1
The impact of preferential trade agreements on governmental repression revisited – Online appendix
To test the robustness of our findings, we also employed other econometric instruments that are not reported in the paper.
- First, Clarke (2005) shows that the inclusion of control variables may actually increase the bias instead of decreasing it. Similarly, the balance statistics from the paper indicate that we were unable to match the samples perfectly for human rights ratification. However, Table 4 in our manuscript demonstrates that making amendments in this regard, i.e., dropping all controls or only human rights ratification, does not affect the substance of our findings.
- We also replaced PTA hard law with a dichotomous variable on any form of human rights reference in PTAs including soft law requirements only. This accounts for Chayes and Chayes (1993) who argue that non-compliance is the exception, since countries may have an inherent interest not to violate agreements (deliberately) to which they have committed themselves. Under those conditions, even soft law may change state behavior. After conducting the matching and estimating the parametric models with the new explanatory variable, however, our models that are summarized in the following Table 1 demonstrate that even those less enforcing human rights standards in PTAs are unlikely to affect states’ compliance with human rights.
Table 1 The impact of any type of human rights law in PTAs on human rights compliance, 1977-2009: Matched sample
Model 1 / Model 2 / Model 3 / Model 4PTA with any type of human rights law / 0.12 / -0.07 / 0.15 / 0.06
(0.24) / (0.21) / (0.21) / (0.19)
Human rights ratification / 0.66 / 0.59
(0.15)*** / (0.14)***
Population density / 0.00 / 0.00 / 0.00
(0.00) *** / (0.00)** / (0.00)**
Political stability / -0.01 / -0.01 / -0.01
(0.01) / (0.01) / (0.01)
Democracy / -0.12 / -0.09 / -0.08
(0.02)*** / (0.02)*** / (0.02)***
GDP per capita / -0.44 / -0.43 / -0.28
(0.11)*** / (0.11)*** / (0.10)***
Trade / -1.01 / -0.90 / -0.83
(0.21)*** / (0.22)*** / (0.20)***
Years since torture / -0.45
(0.04)***
Obs / 3,172 / 3,172 / 3,172 / 3,172
Log pseudolikelihood / -4694.4417 / -3,869.324 / -3,953.9334 / -3,504.8768
Wald2 / 0.26 / 113.59*** / 110.37*** / 452.07***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) included in Model 4, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
- Our dependent variable in the paper only focuses on one type of political repression, i.e., the violation of personal integrity rights, but ignores broader forms of repression such as civil liberties sanctions. Especially these broader forms may be less influenced by the PTAs and may suffer less from selection effects. We therefore replaced the dependent variable with a) the Freedom House civil liberty index and b) in a second step, with the Cingranelli and Richards (1999) measure on government respect for physical integrity rights. Afterwards, we conducted the matching algorithm and estimated our core models again. These robustness tests, which are summarized in the next two tables (Table 2: Freedom House; Table3: Cingranelli and Richards (1999) measure), did not differ from our main results reported in the paper, however.
Table 2 The impact of hard law PTAs on human rights compliance – Freedom House, 1977-2009: Matched sample
Model 5 / Model 6 / Model 7 / Model 8PTA hard law / 0.05 / -0.02 / 0.04 / 0.00
(0.27) / (0.23) / (0.21) / (0.22)
Human rights ratification / 0.16 / 0.26
(0.16) / (0.15)*
Population density / 0.00 / 0.00 / 0.00
(0.00)*** / (0.00)*** / (0.00)
Political stability / -0.02 / -0.02 / -0.02
(0.01)*** / (0.001)*** / (0.01)***
Democracy / -0.42 / -0.42 / -0.38
(0.03)*** / (0.03)*** / (0.03)***
GDP per capita / -0.79 / -0.79 / -0.78
(0.11)*** / (0.11)*** / (0.11)***
Trade / -0.35 / -0.34 / -0.25
(0.24) / (0.24) / (0.23)
Years since torture / -0.34
(0.08)***
Obs / 2,675 / 2,675 / 2,675 / 2,675
Log pseudolikelihood / -4978.5696 / -2828.2473 / -2832.2655 / -2764.5818
Wald2 / 0.04 / 331.55*** / 333.92*** / 412.13***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) included in Model 8, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
Table 3 The impact of hard law PTAs on human rights compliance – Cingranelli and Richards, 1977-2009: Matched sample
Model 9 / Model 10 / Model 11 / Model 12PTA hard law / 0.22 / 0.28 / 0.29 / 0.23
(0.25) / (0.21) / (0.21) / (0.19)
Human rights ratification / -0.49 / -0.50
(0.19)** / (0.17)***
Population density / - 0.00 / - 0.00 / - 0.00
(0.00)** / (0.00)** / (0.00)*
Political stability / -0.01 / -0.01 / -0.01
(0.00)* / (0.00)* / (0.00)
Democracy / 0.11 / 0.09 / 0.11
(0.02)*** / (0.02)*** / (0.02)***
GDP per capita / -0.49 / -0.47 / -0.47
(0.12)*** / (0.12)*** / (0.11)***
Trade / 0.87 / 0.83 / 0.71
(0.22)*** / (0.21)*** / (0.19)***
Years since torture / -0.57
(0.20)***
Obs / 2,764 / 2,764 / 2,764 / 2,764
Log pseudolikelihood / -5792.4431 / -5043.925 / -5080.6657 / -4824.752
Wald2 / 0.78 / 146.36*** / 149.33*** / 459.53***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) included in Model 12, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
- Our paper’s dependent variable’s fifth category, which refers to a level of terror that has expanded to the whole population (see Table 1 in the manuscript), has only 259 observations in the original sample, meaning that this value could be an outlier category that biases our results. In order to address this, we recoded political repression by grouping the fourth and fifth category together. Similarly, we also considered all models using a less stratified dependent variable by employing a dichotomous item for political repression with the value of 0 combining categories 1–2 and the value of 1 combining categories 3-5 of the original variable (note that the ordered logit regression converges to a regular logit here). Both changes did not alter our core results and are summarized in Tables 4 and 5.
Table 4 The impact of hard law PTAs on human rights compliance (four categories only), 1977-2009: Matched sample
Model 13 / Model 14 / Model 15 / Model 16PTA hard law / 0.11 / -0.09 / 0.14 / 0.04
(0.22) / (0.21) / (0.21) / (0.19)
Human rights ratification / 0.69 / 0.62
(0.15)*** / (0.14)***
Population density / 0.00 / 0.00 / 0.00
(0.00) / (0.00) / (0.00)*
Political stability / -0.01 / -0.01 / -0.01
(0.01) / (0.01) / (0.02)
Democracy / -0.12 / -0.09 / -0.08
(0.02)*** / (0.02)*** / (0.02)***
GDP per capita / -0.44 / -0.42 / -0.27
(0.11)*** / (0.11)*** / (0.10)***
Trade / -1.00 / -0.87 / -0.80
(0.22)*** / (0.22)*** / (0.20)***
Years since torture / -0.45
(0.04)***
Obs / 3,172 / 3,172 / 3,172 / 3,172
Log pseudolikelihood / -4332.3484 / -3516.614 / -3606.2196 / -3150.5679
Wald2 / 0.63 / 107.86*** / 105.18*** / 440.23***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) included in Model 16, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
Table 5 The impact of hard law PTAs on human rights compliance (dichotomous measure), 1977-2009: Matched sample
Model 17 / Model 18 / Model 19 / Model 20PTA hard law / - 0.01 / -0.38 / - 0.03 / -0.09
(0.25) / (0.24) / (0.22) / (0.20)
Human rights ratification / 0.90 / 0.84
(0.18)*** / (0.14)***
Population density / 0.00 / 0.00 / -0.00
(0.00) / (0.00) / (0.00)
Political stability / -0.01 / -0.01 / -0.00
(0.01) / (0.01)* / (0.00)
Democracy / -0.13 / -0.09 / -0.08
(0.02)*** / (0.02)*** / (0.02)***
GDP per capita / -0.49 / -0.45 / -0.28
(0.13)*** / (0.12)*** / (0.10)***
Trade / -0.82 / -0.73 / -0.46
(0.27)*** / (0.27)*** / (0.21)**
Years since torture / -0.72
(0.05)***
Obs / 3,172 / 3,172 / 3,172 / 3,172
Log pseudolikelihood / -2190.5891 / -1543.2769 / -1630.9363 / -1114.247
Wald2 / 0.00 / 72.63*** / 66.74*** / 471.34***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Constant and cubic splines (Beck, Katz, and Tucker 1998) included in Models 17-20 and Model 20, respectively, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
- The matching approach in our paper corrects for sample selectivity on observed covariates. For these covariates, we followed the existing literature. However, sample selectivity may also be based on unobserved covariates for which a Heckman selection model (Heckman 1979) might be more appropriate. Hence, a more intuitive design for testing whether hard law PTAs affect states’ human rights practices might be a two-stage process, i.e., a first stage where the treatment is a PTA with any human rights law (1) or none (0), and a second stage where our dependent variable from above constitutes the outcome. For this estimation, the unit of analysis stays the same as in Table 4 of our paper, although we now consider the unmatched sample as observations in our data. Despite this additional check for sample selection based on unobserved covariates, the results do not change. Note, however, that the parameter , which indicates whether the error terms in the selection and the outcome equation are correlated (i.e., an indicator for whether selection matters or not), is statistically significant in Table 6 that is provided below.
Table 6 Heckman selection model
Model 21Outcome equation: Repression
PTA hard Law / 0.07
(0.25)
Human rights ratification / 0.69
(0.11)***
Population density / -0.00
(0.00)
Political stability / -0.01
(0.00)
Democracy / -0.08
(0.01)***
GDP per capita / -0.09
(0.07)
Trade / 0.54
(0.09)***
Years since torture / -0.83
(0.11)***
Selection equation: Human rights clause
Human rights ratification / 0.65
(0.07)***
Trade / 0.57
(0.10)***
GDP / -0.03
(0.05)
Population density / 0.00
(0.00)
Political stability / 0.00
(0.00)
Democracy / -0.01
(0.01)
Years since torture / -0.90
(0.06)***
Constant / -2.02
(0.54)***
Obs (censored) / 4,117 (2,549)
Log pseudolikelihood / -2,960.159
Wald 2 / 1035.71***
/ 0.94***
Robust standard errors clustered on country in parentheses
* significant at 10%; ** significant at 5%; significant at 1% (two-tailed).
- While our approach follows Hafner-Burton (2005) as closely as possible, several other studies in the literature that examine governmental repression as well consider different specifications of the variables we used. For example, Landman and Larizza (2009), Davenport and Armstrong (2004) or de Soysa and Vadlammanati (forthcoming) take the size of the total population instead of population density that we employed. We therefore replaced our item with total population, conducted the matching again, and re-estimated the parametric ordered logit in order to see whether this might have had an impact on our core finding or not. The next table summarizes our results in this regard and clearly shows that our substantial finding is not affected by this change.
Table 7 The impact of hard law PTAs on human rights compliance - population, 1977-2009: Matched sample
Model 22PTA hard law / 0.06
(0.20)
Human rights ratification / 0.40
(0.17)***
Population / 0.00
(0.00)***
Political stability / -0.02
(0.01)**
Democracy / -0.07
(0.02)***
GDP per capita / -0.40
(0.12)***
Trade / -0.37
(0.23)
Years since torture / -0.45
(0.05)***
Obs / 2,754
Log pseudolikelihood / -2992.023
Wald2 / 408.55***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) are included in the model, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
- Similar to the previous point, albeit differently, an anonymous reviewer raised the concern that Hafner-Burton’s (2005) treatment of countries’ general commitment to international human rights agreements via the International Covenant on Civil and Political Rights and the Convention against Torture. Our final variable in the manuscript, the human rights ratification variable, is an ordinal variable ranging from 0 to 2. Those values are derived from thetotal number of the two treaties that a state has ratified into national law by any given year. Against this background, we replaced this operationalization with two dichotomous indicators, i.e., one dummy variable capturing whether a country as signed the International Covenant on Civil and Political Rights or not and one dummy variable capturing whether a country as signed the Convention against Torture or not. Afterwards, we conducted the matching as described in our paper and estimated the ordered logit models again. The findings for our full model are summarized in Table 8.
Table 8 The impact of hard law PTAs on human rights compliance – ICCPR and CAT separately, 1977-2009: Matched sample
Model 23PTA hard law / 0.19
(0.22)
ICCPR / 0.34
(0.38)
CAT / 0.48
(0.30)
Population density / 0.00
(0.00)
Political stability / -0.01
(0.01)
Democracy / -0.06
(0.02)***
GDP per capita / -0.41
(0.11)***
Trade / -0.67
(0.23)***
Years since torture / -0.49
(0.04)***
Obs / 2,764
Log pseudolikelihood / -3056.8904
Wald2 / 485.57***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) included in the model, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
- Recent studies dealing with human rights violations employ the same specification for democracy as we do (e.g., Landman and Larizza 2009; Abouharb and Cingranelli 2009; de Soysa and Vadlammanati forthcoming). Nevertheless, Davenport and Armstrong (2004) essentially argue that democracy might not have linear and negative relationship with human rights violations, but a curvilinear one. We therefore employed their suggested lowess smoother (shown below) and, in fact, our results are virtually identical to Davenport and Armstrong’s (2004: 547) left panel in Figure 1. Against this background, we decided to estimate an additional model that relies on a curvilinear specification of democracy and, thus, avoids the potential misspecification in the actual revised manuscript. Please note, however, that this does not affect our main result.
Fig 1 Lowess smoother: Democracy and its relationship with governmental repression
Table 9 The impact of hard law PTAs on human rights compliance – democracy and democracy squared, 1977-2009: Matched sample
Model 24PTA hard law / 0.18
(0.20)
Human rights ratification / 0.59
(0.17)***
Population density / 0.00
(0.00)
Political stability / -0.01
(0.01)
Democracy / -0.07
(0.02)***
Democracy squared / -0.01
(0.00)**
GDP per capita / -0.22
(0.11)**
Trade / -0.62
(0.23)***
Years since torture / -0.41
(0.04)***
Obs / 2,754
Log pseudolikelihood / -3041.0628
Wald2 / 466.68***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) included in the model, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
- In order to address the impact of financial and market transactions on countries’ tendency to comply with human rights, Hafner-Burton (2005: 617) additionally considers a variable on countries’ inflows and outflows of foreign direct investment. We decided to drop this ‘investment’ variable due to three reasons. First, it theoretically addresses the same concerns as Trade, which we do include. Second, Hafner-Burton’s variable is statistically insignificant throughout any model estimation, rendering it unlikely that this item will crucially affect our results. Finally, the World Bank Development Indicators as the source for the investment variable suffer a lot from missing values. In turn, this makes it difficult to consider this item for our methodological approach that requires ex-ante that missing values do not exist and we lack convincing grounds for imputing the missing information. That being said, we obtained this variable from the World Bank Development Indicators and imputed missing values linearly. Due to this approach, we have data for 3,938 out of 4,117 observations in 1977-2009. Afterwards, we re-calculated the matching procedure before estimating our parametric models again. Our core result remains robust and we summarize the main model in the following table.
Table 10 The impact of hard law PTAs on human rights compliance – FDI included, 1977-2009: Matched sample
Model 25PTA hard law / 0.10
(0.21)
Human rights ratification / 0.38
(0.16)**
Population density / 0.00
(0.00)*
Political stability / -0.01
(0.01)*
Democracy / -0.08
(0.02)***
FDI / 0.04
(0.01)***
GDP per capita / -0.28
(0.10)***
Trade / -0.71
(0.21)***
Years since torture / -0.42
(0.06)***
Obs / 2,674
Log pseudolikelihood / -2928.2618
Wald2 / 430.49***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) included in the model, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
- An anonymous reviewer wondered about the timing of the effects of membership in PTAs with hard conditions. While we did lag all of our explanatory variables by one year in the manuscript, it could also be that the effect is not apparent after a one-year lag, but that it only becomes apparent after a longer lag owing to the types of changes that need to be made to improve human rights practices. For addressing this valid concern, we re-estimated all procedures and models with a two-year lag and a five-year lag, respectively. Our final models based upon these amendments are summarized below (first model: two-year lag; second model: five-year lag). Again, our core finding remains robust and it seems unlikely that PTAs exert an impact after a longer period of time has elapsed. Instead, the self-selection mechanism holds and seems to be the driving force in these circumstances.
Table 11 The impact of hard law PTAs on human rights compliance – FDI included, 1977-2009: Matched sample
Model 26 / Model 27PTA hard law / 0.04 / 0.08
(0.18) / (0.20)
Human rights ratification / 0.57 / 0.52
(0.16)*** / (0.13)***
Population density / 0.00 / 0.00
(0.00)** / (0.00)**
Political stability / -0.01 / -0.01
(0.01)* / (0.01)*
Democracy / -0.08 / -0.07
(0.02)*** / (0.02)***
GDP per capita / -0.35 / -0.34
(0.10)*** / (0.10)***
Trade / -0.85 / -0.70
(0.23)*** / (0.20)***
Years since torture / -0.43 / -0.58
(0.04)*** / (0.05)***
Obs / 2,718 / 2,600
Log pseudolikelihood / -2958.9632 / -2791.1187
Wald2 / 553.67*** / 433.36***
Table entries are coefficients. Robust standard errors clustered on country in parentheses. Cubic splines (Beck, Katz, and Tucker 1998) included in the models, but omitted from table
* significant at 0.1 level, ** at 0.05 level, *** at 0.01 level (two-tailed)
- Hafner-Burton’s (2005) models omit a series of control covariates that are now standard explanatory variables of repression and human rights compliance. Due to our rationale to follow Hafner-Burton (2005) as closely as possible, we also omitted these variables. Nevertheless, we also concur with the most recent literature that neglecting these items may also induce severe bias in our models, ultimately leading to the outcome that our core finding does not stem from the proclaimed selection process, but misspecification. Therefore, in a first step, we consulted the relevant literature and identified the following variables as particularly robust and influential:
- Domestic and international conflict involvement: for capturing these two concepts, we rely on data from Abouharb and Cingranelli (2009) that was originally compiled by Gleditsch et al. (2002). These two variables measure the level of national and international conflict, respectively, a country is involved in on an ordinal 0-3 scale. The variables are denoted civilwarintensitylag1 and interstatewarlag1, respectively.
- Regional norms: countries’ positions toward compliance with human rights might be affected by the norms and behavior of neighboring countries. This rationale is in line with, e.g., Gleditsch (2007) who claims that conflicts in a state’s regional contextcan increase the risk of conflict in one’s own country. Accordingly, we take one variable from Gleditsch (2007: 302) that measures regime types in connected states. This is measured “by the average level of democracy among states in a region surrounding a country. This variable can range from a low of –10 to a high of 10, in the event that all neighbors are considered full democracies.” The variable is denoted nbsdma_i.
- Several studies emphasize the importance of the degree of judicial independence (e.g., Cingranelli and Richards 1999). We obtained a variable from Cingranelli and Richards (1999) that indicates the extent to which the judiciary is independent of control from other sources, such as another branch of the government or the military. A score of 0 indicates “not independent”, a score of 1 indicates “partially independent” and a score of 2 indicates “generally independent.” We interpolated missing values linearly and the final variable is denoted injud_interpolated.
- Finally, in a related fashion to the variable on regional norms, one might consider two other factors in this regard. First, international norms that emerge from the participation in global governance efforts and the interaction with other states in the international arena. Aaronson and Abouharb (2011) consider the same rationale and include an annual count of inter-governmental organizations a state has joined until this year. We take the data from these scholars, which have been taken originally from Pevehouse et al. (2003). The variable is denoted igosjoined. Second, norms may not emerge from contacts with “official actors” in a regional or global context, but also due to pressure exerted by civil society groups. Murdie and Bhasin (2011), for example, consider here the count of permanent offices that “the 432 human rights organizations identified by Murdie and Davis (2008) have in a county in a given year. We take this variable, impute missing values linearly, and include it as the final additionally considered control.
Against the background of the operationalization of these additional variables, we conducted the matching as described in our paper. The only change to the procedure described there is that we also match on these further controls as this resulted in the optimally achieved balance for all control covariates. The following statistics summarize the balance statistics for our covariates after the matching: