Overall

• Clearly written
• Clarifies important concepts that have been used very sloppily
• Answeres a big, elusive question
• Teaches us about stats, as well as primaries
  • propagates measurement uncertainty through model stages

“All ideal point estimates are uncertain, but not many studies confront this uncertainty.”

Summary of comments:

• Clarify terminology and scope
• Framing (strategic positioning vs. representation|DIME as the DV)
• Do I understand the measures, model, & results?
• Robustness checks
• Am I updating my priors correctly?
• Thoughts on ontology and causality (possibly nonsense)

???

Mike has clearly thought about this so carefully that I sincerely doubt that I can offer anything thing that pushes the thinking in any way. My comments are really about framing.

I see my role here as commenting on what is more or less clear as a reader. The thoughts it triggered in my mind and how well that aligns with how you intend me to be updating my priors.

I’ll suggest some robustness checks because I know you are into that kind of thing.

I also just want to reiterate some of the important contributions.

Summary

People blame primaries for - polarization - selecting general-election losers - populism

–

BUT, we have yet to observe the preferences of local partisans.

–

Mike estimates ideal points for each districts’ partizans.

–

So, do primaries work?

–

Theory

Primaries work if the 1. local partisan voters’ 1. policy preferences/ideologies 1. are represented in/transmitted into government

Implication: If primaries work, a district’s partizans’ preferences should affect - Chapter 4: Who runs (and their positions) - Chapter 5: Who wins

???

Mike helpfully gives us a clear definition that allows us to contrast theory and findings.

When should primaries not work?

1. Unchallenged candidates (?)
2. Uncompetitive districts (?)
3. Uncompetitive candidates (✓)

Are a-typical candidates/donors responsible for the higher variance in DIME scores for challengers and open seats? Are these outliers long-short candidates?

???

Lots of donors only give to conventional candidates they think can win (esp. incumbents), so the higher correlation here may be more a function of time than your measurement.

Measurement-DIME Scores

(How to frame this chapter)

DIME ~ Strategic candidate positioning? If not, not so informative regarding the strategic positioning dilemma per se (but this is fine)

DIME ~ Credible candidate policy commitments? If so, good for testing the core representation hypotheses (do primaries work?)

DIME ~ Informal party networks? If so, are we recovering the correlation between district ideology and party networks?

background-image: url(mgd-scores.png) background-size: contain

Measurement-DeCrescenzo Scores

I would be interested in Figure 4.1 with uncertainty around your measure.

That level of separation! (Much more than DIME scores). Do the party mean estimates have a strong pull or that the battery is fairly partisan? The most conservative local Democrats are way more liberal than the most liberal local Republicans.

Are cutpoints generally less than 0?

???

You may not want to put this in the paper, but I would be interested in figure 4.1 with uncertainty around your measure–I imagine this would show a lot less separation.

I thought they averaged to 0

Model

The unit of observation is \(DIME_i\) (4.4-4.7), but only \(\overline{DIME}_g\) is identified (4.2.1), correct?

–

the variables that could confound the relationship between citizen and candidate ideology could differ dramatically across parties

Thus separate models (latent space?) by party. ✓

incumbency status appears to be a substantial modifier of the relationship between citizen and candidate ideology

By this logic, must we estimate incumbents separately? (Say more about why this is not just a candidate trait.)

Mediation

" it is important to demonstrate that it affects candidate positioning above and beyond its intermediate effect on district voting."

A district’s past vote share likely influences candidate positions (and decisions to run). However, it may do so primarily because primaries exist. Thus, 1. If the effect of ideology is 100% mediated, primaries might still “work,” and thus, the total effect may be more relevant. 1. What we really want to block is the portion of the effect of past presidential vote share that is not due to its signal of district ideology (i.e., the mediated effect). The way you calculated the average total effect does this, right? It is just a matter of emphasis on what is blocked and why.

“\(z_i\) currently contains an order-3 polynomial function of candidate i’s total campaign receipts.”

Why?

Is this an intermediate confound between past presidential vote share and candidate DIME score?

Background features

“we should be concerned that district features that promote conservative voters also promote conservative candidates.”

But the role of background features depends on the candidate selection process. For example, does this mean we want to exclude the effect of the fact that the candidate is a draw from the partisan voter pool? Primaries work if the draw is centered on the mean of this group–not more likely to come from the left or right tails. (Approximated by estimating the relationship between \(\bar{\theta}_g\) and \(\overline{DIME}_g\), because we can’t scale DIME and MGD scores.) Is this what we mean when we are estimating the mediated effect of district partizans?

Vote choice (vote share) varies, but voters’ policy preferences do not?

4.3.1 Does this over-estimate the effect of previous vote share on donor patters (DIME scores vary over time as candidates come and go)?

Interpretation

Figure 4.4: Past vote share is only an informative for Dem Challengers. The more Republican the district, the more conservative a Democrat challenger’s donors are, right?

Related questions about representation

Under what conditions do a district’s partizans policy preferences diverge from the policy preferences (donation patterns) of candidate slate?

–

Ontology & Causality

“Ideology affects responses”

“Items reveal different information about the latent construct.”

Latent variable: Brains can be parameterized (true idology exists, is a “cause”, a “force”)

vs.

“Data depend on the construct.”

“IRT lets us summarize an individual’s policy preferences.”

Latent construct: We construct ideology from responses (true ideology is responses to an infinite number of questions? )

???

I would not normally go down this path, but I figure that you might actually care, have answers for these such obtuse questions that help me make sense of them.

The first aligns with how we model things

The later alings with how we measure things

Maybe we can align these, may it is nonsense?

Clarifications

In a policy choice context, the difficulty parameter is better understood as the “cutpoint” parameter, the midpoint between two policy choices where the respondent is indifferent between the choice of Left or Right on item j.

This phrasing could track the above and below sentences better to stress thas cutpoints, like difficulty, are features of a item, not a respondent. Perhaps this is as simple as saying the “mean respondent is indifferent.” I would actually introduce the cutpoint language in the educational paragraph above, so you don’t have to intoduce two things at once here, and so you can just use cutpoint going forward.

I like “midpoint” better than “cutpoint” or “difficulty”

“how much to cut capital gains taxes will have a more conservative cutpoint than a question of whether to cut capital gains taxes at all.”

This example is confusing because the response scales also seem to differ. Try “whether to cut capital gains tax by 50% will have a more conservative cutpoint than whether to raise capital gains tax by 50%” or “whether to cut capital gains tax by 50% will have a more conservative cutpoint than whether to cut capital gains tax by 1%”

Figure 2.1 could have shading/color for difficulty and line type for discrimination (solid lines for ι = 1)

Instead of “behave identically,” you could say “be as difficult/conservative and as discriminating…”

I like “their ideal point” rather than “her ideal point” for the generic person.

“standard Normal utility error from εi j above”

nbd but I don’t see any equasion with εi j above

Not important, but given how nice you walk through every other step, you could explain the move from individual to population standard deviation in Equasion 2.8

Might as well say “partizan voters in a district” rather than partisan citizens in a district"

The discription of 4.2 uses different notation than the figure.

Figure 4.4 is pretty, but could be a little clearer if there were boxes around the facets (it takes a second to realize that there are 3 x axes). Also the 0 line from 4.5 is nice (this might actually be all you need).

Possible typos

where groups are define as

policy ideology with a partisan constituency

the policy ideology of a partisan constituency?

The strength of these relationships vary

That setup violates likely violates the

whereas most Dime scores

Figures 4.4 and 4.5, DIME not Dime, meh

simulation-based statistics they are more stable estimators

All ideal point estimates are uncertainty