His secret? The ability to produce a high percentage of quality starts, high quality starts, and elite starts. But more on that later. First let's look at how the Mets and Bartolo have fared in his starts and compare them to our three hypothetical pitchers: Mr. Fantastic, Mr. Reliable, and Mr. Meh. Each of these three pitchers share the talent of pitching the same exact number of innings as Bartolo in the exact same game situations as Bart did - amount of direct run support (runs that count towards a decision, aka those that were on the board when he left the game or scored in the half inning following his exit as long as he finished his half inning), total run support, unearned runs surrendered, and bullpen runs surrendered. These three pitchers also ALWAYS pitch consistently to their ERAs, giving up the number of runs in each outing that would most closely match their ERA rounded to the nearest run. Mr. Fantastic is a 3 ERA pitcher, Mr. Reliable is a 3.5 ERA pitcher, and Mr. Meh is a 4 ERA pitcher. Here's how they stack up when substituted into Bartolo's starts for the Mets:

Bartolo's actual stats: 45 games, 4.31 ERA, team 26-19 record, pitcher 24-18 record, team 57.8 win%
Mr. Fantastic's stats: 45 games, 2.97 ERA, team 25-17 record (3 undetermined outcomes, extra innings), pitcher 22-11 record, team 59.5 win%
Mr. Reliable's stats: 45 games, 3.53 ERA, team 23-17 record (5 undetermined outcomes, extra innings), pitcher 17-13 record, team 57.5 win%
Mr. Meh's stats: 45 games, 4.06 ERA, team 22-19 record (4 undetermined outcomes, extra innings), pitcher 16-15 record, team 53.7 win%

THE DATA

As we can see, the Mets would have won games at the same rate with Bartolo Colon on the mound with his 4.31 ERA as with Mr. Reliable on the mound with his 3.53 ERA. This reveals an interesting insight into successful pitching in the major leagues (if we define success for a starting pitcher as helping your team win games): If you're going to lose it doesn't matter how big you lose, but the best way to win is to rack up starts where you're good/great even if you're terrible some other times. Let's take a look at how Bartolo has succeeded in this regard:

Quality starts - 6.0 IP with 3 or fewer ER: 29 of 45 (64.4%).
Team record in those starts: 23-6

High quality starts - 7.0 IP with 2 or fewer ER: 17 of 45 (37.8%)
Team record in those starts: 15-2

Elite starts - 8.0 IP with 1 or fewer ER: 3 of 45 (6.7%)
Team record in those starts: 3-0

Here we see the key to how a binary pitcher like Bartolo Colon can provide value that far exceeds what his ERA suggests. In the 26 games that the Mets have won that he's started he provided a quality start in 23 of those. In games that he didn't provide a quality start it really didn't matter how badly he blew up as long as he kept his season percentage of quality starts high. His performance in non-quality starts was mostly irrelevant, since once he blew up the game was essentially over - the team went 3-13 in his non-quality starts. Pitchers who consistently pitch to their ERA may blow up less often, but their consistency may actually be their worst enemy since the example of Bartolo Colon seems to suggest that a high percentage of quality/high quality/elite starts allows a pitcher to generate team wins at a rate that exceeds the expected rate for their ERA. It seems that for a starting pitcher perhaps it is better to be sometimes good and sometimes terrible than to be always average.

Some clarification on the data set:
The score for each game is determined by earned runs for pitcher unearned runs for pitcher bullpen runs. Unearned runs for pitcher and bullpen runs stay the same regardless of which pitcher is being compared, only the earned runs change.

The earned runs for each pitcher in each start is calculated as (Bart IP/9)*(Pitcher ERA) rounded to the nearest run.

Decisions are determined by the number of runs given up by the pitcher at the time of exit (ER unearned runs) compared to direct run support, and then converted to W/L if matching game outcome or ND if pitcher left in a tie game or lead changed.

Simulated games that ended in a tie are not counted in the W/L% or decisions as their outcome is undetermined (would have gone to extra innings).