Skip to main content

Analyzing Stimulus Funds Data

A report from Brito & De Rugy at the Mercatus Center at GMU from earlier this year reports that (emphasis added):
"An OLS regression analysis controlling for the district representative’s political party, tenure in office, leadership position, membership on the appropriations committee, as well as for the district’s unemployment, mean income (i.e., the average income of a given wage earner in the district), and the percentage of employed persons working in the construction sector in 2008 finds that having a Republican representative decreases a district’s stimulus award by 24 percent.   This effect is statistically significant at the p < .001 level."
It's an interesting and useful paper.  They've put a lot of effort into compiling some data on the federal stimulus outlays and some other political variables as well as even correcting & error-checking the data they got from other sources. I like that they have provided the actual Stata dataset that they compiled (and they've even copied/pasted the Stata output tables in the document), but I thought that it would be interesting to run the models again with some interaction terms and clustering as well as adding in a couple of other political variables commonly used in some of the literature on Congressional pork/spending.    Looking past the chatter about the alleged political leanings of the Mercatus Center , it seemed to me that a regression model concluding that the stimulus funds where doled out by partisan lines should probably be examined a bit more closely.

First of all, I wanted to re-run their model with an interaction term for being a republican in a leadership position in a marginal seat and on the appropriations committee (republican*marginally*leadership*accountability).  My guess is that the effect of being republican might be different for those in leadership, marginal, and/or appropriations committee positions and if Democrats do have an advantage, which democrats got the biggest share (leadership, safe seats, appropriations committee members ?)-- that is, there's a bit more to the political calculus of stimulus spending than simple party lines.  
Second, I wanted to add a couple of other political variables, such as: 
  • the number and size ( in dollars) of earmarks secured by each member (No. of Earmarks & log_ttlearmarks)
  • the number of solo-authored legislation by each member (solo)
  • the number of votes and % of Democratic vote in each member's district in the prior election (log_ttlvotes & democrat_pctvote)
  • whether the governor, upper, and lower house of the member's state delegation are Republican controlled or not (governor_republican2, lower/upper_legisl_republican2)

to the variables already in the Brito & De Rugy model:
  • tenure - member's tenure
  • republican 
  • leadership  
  • marginaly - dummy = 1 if seat decided by margin less than 5% of votes
  • appropriations - member on appropriations comm.
  • construction - % of the civilian labor force over the age of 16 employed in the construction industry
  • meaninc - Mean household income (dollars) as estimated by the U.S. Census Bureau
  • unemployment - % of the civilian labor force over the age of 16 that was unemployed as of  2008
Note:  There appear to be other variables in their dataset that aren't in the Stata dataset they provide [see appendix of this paper].  I'd also like to get into their raw stimulus data and break out the types of stimulus fund payouts (grants vs. contracts).  
Finally, the data/variables in the models reported below haven't been checked closely -- this is just a quick & dirty analysis.   For instance, I know a couple of cases were dropped while merging in the extra data, so I'll need to go back and clean up the merges as some point.  Email me if you want a copy of the data I collected and the Stata do-file.

In the first table below, I've re-run the model for predicting the amount of stimulus dollars as described in the authors' original paper (GMU MODEL) and then built the model to include the interaction terms I described (Model2) as well as a couple more political variables on earmarks, authored bills, votes, and state government control (M3 - M7).  Note that the estimates I get for the GMU Model and the original table in the authors' paper are slightly different -- I clustered by state for all of these models.

A couple of initial observations:
  • The interactions in Models 2-7 show some interesting patters for being Republican on stimulus payout.   Controlling for the interaction of being a Republican*leader*marginalseat*appropriations member, the effect of Republican is larger than in the GMU Model (29% decrease compared with the base category of a Democrat Leader in a marginal seat and on the appropriations committee).  However, when you start to add in some more political variables I collected in Models 3-7 (esp. Earmarks, Solo Authored Bills, and State Legislator Party Control), the impact of being Republican starts to diminish.  The authors mention running some checks for collinearity, correlation, and heteroscedasticity in their original paper, and while there's some concern with the correlation (via -corr-) for the % democratic vote and Republican variables, I didn't find anything too troubling.  
  • Being a Republican leader in a marginal district, but not on the appropriations committee, decreases your stimulus payout by 33%, even when controlling for the political variables I added.  This might be more interesting than the original finding of the effect of just being Republican.  If the Democrats are punishing Republican members of Congress with lower payouts, they are really punishing the vulnerable Republican leadership who do not have the ability to protect themselves via appropriations committee representation.  
  • After controlling for all these variables, the MCs with the least stimulus dollars are Democratic leaders who are in safe seats and on the appropriations committee.  These MCs experience a 52.5% decrease in fund allotment.  These are probably the MCs who were in the best position to grab stimulus dollars, but didn't do so.
  • Finally, Democratic leaders who are in marginal seats & not on the appropriations committee also had a 27% decrease in their stimulus share.
  • So, Democrats did get a bigger share and it went to those with leadership positions, in marginal seats, and on the appropriations committee.  I think this is expected, but it wasn't evident from the GMU paper.
  • MCs who have a passed a high number of past earmarks were also successful in garnering stimulus dollars.  
  • MCs who have solo authored more bills (and are therefore probably less collaborative) were less likely to get stimulus funds in their district.
  • Curiously, Republican control of the lower house in the MC's state increases stimulus funds by 17% while Republican control of the upper house in the MC's state decreases stimulus funds by 13%.

(NOTE:  You may need to enlarge/widen your browser window to see all 7 models in the table below, or you can download an Excel or PDF version of this table)

DV = Log(Dollars)GMU MODELm2m3m4m5m6m7
No. of Earmarks0.012*0.012*0.012*
+ p<0.10, * p<0.05

I'll try to go back and improve the analysis and interpretations at some point, but feel free to email me your suggestions or questions. 


  1. Interesting post. Where did you get the data on the number and cost of earmarks?



Post a Comment