The Wild West is Wild: The Homicide Resource Curse Online Appendix Mathieu Couttenier∗
Pauline Grosjeana
Marc Sangnierb
February 2016
Missing discovery dates in the MRDS The US part of the MRDS contains 267, 072 distinct information points. Out of these observations, 152, 477 points correspond to places where mining has been or is still operated. Other information points are worthless ground geology information from the econometrician’s point of view. Among the 152, 477 that might contain useful information, the discovery year is available for 17, 595 observations. To the best of our knowledge, observations for which the discovery year is not available correspond to subsequent detailed information on characteristics of the area that immediately surrounds places where mining has been operated (e.g. discovery of a new mineral vein next to the one already exploited) and/or to deposits that are not really valuable, i.e. not worth being exploited. We present below the different approaches we use to illustrate this claim. First, we map undated discoveries and obtain Figure A2(a). Undated discoveries appear to be evenly distributed across space and located in 2, 736 out of the 3, 108 counties, a figure that is completely at odds with all available information ∗
University of Lausanne School of Economics, University of New South Wales b Aix-Marseille Univ. (Aix-Marseille School of Economics), CNRS & EHESS
a
1
about mineral resources exploitation in the US. The 849 counties considered as having mineral resources when we use dated discoveries portray a much more reasonable distribution of mining history and activity as shown by Figure A2(b) and Figure 1 in the main text. The fact that undated discoveries do not exhibit a particular spatial distribution can also be illustrated thanks to a simple regression. For instance, we estimate a linear probability model where the dependent variable is equal to 1 if the discovery year is unavailable and the independent variables are each discovery’s latitude and longitude. The R-squared from this model equals 0.003, i.e. it is very close to 0, meaning that the probability that the discovery date is missing does not vary substantially when moving on the East-West and North-South axis. Second, it is also possible to take advantage of some information about commodities provided by the MRDS and investigate differences in observable characteristics between dated and undated discoveries. It appears that 41% (3%) of undated (dated) discoveries contain sand, gravel or stone as primary commodities. In contrast, 38% (22%) of dated (undated) discoveries contain gold or silver as primary commodity. These simple figures suggest that undated discoveries contain less valuable ores. The MRDS also contains vague and frequently missing information about the size of each deposit. It turns out that this information is available for 61% (18%) of dated (undated) points. This suggests, under the hypothesis that size information is less likely to be reported for negligible deposits, that undated discoveries are likely not to correspond to anything important in terms of exploitation. Further comparing dated and undated discoveries within points for which some size information is available, we find that 87% of dated discoveries are categorized as “small”—as opposed to “large” or “medium”.1 A similar share, i.e. 90%, of undated discoveries are identically categorized. This means that, conditional on size information being available—which we interpret as a signal of non-negligibility— dated discoveries are not systematically larger than undated ones. Finally, we re-estimate equation (2) from the main text including undated discoveries as a supplementary explanatory variable. Estimated coefficients are 1
According to the data provider itself, “the precise meanings of [size information] have changed over time and are lost to history [...].”
2
presented in Table A7. Columns 1 and 2 reproduce columns 1 and 5 of Table 3. Columns 3 and 4 display the estimated coefficients when adding undated discoveries as a right-hand side variable. The magnitude and statistical significance of our coefficient of interest remains remarkably stable. In contrast, the estimated coefficient of undated discoveries is small and hardly statistically significant, which suggests that this variable has little explanatory power. Columns 5 and 6 further present estimated coefficients when our main variables of interest are interacted with the county-level undated discoveries, i.e. as if pre- and post-statehood discoveries were weighted by the extent of local missing information. Interaction terms turn out to be hardly statistically significant, while estimated coefficients of variables of interest remain fairly unchanged. This last finding and the above mentioned points make us confident that no systematic or relevant bias arises from the fact that numerous discovery dates are missing from the MRDS.
3
Figure A1: T-statistics of pre-statehood mines when discoveries are randomized across space or time.
The figure plots the distributions of the t-statistics of pre-statehood mines obtained when randomizing discoveries across space or time. Randomization across time is achieved by randomizing the share of mines discovered in a county before or after that county’s land was organized, taking the number of mines on the county’s surface as given. Randomization across space is achieved by randomly allocating mines to counties, keeping constant the overall distribution of mines across countries. The estimated specification is the same as the one displayed in Table 3, column 5. Each randomization and its subsequent estimation has been performed 1, 000 times. The vertical line corresponds to the t-statistic of the estimate of pre-statehood mines displayed in Table 3, column 5.
4
Figure A2: Dated and undated mineral resources information points.
(a) Undated mineral resources points.
(b) Dated mineral resources points. Sources: Mineral Resources Data System. Each point corresponds to a distinct MRDS entry. Sub-figure (a) points are observations for which the discovery year is not available. Sub-figure (b) points are observations for which the discovery year is available.
5
Table A1: Historical data on homicide. State
Arizona California Colorado Florida Georgia
Period
1890 1850 1880 1830 1790
-
1900 1900 1900 1860 1900
# counties
State
1 8 1 23 7
Period
Illinois Nebraska Ohio South Carolina Virginia
1830 1880 1800 1790 1790
-
# counties
1890 1900 1900 1900 1900
3 1 3 2 14
Data for counties in Arizona, Nebraska, Colorado, and California (except San Francisco) are from McKanna (2002), data for San Francisco are from Mullen (2005); Data for counties in Georgia, Ohio, and Virginia are from Roth (2009); data for counties in Florida are from Denham (1997); data for counties in Illinois are form Allaman (1989) and Erwin (1976); data for South Carolina are from Eckberg (2001).
Table A2: Historical evidence on mineral discoveries and homicide: Robustness checks. Dependent variable: yearly rate of homicide per 1, 000 inhabitants
Discovery Discovery × pre-statehood mines
County fixed effects Year fixed effects Spatial correlation adjustment County specific time trend Observations R-squared
(1)
(2)
(3)
(4) Log of dependent variable
(5) Excluding 2-sigma outliers
(6) Excluding 3-sigma outliers
-0.14 (0.09) 0.22** (0.11)
-0.14 (0.09) 0.22* (0.12)
-0.13 (0.09) 0.18* (0.09)
-0.05 (0.03) 0.10** (0.05)
-0.05 (0.04) 0.11** (0.05)
-0.04 (0.02) 0.10** (0.04)
Yes Yes 100 km
Yes Yes 500 km
Yes Yes
Yes Yes
Yes Yes
Yes Yes
3,588 0.41
3,473 0.56
3,539 0.49
Yes 3,588 0.36
3,588 0.36
3,588 0.43
*** p
