I expanded on my post of the other day to include first-order autocorrelation when calculating linear temperature trends.
Not surprisingly, the percentage of times the HadCRUT4 temperature series is significant, for any interval of N years, dropped significantly:
Even, for example, some 60-year linear trends aren't signficant at the 95% level -- the last being December 1877 to November 1937, when it did a U-turn about 1910. The statistical significance of that 60-year interval is 93.7%.
BTW, a useful and easy-to-read introduction to calculating trends and their uncertainities is Statistical Issues Regarding Trends by Tom Wigley.