Naidu (2010)
| Article | |
|---|---|
| Has bibtex key | |
| Has article title | |
| Has author | Naidu |
| Has year | 2010 |
| In journal | |
| In volume | |
| In number | |
| Has pages | |
| Has publisher | |
| © edegan.com, 2016 | |
Contents
Empirical Questions
What's the author's research question and hypothesis?
The author is studying the effect of disenfranchisement of Southern blacks (through poll taxes and literacy tests) on:
- (a) Voter turnout,
- (b) The Democratic party vote share,
- (c) The teacher/child ratio for blacks,
- (d) The teacher/child ratio for whites.
- (e) Land values in counties with the poll taxes and literacy tests.
- (f) Migration of blacks.
... in the counties/states where this disenfranchisement was implemented.
For the statistical tests, the null hypothesis is that the effects were zero. The author does not give a sense of his priors, but he does say that his findings (all null hypotheses rejected except for (d)) are "[C]onsistent with historical evidence that these disenfranchisement laws independently lowered black political participation."
In particular, the author notes that the fall in black educational inputs (ie, the teacher/student ratio) is consistent with theoretical political economy models including the one developed later in this paper.
All of this is on page 2 and 3 of the paper.
How it is tested?
The author compares adjacent county-pairs that straddle borders.
Note that the same county can be in multiple pairs, and therefore is in the sample multiple times. This creates the need for multidimensional clustering as developed by Cameron et al, 2006. Standard errors are clustered for both within-state over time correlations of county residuals, as well as within borer-segment. Discussion of these in first paragraph of Section 5 on page 22.
- COMMENTARY ABOUT EQUATION 11 ON PG 22: Why are the two [math]D[/math] dummy variables summed?
The author goes on to state that [math]\beta[/math] is supposed to represent the effect of an additional poll tax or literacy test within a state. He also notes that the laws are highly collinear -- ie, they tend to co-exist together -- and there is little information in the independent dummies.
Given this, I would think that he'd want to estimate the effect of having (either) a poll tax or a literacy test, not the sum of both.
The author does note that in an earlier version of the paper he estimated the effects of poll taxes and literacy tests separately and reached similar qualitative conclusions with slightly smaller standard errors.
What do the tests achieve?
The tests reject all null hypotheses listed above except for (d). With additional disenfranchisement laws:
- (a) Voter turnout in all elections decreases. Presidential turnout decreases by 8%-11% and gubernatorial turnout decreases by 23%. ~10%-12% decrease in Congressional election turnout, less precisely estimated.
- (b) The Democratic party vote share increases. 5.8% increase in Democratic presidential share, and 10% increase in congressional Democratic share. Positive but insignificant effects on gubernatorial Dem voteshare. Discussion, pg 26.
- (c) The teacher/child ratio for blacks decreases by 50% (!).
- (d) The teacher/child ratio for whites don't change (bottom of pg 27).
- (e) Land values in counties increase by 7%, and the number of farms increase by 6%.
- (f) Blacks leave the counties in question.
How could the tests be improved?
The treatment/control spillover concern (addressed below) could be improved upon by relaxing the adjacency requirement. For example: A nearest-neighbor matching algorithm could select counterfactual counties that are similar but not contiguous.
What are the tests' strengths and weaknesses?
The most obvious source of confounds to the empirical test would be other legislation or developments that happened simultaneously with disenfranchisement.
Another issue is related to spillover. Because the strategy involves treatment/control groups that are contiguous, its possible that the control groups are getting some form of spillover treatment. IE, the control groups in this study may actually be bad counterfactuals for the treatment groups -- because their spacial proximity means that they're being partially (if not fully) treated.
These issues discussed explicitly on pg 23, "Threats to Identification."
Can you think of any alternative empirical tests?
Note a few tweaks above to this analysis involving a different type of matching. Similarly, one could use p-score matching to match counties.
Is there randomness in the timing of compliance with the state-level disenfranchisement law?
One could also do a similar analysis around the times that the disenfranchisement was rolled back.
