Estimating the Dose–Response Function through a Generalized Linear Model Approach

Abstract

In this article, we revise the estimation of the dose-response function described in Hirano and Imbens (2004, Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives, 73-84) by proposing a flexible way to estimate the generalized propensity score when the treatment variable is not necessarily normally distributed. We also provide a set of programs that accomplish this task. To do this, in the existing doseresponse program (Bia and Mattei, 2008, Stata Journal 8: 354-373), we substitute the maximum likelihood estimator in the first step of the computation with the more flexible generalized linear model.

Is the model “loans‐plus‐savings” better for microfinance in Eastern Europe and Central Asia? A propensity score matching comparison

“…Since the data seem to indicate that either offering savings is an important part of MFI operations or it is just a small portion of the overall liabilities, it is prudent to address this heterogeneity of “treatment.” To do that, instead of using a propensity score for a categorical variable, we estimate generalized propensity score appropriate for continuous treatment cases. We follow Guardabascio and Ventura () who apply a generalized linear model instead of a maximum likelihood estimator. This approach complements the one used by Bia and Mattei () and allows for nonnormal distribution of the treatment variable.…”

“…In addition to the conventional propensity score matching method we apply also dose–response function to control for possible effects of nonhomogenous treatment across the treated units (Hirano & Imbens, ). The generalized propensity score is estimated using a linear model (Guardabascio & Ventura, ), which allows for continuous treatment modeling. In our particular case the MFIs vary in the level of deposits they collect, which makes this approach a valid way to test robustness of our results.…”

“…We apply semiparametric quasi‐experimental PSM, as well as dose–response function approaches to estimate the impact of offering savings on financial sustainability and credit outreach of MFIs. Estimation of the dose–response function (Guardabascio & Ventura, ) provides an opportunity to take into account the continuous nature of treatment, which in our case is the relative share of deposits. In both settings we use data on capital structure of MFIs, operations profile (age, outreach direction, profit status, risk profile), and country level macro data.…”

Are local public services better delivered in more autonomous regions? Evidence from European regions using a dose‐response approach

“…In this paper, we are interested in studying the causal effect of regional autonomy ( rai ) on the overall performance of local public goods ( regqog ). To this purpose, we rely on the econometric literature on treatment effects estimation (Imbens and Wooldridge ), and more specifically on that estimating dose‐response models (Hirano and Imbens ; Adorno et al ; Bia and Mattei ; Guardabascio and Ventura ; Cerulli ). Dose‐response models are well suited in socio‐economic contexts where a ‘cause’ takes the form of a continuous exposure to a certain treatment.…”

Geography of plant breeding systems, agroclimatic similarity, and agricultural productivity: evidence from Nigeria

“…As is shown later, gamma distribution is found to better characterize than the normal distribution. The estimation of GPS with the assumption of the gamma distribution is conducted using STATA's gpscore2 command (Guardabascio and Ventura, ).…”