Published papers

We propose a new semiparametric Bayesian model for causal inference in which assignment to treatment depends on potential outcomes. The model uses the probit stick-breaking process mixture proposed by Chung and Dunson (2009), a variant of the Dirichlet process mixture modeling. In contrast to previous Bayesian models, the proposed model directly estimates the parameters of the marginal parametric model of potential outcomes, while it relaxes the strong ignorability assumption, and requires no parametric model assumption for the assignment model and conditional distribution of the covariate vector. The proposed estimation method is more robust than maximum likelihood estimation, in that it does not require knowledge of the full joint distribution of potential outcomes, covariates, and assignments. In addition, the method is more efficient than fully nonparametric Bayes methods. We apply this model to infer the differential effects of cognitive and noncognitive skills on the wages of production and nonproduction workers using panel data from the National Longitudinal Survey of Youth in 1979. The study also presents the causal effect of online word-of-mouth onWeb site browsing behavior. Supplementary materials for this article are available online. © 2013 American Statistical Association.

Nowadays, many supermarkets implement a customer tier program to increase their profits because it is expected to raise customers' willingness to purchase by setting thresholds. However, designing an appropriate program is difficult because each customer's heterogeneous purchase behavior is difficult to capture. Therefore, we simultaneously modeled purchase frequency and amount through a marked point process approach while considering program effects and customer characteristics. The results clarified that the points pressure effect was particularly strong among customers who originally visited the store infrequently and had not attained the threshold set up in the customer tier program many times in the past. In addition, we found that the three-tier customer tier program was superior to the two-tier program with respect to the operating income in supermarkets.

We show the identification of the joint distribution of the potential outcomes under various parametric specifications. The key factor of the identification is the nonnormality of the distribution of the observed variables, with which we can obtain information of higher order moments that are not determined only by mean and variance. In particular, we show the identification of the joint distribution of the potential outcomes when it is specified by a normal mixture. Because any continuous distribution can be well approximated by a finite mixture distribution, our result may cover a wide class of distributions. The identification results derived are useful for estimating quantile treatment effects, causal mediation effects, and heterogeneous treatment effects, which cannot be estimated even if the unconfoundedness assumption is satisfied.

In this paper, we consider a novel framework of positive-unlabeled data in which as positive data survival times are observed for subjects who have events during the observation time as positive data and as unlabeled data censoring times are observed but whether the event occurs or not are unknown for some subjects. We consider two cases: (1) when censoring time is observed in positive data, and (2) when it is not observed. For both cases, we developed parametric models, nonparametric models, and machine learning models and the estimation strategies for these models. Simulation studies show that under this data setup, traditional survival analysis may yield severely biased results, while the proposed estimation method can provide valid results.

The mRNA-1273 Moderna COVID-19 vaccine was introduced to combat the COVID-19 global pandemic in 2020. Although the safety of the vaccine has been investigated worldwide, real-world safety data is scarce in Japan. An online, real-time survey of adverse events following immunization (AEFIs) with mRNA-1273 was conducted in the setting of a workplace vaccination program at the School of Pharmacy, Keio University from 26 June 2021, to 11 June 2022. Participants were requested to take four surveys during a seven-day follow-up period after each of the first, second, and third booster doses. The maximum number of responses, from 301 respondents, was obtained on day 0 (vaccination date) for the first dose. 98% of respondents reported local and systemic AEFIs for the second dose on day 1. No noticeable difference in local reactions was seen among the three doses. Females reported more AEFIs than males, and the young group (18–29 years) reported a higher rate than the middle age group (≥30 years) after the first dose. Age and gender differences in rates decreased at the second and third doses. This survey confirmed that the safety profile of mRNA-1273 in a real-world setting was similar to that derived from the clinical trials, and that the agent was well-tolerated.

We develop a new semiparametric Bayes instrumental variables estimation method. We employ the form of the regression function of the first-stage equation, and the disturbances are modeled nonparametrically to achieve better predictive power of the endogenous variables, whereas we use parametric formulation in the second-stage equation, which is of interest in inference. Our simulation studies show that under small sample sizes, the proposed method obtains more efficient estimates and very precise credible intervals compared with existing IV methods with smaller mean squared error. We applied our proposed method to a Mendelian randomization dataset where a large number of instruments are available and semiparametric specification is appropriate. We obtained statistically significant results that cannot be obtained by the existing methods, including standard Bayesian IV.

© 2020, © 2020 American Statistical Association. We propose an econometric two-stage model for category-level purchase and brand-level purchase that allows for simultaneous brand purchases in the analysis of scanner panel data. The proposed model formulation is consistent with the traditional theory of consumer behavior. We conduct Bayesian estimation with the Markov chain Monte Carlo algorithm for our proposed model. The simulation studies show that previously proposed related models can cause severe bias in predicting future brand choices, while the proposed method can effectively predict them. Additionally in a marketing application, the proposed method can examine brand switching behaviors that existing methods cannot. Moreover, we show that the prediction accuracy of the proposed method is higher than that of existing methods.

We propose an estimation method of population moments or parameters in “biased sampling data” in which for some units of data, not only the variables of interest but also the covariates have missing observations, and the proportion of “missingness” is unknown. We use auxiliary information such as the distribution of covariates or their moments in random sampling data in order to correct the bias. Moreover, with additional assumptions, we can correct the bias even if we have only the moment information of covariates. The main contribution of this paper is the development of a doubly robust-type estimator for biased sampling data. This method provides a consistent estimator if either the regression function or the assignment mechanism is correctly specified. We prove the consistency and semi-parametric efficiency of the doubly robust estimator. Both the simulation and empirical application results demonstrate that the proposed estimation method is more robust than existing methods.

It is important to estimate the local average treatment effect (LATE) when compliance with a treatment assignment is incomplete. The previously proposed methods for LATE estimation required all relevant variables to be jointly observed in a single dataset; however, it is sometimes difficult or even impossible to collect such data in many real-world problems for technical or privacy reasons. We consider a novel problem setting in which LATE, as a function of covariates, is nonparametrically identified from the combination of separately observed datasets. For estimation, we show that the direct least squares method, which was originally developed for estimating the average treatment effect under complete compliance, is applicable to our setting. However, model selection and hyperparameter tuning for the direct least squares estimator can be unstable in practice since it is defined as a solution to the minimax problem. We then propose a weighted least squares estimator that enables simpler model selection by avoiding the minimax objective formulation. Unlike the inverse probability weighted (IPW) estimator, the proposed estimator directly uses the pre-estimated weight without inversion, avoiding the problems caused by the IPW methods. We demonstrate the effectiveness of our method through experiments using synthetic and real-world datasets.

In online display advertising, selecting the most effective ad creative (ad image) for each impression is a crucial task for DSPs (Demand-Side Platforms) to fulfill their goals (click-through rate, number of conversions, revenue, and brand improvement). As widely recognized in the marketing literature, the effect of ad creative changes with the number of repetitive ad exposures. In this study, we propose an efficient and easy-to-implement ad creative selection algorithm that explicitly considers user's psychological status when selecting ad creatives. The proposed system was deployed in a real-world production environment and tested against the baseline algorithms. The results show superiority of the proposed algorithm.

Intermittent missingness in repeated duration analysis is common in applied studies, but has not been rigorously considered in statistics. Under intermittent missingness, whether any missing events exist between two observed events is unknown. In other words, the missing indicators are never observed. Thus, if there exist any missing events between two observed events, researchers observe only the cumulative duration of the two or more events. A quasi-Bayes estimation method that utilizes population-level information is used to appropriately estimate the parameters under unobserved intermittent missingness. The proposed model consists of the following: (1) a latent variable model, (2) a latent missing indicator model separating the true and composite durations, (3) mixtures of duration models, and (4) moment restriction from population-level information to deal with nonignorable intermittent missingness. A new estimation procedure is used to simultaneously combine likelihood and the objective function of GMM with the latent variables this is called Bayesian data combination. The model is applied to the interpurchase duration in database marketing using the purchase history data of Japan these data capture the purchase incidences and stores.

A foreign language (a language not spoken in one's community) is difficult to master completely. Early introduction of foreign-language (FL) education during childhood is becoming a standard in many countries. However, the neural process of child FL learning still remains largely unknown. We longitudinally followed 322 school-age children with diverse FL proficiency for three consecutive years, and acquired children's ERP responses to FL words that were semantically congruous or incongruous with the preceding picture context. As FL proficiency increased, various ERP components previously reported in mother-tongue (L1) acquisition (such as a broad negativity, an N400, and a late positive component) appeared sequentially, critically in an identical order to L1 acquisition. This finding was supported not only by cross-sectional analyses of children at different proficiency levels but also by longitudinal analyses of the same children over time. Our data are consistent with the hypothesis that FL learning in childhood reproduces identical developmental stages in an identical order to L1 acquisition, suggesting that the nature of the child's brain itself may determine the normal course of FL learning. Future research should test the generalizability of the results in other aspects of language such as syntax. © 2010 Massachusetts Institute of Technology.

© 2015 John Wiley & Sons A/S. Published by John Wiley & Sons Ltd Objective: Secular change in marginal bone loss (MBL), which is the index adopted for implant success criteria, has often been used to evaluate risk factors. However, the need to revise these criteria has recently been indicated due to rapid developments in implant treatment. The purpose of this study was to evaluate risk factors by analyzing MBL with an alternative statistical method. Material and methods: The analyses were performed on the outcomes of 366 patients with 1,902 implants during an average follow-up period of 84.8 months (with a maximum follow-up of 258 months). Instead of evaluating annual MBL, time was calculated as one of the explanatory variables because the correlation between MBL and time was small (correlation coefficient of 0.09010). Analysis of covariance (ANCOVA) was used for exploratory assessment of each factor, and multiple regression analysis was then utilized to identify risk factors. The multiple regression analysis was performed twice, once among all implants and another in which one implant per patient was randomly selected. Results: As a result of multiple regression analysis, smoking habits showed a significant effect on MBL. Age, sex, diabetes mellitus, implant positions, guided bone regeneration, and sinus floor elevation did not affect MBL. IMZ ® implants were associated with significantly higher MBL than were ANKYLOS ® and SPI ® implants. There was no significant difference between Straumann ® and other implants. Conclusion: Our results showed that another statistical process, which eliminated the effect of time rather than comparing annual MBL, could be applied to evaluate MBL because the correlation between MBL and time was small.

© 2019 The Author(s). Background: Influenza vaccination is recognized as a primary public health intervention which prevents the illness of patients and relieves the societal burdens of influenza for medical community as well as the economy. To date, no effectiveness study of influenza vaccination has been conducted including a large population with a wide age span, in Japan. Here, we evaluated the clinical effectiveness of influenza vaccination in a large Japanese population. Methods: We conducted a cohort study using a large-scale claims database for employee health care insurance plans. Vaccination status was identified using plan records for influenza vaccination subsidies. We excluded people aged 65 years or more because of the unavailability of vaccination records. Effectiveness of vaccination in preventing influenza and its complication was evaluated with doubly robust methods using inversed probability treatment weighting to adjust health conscious behaviours and other confounders. Results: During the 2013/2014 influenza season, 369,425 subjects with age range from 1 to 64 years were eligible. Vaccination rate was 39.5% and an estimated odds ratio (OR) for influenza onset was 0.775 after doubly robust adjustment. Age-stratified ORs were significantly reduced in all age groups; lowest in subjects aged 1 to 4 years (0.600) and highest in those aged 13 to 19 (0.938). ORs for all the influenza complication outcomes were also statistically significant (0.403-0.709). Conclusions: We confirmed the clinical effectiveness of influenza vaccination in people aged 1 to 64 years. Influenza vaccination significantly prevented influenza onset and was more effective in reducing secondary risks of influenza complications.