Data are stored in ASCII.

Temperatures are stored as degrees C

Land squares and missing data are set to -99.99

The month and year are stored at the start of each month.

Data Array (72x36)

Item ( 1, 1) stores the value for the 5-deg-area centred at 177.5W and 87.5N

Item (72, 36) stores the value for the 5-deg-area centred at 177.5E and 87.5S

----- ----- | | | | MON | YR | |_____|_____|__________________________________ 90N |(1,1) | | | | | | | | | |(1,18) | Equ | | |(1,19) | | | | | | | | | 90S |(1,36)_________________________________(72,36)| 180W 0 180E

The observations that make up this dataset are taken from the International Comprehensive Ocean-Atmosphere DataSet, ICOADS (see http://www.cdc.noaa.gov/coads/), until 1997 and from the NCEP GTS archive thereafter. Individual observations must first pass a series of quality checks (track check, reality check, positional check, climatology check, buddy check, duplicate check). The quality-checked observations in each 1degree longitude X 1degree latitude X pentad gridbox are then averaged using a winsorised average. The pentad climatology is then subtracted from these pentad superobs and the resulting anomalies are averaged to 5degree X 5degree X monthly resolution. The data are then bias-corrected for the use of buckets in the period 1850-1941.

Measurement and sampling error refers to the random error caused by estimating an area-averaged quantity from a finite number of noisy observations. These errors are calculated directly from the gridded data. Measurement and sampling errors are uncorrelated between gridboxes.

The bias bound represents the low (2.5 percent) error bound on the bias corrected data as calculated from 1000 physically plausible realisations of the bias corrected dataset. The bias errors are completely correlated between gridboxes.

The bias bound represents the high (97.5 percent) error bound on the bias corrected data as calculated from 1000 physically plausible realisations of the bias corrected dataset. The bias errors are completely correlated between gridboxes.

If you would like to create combined 2 sigma uncertainties at each gridbox, it is suggested that you subtract the "low_bias" field from the "high_bias"field, divide by two and then add to two times the sampling and measurement errors in quadrature, as they are independent uncertainties. This is because the bias uncertainties are 95% confidence intervals, but the sampling and measurement errors are 1 standard error. Note that, although the sampling and measurement errors are uncorrelated between grid boxes, the bias correction uncertainties are completely correlated between grid boxes. This will have implications for the way youcalculate average uncertainties for any averaged time series that you create, e.g. your Nino3.4 index.

Files showing the corrections applied to the data are:

- HadSST2_bucket_correction_median.txt - the corrections applied to the data 1850-1941.
- HadSST2_bucket_correction_2.5pc.txt - the lower bound of the 95% confidence range of the uncertainties 1850-1941.
- HadSST2_bucket_correction_97.5pc.txt - the upper bound of the 95% confidence range of the uncertainties 1850-1941.