# Explaining discrepancy between ANES data and actual results for 2012 US presidential election

I am looking at the ANES 2012 Time Series Study. The three variables that (I believe) tell us about who survey respondents voted for are:

1. prevote_presvtwho: Pre-election survey (early voters).
2. postvote_presvtwho: Post-election survey.
3. presvote2012_x: Basically the above two variables added together.

Data below. From the third variable, we see that there were 2,496 respondents who reported voting for Obama vs only 1,692 for Romney.

I was surprised by this: Obama had 2496/1692 - 1 = 47.5% more votes than Romney in this ANES sample, vs only 8.2% more votes in the actual national popular vote (NPV).

What explains this huge discrepancy between ANES and the actual NPV? Or am I just totally messing up somewhere?

``````>     . tab prevote_presvtwho
>
>     PRE: For whom did R vote for |
>                        President |      Freq.     Percent        Cum.
>     -----------------------------+-----------------------------------
>                      -9. Refused |          9        0.15        0.15
>                 -1. Inapplicable |      5,519       93.32       93.47
>         1. [preload: dem_pcname] |        222        3.75       97.23
>         2. [preload: rep_pcname] |        153        2.59       99.81
>     5. Other candidate {SPECIFY} |         11        0.19      100.00
>     -----------------------------+-----------------------------------
>                            Total |      5,914      100.00
>
>     . tab postvote_presvtwho
>
>     POST: For whom did R vote for President |      Freq.     Percent        Cum.
>     ----------------------------------------+-----------------------------------
>                                 -9. Refused |         43        0.73        0.73
>     -7. Deleted due to partial (post-electi |        152        2.57        3.30
>     -6. Not asked, unit nonresponse (no pos |        252        4.26        7.56
>                            -1. Inapplicable |      1,547       26.16       33.72
>                    1. [preload: dem_pcname] |      2,274       38.45       72.17
>                     2. [preload: rep_pcname |      1,539       26.02       98.19
>                5. Other candidate {SPECIFY} |        107        1.81      100.00
>     ----------------------------------------+-----------------------------------
>                                       Total |      5,914      100.00
>
>     . tab presvote2012_x
>
>        PREPOST SUMMARY: For whom did R vote |
>                       for President in 2012 |      Freq.     Percent        Cum.
>     ----------------------------------------+-----------------------------------
>                                          -9 |         52        0.88        0.88
>                                          -6 |          1        0.02        0.90
>     -2. R did not vote for Pres or did not  |      1,555       26.29       27.19
>                             1. Barack obama |      2,496       42.20       69.39
>                              2. Mitt romney |      1,692       28.61       98.00
>                                    5. Other |        118        2.00      100.00
>     ----------------------------------------+-----------------------------------
>                                       Total |      5,914      100.00
``````
• Did you remember to weight the responses? electionstudies.org/resources/HowToAnalyzeANESData.pdf Commented May 25, 2017 at 4:58
• @indigochild: That must be it. Thanks. I'll read that and try to answer my own question if I figure it out.
– user2212
Commented May 25, 2017 at 6:14

As pointed out by indigochild, this discrepancy is largely because I forgot to weight the responses. (ANES does not use a simple random sample and so to generalize from the sample to the population, one must weight them appropriately.)

Without weighting, our sample suggests that Obama's vote share is Obama / (Obama + Romney) = 2496 / (2496 + 1692) = 59.6%. And Romney's is R / (O + R) = 40.4%.

With weighting, Obama's share falls to 53.6% and Romney's rises to 46.4%. There is still a slight discrepancy with the true figures,* but this can be explained by sampling error.

*Obama vote share = 65,915,795/(60,933,504+65,915,795) = 52.0% and Romney vote share = 60,933,504/(60,933,504+65,915,795) = 48.0%.

The following is the code I used in Stata on the file anes_timeseries_2012.dta to get the above 53.6% and 46.4% figures. (Please let me know if I'm still doing it wrong.)

``````*Create variable Obama. Voted Obama = 1. Voted Romney = 0.

gen obama=.
replace obama=0 if presvote2012_x==2
replace obama=1 if presvote2012_x==1

reg obama
reg obama [pw=weight_full]

*Create variable Romney. Voted Romney = 1. Voted Obama = 0.

gen romney=.
replace romney=0 if presvote2012_x==1
replace romney=1 if presvote2012_x==2

reg romney
reg romney [pw=weight_full]
``````