Big Data Analytics
Shein-Chung Chow in Innovative Statistics in Regulatory Science, 2019
In clinical research, one of the major concerns of a case-control study is selection bias, which often caused by significant difference or imbalance between the case and control groups, especially for large observational studies (Rosenbaum and Rubin, 1983, 1984; Austin, 2011). In this case, the target patient population under study in the control group may not be comparable to that of the case group. This selection bias could alter the conclusion of the treatment effect due to possible confounding effect. Consequently, the conclusion may be biased and hence misleading. To overcome this problem, Rosenbaum and Rubin (1983) proposed the concept of propensity score as a method to reduce selection bias in observational studies. Propensity score is a conditional probability (or score) of the subject being in a particular group when given chosen characteristics. That is, consider the case group as those who received a certain treatment (T = 1), and the control group as those who did not receive this treatment (T = 0). Let X be a vector represents baseline demographics and/or patient characteristics that are important (e.g., possible confounding factors) for matching the case and control populations for reducing the selection bias. The propensity score is then given by
Analysis of Vaccine Studies and Causal Inference
Leonhard Held, Niel Hens, Philip O’Neill, Jacco Wallinga in Handbook of Infectious Disease Data Analysis, 2019
Most studies are not randomized at two stages. In fact, we do not know of any vaccine studies to date that have been randomized at two stages. A study could be randomized at the individual level, at the group (cluster) level, or neither. Then the estimators described herein in general would be biased or inconsistent. For the observational setting where the treatment assignment mechanism is not known and there is no interference, propensity scores are one method to adjust the analysis to resemble results that might be obtained from a randomized trial (Rosenbaum and Rubin, 1983). The propensity score is the probability that an individual receives a treatment assignment based on a function of the observed covariates. The propensity score can be used in different ways to adjust for measured confounders, including weighting by the inverse of the propensity score, called inverse probability weighting (IPW), or stratifying on them (Hong and Raudenbush, 2006).
Real-World Data and Real-World Evidence
Wei Zhang, Fangrong Yan, Feng Chen, Shein-Chung Chow in Advanced Statistics in Regulatory Critical Clinical Initiatives, 2022
The propensity score was proposed by Rosenbaum and Rubin in 198361. The propensity score is a function of the covariates, which is defined as the probability of receiving treatment (or exposure) conditional on observed covariates. In conventional analysis, the propensity score was typically estimated from regression models, such as a logistic regression or a probit regression of the treatment conditional on the covariates. In practice analysis, treatment variables are often multiple categorical or continuous, and hence generalized propensity scores should be used instead, the methods for which have been proposed by Imbens 62 and Hirano&Imbens 63, respectively. Generalized propensity scores can be estimated by regression model (such as logistic and probit regression model), machine learning (such as gradient boosting machine), the covariate balancing generalized propensity score based on generalized method of moments, and so forth.
Effects of vitamin C supplementation on blood pressure and hypertension control in response to ambient temperature changes in patients with essential hypertension
Published in Clinical and Experimental Hypertension, 2019
Xiaojie Yuan, Xiaochun Li, Zhaohua Ji, Jing Xiao, Lei Zhang, Weilu Zhang, Haixiao Su, Kanakaraju Kaliannan, Yong Long, Zhongjun Shao
This study aimed at studying the effect of VC on BP reduction in hypertensive subjects. We recruited 512 hypertensive subjects in total. Among them, 480 subjects were able to finish the trial successfully. Some confounding variables were suspected to introduce a serious bias in the study results so we wished to quantify this effect and build up a subsample of it that reduces this bias. This was achieved by pairing similar individuals in terms of confounding variables using PSM. The propensity score, defined as the conditional probability of being treated given the covariates, can be used to balance the covariates in the 2 groups and therefore reduce this bias. There were 24 variables used in this study. The first one was binary variable entitled participation in the VC treatment that indicates if the participant did participate in the VC treatment (1) or not (0). This was our group variable. The suspected major confounding variables were age, sex, and DBP according to a chi-square test comparison using all available variables. Next, a propensity score (range of 0–100%) was calculated using a logistic regression with the aforementioned confounding variables. Finally, subjects were matched according to propensity scores, leading to an even distribution of potential confounding factors between 2 groups. In our propensity analyses, a 2:1 matching ratio, optimal algorithm and Mahalanobis distance were used. PSM analysis was performed using the software SAS version 9.3 (SAS Institute Inc).
Comparable safety of ERCP in symptomatic and asymptomatic patients with common bile duct stones: a propensity-matched analysis
Published in Scandinavian Journal of Gastroenterology, 2021
Lina Xiao, Chong Geng, Xiao Li, Yanni Li, Chunhui Wang
We performed propensity score matching (PSM) to reduce the effects of selection bias and potential confounding factors on measured outcomes due to differences in baseline characteristics between the groups. The propensity score was calculated using a multivariable logistic regression model. The following patients’ characteristics that were reported as risk factors for ERCP-related complications were included in the model: age, gender, periampullary diverticulum, nondilated CBD, history of pancreatitis, precut sphincterotomy, EST, ERBD, difficult cannulation, and use of anticoagulants before ERCP [14–19]. Using these propensity scores, patients in the asymptomatic CBD stones group were matched with those in the symptomatic CBD stones group at a 1:4 ratio. PSM analyses were performed using R version 3.6.1 (free to download from http://www.r-project.org).
The incremental cost of traumatic brain injury during the first year after a road traffic accident
Published in Brain Injury, 2019
Helena Van Deynse, Griet Van Belleghem, Door Lauwaert, Maarten Moens, Karen Pien, Stefanie Devos, Ives Hubloue, Koen Putman
Matching was conducted by means of a greedy matching algorithm, using the ‘gmatch’ SAS macro (22,23). In order to be eligible as a control, two conditions should both be fulfilled. First, an exact match is required on the presence of each of the five most frequent non-head injury groups in cases. Based on the Barell Injury Diagnosis Matrix (21), these were identified to be the following: internal injury of torso, fracture of torso, fracture of upper extremities, fracture of spine and back, fracture of lower extremities. As a second condition, a logit propensity score representing other relevant health factors, should be within an interval (caliper) of 0.2 standard deviations around that of the case, as recommended by Austin (24). The use of a propensity score reduces bias by producing two samples (one containing cases and the other containing controls) in which the variables of interest are distributed similarly, without requiring exact matching on these variables (25). This was done by means of a logistic regression with ‘having a TBI’ as the outcome variable and the following variables as covariates: age, ICISS excluding TBI, type of roadway user, the presence of acute events, the presence of chronic conditions, survival during as well as after hospitalisation and the aforementioned injury categories. The balance between the obtained samples was verified by the calculation of the standardised difference, where a value greater than or equal to 0.10 was considered to indicate an inadequate balance (26,27).
Related Knowledge Centers
- Causal Inference
- Observational Study
- Matching
- Dependent & Independent Variables
- Confounding
- Statistical Unit
- Treatment & Control Groups
- Average Treatment Effect
- Randomized Experiment
- Nearest Neighbor Search