this is for holding javascript data
Yoav Ram renamed Discussion1.md to Discussion2.md
about 9 years ago
Commit id: d8223daf145f8ac0c79af6d8caefc097836337cf
deletions | additions
diff --git a/Discussion2.md b/Discussion2.md
new file mode 100644
index 0000000..1b0dac8
--- /dev/null
+++ b/Discussion2.md
...
#Discussion
## Errors and Biases
When compared to major polling companies and their ongoing polls published in the media, it seems that several parties are under-represented while others are over-represented, even after introducing our statistical controls.
Some notable examples are:
- The left-wing party Meretz gets 9 seats but only 4-5 seats in most polls.
- The right-wing party Yachad gets 9-10 seats but only 4-5 in most polls.
- The separadic-orthodox party Shas gets only 4-5 seats but 8 in most polls.
- The right-wing party Israel Beytenu gets only 0 seats but 4-5 in most polls.
One possible source of bias is the influence of abstention (non-voting). Our approach doesn't include a mechanism to assess changes in turn-out. While we did ask our respondents whether they abstained in 2013, it would be naïve to assume that they represent the non-voting population. A non-voter is presumably indifferent and would not participate in our poll. Those who do participate, probably intend to vote in 2015. We could therefore easily reach the false conclusion that turn-out will increase to nearly 100% giving those users who reported abstention in 2013 unreasonably high weights. Eventually we chose to ignore possible changes in abstention, implicitly assuming that the voting population is constant and that we need only to infer if and how they will change their vote. In particular, there are media reports that turn-out will increase dramatically in the Arab population, which could increase the number of seats for the Arab Union.
Other possible sources of bias are several demographic variables that we did not control. Since we are already control for the 2013 election results (_i.e._ our weighted sample agrees with the official 2013 results), the following question is of relevance to the sources of bias: In what way are our respondents that voted for party \(i\) in 2013 different from the actual voting population that voted for party \(i\) in 2013? Some variables that may have played a role in biasing our sample could be:
- Age. It is reasonable to assume that young voters are more likely to vote for new, small, niche or extreme parties than are older voters. For example, older voters who voted for Ha'Likud in 2013 are more likely to vote again to Ha'Likud than are youngers voters, who are in turn more likely to switch to Yachad or Kulanu. This same bias could explain the high number of seats projected for Merez and the low number of seats projected for Israel Beytenu.
- Sex. It is possible that men changed their minds (between 2013 and 2015) in a manner different from women. Ha'Mahane Ha'Zioni has a high level of female representation, including Livni who is set to become prime minister through rotation with Herzog. As an example, women who voted for Yesh Atid in 2013 might be more likely to switch to Ha'Mahane Ha'Zioni than men who voted for Yesh Atid in 2013.
Another source of error could be the sample size. 2013 votes were only collected from respondents after February 13th, and only in Android devices. Therefore, the 2013 dataset contains only ~2,400 votes, roughly a third of our entire sample. Due to our methodology, it is imperative to have a reasonable sample size for each voter 'weight class', which is decided in our case by the voter's 2013 vote. For instance, we might have 400 respondents who voted Ha'Likud in 2013, but only 5 that voted Shas. Those few respondents in the 2013-Shas weight class will get high weights, likely leading to large errors. Due to time limitations, we did not have the time to estimate these errors, so it is likely that they explain at least some of the deviance in our forecast. In particular, this could explain the low number of seats projected for Shas, Yahadut Hatora or Israel Beytenu.
