Abstract:This article suggests a doubly robust method of estimating potential outcome means for multivariate fractional outcomes when the treatment of interest is unconfounded and can take more than two values. The method involves maximizing a propensity score weighted multinomial quasi-log-likelihood function with a multinomial logit conditional mean. We show that this estimator, which we call weighted multivariate fractional logit ( wmflogit ), consistently estimates the potential outcome means if either the propensity score model or the conditional mean model is misspecified. Our simulations demonstrate this double robustness property for the case of shares generated using a Dirichlet distribution. Finally, we advocate for the use of wmflogit by applying it to estimate time-use shares of women participating in the Mexican conditional cash transfer program, Progresa, using Stata's fmlogit command developed by Buis.

What problem does this paper attempt to address?

This paper attempts to solve the problem of estimating the potential outcome means of multivariate fractional outcomes in the case of multi - valued treatments. Specifically, when the treatment variable can take multiple values and the outcome variable is in the form of a multivariate fraction, how to estimate the potential outcome means of these outcome variables. The paper proposes a doubly - robust method, namely the weighted multivariate fractional logit model (wmflogit), to estimate these potential outcome means. This method is particularly useful when the treatment variable is unconfounded and can take multiple values. ### Main problems 1. **Multivariate fractional outcomes**: The outcome variable is in the form of a multivariate fraction, each outcome variable takes values within the unit interval, and the sum of all outcome variables is 1. For example, time allocation ratios, expenditure shares, etc. 2. **Multi - valued treatments**: The treatment variable can take multiple values, not just binary treatments. For example, different years of education, different policy combinations, etc. 3. **Estimation of potential outcome means**: Under the assumption that the treatment variable is unconfounded, estimate the potential outcome means at different treatment levels. ### Solutions The paper proposes a doubly - robust estimation method - the weighted multivariate fractional logit model (wmflogit). This method has the following characteristics: - **Doubly - robustness**: If either the propensity score model or the conditional mean model is correctly specified, then the estimator can consistently estimate the potential outcome means. - **Wide range of applicability**: Applicable to the cases of multi - valued treatment variables and multivariate fractional outcomes. - **Model form**: Use the multivariate fractional logit model to model the conditional mean, ensuring that the predicted values are within the unit interval and satisfy the aggregation constraint. ### Method overview 1. **Model specification**: - **Conditional mean model**: Use the multivariate fractional logit model (mflogit) to model the conditional mean. - **Propensity score model**: Use the multivariate logistic regression model (MNL) to estimate the propensity score. 2. **Estimation process**: - **Weighted quasi - log - likelihood function**: Estimate the parameters by maximizing the weighted multivariate quasi - log - likelihood function. - **Doubly - robustness**: Even if one of the models is mis - specified, the estimator is still consistent. 3. **Theoretical properties**: - **Consistency**: In the case where the treatment variable is unconfounded and the overlap assumption holds, the estimator is consistent. - **Asymptotic normality**: Under standard regularity conditions, the estimator is asymptotically normal. ### Simulation study The paper verifies the doubly - robustness of wmflogit through simulation studies. The simulation results show that wmflogit performs better than other estimation methods in various situations, such as linear regression adjustment (lra), inverse probability weighting (ipw), weighted linear regression adjustment (ipwlra) and multivariate fractional logit model (mflogit). ### Empirical application The paper uses the wmflogit method to estimate the time allocation ratios of women in the Mexican Conditional Cash Transfer Program (Progresa). The results show that wmflogit can effectively estimate the time allocation ratios at different treatment levels. ### Conclusion The wmflogit method proposed in the paper provides a doubly - robust estimation method in the case of dealing with multi - valued treatment variables and multivariate fractional outcomes. This method not only has consistency and asymptotic normality in theory, but also performs well in practical applications.

Doubly robust estimation of multivariate fractional outcome means with multivalued treatments

Doubly weighted M-estimation for nonrandom assignment and missing outcomes

Inverse probability of treatment weighting with generalized linear outcome models for doubly robust estimation

Marginal mean weighting through stratification: A generalized method for evaluating multivalued and multiple treatments with nonexperimental data.

Robust and Efficient Estimation of Potential Outcome Means under Random Assignment

Double robust estimation of partially adaptive treatment strategies

High-dimensional model-assisted inference for treatment effects with multi-valued treatments

Robust propensity score weighting estimation under missing at random

Double robust estimation of functional outcomes with data missing at random

Dynamic treatment effects: high-dimensional inference under model misspecification

Robust weights that optimally balance confounders for estimating marginal hazard ratios

Modified generalized method of moments for a robust estimation of polytomous logistic model

Doubly Robust Estimation of Local Average Treatment Effects Using Inverse Probability Weighted Regression Adjustment

Estimation of Heterogeneous Treatment Effects Using a Conditional Moment Based Approach

Doubly Robust Semiparametric Difference-in-Differences Estimators with High-Dimensional Data

Multiply robust estimation of the average treatment effect with missing outcomes

Doubly robust estimation of average treatment effect revisited

Intervention treatment distributions that depend on the observed treatment process and model double robustness in causal survival analysis

Doubly robust matching estimators for high dimensional confounding adjustment

Multiply robust estimation of marginal structural models in observational studies subject to covariate-driven observations

Doubly robust treatment effect estimation with missing attributes