Call for Papers: TPM Special Issue on Peer Review

The Political Methodologist is calling for papers for a special issue of TPM concerning the process of peer review in political science!

Peer review is something that every political scientist in the field will be subject to and asked to perform, but is a task for which almost no one receives formal training. Although some helpful guidelines exist in the literature, I believe there is still considerable heterogeneity in how people think about the peer review process and that the community would benefit from discussing these views. Moreover, new developments in the discipline raise new questions about the review process (e.g., the degree to which journals and reviewers have a responsibility to ensure replicability and reproducibility).

A wide variety of topics would fit well into this special issue, including (but not exclusive to):

  • how one should write a review, including and especially what constitute fair criteria for evaluation, and what criteria are unfair
  • what is the reviewer’s role in the process: Quality Assurance? Error Checking? Critical Commentary? or what?
  • how one should respond to a review when invited to revise and resubmit (or rejected)
  • the role that peer review should play in error checking / replication / verification
  • the “larger view” of how peer review does or should contribute to (political) science
  • the role of editorial discretion in the process, and how editors should regard reviews

Submissions should be between 2000-4000 words (although shorter submissions will also be considered), and should be sent to by December 1, 2015. Accepted articles will be featured on our blog, and also in the print edition of TPM.

If you’re interested in contributing to the special issue and would like to talk about prospective contributions before writing/submitting, please feel free to contact me (

Posted in Uncategorized | Leave a comment

Mike Ward IMC Talk Friday 3/20 @ Noon Eastern: “Irregular Leadership Changes in 2015: Forecasts using ensemble, split-population duration models”

Our final talk for Spring 2015 will be given by Mike Ward (Duke University); his presentation is entitled “Irregular Leadership Changes in 2015: Forecasts using ensemble, split-population duration models.” Everyone is welcome to join us this Friday, March 20th, at 12:00 PM Eastern time for the presentation and interactive Q&A with participants from around the world. The entire seminar will last approximately one hour.

To sign up for the presentation (or to join it while it is in progress), click here:

You can participate from a Mac, PC, tablet, or smartphone from anywhere with a reliable internet connection; please see here for more details on how to participate in the IMC as an audience member.

Posted in Uncategorized | Leave a comment

Graphical Presentation of Regression Discontinuity Results

[Editor’s note: this post is contributed by Natalia Bueno and Guadalupe Tuñón.]

During the last decade, an increasing number of political scientists have turned to regression-discontinuity (RD) designs to estimate causal effects.  Although the growth of RD designs has stimulated a wide discussion about RD assumptions and estimation strategies, there is no single shared approach to guide empirical applications. One of the major issues in RD designs involves selection of the “window” or “bandwidth” — the values of the running variable that define the set of units included in the RD study group. [i]

This choice is key for RD designs, as results are often sensitive to bandwidth size. Indeed, even those who propose particular methods to choose a given window agree that “irrespective of the manner in which the bandwidth is chosen, one should always investigate the sensitivity of the inferences to this choice. […] [I]f the results are critically dependent on a particular bandwidth choice, they are clearly less credible than if they are robust to such variation in bandwidths.” (Imbens and Lemieux, 2008,p. 633) Moreover, the existence of multiple methods to justify a given choice opens the door to “fishing” — the intentional or unintentional selection of models that yield positive findings  (Humphreys, Schanchez de la Sierra, and van der Vindt, 2013).

In this note, we propose a simple graphical way of reporting RD results that shows the sensitivity of estimates to a wide range of possible bandwidths.[ii] By forcing researchers to present results for an extensive set of possible choices, the use of these plots reduces the opportunities for fishing, complementing existing good practices in the way in which RD results are presented. Some empirical applications of RD designs have presented their results using plots that are in some ways similar to the ones we propose. However, this note explores the virtues of the plots more systematically (e.g., in connection with balance tests and the discussion of the RD estimator) and provides code so that scholars can adapt them to their own applications. The following paragraphs describe how RD results are usually reported in two top political science journals and the ways in which the graphs we propose can improve on current practices. An R function to construct these plots for a wide set of applications is publicly available on the online Appendix.

Reporting RD Analyses: Current Practice

How do political scientists report the results of regression-discontinuity designs? We reviewed all papers using RDDs published in the American Political Science Review and American Journal of Political Science. We surveyed these papers and coded (1) their choice of estimators, (2) whether they present any type of balance test and, (3) if they do, the window(s) chosen for this.

Out of a total of twelve RD papers published in these journals, five report results using a single estimator.[iii] Four articles present results for a single window — usually the full sample. The remaining papers present results using multiple windows, but the number and selection of windows are neither systematic nor extensive.[iv] Seven papers present some type of balance test, but while researchers often report their main results using a handful of windows, they do not report balance tests for different windows to the same extent.[v]

A Graphical Alternative: An Example

We use electoral data from the U.S. House of Representatives from 1942 to 2008 collected by Caughey and Sekhon (2011) and a regression discontinuity design examining incumbency advantage to illustrate how a researcher can use plots to present RD results in a transparent and systematic way. This application has two advantages for illustrating the use of these plots. First, close-race electoral regression-discontinuity designs are one of the main applications of this type of design in political science — we are thus presenting the plot in one of the most used RDD setups, although researchers can use this type of graph in all sorts of RD designs. Second, the use of close-race RDDs to measure the effect of incumbency advantage has sparked a vigorous debate about the assumptions and validity of these designs in particular settings. Using this application allows us to show an additional advantage of the plots we propose: tests of balance and other types of placebo tests.

Figure 1 plots the estimates of local average treatment effects as a function of the running variable, here vote margin. For example, the first solid back circle represents the average difference in vote share between an incumbent party that won by 0.45% or less and an incumbent party that lost by 0.45% or less is about 11 percentage points. It also reports the average difference in vote share between incumbent parties that won and lost by sequential increases of 0.2% in vote margin between 0.45% to 9.85%. Figure 1 has an additional feature: the solid gray line represents the results using an additional estimator that enables us to compare the effects estimated from two different estimators, across different windows, in the same plot. In this case, we present the estimated effects of party incumbency on vote share using a local linear regression with a triangular kernel, across different bandwidths. Researchers could use different estimators, such as a polynomial regression model.[vi] Note, however, that the black circles and the solid gray line represent different quantities. The black circles represent estimates of the average causal effect for the RD study group N. The difference of means estimator (Y^T - Y^C) — the difference between average vote share for an incumbent party minus the average vote share for a non-incumbent party — is unbiased for the average causal effect (\tau_{ACE}), represented in equation (1).[vii] The gray line presents estimates of the average causal effect precisely at the point of discontinuity (c). We fit a local linear regression with a triangular kernel, within the RDD bandwidth, to estimate this limit parameter (\tau_{lim}), represented in equation (2).[viii]

\tau_{ACE} = \frac{1}{N}\sum_{i=1}^{N} [Y_{i}(1)-Y_{i}(0)] (1)

\tau_{lim} = \lim_{r \Downarrow c} [\overline{Y}_i(1)|R_i=r] - \lim_{r \Uparrow c} [\overline{Y}_i(0)|R_i=r] (2)

Figure 1: Mean vote share difference between winners and losers by Democratic margin of victory in previous election, U.S. House of Representatives from 1942 to 2008.
Note: Dashed gray line at the optimal bandwidth estimated by the method of Imbens and Kalyanaraman (2011). Difference of means is the average difference in vote share for an incumbent party minus the average vote share for a non-incumbent party. The local linear regression uses a triangular kernel.

The key part of the plotting function is the section that produces the estimate for each window. For this, we first pre-specify functions for each estimator. For example, for the difference of means we have:[ix]

#Difference of mean (using OLS with robust standard errors)
 dom <- function(rescaled, treat, outcome){
 model <- lm(outcome~treat)
 est <- NA
 est[1] <- model$coefficients[2]
 est[2] <- sqrt(diag(vcovHC(model,type="HC3"))[2])

We then include a loop which takes a window value and subsets the data to keep only the observations within that window.

ests <- matrix(NA,length(windows),3) #object to store loop output
 for (i in 1:length(windows)){
 # select data
 temp <-[abs_running<=windows[i],])

We take this subset of the data to calculate the estimates of interest for that window, here the difference of means. The plot requires that we calculate both the point estimates and the confidence intervals. In these figures, confidence intervals are calculated using a normal approximation and unequal variances are assumed for standard errors in the treatment and control groups. If the researcher wanted to include an additional estimator, the calculation of the estimate for a particular window would also be included in the loop.[x]

ests[i,1:2] <- with(temp, dom(rescaled=rescaled, treat=treat,

if (ci=="95") CI <- cbind(ests[,1]+1.96*ests[,2],ests[,1]-1.96*ests[,2])
 if (ci=="90") CI <- cbind(ests[,1]+1.64*ests[,2],ests[,1]-1.64*ests[,2])

As expected, the confidence intervals in figure 1 become increasingly smaller for results associated with larger vote margins because the number of observations is larger. This increase in the number of observations associated with larger windows can also be reported in the plot, which we do in the upper axis of figure 1. To include the number of observations as a function of window size, we order the observed values of the outcome of interest according to their value for the running variable and allow the user to set the different number of observations that she would want to show in the plot. We calculate the number of observations at the specified values of the running variable — then, we add the number of observations to an upper axis in the plot.

# as an argument in the function, the user defines nr_obs_lab,
# the labels for the number of observations she would like to include
# in the plot

# ordering the vata by the values for the running variable
data <-[order(abs_running),])

if (nr_obs==T) {
# binding the labels with the corresponding value for the running variable
nr_obs_lab<- cbind(nr_obs_lab, data$abs_running[nr_obs_lab])
# Finally, we include an additional axis in the plot
axis(3, at=c(nr_obs_lab[,2]), labels=c(nr_obs_lab[,1]), cex=.6, col="grey50",
lwd = 0.5, padj=1, line=1, cex.axis=.7, col.axis="grey50")
mtext("Number of observations", side=3, col="grey50", cex=.7, adj=0)

Figure 2 follows the same logic described for Figure 1 but reports balance tests: each plot shows the effects of incumbency on a pre-treatment covariate as a function of the running variable.[xi] These plots allow for an extensive examination of the sensitivity of balance to different windows, reporting the magnitude of the difference between treatment and control and its confidence interval. For a given identifying assumption (such as continuity of potential outcomes or as-if random assignment near the threshold), Figures 1 and 2  help the reader to evaluate whether or not — or for which window — these assumptions are plausible.


(a) Democratic Win in t-1                                       (b) Voter Turnout %

Figure 2: Tests for balance: Standardized difference of means of pre-treatment covariates by Democratic margin of victory (95% confidence intervals), U.S. House of Representatives, from 1942 to 2008.
Note: Dashed gray line at the optimal bandwidth estimated by the method of Imbens and Kalyanaraman (2011). Difference of means is the average difference in vote share for an incumbent party minus the average vote share for a non-incumbent party.

For instance, panel (a) in Figure 2, reports the difference in previous Democratic victories between the incumbent and non-incumbent party and shows a large imbalance, strikingly larger for smaller windows — a point made by Caughey and Sekhon (2011). Panel (b) in Figure 2 shows the difference in voter turnout between treatment and control districts. For the entire set of windows covered by the plot the difference is never statistically different from zero, suggesting that the groups are balanced in terms of this covariate. Note that the size of the difference between treatment and control is much smaller in panel (b) than in panel (a) — also, relatively to the size of the effect of incumbency, the imbalance in panel (a) is substantial.[xii]

For ease of presentation, here we present plots for only two pre-treatment covariates. However, analysts should be encouraged to present plots for all pre-treatment covariates at their disposal.[xiii] Some plots may be more informative than others, for instance, because some pre-treatment covariates are expected to have a stronger prognostic relationship to the outcome. However, presentation of balance plots for the full set of available pre-treatment covariates may reduce opportunities for intentional or unintentional fishing.

Concluding Remarks

We believe that these simple plots are a useful complement to the standard way in which scholars report results and balance tests from regression-discontinuity designs. They provide a simple visual comparison of how different estimators perform as a function of different windows, communicating the robustness of the results. Moreover, using these plots both for analysis of effects and placebo tests enables an informal visual inspection of how important confounders may be, relative to the size of the effect — this is particularly informative when researchers use pre-treatment values of the outcome variable as a placebo test. However, researchers may also compare the size of treatment effects relative to other placebo tests by using standardized effect sizes across different windows, so that the scale of all plots is comparable. In summary, these plots improve the communication of findings from regression-discontinuity designs by showing readers the results from an extensive set of bandwidths, thus reducing researchers’ latitude in presentation of their main findings and increasing the transparency of RD design applications.


[i] Formally, the window is an interval on the running variable, W_0=[\underline{r},\overline{r}], containing the cutoff value r_0. The analysis then focuses on all observations with R_i within this interval. Scholars have developed numerous tools to determine the right window for a given application and estimator (Imbens and Lemieux, 2008; Imbens and Kalyanaraman, 2011, Calonico, Cattaneo and Titiunik, 2015).

[ii] This choice is posterior to defining the estimand and choosing an estimator for the causal effect of treatment. While the plots we propose focus on presenting the sensitivity of results to bandwidth choice, we also show how they can be used to explore the sensitivity of results to these other choices. For discussions of estimands in RD design see Duning (2012) and Calonico et al. (2015).

[iii] Polynomial regression is the most popular model: nine papers use a type of polynomial regression, five employ a local linear regression, and three use a difference of means (via OLS). Only one presents results using all three estimators.

[iv] Gerber and Hopkins (2011), Fewerda and Miller (2014) and Eggers et al. (2015) are exceptions—they show the robustness of the main result using plots similar to the one we suggest here.

[v] All of the papers that use local linear regressions also use a type of standard procedure to choose the “optimal” bandwidth — either Imbens and Lemieux (2008) or Imbens and Kalyanaraman (2011).

[vi] We present an example of this plot using an fourth-degree polynomial regression in Figure A.1 available at the online Appendix.

[vii] See Dunning (2012) and Bueno, Dunning, and Tuñón (2014) for a discussion and proofs.

[viii] We use the terms “window” and “bandwidth” interchangeably, since both denote the values of the running variable (r) that define the set of units included in the RD study group. However, in local linear regression with kernel smoothers, bandwidth refers to the width of the kernel.

[ix] Note that we compute the difference of means by regressing the outcome variable on a dummy for treatment assignment, with robust standard errors allowing for unequal variances, which is algebraically equivalent to the t-test with unequal variances.

[x] Our plot, and the accompanying documented R code, is flexible to incorporating different ways of estimating standard errors and constructing confidence intervals. For a detailed discussion of standard errors and confidence intervals in RD designs, see Calonico et al. (2015).

[xi] In these plots, we chose to omit the axis with the number of observations because even though there are different rates of missing observations for covariates, the number of observations for the windows we were mostly interested in did not vary substantially from those in Figure 1.

[xii] See Figure A.2 in the online Appendix for a the standardized effect of incumbency on vote share.

[xiii] An extensive set of balance plots for pre-treatment covariates in Caughey and Sekhon (2011) can be found in Figure A.3 of the online Appendix.


Bueno, Natália S., Thad Dunning and Guadalupe Tuñón. 2014. “Design-Based Analysis of Regression Discontinuities: Evidence From Experimental Benchmarks.” Paper Presented at the 2014 APSA Annual Meeting.

Calonico, Sebastian, Matias D Cattaneo and Rocio Titiunik. 2015. “Robust nonparametric confidence intervals for regression-discontinuity designs.” Econometrica 86(2):2295–2326.

Caughey, Devin and Jasjeet S. Sekhon. 2011. “Elections and the Regression Discontinuity Design: Lessons from Close US House Races, 1942–2008.” Political Analysis 19(4):385–408.

Dunning, Thad. 2012. Natural Experiments in the Social Sciences: a Design-Based Approach. Cambridge University Press.

Eggers, Andrew C., Olle Folke, Anthony Fowler, Jens Hainmueller, Andrew B. Hall and James M. Snyder. 2015. “On The Validity Of The Regression Discontinuity Design For Estimating Electoral Effects: New Evidence From Over 40,000 Close Races.” American Journal of Political Science 59(1):259–274.

Ferwerda, Jeremy and Nicholas L. Miller. 2014. “Political Devolution and Resistance to Foreign Rule: A Natural Experiment.” American Political Science Review 108(3):642–660.

Gerber, Elisabeth R. and Daniel J. Hopkins. 2011. “When Mayors Matter: Estimating the Impact of Mayoral Partisanship on City Policy.” American Journal of Political Science 55(2):326–339.

Humphreys, Macartan, Raul Sanchez de la Sierra and Peter van der Windt. 2013. “Fishing,
Commitment, and Communication: A Proposal for Comprehensive Nonbinding Research
Registration.” Political Analysis 21(1):1–20.

Imbens, Guido and Karthik Kalyanaraman. 2011. “Optimal Bandwidth Choice for the Regression Riscontinuity Estimator.” The Review of Economic Studies 79(3):933–959.

Imbens, Guido W. and Thomas Lemieux. 2008. “Regression Discontinuity Designs: A guide to Practice.” Journal of econometrics 142(2):615–635.

Posted in Uncategorized | 2 Comments

Mark Nieman IMC Talk Friday 3/13 @ 12 Eastern: “Statistical Analysis of Strategic Interaction with Unobserved Player Actions”

This week, Mark Nieman (University of Alabama) will be giving a presentation entitled “Statistical Analysis of Strategic Interaction with Unobserved Player Actions.” Everyone is welcome to join us this Friday, March 13th, at 12:00 PM Eastern time for the presentation and interactive Q&A with participants from around the world. The entire seminar will last approximately one hour.

To sign up for the presentation (or to join it while it is in progress), click here:

You can participate from a Mac, PC, tablet, or smartphone from anywhere with a reliable internet connection; please see here for more details on how to participate in the IMC as an audience member.

On Friday, March 20th, our final talk of the semester will be given by Mike Ward (Duke University).

Posted in Uncategorized | Leave a comment

Some code for estimating clustered SEs in mlogit models

There’s a well-known bit of code for estimating Liang and Zeger (1986) type cluster robust standard errors for GLM models in R (see also Rogers 1993), but it doesn’t work exactly right off-the-shelf for multinomial models estimated in the mlogit package. In support of another project, I’ve modified it slightly to work with mlogit models. I thought it might be nice to provide this code to anyone interested in using it:

# load in a function to create clustered standard errors for mlogit models
# initial code by Mahmood Ara:
# slightly modified for mlogit models by Justin Esarey on 3/3/2015

cl.mlogit   <- function(fm, cluster){

# fm: a fitted mlogit model
# cluster: a data vector with the cluster
#          identity of each observation in fm

require(sandwich, quietly = TRUE)
require(lmtest, quietly = TRUE)
M <- length(unique(cluster))
N <- length(cluster)
K <- length(coefficients(fm))
# edit 6/22/2015: change dfc
# dfc <- (M/(M-1))*((N-1)/(N-K))
dfc <- (M/(M-1))
uj  <- apply(estfun(fm),2, function(x) tapply(x, cluster, sum));
vcovCL <- dfc*sandwich(fm, meat.=crossprod(uj)/N)
coeftest(fm, vcovCL) }

And here’s a little example you can test it on:

vtinpat$hos.num <- as.numeric(vtinpat$hospital)
vtinpat$age <- as.numeric(vtinpat$
vtinpat.mlogit <-, choice = "admit", shape = "wide")
vt.mod <- mlogit(admit ~ 0 | age + sex, data = vtinpat.mlogit)

cl.mlogit(vt.mod, vtinpat$hos.num)

You can verify that this yields the same results as mlogit with cluster in Stata by writing the data file out with library(foreign) and read.dta, then running an identical model.

The reason for writing this code is to eventually use it as a part of another project:

Pairs cluster bootstrapping seems to work much better when you use cluster-robust standard errors for the replicates, and until now I couldn’t estimate them in R (had to do it in Stata). With this code, I should be able to adapt the pairs cluster bootstrap procedure for mlogit models in R.

[Update 6/22/2015]: Further checking indicated that Stata actually uses a slightly different degrees of freedom correction for mlogit (and MLE models in general) than was initially included in this code. The commented (original) code includes the “regression-like” correction while the uncommented (updated) code includes the “asymptotic-like” correction referenced on pp. 1865-1866 of the Stata 13 [R] manual; see also p. 54 here. The answers with the updated dfc code are closer to Stata’s answers for mlogit models.


Liang, Kung-Yee and Scott L. Zeger. 1986. “Longitudinal data analysis using generalized linear models.” Biometrika 73(1):13-22.

Rogers, William. 1993. “Regression standard errors in clustered samples.” Stata Technical Bulletin 13: 19-23.

Posted in Software | Leave a comment

Corrections and Refinements to the Database of Political Institutions’ yrcurnt Election Timing Variable

The yrcurnt variable in the Database of Political Institutions (Beck et al. 2001, updated in 2013)1–DPI–is a regularly used measure of government election timing. For example, Alt, Lassen, and Wehner (2014) use the variable in their recent study of fiscal gimmickry in Europe. They find that fiscal gimmickry–straying from accepted accounting standards–is more common directly before elections (and in countries with weak fiscal transparency).

Because the DPI’s yrcurnt variable is so regularly turned to for testing how election timing affects governments’ choices, it is especially important that it be reliable and valid. However, the variable in the current (2013) DPI release has a number of issues that this note aims to correct.2

Variable definition

Before looking at problems in the yrcurnt variable, let’s reiterate how the variable is defined. The DPI Codebook3 (p. 4) defines the yrcurnt variable as the years left in a country’s chief executive’s current term such that:

“a ‘0’ is scored in an election year, and n-1 in the year after an election, where n is the length of the term. In countries where early elections can be called, YRCURNT is set to the de jure term limit or schedule of elections, but resets in the case of early elections.”


The original variable has a number of issues that make it problematic for studying how election timing impacts government policymaking. The first set of concerns are clearly errors that make the variable a less reliable measure of the concept defined above. These errors can be straightforwardly corrected. The second set of concerns have to do with issues that bring into question the variable’s validity in a number of cases for studying how election timing shapes policymaking. I argue that simple refinements can be made to improve the variable’s validity in these cases.

In this note I focus on the EU-28 countries (all 28 current European Union member states). Problems with the variable may exist for a wider range of countries. Given the depth of problems with the yrcurnt variable discussed here, it would certainly be worthwhile in future work to reevaluate the variable for the full breadth of countries covered by the DPI.

I used the European Election Database4 to find correct election years. This information was cross-checked and supplemented with individual country-election articles on Wikipedia.5


There are many instances in the DPI where election years are not recorded as 0. In other cases non-election years were mis-coded as 0. In some cases incorrect statutory election timing was also given. The following table lists these errors:

Country Errors in the DPI yrcurnt variable for the EU-28
Belgium Missing the 2010 election.
Croatia Incorrect election timing for the 1995 and 2000 elections. Also, the DPI incorrectly classifies 1991 as 4 years from the next election. Croatia gained independence in 1991 and its first election as an independent country was scheduled for the following year. So, 1991 should be coded as 1 year from the next election.
Denmark Missing the 2001 and 2007 elections.
Estonia Incorrect election timing for the 1995, 1999, 2003, 2007, and 2011 elections. Also counting originally started at 4, but should start at 3 as there is a 4 year term limit (not 5).
Germany Missing the 2005 election.
Greece Missing the 2007, 2009, 2012 election years.
Ireland Missing the 2011 election.
Italy Missing the 2008 election.
Latvia Missing the 2006, 2010, 2011 election years.
Netherlands Missing the 2003 and 2006 elections.
Portugal Missing the 1979, 1999, and 2011 elections.
Slovakia Missing the 2012 election.
Spain Missing the 1989, 1996, and 2011 elections.
United Kingdom Missing the 2001 and 2005 elections.


For a number of countries the elections recorded are for largely figurehead presidents. This can affect both when elections are recorded and how many years are given until the next election as figurehead presidents often have longer terms than parliaments. In these cases the current yrcurnt variable is not a valid measure of government election timing.

Some countries are less clear-cut because they have semi-presidential systems. Nonetheless, in a number of these cases the prime minister is the leader of the government and largely sets the domestic policy agenda. These powers are most relevant for studying topics such as public budgeting.

The following refinements should be made to create a more valid indicator for research on how election timing affects policymaking:

Country Refinements to make to the DPI yrcurnt variable for the EU-28
Austria Use parliamentary rather than (figurehead) presidential elections.
Lithuania Use parliamentary rather than presidential elections. Lithuania is a semi-presidential system where the president appoints the PM, the legislature’s approval is needed. The PM is more responsible for domestic policy.
Romania Romania is semi-presidential where the president appoints the PM, but the PM must be approved by the parliament. The PM is both the head of government and sets the legislative agenda. Before 2008, presidential and parliamentary elections occurred in the same year. Since then they have diverged, from which point the parliamentary elections should be used.
Slovenia Use parliamentary rather than (figurehead) presidential elections.

Available updated data

In sum, of the EU-28 countries there are clear errors for 14 countries’ election timing in the DPI yrcurnt variable. There are a further four countries where a more valid measure of election timing could be created by focusing on parliamentary rather than presidential elections.

A data set that implements these corrections and refinements for the EU-28 countries can be found on GitHub.6 The data (including both the original and corrected versions of the variable) can be downloaded directly into R using the repmis package (Gandrud 2015) with the following code:

yrcurnt_corrected <- repmis::source_data('')

This file also has updated election timing data through 2013.


Alt, James, David Dreyer Lassen, and Joachim Wehner. 2014. “It Isn’t Just About Greece: Domestic Politics, Transparency and Fiscal Gimmickry in Europe.” British Journal of Political Science 44 (04): 707–16.

Beck, Thorsten, George Clarke, Alberto Groff, Philip Keefer, and Patrick Walsh. 2001. “New Tools in Comparative Political Economy: The Database of Political Institutions.” World Bank Economic Review 15 (1): 165–76.

Gandrud, Christopher. 2015. repmis: Miscellaneous Tools for Reproducible Research with R.

  1. See: Accessed February 2015. All corrections and refinements discussed here refer to this version of the database.
  2. Note that Alt, Lassen, and Wehner’s (2014) substantive findings hold up when using the correct data presented here, though the estimated magnitudes of the effects are reduced. See the replication repository for more details:
  3. The DPI Codebook can be found at: Accessed February 2015.
  4. See Accessed February 2015.
  5. See Accessed September 2014 and February 2015.
  6. See
Posted in Data, Replication | Tagged , , | 1 Comment

The 2015 Asian Political Methodology Conference

Editor’s Note: this post is contributed by Fang-Yi Chiou, research fellow of the Institute of Political Science, Academia Sinica.

In the political science research communities across Asian countries, the use of quantitative methodology is a relatively recent phenomenon. However, quantitative political science has a potential to greatly inform debates and enhance our understanding of long-standing questions that are of interest to political scientists and decision makers in the region.

In January 2013, Professor Kosuke Imai of Princeton University initiated the idea of having an annual political methodology meeting in Asia and organized a program committee that is composed of methodologists from various parts of Asia. The committee then set forth to promote the advancement of quantitative social science research in Asia by holding a two-day annual conference on political methodology in Asia. The inaugural meeting was held in Tokyo in January 2014, jointly sponsored by the Tokyo Institute of TechnologyGakushuin University, and Princeton University. It attracted about 70 participants from 13 countries, and roughly 40% of them traveled from outside of Japan.

With the same spirit, the 2015 Asian political methodology conference took place in Taipei, Taiwan. We were very honored to have three distinguished invited speakers: Professor Kevin Quinn of University of California, Berkeley who is the current president of the Society for Political Methodology (SPM), Professor Simon Jackman of Stanford University who is a former president of the SPM, and Professor Chi Huang who is one of the most respected political methodologists in Taiwan. In addition, there were about 120 conference attendees, with more than 30 from regions other than Taiwan.  The conference created many opportunities for scholarly exchanges among participants who would otherwise have had no opportunity to meet.

This conference program included 3 invited talks, 12 paper presentations, and 9 poster presentations. While some presenters were domestic scholars, other presenters came from South Korea, Japan, Hong Kong, Australia, Ireland, Britain, Germany, and the U.S. These papers not only addressed some of the most interesting substantive topics in political science, but also spoke to the most important research areas in political methodology, including survey design, computation methods, measurement problems, Bayesian data analysis, and causal inference (please see for the conference program). Our presenters shed new light on existing methodological issues and highlighted areas for future research. The conference was large enough to cover various methodological issues, and yet was small enough to allow numerous opportunities for attendees to interact with each other.

This conference was organized by the Institute of Political Science at Academia Sinica, and cosponsored by the Institute of Political Methodology and Princeton University (Program for the Quantitative and Analytical Political Science). Academia Sinica has played a significant role in contributing to the development of political methodology in Taiwan and across Asia by holding international conferences on positive political economy, workshops on methodology, and summer methodology courses to promote formal modeling and political methodology in Taiwan. This conference was the first-ever large-scale international conference in Taiwan focusing on the advancement of political methodology. We believe that it will have a substantial impact on Taiwan’s next generation of political scientists, as many of the attendees were graduate and undergraduate students interested in pursuing careers in the discipline.

The program committee for this conference was composed of Fang-Yi Chiou (Academia Sinica, Taiwan, committee chair and local host), Kentaro Fukumoto (Gakushuin University, Japan), Benjamin Goldsmith (University of Sydney, Australia), Kosuke Imai (Princeton University, USA), Xun Pang (Tsinghua University, China), and Jong Hee Park (Seoul National University, South Korea). The administrative and funding support from Academia Sinica and Princeton University contributed to the conference’s success.

We will hold this conference again in Beijing in 2016, in Seoul in 2017, and in Sydney in 2018. We hope to see you in Beijing next year!

Posted in Call for Papers / Conference, The Discipline | Leave a comment