Comparing Lung Cancer Screening Strategies in a Nationally Representative US Population Using Transportability Methods for the National Lung Cancer Screening Trial

Key Points Question What is the comparative effectiveness of low-dose computed tomography (CT) vs chest radiography screening strategies evaluated in the National Lung Screening Trial (NLST) in a nationally representative target population of US adults who meet the NLST eligibility criteria? Findings In this comparative effectiveness study, transportability analysis was used to reweight NLST data to resemble a nationally representative target population. Estimates of the comparative effectiveness of low-dose CT screening compared with chest radiography on lung cancer–specific and all-cause mortality in the target population of 5.7 million adults were similar to estimates from unweighted NLST analyses, but differences in baseline characteristics between the NLST and the target population resulted in increased uncertainty. Meaning These results suggest that low-dose CT screening resulted in improved outcomes compared with chest radiography in a nationally representative population NLST-eligible individuals, indicating that the trial findings are transportable to this target population.


eMethods.
Transportability weights: We obtained weights for the transportability analyses by estimating the probability of trial participation and the probability of assignment to each screening strategy in the trial, conditional on the covariates in Table 1 of the main text.To avoid extreme weights and improve balance, we chose to collapse the following multi-category variables from Table 1 into binary variables: race was coded as white versus non-white, marital status was coded as married/living as married versus all others, and education was coded as any college versus no college.To form the transportability weights, we calculated the inverse of the estimated odds of trial participation. 1 To estimate the probability of trial participation and the probability of assignment to each screening strategy in the trial, we used logistic regression models.Specifically, the probability of trial participation was estimated using a weighted logistic regression with weights for the NHIS participants set to the survey sample weights and weights for the NLST participants set to 1.The probability of assignment to each screening strategy in the trial was estimated using unweighted logistic regression using the trial data alone.To estimate measures of incidence and measures of association in the target population, we used outcome regressions with treatment as the only predictor in the model and obtained standard errors for the model parameters using the robust estimator of the sampling variance. 1amining the transportability weights: To examine the transportability weights, we calculated the ratio of the sum of the weights over all individuals in the composite dataset, divided by the number of individuals in the target population (after using the survey weights). 1Values of this statistic that are meaningfully different from 1 suggest violations of the assumption that every pattern of these covariates in the target population has a non-zero probability of being observed in the trial (i.e., positivity over trial participation status) or misspecification of the model for the probability of participation.In our main analysis (using the fully adjusted model with weights trimmed to the 99 th percentile of their empirical distribution), the value of the diagnostic was 1.05, suggesting a lack of violations.
Handling missing data: The analyses presented in the main text used observations with complete information on the baseline covariates listed in Table 1 of the main text.Missingness was limited for covariates; no assignment data was missing in the trial.As a stability analysis, we used inverse probability of missingness weighting to adjust for missingness in any baseline covariates. 2 In the model for the probability of missingness among NLST participants, we included the screening intervention, outcome, and all of the baseline covariates used in the fully adjusted model with complete observations (age, sex, pack-years, current smoking status) as predictors.In the model for the probability of missingness among NHIS participants, we included baseline covariates used in the fully adjusted model with complete observations (age, sex, ethnicity, pack-years, current smoking status, body mass index, race, and diabetes) as predictors.We repeated the trial-only analysis and transportability analyses incorporating the missingness weights.

Sensitivity analyses:
We conducted sensitivity analyses to examine how violations of the transportability assumption would affect our results.These sensitivity analyses estimate the comparative effectiveness of screening strategies under different magnitudes of assumption violations by varying the strength of association between participation in the NLST and the potential outcomes under each screening strategy, conditional on the observed covariates.We used the NLST and 2010 NHIS data to examine the sensitivity of the transportability analysis for lung cancer-specific mortality on the rate ratio scale.We modified a global sensitivity analysis approach 3,4 that uses an odds of selection model with sensitivity parameters that express the conditional association between trial participation and the (unobservable) potential outcomes under the two screening strategies compared in the trial, given the measured covariates. 5,63][4] We estimated the effectiveness of screening in the target population using the approach described in the main text -Poisson regression, weighted by the transportability weights -with the transportability weights were re-estimated for different values of the sensitivity parameters.As in the main analysis, we used the fully adjusted model when forming the transportability weights and trimmed the weights to the 99 th percentile.
Different values of the sensitivity parameters express different beliefs about the association between trial participation and the potential outcome under each treatment (separately for each screening group in the trial) on the logit scale, given the measured covariates included in the analysis.We examined values of the sensitivity parameters ranging from -0.30 (denoting that lung cancer-specific mortality under the screening or control intervention made participation more likely) to 0.30 (denoting that lung cancer-specific mortality under the screening or control intervention made participation less likely).
The results of the sensitivity analysis are shown in the contour plot of eFigure 1.In this plot, the black "x" corresponds to the point estimate of the main analysis, when the sensitivity parameters are set to 0 (i.e., when the transportability assumption holds).The estimated effectiveness of screening varies according to the value of the sensitivity parameters.For sensitivity parameter values that correspond to the diagonal that runs from the bottom left to the top right (i.e., the diagonal where the "x" falls; teal blue color), the results are close to those from our main analysis for lung cancer-specific mortality, even for fairly strong violations of the assumption.That is the case because for this set of potential violations of the assumption of selection into the trial is not differentially associated with the potential outcomes for lung cancerspecific mortality under each screening strategy (e.g., they reflect situations in which the risk of the outcome in the target population may be lower or higher than the population enrolled in the trial, but not differentially between different screening strategies).
For sensitivity parameter values corresponding to the region below the diagonal (blue color), the main analysis results overestimate the benefit of low-dose CT screening.This would happen, for example, if the trial preferentially enrolled people who have reduced lung cancerspecific mortality with low-dose CT screening versus chest radiography screening (because of unmeasured variables) compared to the national population.For sensitivity parameter values corresponding to the region above the diagonal (yellow color), our main analysis results would underestimate the benefit of low-dose CT screening.This would happen, for example, if the trial preferentially enrolled people who had increased lung cancer-specific mortality with low-dose CT screening versus chest radiography screening (because of unmeasured variables) compared to the national population.In summary, under violations of the transportability assumption reflecting differential selection into the trial, our main analysis results could either underestimate or overestimate the benefit from low-dose CT screening compared with chest radiography screening.To change our qualitative conclusions about the comparative effectiveness of these screening strategies, one would have to assume moderately strong and differential selection into the trial (compared to the target population) of individuals who would benefit more (or be harmed less) by low-dose CT screening compared with chest radiography screening.The black x corresponds to the base case of the main analysis, when the sensitivity parameters are set to 0. The estimated effect of screening (indicated by the color gradient levels) varies according to the value of the sensitivity parameters.

eTable 1 .
List of Eligibility Criteria from the NLST Protocol and Corresponding Implementation in NHIS All-Cause Mortality Results for the Trial Only (Unadjusted) and Transportability Analyses (Adjusting for Different Covariate Sets) for Low-Dose CT vs Chest Radiography, With Weights in the Adjusted Analysis Trimmed to the 99 th Percentile Among Trial Participants Lung Cancer-Specific Mortality Results for the Trial Only (Unadjusted) and Transportability Analyses (Adjusting for Different Covariate Sets) for Low-Dose CT vs Chest Radiography, Using a Cox Proportional-Hazards Model All-Cause Mortality Results for the Trial Only (Unadjusted) and Transportability Analyses (Adjusting for Different Covariate Sets) for Low-Dose CT vs Chest Radiography, Using a Cox Proportional-Hazards Model All-Cause Mortality Results for the Trial Only (Unadjusted) and Transportability Analyses (Adjusting for Different Covariate Sets) for Low-Dose CT vs Chest Radiography When Adjusting for Missingness, With Weights in the Transportability Analysis Trimmed to 99% Lung Cancer-Specific Mortality Results for the Transportability Analyses Using the Fully Adjusted Model, With Weights in the Transportability Analysis Trimmed to 99% All-Cause Mortality Results for Using the Fully Adjusted Model, With Weights in the Transportability Analysis Trimmed to 99% Sensitivity Analysis Results, on the Rate Ratio Scale for Lung Cancer Mortality, for Violations of the Transportability Assumption eFigure 2.