So how’d we do? As I wrote before, Senate poll medians did well, with Nevada as one glaring exception, leading to a 1-seat deviation from the 95% CI. There’s been talk about Hispanic voter turnout in Nevada, but I believe that’s not supported by exit polling data. It’s more likely that undecideds simply made a last-minute decision to stay with Reid.

In the House, we were off by about 13 seats, much more than the nominal error, +/-2 seats. Therefore there’s a systematic error. One possibility is bad time sampling – at the district level, polls are more spaced in time. However, we did just about as well as the famous guy, Nate Silver. In my case the tools were the Pollster database and a sheet of paper. By including national polls one could do better, not to mention other sources. Using these, a better topline (total # of seats) estimate than either of us came from Stochastic Democracy.

If you like district-level predictions, go read Stochastic Democracy. David Shor’s model that is theoretically fairly solidly based, enough to give true probabilities (for instance, House takeover 98%). His r^2 value for exact district-level vote share was much better than FiveThirtyEight. He needs to write all that up!

I have more to say about the subject of “snapshots,” “forecasts,” how they’re fundamentally the same thing, and what they’re good for. I have had interesting conversations with David (who is visiting Princeton this year) as well as correspondence with Andrew Gelman and Nate Silver. I am digesting it all, and will comment in a few days.

The Senate medians got the outcome of three races correct (WA/IL/WV) but not the closest race, CO. But then there’s this:

This error is so statistically glaring. I am not sure what to think about it. More discussion of this, and the House result, later.

In addition to random variation, Democrats seem to have a hidden advantage relative to national House preference polls, which are averaging R+7%. But district-level polls (which indicate 230R-205D) are said to be equivalent to R+4%. Can both be right? District polls tend to be concentrated in key districts. If the Republicans’ final House seat advantage matches district polls (a majority of 25 seats), it’s an underperformance relative to national surveys. Lucky for the Democrats – though not lucky enough. Maybe they have a better machine / strategy. Or maybe it’s systematic error when the question is asked generically.

Based on the excellent performance of direct polls in 2004/2006/2008, the hypothesis to beat is listed above. If the House is outside [228, 232] R or the Senate is outside [50, 52] D, then I’ll want to know why.

In the House: The Republicans will take control. By the same methods as before, we’re at 230R-205D. The 95% confidence interval for a gain of 50-54 seats to 228-232R. Pollster currently has 27 “toss-up” races with a total of over 100 polls, an amazing number.

P.S. For the record, as of 10PM tonight, two model-intensive comparables are FiveThirtyEight Senate 51.7D, House 233.1R, and StochasticDemocracy Senate 52.2D, House 237.7R. Both House results are outside my 95% confidence interval. I am interested in how far my simple-minded approach can go toward top-line characterization. But who knows, maybe there’s unseen complexity.

*This goes down to 90% if there is a 50-50 split and Lieberman jumps ship again. 100% if he stays with the Democrats this time.

I should express the idea of the previous post more carefully. Let us pose a hypothesis: Pollsters sample voters with no average bias. Their errors are small enough that in large numbers, their accuracy approaches perfect sampling of real voting. This hypothesis held up well in 2004, 2006, and 2008.

The 2010 test occurs on Tuesday. Polls predict an outcome (95% confidence) of a House Republican majority of 227-233 seats. For the Senate Democrats it’s 50-52 seats (assuming Lieberman continues to caucus with them). If either outcome falls outside this range, it’s time to re-examine the hypothesis.

As I wrote in 2008, despite the fuss about fancy modeling, for topline estimates, models such as FiveThirtyEight essentially add noise – needless uncertainty. On average, pollsters know what they are doing. This can focus your efforts: donations, critical Senate GOTV (indicated in yellow), and critical House GOTV.

Historically, a simple meta-analysis of existing polls gives the most accurate predictions. In 2008, this worked beautifully, when my pre-election estimates were off by only 1 EV for President, 0 seats for House (an exact match!), and 1 seat in the Senate. So what about 2010? First off: a Republican-controlled House and a Democratic-controlled Senate are essentially certain. Here’s why.

Based on the easily inspected database at Pollster.com, it is possible to assign a binary outcome in each of 30 swing districts based on who’s ahead. This gives a split of 18R-12D. Added to the noncompetitive races which give 212R-193D, the total is House 230R-205D.

For the Senate, the number of races is small enough that it’s necessary to be probabilistic, at least in part. Democrats will control at least 49 seats (47 Dems + Lieberman-CT + Sanders-VT). Of the five borderline races (WA, NV, CO, IL, and WV), Democrats lead in two (WA, WV), a Republican leads in one (NV), and two are within 1 percentage point (CO, IL). Splitting the last two gives Senate 52D-48D. Note that I am not making specific predictions on these five races – my meta-analytic goal is to estimate the total number of seats, not individual outcomes.

So what’s the action item? If you’re a committed activist, I suggest you 1) give to ActBlue, 2) Get Out The Vote (GOTV) in the borderline states for Senate races (indicated in yellow), or 3) GOTV in swing House districts.

What should we make of the less-certain projections by Nate Silver, Pollster.com, and others? They give probabilities around 80% for one party or the other controlling each chamber? That’s the uncertainty they add by making their assumptions. My probability is a nominal one: conditional on the idea that pollsters are on average unbiased, what do the polls tell us? Given that assumption, which worked so well in 2004, 2006, and 2008, the prediction is clear.

For the true geeks only: What is the nominal uncertainty on these estimates? “Nominal” means: what’s the minimum uncertainty assuming that polls are all perfect samples of likely voters? Here is one argument. In the House, there are about 100 polls for 30 races. Imagine that Democrats lead in D of them. Polls give an estimate of d=D/30=0.4 and for Republicans, r=0.6. 100 trials sampling that probability should give an uncertainty in number of seats equal to 30*sqrt(d*r/100). So: 30*sqrt(0.4*0.6/100)=1.5 seats. So the 68% confidence interval is R = [228.5,231.5] seats. For the Senate the uncertainty is about 1 seat: a 68% CI of D = [51,53] seats.

Efficient resource allocation is directed at the idea that you want the maximum benefit for a fixed amount of resources. In this case, if you have 50 hours (or $50) to spend, where should it go? As longtime readers of this site know, this is the point of the jerseyvotes calculation. Based on this reasoning, in the case of Princeton’s local Congressional district (NJ-12), perhaps the most effective GOTV action is to go to Pennsylvania for the Senate race. Sestak (D) and Toomey (R) are neck-and-neck.

The choice is not between voting and not voting. The point is that you’re probably reading this site because you care about politics, and want to maximize the return for your time or money spent. This is especially important if you are worried, since too much emotion can cloud judgment.

If you’re a Democrat, give at ActBlue. Republicans are welcome to leave suggestions for donation targets in comments.

My recommendations are based on the jerseyvotes principle (for explanations see here, here, and here).

The power of your donation increases steeply as a function of 1/(M*P), where M is the current polling margin and P is the population of the state or district. The following states have M less than 2.0%: Illinois (D+0.1%), West Virginia (D+0.9%), Nevada (D+1.2%), Colorado (R+0.9%), and Pennsylvania (R+1.2%). See state populations here.

Polling analysis is harder this year because results show more systematic (as opposed to random sampling-based) variability than in 2008. A notable cause is new variation in methods among pollsters. Some pollsters are starting to incorporate cell phones in their surveys. Others are not because of the expense of manual dialing (federal law prohibits autodialing). The principle of averaging multiple polls is more important than ever for getting an accurate picture – but also harder.

Many of you have written me asking where campaign donations are most efficiently given. As I have written in the past, donations are most effective in the closest situations. In other words, don’t give to a race where your favored candidate is 10 points behind – or 10 points ahead. For example, Rep. Rush Holt (D-NJ12) is likely to win by 15 points. Delaware Senate candidate Coons (D) is likely to defeat Christine O’Donnell (R) by over 20 points.

A second factor is population of the district or state. States with large populations are harder to sway. I call this the jerseyvotes principle: in New Jersey, my vote is not influential because my district and state are overwhelmingly Democratic – the probability of affecting the outcome is vanishingly small.

Note that I am not recommending donations to House races. At this point the House is nearly certain to switch to Republican control. Effectively, P is very large. If you care about the exact vote count, you can still give to an organization like the DCCC (or its Republican counterpart). But I wouldn’t bother.