A couple of late polls in the Senate race in Massachusetts indicate that Scott Brown continues to lead Martha Coakley and is positioned for a game-changing upset in the special election.  According to a Coakley campaign internal poll, not much has changed since the beginning of the week:

I’ve been told reliably that Martha Coakley’s internal poll for Friday night showed Republican Scott Brown leading by two points, 47 to 45 percent. Her campaign’s three-night average for Friday, Thursday and Wednesday is the same — a 47 to 45 lead for Brown.

This is, obviously, not great news for the Democratic nominee. But it does suggest that Brown’s momentum — which took him from a double-digit gap to the lead in Coakley’s poll in about a week’s time — has been arrested. On Wednesday night, Coakley’s poll put her ahead by two, 46 to 44. On Thursday night, Brown surged ahead by three, 48 to 45. And on Friday, it was back to a two-point race. In other words, a nail-biter on Tuesday looks likely.

Well, it depends on the margin of error, the sampling technique, and the timing.  Steve Kornacki notes in an update that this survey took place before Coakley’s jaw-dropping assertion that Curt Schilling was a closet Yankee fan, which may not have much to do with policy but has everything to do with arrogance and being in touch with one’s constituents.  Even without that, a one-point shift in a poll does not show much of anything except a little statistical noise.  It also doesn’t make much difference to “arrest” Brown’s momentum if Brown’s still ahead with only two days left in the race.

Meanwhile, American Research Group conducted its final poll of the race and reported Brown up by three:

Republican Scott Brown leads Democrat Martha Coakley 48% to 45% in the special Massachusetts US Senate race to replace Senator Ted Kennedy in a telephone survey conducted January 12-14 among 600 likely voters in Massachusetts saying they will definitely vote in the special election on January 19.

Brown leads Coakley 94% to 1% among registered Republicans and he leads 58% to 37% among unenrolled voters. Coakley leads Brown 71% to 20% among registered Democrats. A total of 8% of Democrats and 5% of Republicans remain undecided.

If the Democrats hope to have a better time with absentee voters who cast ballots before Coakley’s collapse, the ARG survey has bad news for them there, too:

A total of 9% of likely voters say they have already voted by absentee ballot, with Brown leading Coakley 58% to 42%.

ARG provides the sampling for their likely voter survey (which we don’t have from the Coakley campaign):

Dems 263 43.8%
GOP 122 20.3%
Indys 213 35.5%

This looks like an oversample for Democrats at the expense of Republicans — and a seriously incorrect model for a special-election turnout in this election.  Pollster.com looked at this very issue in September, when Massachusetts scheduled the special election.  They concluded that a poll of MA adults gave Democrats a +17, and likely voters a +19 for next November’s gubernatorial race.  This sample gives Democrats a +23.5, almost the same gap by which Barack Obama beat John McCain in a heavy turnout election (+26).  Given the lack of voter intensity by Democrats and the escalating intensity of the Republicans and anti-incumbent independents, the actual gap for a turnout model should be significantly lower — perhaps by as much as a third.

In short, this is probably more than a three-point lead in actuality by Brown — and I’d bet the same modeling problems are in Coakley’s internal polls as well.

Update: I’ve run the numbers through a couple of permutations using the ARG survey answers.  Instead of using the +23.5 Dems survey ARG has, I’ll use the numbers from Pollster.com’s likely-voter scenario and rerun the survey percentages:

Brown Coakley
Dems 300 60 213
GOP 192 180.48 1.92
Indys 108 62.64 39.96
303.12 254.88
50.52% 42.48%

In this case, Brown wins by almost eight points, and wins a majority of the vote before the undecideds make up their minds.  Now lets look at a sample that uses the 35% independent vote and gives the Dems a +15 turnout in the election:

Dems 240 48 170.4
GOP 150 141 1.5
Indys 210 121.8 77.7
310.8 249.6
51.80% 41.60%

Now we see a ten-point lead for Brown.  The turnout model used by ARG and probably Coakley is almost certainly far too optimistic about Democratic turnout, and the race probably has more separation than either poll suggests.

Update II: Reader S averaged the exit-polling data from the 2006 and 2008 races — in which Dems were highly motivated and independents tilted against the GOP — and came up with a 41-18-41 split.   In that best-case turnout model, Brown leads by 4.5 points using the ARG survey responses:

2006-8 Average Model Brown Coakley
Dems 246 49.2 174.66
GOP 108 101.52 1.08
Indys 246 142.68 91.02
293.4 266.76
48.90% 44.46%

That’s high-end best-case for Coakley, and she’s still losing.