumor size	None	None	Multivariable logistic regression
3 [1]	Race	Treatment receipt	Black: 1.58	Race Gender Marital Status Age Income Insurance status	None	None	Multivariate logistic regression
4 [2]	Geographic region	Treatment receipt	Rural: 0.04 Urban: 0.24	Gender Race Insurance status Income Age Health status	None	None	Multivariate logistic regression
5 [16]	Geographic region	Treatment receipt	Urban: 0.92 (0.85–1.01) Rural: 1 (0.82–1.22)	Age Race Lower education Lower median income	None	None	Multivariate logistic regression
6 [4]	Race	Treatment receipt	Black: 0.59	Insurance status Education Race Age Gender Comorbidity score Tumor size Geographical area of residence	None	None	Multivariate logistic regression
7 [17]	Race	Treatment receipt	Black: 0.61 (0.58–0.64)	Gender Geographic region Type of treatment facility Rural urban region Insurance Status Comorbidity score	None	None	Multivariate logistic regression
8 [18]	Disability status	Treatment receipt	Disabled: 0.27	Age	None	None	Bivariate logistic regression

Table 2. Factors affecting survival outcomes mapping stage of ESC-DAG.

	Exposure Variable Group	Outcome Variable	Odds Ratio (95% CI)	Hazards Ratio (95% CI)	Confounders Based on Study Conclusion	Identified Mediators	Identified Colliders	Statistical Analysis Approach
1 [19]	Treatment type	Survival		1.38 (0.70–2.73)	Tumor size Tumor histology Patient performance status Age	None	None	Propensity Score Matching
2 [20]	Treatment type	Survival		1.66 (1.51–1.83)	Age Gender Tumor size Histology type	None	None	Cox regression
3 [21]	Treatment type	Survival		Surgery: 0.18 Radiation: 0.51 Both: 0.36	Age Gender Tumor histology T staging	None	None	Cox regression
4 [22]	Treatment type	Survival	Surgery: 3.65 Radiation: 7.43		Age Histology type	None	None	Kaplan Meier and Log rank test
5 [23]	Histologic type	Survival		Lobectomy: 0.92 (0.83–1.02)	Treatment type Age Gender Tumor grade Number of resected lymph nodes Tumor size	None	None	Propensity Score Matching Cox regression
6 [3]	Treatment type	Survival		Surgery: 0.91 (0.86–0.96) Radiotherapy: 0.77 (0.71–0.83)	Year of diagnosis Gender Race Age	None	None	Cox regression
7 [24]	Marital status	Survival		Married: 0.85 (0.82–0.89) Divorced: 1.08 (1.02–1.15)	Gender Race Tumor grade Age Tumor size	None	None	Cox regression
8 [25]	Treatment type	Survival		Segmentectomy: 0.83 (0.71–0.96) Radiotherapy: 0.65 (0.52–0.81)	Age Gender Tumor size Tumor grade Adjuvant therapy	None	None	Propensity Score Matching Cox regression
9 [26]	Treatment type	Survival	Radiofrequency ablation (RFA): 1.25		Hospital region Year of diagnosis Tumor size	None	None	Propensity score matching
10 [27]	Treatment type	Survival	RFA: 1.23SBRT: 0.13		None	None	None	Propensity score matching
11 [28]	Treatment type	Survival		RFA: 0.97 (0.86–1.11)	Gender Age Tumor size Histologic type Tumor grade Insurance status	None	None	Propensity score matching and Cox regression
12 [29]	Treatment type	Survival		Lobectomy: 0.78 (0.41–1.48) Adjuvant radiotherapy: 0.14 (0.03–0.64)	Gender Number of Lymph nodes examined	None	None	Propensity score matching and Cox regression
13 [30]	Treatment type	Survival		Sublobectomy: 1.40 (1.25–1.58)	Age Gender Tumor grade Histologic type Tumor size Number of lymph nodes sampled	None	None	Propensity score matching and Cox regression
14 [31]	Treatment type	Survival		Radiotherapy: 2.42 (2–3)	Age Gender Histologic type Number of Lymph nodes examined Tumor grade Year of diagnosis Tumor size	None	None	Propensity score matching and Cox regression
15 [32]	Treatment type	Survival		Lobectomy: 0.76 (0.60–1) Segmentectomy: 0.80 (0.54–1.18)	Age Gender Lymph nodes examined status Tumor grade	None	None	Cox regression
16 [33]	Treatment type	Survival		Segmentectomy: 1.44 (1.11–1.86)	Age Tumor grade	None	None	Propensity score matching and Cox regression
17 [34]	Treatment type	Survival		Lobectomy: 0.82 (0.77–0.87)	Race Tumor size Gender Tumor grade Age	None	None	Cox regression
18 [35]	Treatment type	Survival		SBRT: 1.56 (1.50–1.62) RFA: 1.91 (1.73–2.10) VATS: 0.55 (0.52–0.60)	Age Gender Race Treatment facility type Income Treatment facility location Comorbidity score Tumor size Tumor grade Tumor histology	None	None	Propensity score matching and Cox regression
19 [36]	Treatment type	Survival		Segmentectomy: 0.89 (0.54–1.46) Wedge resection: 1.29 (0.97–1.72)	Lymph node dissection	None	None	Cox regression
20 [37]	Treatment type	Survival	Segmentectomy: 0.88 (0.76–1.02)		Age Tumor grade Tumor histology Number of Lymph nodes dissected Gender Tumor size	None	None	Kaplan Meier Log rank and Multivariate analysis
21 [38]	Treatment type	Survival		Segmentectomy: 1.35 (1.18–1.54) Adjuvant radiotherapy: 1.91 (1.58–2.30)	Age Year of diagnosis Gender Tumor size Marital status Insurance status Tumor grade Histologic type Number of lymph nodes dissected	None	None	Cox regression
22 [39]	Treatment type	Survival		Thermal ablation: 1.40 (1.04–1.86) Adjuvant radiotherapy: 1.68 (1.40–2.05)	Race Tumor size	None	None	Propensity score matching and Cox regression
23 [40]	Treatment type	Survival		Sublobectomy: 1.45 (1.35–1.56)	Gender Tumor size Number of lymph nodes sampled Age	None	None	Cox regression
24 [41]	Treatment type	Survival		Lobectomy: 1.12 (0.93–1.35)	None	None	None	Propensity score matching and Cox regression
25 [42]	Radiation therapy	Survival		Radiotherapy: 0.90 (0.84–0.97)	Number of Lymph nodes examined Age Gender Tumor size Tumor grade Tumor histology	None	None	Cox regression
26 [43]	Treatment type	Survival		Lobectomy: 1.01 (0.93–1.11)	Age Gender Tumor size Tumor histology Tumor grade Number of lymph nodes examined	None	None	Cox regression
27 [44]	Treatment type	Survival		SABR: 0.86 (0.45–1.65)	Age Gender Race Region of enrollment Tumor histology Tumor size Mediastinal lymph node examination	None	None	Kaplan Meier
28 [45]	Treatment type	Survival	RFA: 0.21 (0.00–0.44)		Age Gender Region of enrollment Performance status Vital capacity Lung function	None	None	Kaplan Meier

3.2. Translation Stage

At this stage, the extracted implied DAGs for each study were utilized to build a DAG edge index (Appendix A, Table A2 and Table A3) to determine the arrow directionality decision-making between an implied set of variables. To reach objective decisions, the proposed theoretical frameworks for each component were utilized to determine whether the arrow directionality was accurate. While deciding to remove or retain the edge, the construct validity and temporality of the edge direction were determined using the theoretical framework. Bi-directionality was determined for each study individually by utilizing their implied graphs to determine if the edge direction of the arrow was bi-directional, given the set of variables in the existing study context. For factors affecting treatment receipt (Appendix A, Table A2), Andersen and Aday’s [12] theoretical framework was used to guide the construct validity and temporality of a particular arrow direction in a given set of variables. Likewise, Shi and Steven’s [13] theoretical framework for vulnerable populations was utilized to identify the factors affecting survival outcomes (Appendix A, Table A3).

3.3. Integration Stage

3.3.1. Factors Affecting Treatment Receipt

The outcome variable of interest was treatment receipt (Figure 2). Studies with no specific exposure variable were inconclusive regarding the back door causal pathways in iDAGs. There were 10 causal paths and 13 covariates in the iDAG for the exposure variables of interest on the outcome variable. The total effect adjustment for the given effect of interest suggests controlling for only the following necessary variables to close all the back door paths (purple lines): age, chronic obstructive pulmonary disease (COPD), comorbidity score, coronary artery disease, education, sex, health status, income, insurance status, marital status, patient preference, physician preference, and tumor size. The front door paths (green lines) represent the effects of interest in the extracted studies.

Figure 2. Integrated DAG for factors affecting treatment receipt.

The following conditional independence testable implications are identified by the iDAG results for total effect adjustment. After adjusting for age and type of treatment facility, the comorbidity score was not related to disability status. In addition, the comorbidity score was unrelated to geographic region, insurance status, marital status, patient preference, physician preference, tumor size, education, sex, income, and race.

Adjusting for age and race, coronary artery disease was unrelated to disability status, geographic region, and type of treatment facility. After adjusting for age and type of treatment facility, coronary artery disease was not related to disability status. It is also unrelated to insurance status, marital status, patient preference, physician preference, tumor size, type of treatment, age, COPD, education, gender, or income.

Adjusting for age, sex, geographic region, insurance, and race, disability status was not related to health status, marital status, or income, while adjusting for age, race, and disability status was unrelated to COPD. Adjusting for age and type of treatment facility, disability status was unrelated to income, race, COPD, education, sex, geographic region, health status, insurance status, tumor size, and marital status.

Hence, future meta-analysis research from the U.S. context attempting to identify factors affecting treatment receipt might benefit by individually investigating each causal relationship between geographic region, type of treatment facility, race, or disability status and treatment receipt. Accounting for individual independent factors in a statistical model, future research must consider differential causal pathways, i.e., while statistically measuring true causal relationship estimate between geographic region and treatment receipt, one must account for significant identified confounders like age, gender, insurance status, marital status, income, and health status. Moreover, a comprehensive approach might include all the independent factors together. However, as identified through integrated DAGs (Figure 2), several confounders, if adjusted together in a statistical model in such an approach, might pose multicollinearity problems.

3.3.2. Factors Affecting Survival Outcomes

The iDAG (Figure 3) has two causal pathways and 32 covariates for the two exposures of interest, marital status, treatment type, and the outcome variable (survival). The two exposure variables are informed by the extracted studies, and Shi and Steven’s [13] theoretical framework verifies its temporality. The total effect adjustment for the given effect of interest suggests controlling for only the following necessary variables to close all the back door paths (purple lines): access to care, adjuvant therapy, age, cardiopulmonary function, comorbidities, enrollment bias, sex, hospital region, imaging information, insurance status, lung function, mediastinal lymph node examination, number of lymph nodes examined, number of lymph nodes resected, patient functional status, patient preference, provider bias, quality of life, race, recurrence rate, region of enrollment, smoking status, surgeon expertise, T staging, treatment facility location, treatment facility type, treatment selection criteria, tumor grade, tumor histology, tumor markers, tumor size, and year of diagnosis.

Figure 3. Integrated DAG for factors affecting survival outcomes.

The following conditional independence testable implications were identified by DAGitty diagnostics for the total effect adjustment of the developed iDAG. Access to care was not related to enrollment bias, imaging information, lung function, marital status, mediastinal lymph node examination, number of lymph nodes examined, number of lymph nodes resected, patient functional status, patient preference, provider bias, quality of life, recurrence rate, region of enrollment, smoking status, surgeon expertise, T staging, treatment facility type, adjuvant therapy, treatment selection criteria, tumor grade, tumor histology, tumor markers, tumor size, year of diagnosis, age, comorbidities, sex, race, and cardiopulmonary function.

Cardiopulmonary function was not related to enrollment bias, hospital region, imaging information, insurance status, lung function, marital status, mediastinal lymph node examination, number of lymph nodes examined, number of lymph nodes resected, patient preference, provider bias, quality of life, recurrence rate, region of enrollment, smoking status, surgeon expertise, T staging, treatment facility location, treatment facility type, tumor grade, tumor histology, tumor markers, tumor size, year of diagnosis, age, comorbidities, sex, and race.

Insurance status was not related to lung function, marital status, mediastinal lymph node examination, number of lymph nodes examined, number of lymph nodes resected, patient functional status, patient preference, provider bias, quality of life, recurrence rate, region of enrollment, smoking status, surgeon expertise, T staging, treatment facility location, treatment facility type, treatment selection criteria, tumor grade, tumor histology, tumor markers, tumor size, year of diagnosis, comorbidities, sex, or race.

The number of lymph nodes resected was not related to patient functional status, patient preference, provider bias, quality of life, recurrence rate, region of enrollment, smoking status, surgeon expertise, T staging, treatment facility location, treatment facility type, treatment selection criteria, tumor grade, tumor histology, tumor markers, tumor size, year of diagnosis, age, comorbidities, sex, and race.

Tumor grade was unrelated to tumor histology, tumor markers, tumor size, year of diagnosis, age, comorbidities, sex, and race. Tumor histology is unrelated to tumor markers, tumor size, year of diagnosis, age, comorbidities, sex, or race. Tumor markers were unrelated to tumor size, year of diagnosis, age, comorbidities, sex, and race. The tumor size was not related to the year of diagnosis, age, comorbidities, sex, or race. The year of diagnosis was not associated with age, comorbidities, sex, or race. Age was not related to comorbidities, sex, or race. Sex is not related to race, and comorbidities are not related to sex or race.

Hence, future meta-analysis research from the U.S. context attempting to identify factors affecting survival outcomes might benefit by individually investigating each causal relationship between marital status or treatment type and treatment receipt. Accounting for individual independent factors in a statistical model, future research must consider differential causal pathways, i.e., while statistically measuring true causal relationship estimate between treatment type and survival, one must account for significant identified confounders like access to care, provider bias, hospital region, tumor grade, quality of life, comorbidities, smoking status, etc. Moreover, a comprehensive approach might include all the independent factors together. However, as identified through integrated DAGs (Figure 3), several confounders, if adjusted together in a statistical model in such an approach, might pose multicollinearity problems.

4. Discussion

To the best of our knowledge and according to searches run on Google, PubMed, Embase, and Web of Science, this is the first study that utilized ESC-DAG for early-stage lung cancer. Hence, it is difficult to provide context with reference to other similar previous literature. However, the results of this study provide commonly identified confounders that corroborate with the final study sample and statements described in their results/conclusion sections. Eight studies provided information on the confounding factors affecting treatment receipt. In comparison, 28 studies provided information on the factors affecting survival outcomes. The confounding factors that affect treatment receipt are age, comorbidity score, education, sex, income, insurance status, marital status, patient preference, physician preference, tumor size, geographic location, and treatment facility type, which aligns with the existing literature. Therefore, adjusting for these confounding factors in a regression model can help improve the prediction abilities of the model in determining close to true causal relationships and the direct effect on the treatment receipt outcome variable.

The confounding factors that affected survival outcomes were access to care, adjuvant therapy, age, cardiopulmonary function, comorbidities, enrollment bias, sex, hospital region, imaging information, insurance status, lung function, mediastinal lymph node examination, number of lymph nodes examined, number of lymph nodes resected, patient functional status, patient preference, provider bias, quality of life, race, recurrence rate, region of enrollment, smoking status, surgeon expertise, T staging, treatment facility location, treatment facility type, treatment selection criteria, tumor grade, tumor histology, tumor markers, tumor size, and year of diagnosis, which can help determine close to true causal relationship and the direct effect of exposure variables on the outcome variable.

The present study has its limitations. Since the final study included a sample that is only focused on the U.S. population, the DAGs might differ when investigated under different study geography conditions, i.e., countries like the UK, Canada, Australia, and others have universal healthcare coverage irrespective of the socioeconomic status, while other countries have a different healthcare services infrastructure. Hence future studies might be able to account for several contextually different confounders, depending on the study setting. Moreover, results from the present study cannot be generalized to other country populations, as several confounders identified in this study, i.e., racial makeup, quality of life, inherent data collection, and country-specific registry limitations, as well as other country-specific health policies might differ that affect the study DAG pathways. Finally, there exists a publication bias, in general, irrespective of the country where original similar studies were conducted, which might affect the included study sample’s unreported factors, thereby affecting the causal pathways. Future studies might want to account for these factors while investigating a causal pathway.

5. Conclusions

The integrated DAGs developed in this study might serve as a supplement to inform statistical modeling decision-making for including confounding covariates in future studies to determine the factors affecting treatment receipt and survival outcomes among patients with stage 1 NSCLC TN0M0. The results of this study are not a substitute for other relevant regression diagnostics, such as correlations or multicollinearity. DAG is a graphical representation that helps identify all possible backdoor pathways to evaluate the total effect of exposures on outcome variables. Several factors, such as sample size, time trend, statistical modeling, composition of the study sample, sample selection bias, age group, type of data, study design, and type of intervention, contribute to the significance of confounders in the study context. Further implication testing can be carried out by future studies through statistical analysis to determine the effect of significance in a given study setting using meta-analysis and regression.