The German Stata Users Group Meeting was Friday, 23 June 2017 but you can view the program and presentation slides below.
Key Note I: Why propensity scores should be used for matching
Abstract: In their paper titled Why Propensity Scores Should Not Be Used for Matching, Gary King and Richard Nielsen suggest that propensity-score matching (PSM) is inferior to other matching procedures such as Mahalanobis matching (King and Nielsen 2016). They argue that PSM approximates complete randomization, whereas other techniques approximate fully blocked randomization, and that fully blocked randomization dominates complete randomization in terms of statistical efficiency. They illustrate their argument using constructed examples, simulations, and applications to real data. Overall, their results suggest that PSM has dramatic deficiencies and should best be discarded. Although the claim about the superior efficiency of a fully blocked design over complete randomization is true (given a specific sample size), the problems King and Nielsen identify apply only under certain conditions. First, the complete randomization argument is valid only with respect to covariates that are not related to the treatment. Second, and more importantly, King and Nielsen's "PSM paradox" occurs only for specific variants of PSM. I will explain why this is the case, and I will show that other variants of PSM compare favorably with blocking procedures such as Mahalanobis matching. I will illustrate my arguments using a new matching software called "kmatch".
University of Berne
Key Note II: Connecting Stata to the rest of the world via SWire: Several applications including SWordy, an Office add-in to facilitate interaction between Microsoft Word and Stata
Abstract: SWire is a plugin that connects Stata to other software programs, thus permitting the interaction between Stata and other client applications. Software programs relying on SWire can exchange data with Stata or request that Stata execute basic data management operations. Client applications can be developed in many programming languages, and even web pages can communicate with Stata via Swire. The only requirement is that client applications communicate with Stata via the SWire protocol, which is based on HTTP. Software programs like R, QGis, and Office applications can be extended to interact with Stata. Several software programs have been developed to demonstrate how SWire can be usefully employed for connecting Stata to other software programs. One of these is the new SWordy add-in for Microsoft Word, which will be presented here. It allows for the retrieval of data from Stata to Word and the creation of automatic reports, namely, Word documents with numerical data and tables that can be automatically obtained from Stata. Automatic reports are useful for saving time when presenting results in addition to providing a way for documenting the final part of data analysis workflow.
Giovanni Luca Lo Magno
University of Palermo
Sequential (two-stage) estimation of linear panel-data models
Abstract: I present the new Stata command xtseqreg, which implements sequential (two-stage) estimators for linear panel-data models. In general, the conventional standard errors are no longer valid in sequential estimation when the residuals from the first stage are regressed on another set of (often time-invariant) explanatory variables at a second stage. xtseqreg computes the analytical standard-error correction of Kripfganz and Schwarz (2015), which accounts for the first-stage estimation error. The command can be used to fit both stages of a sequential regression or either stage separately. OLS and 2SLS estimation are supported, as well as one-step and two-step "difference"-GMM and "system"-GMM estimation in the spirit of Arellano and Bond (1991), Arellano and Bover (1995), and Blundell and Bond (1998), with flexible choice of the instruments and weighting matrix. Available postestimation statistics include the Arellano–Bond test for absence of autocorrelation in the first-differenced errors and Hansen's J-test for the validity of the overidentifying restrictions. While I do not intend to introduce xtseqreg as a competitor for existing commands, it can mimic part of their behavior. In particular, xtseqreg can replicate results obtained with xtdpd and xtabond2. In that regard, I will illustrate some pitfalls in the estimation of dynamic panel models.
Arellano, M., and S. R. Bond. 1991. Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Review of Economic Studies 58: 277–297.
Arellano, M., and O. Bover. 1995. Another look at the instrumental variable estimation of error-components models. Journal of Econometrics 68: 29–51.
Blundell, R., and S. R. Bond. 1998. Initial conditions and moment restrictions in dynamic panel data models. Review of Economic Studies 87: 115–143.
Kripfganz, S., and C. Schwarz. 2015. Estimation of linear dynamic panel data models with time-invariant regressors. ECB Working Paper 1838. European Central Bank.
University of Exeter Business School
A case study in efficient programming in Stata and Mata: Speeding up the ardl estimation command
Abstract: The user-written package ardl, first released in 2014, estimates autoregressive distributed lag (ARDL) time-series models and provides the popular Pesaran, Shin, and Smith (2001, Journal of Applied Econometrics) bounds testing procedure for a long-run relationship. In this presentation, the statistics and application side of the command take a back seat and give way to a discussion of the algorithms used under the hood of ardl. Efficient programming is critical for ardl for two reasons: optimal lag selection and for obtaining critical values via simulation. This presentation will use the "case study" of the ardl estimation command to discuss efficient programming in Stata and Mata. Various programming concepts (compilation, argument passing, data types, pointer variables, etc.) and their implementation in Stata/Mata will be explained, as well as various finer Mata-specific topics (fast matrix indexing, matrix inversion, etc.). The overall message is that coding based on common sense, knowledge of the workings of Stata/Mata, and knowledge of linear algebra goes a long way when trying to write high-performance code and in many cases is to be preferred to the tedium of moving to a lower-level programming language like C/C++.
Daniel C. Schneider
Max Planck Institute for Demographic Research
How to assess the fit of multilevel logit models with Stata?
Abstract: Stata 14 includes the multilevel model for binary (melogit) and ordinal logits (meologit). Unfortunately, except for the global Wald test of the estimated fixed effects, both models do not provide any fit measure to assess its practical significiance. Therefore, I developed an ado-file to calculate McFadden's and McKelvey and Zavoina's pseudo-R²s. It estimates the intraclass correlation (ICC) of the dependent variable for the actual sample to assess the maximum of the contextual effect. Since the early 1990s, a lot of Monte Carlo simulation studies (Hagle and Mitchell 1992; Veall and Zimmermann 1992, 1993, 1994; Windmeijer 1995; DeMaris 2002) proved that McKelvey and Zavoina pseudo-R² is the best one to assess the fit of binary and ordinal logit models. My ado-file calculates this fit in two complementary ways: first, for the fixed effects only, and second, for the fixed and random effects together. The estimation of McFadden's pseudo-R² uses two different zero models: first, the random-intercept-only model (RIOM) knowing the contextual units, and second, the fixed-intercept-only model (FIOM) ignoring the contextual units completely. For each of them, it calculates the global likelihood-ratio-chi2 test statistic whether all fixed effects or all fixed and random effects are zero in the population. An empirical study of drug consumption in European countries demonstrates the usefulness of my fit_meologit_2lev.ado or fit_meologit_3lev.ado files for multilevel binary and ordinal logit models.
DeMaris, A. 2002. Explained variances in logistic regression. A Monte Carlo study of proposed measures. Sociological Methods & Research 11: 27–74.
McFadden, D. 1979. Quantitative methods for analysing travel behaviour of individuals: Some recent developments. Behavioural travel modelling, ed. D.A. Hensher and P.R. Stopher, 279–318. London: Croom Helm.
McKelvey, R., and W. Zavoina. 1975. A statistical model for the analysis of ordinal level dependent variables. Journal of Mathematical Sociology 4: 103–120.
Hagle, T. M., and G. E. Mitchell II. 1992. Goodness of fit measures for probit and Logit. American Journal of Political Science 36: 762–784.
Veall, M.R. & Zimmermann, K.F. (1992): Pseudo-R² in the ordinal probit model. Journal of Mathematical Sociology, 16, 4, pp. 333–342.
Veall, M. R., and K. F. Zimmermann. 1994. Evaluating pseudo-R²'s for binary probit models. Quality & Quantity 28: 151–164.
Windmeijer, F. A. G. 1995. Goodness-of-fit measures in binary choice models. Econometric Reviews 14: 101–116.
Zimmermann, K. F. 1993. Goodness of fit in qualitative choice models: Review and evaluation. In Studies in Applied Econometrics, ed. H. Schneeweiß and K. Zimmermann, 25–74. Heidelberg: Physika.
Martin Luther University of Halle-Wittenberg
Twin data analysis with ACE-decomposed explanatory variables using Stata
Abstract: Several authors have introduced different methods for decomposing the variance of a variable into an additive genetic (A), a shared environmental (C), and a unique environmental (E) component using twin data and multilevel mixed-effects (MME) models; Guo and Wang 2002; McArdle and Prescott 2005; Rabe-Hesketh, Skrondel, and Gjessing 2008, who used Stata). In recent years, the focus of behavioral genetic research has increasingly shifted toward analyzing the causal influence of these genetic and environmental components of traits on the development of inequalities. Regarding methods, this implies estimating the effects of ACE components, that is, estimating models with ACE-decomposed explanatory variables. This presentation compares different MME implementations of such models using the meglm and the gsem packages of Stata: A bivariate ACE decomposition (McArdle and Prescott 2005), a one step-estimator for the ACE decomposition and its effects, and a more flexible two-step estimator based on plausible values for the ACE components. Conceptually, these models are extensions of hybrid MME models (Allison 2009), which replace the within-between-group-decomposition of explanatory variables with an ACE-decomposition. To demonstrate how these models facilitate the causal analyses of inequalities, the presentation uses examples based on data of TwinLife, the new German twin family panel.
Allison, P. D. 2009. Fixed Effects Regression Models. Quantitative Applications in the Social Sciences 160. Thousand Oaks, CA: SAGE publications.
Guo, G., and J. Wang. 2002. The mixed or multilevel model for behavior genetic analysis. Behavior Genetics 32: 37–49.
McArdle, J. J., and C. A. Prescott. 2005. Mixed-effects variance components models for biometric family analyses. Behavior Genetics 35: 631–652.
Rabe-Hesketh, S., A. Skrondel, and H.K. Gjessing. 2008. Biometrical modeling of twin and family data using standard mixed model software. Biometrics 64: 280–288.
Assessing inter-rater agreement in Stata
Abstract: Despite its well-known weaknesses and existing alternatives in the literature, the Kappa coefficient (Cohen 1960: Fleiss 1971) remains the most frequently applied statistic when it comes to quantifying agreement among raters. It is also the only available measure in official Stata that is explicitly dedicated to assessing inter-rater agreement for categorical data. In this presentation, I briefly review Cohen's Kappa and five related statistics within a general framework of chance-corrected agreement coefficients, discussed in Gwet (2014). The presentation covers the generalization of all measures to multiple raters, weights for partial disagreement that are suitable for any data level of measurement, the treatment of missing ratings, and a new probabilistic method for benchmarking the estimated coefficients. I introduce the kappaetc command, which implements these concepts.
Cohen, J. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement 20: 37–46.
Fleiss, J. L. 1971. Measuring nominal scale agreement among many raters. Psychological Bulletin 76: 378–382.
Gwet, K. L. 2014. Handbook of Inter-Rater Reliability. Gaithersburg, MD: Advanced Analytics, LLC.
University of Kassel
Nonlinear mixed-effects models
Abstract: Stata 15 introduces the new estimation command menl for fitting nonlinear mixed-effects models, also known as nonlinear multilevel models and nonlinear hierarchical models. These models can be thought of in two ways: as nonlinear models containing random effects or as linear mixed-effects models in which some or all fixed and random effects enter nonlinearly. The overall error distribution is assumed to be Gaussian. Nonlinear mixed-effects models have been used to model drug absorption in the body, intensity of earthquakes, and growth of plants, to name a few.
In my presentation, I will demonstrate how to use the new menl command to fit nonlinear mixed-effects models in a variety of applications, including population pharmacokinetics and macroeconomics.
New in Stata 15!
Wishes and grumbles
Workshop: 22 June
Generalized propensity-score matching and its implementation in Stata
Presented by Michaela Bia, Luxembourg Institute of Socio-Economic Research (LISER)
This workshop examines advanced techniques for causal inference, with a focus on generalized propensity score-based methods. Much of the work on propensity-score analysis has focused on the case where the treatment is binary, but in many empirical studies, treatments may take on many values, implying that participants in the study may receive different treatment levels. In such cases, focus is on assessing the heterogeneity of treatment effects arising from variation in the amount of treatment exposure, that is, on estimating a dose–response function (DRF). In this workshop, we build on the work by Hirano and Imbens (2004), who introduced the concept of the generalized propensity score (GPS) and employed it to estimate the DRF of a continuous treatment, within the potential outcome approach to causal inference (Rubin 1974, 1978). In particular, we will focus on parametric (Hirano and Imbens 2004; Bia and Mattei 2008) and semiparametric techniques (Bia, Flores-Lagunes, Flores, and Mattei 2014) to estimate the DRF.
The logistics organizer for the 2017 German Stata Users Group meeting is Dittrich & Partner Consulting GmbH, the distributor of Stata in Germany, the Netherlands, Austria, the Czech Republic, and Hungary.
View the proceedings of previous Stata Users Group meetings.