Effect Estimation And Comparative Study Of Instrumental Variable In Clinical Research

ObjectiveRCTs were well recognized since the first employment in streptomycin treatment of pulmonary tuberculosis. Well-designed blind RCTs were golden standard for effect estimation in intervention studies. The traditional analysis strategy such as ITT, PP and AT couldn’t keep the randomization randomize scheme or take compliance in to consideration at the same time due to the frequently observed protocol deviation and noncompliance. While stratified analysis, matching, multiple regression and propensity score method can be used to control confounding factors we known in observational studies, for the unobserved confounding factors, they can do nothing. Recently, instrumental variable method was more frequently used in clinical research, but it’s still scattered and vague, and no systematic research conducted from methodology viewpoint.This paper aims to give a systematic research on selection method, evaluation criterion, and effect estimation of instrumental variable in clinical research from the methodology viewpoint, and to compare IV, ITT, PP and AT estimation in RCTs through Monte Carlo stochastic simulation study.MethodsLiterature review method was used for selection method and evaluation criterion. Of instrumental variable in clinical research. Pubmed, Emabse, CNKI, CMJD and WANGFANG Data were searched with the keywords of "Instrumental variable".We conducted 1000*1000 times of Monte Carlo stochastic simulation, frequency distribution plot of regression coefficient, mean, standard error and 95%CI, PRB, and ECP will be used to evaluation and compare the results of IV, ITT, PP and AT analysis strategy under all kinds of scenarios for mean, rate or time to event data as primary endpoint.ResultsA total of 1401 English paper and 260 Chinese paper were identified. We selected 205 paper after evaluation. From result of literature review, we proposed a simple method to find IV. Random assignment variable can be excellent IV in RCTs, and in observational studies, IV can be found from three ways:1. Characteristic of geography location, such as distance between homes to clinic center.2. Characteristic of clinic center, such as the use proportion of some specific drug or operation.3. Characteristic of calendar time, such as time to some specific drug or operation has taken. There are three assumptions for IV methods:1. The IV must be related to the exposure individually assigned.2. The IV must be unrelated to observed and unobserved prognostic factors.3. The IV must be unrelated to outcome, except through pathways that operate via the exposure individually assigned. The first assumption can be tested through significant test of correlation coefficient, and the second and third assumption can only be persuade by logic inference.Simulation results of effect estimation and comparison are summarized as follows:1. For the primary endpoint is mean value, under scenario of stochastic non-compliance, ITT, PP, AT and IV were all very close to true effect value when the true effect value was 0. ITT estimator was lower than true effect value if the true effect value was not 0. Under the scenario of non-stochastic non-compliance, ITT estimator was easy to lower, and PP, AT estimator were easy to higher than true value, and IV estimator was easy to close to true value when covariate was used, or it was easy to lower than true value under low-value non-compliance scenario and higher than true value under high-value non-compliance scenario.2. For the primary endpoint is rate value, under scenario of stochastic non-compliance, ITT, PP, AT and IV were all very close to true effect value when the true effect value was 0. ITT estimator was lower than true effect value if the true effect value was not 0. Under the scenario of non-stochastic non-compliance, ITT estimator was easy to lower, and PP, AT estimator were easy to higher than true value, and IV estimator was easy to close to true value when the true value is 0, or it was easy to lower than true value under low-value non-compliance scenario and higher than true value under high-value non-compliance scenario3. For the primary endpoint is time to event, under scenario of stochastic non-compliance, ITT, PP, AT and IV were all very close to true effect value when the true effect value was 1. ITT estimator was lower than true effect value if the true effect value was not 1. Under the scenario of non-stochastic non-compliance, ITT estimator was easy to lower than the true value is the true value was not 1, and PP, AT estimator were easy to lower than true value while ITT estimator was easy to higher than true value, no matter the true value was 1 or not.ConclusionsResults from literature review showed that we can find some methods and criteria for IV selection and evaluation. Simulation study of RCTs demonstrated that IV estimation has its irreplaceable advantages compared to ITT, PP and AT estimation. IV method should be paid more attention and should be more frequently applied in clinical research.