The type and detail of information in CISDA, particularly floe size and stage of development, far exceeds what is attainable from a single satellite source. However, because the ice charts are an integration of ice information from a variety of sources, the greatest shortcoming of the data set with respect to statistical analysis is that errors in estimates are nonuniform in both space and time. In Appendix A of this paper the early government reports [CIS, 2007a, 2007b; Crocker and Carrieres, 2000a, 2000b] that describe the database are synthesized; the main sources of uncertainty in the database are reviewed and discussed along with estimates of their contribution to the overall error in ice position, ice concentration and stage of development. In this section, the rationale behind the choice of sea ice parameter and time period used for the statistical analysis in this study is first presented. Next, the time series are checked for evidence of systematic biases that could be introduced from known and suspected biases in CISDA (see Appendix A).
2.3.1. Sea Ice Parameter
 The choice of an ice parameter for statistical analysis is not trivial. The lack of continuity in the data set has the potential to degrade data quality. Temporal changes in any specific ice parameter could be due to changes in the quality of the input data or interpretative procedures and not to real changes in the state of the ice environment. The parameters used in this study are the average sea ice coverage for all ice types combined (all ice coverage (AIC)), multiyear ice (multiyear ice coverage (MYIC)), and first-year ice (first-year ice coverage (FYIC)) in each CISSIR region calculated using
Awater is the total sea surface area and Aice is the total sea ice area. Aice can be subdivided into FYI or MYI and N is any number of weeks that define a season of interest. The summer season is of interest and for the eastern and western Arctic (Figure 1) it is chosen as the 17 week period from 25 June to 15 October. For Hudson Bay (Figure 1) it is chosen as the 23 week period from 25 June to 19 November. These season definitions have been adopted by CIS because they maximize the use of the regional charts which are only created weekly in the Arctic during the summer shipping season [CIS, 2007a]. For any given region there is a single value for ice coverage that defines the sea ice severity during the summer shipping season.
 A seasonal parameter for the CISIRR overcomes random errors in the data set (see Appendix A) and minimizes the influence of gross errors in a single ice chart. However, the impact of spatial and temporal discontinuity in the data on statistical analysis, particularly trends, needs to be addressed. This is handled in two steps. First, a subjective assessment of data quality in each CISIRR region is used to determine the most appropriate data and time period for statistical analysis. Second, the potential impact of known and suspected systematic bias in the data set on statistical analysis is assessed.
2.3.2. Addressing Nonhomogeneity in Data Quality
 Quality Index (QI) scores (see Appendix A) were used to categorize and assess the data quality of the AIC and MYIC time series to determine an appropriate start year for statistical analysis. Data over a time period classified as “not considered for trends” or “not considered” are unreliable for trend analysis because observations were too scarce at the beginning of the record. A computed trend in this circumstance could be caused by a change in data quality over time, not real environmental change. QI scores are based on a detailed analysis of the AIC and MYIC time series in each region with particular attention to homogeneity in the time series. Regions in the fair, good and excellent categories are considered of high enough quality for any statistical analysis.
 The data quality in AIC/MYIC for each region derived from the CISDA regional ice charts over the 1968–2007 period is shown in Figure 5. In general, data quality is higher in the operational regions. The AIC/MYIC time series in the northern regions of the CAA and along the northern coast of the CAA are not reliable for trend analysis. If the regional charts are replaced by the historical charts for the overlapping period of 1968–1974, the data quality increases for these remote regions allowing for trend analysis (Figure 5). Figure 5 (bottom) shows the data quality in AIC only for the time series from the blended data sets taken back to 1960. It is apparent that trend analysis beginning in 1960 for AIC is only reliable along the main shipping routes. In summary, we have used a combination of the regional ice charts and the historical ice charts (1968–1974) in this study in order to have the best possible estimate of ice conditions and for the 1960–2008 period we have only calculated trends along the main shipping routes.
Figure 5. (top) Qualitative quality index scores in each subregion for AIC/MYIC for the regional ice charts, 1968–2008; (middle) AIC/MYIC for the blended historical and regional ice charts, 1968–2008; and (bottom) AIC for the blended historical and regional ice charts, 1960–2008.
Download figure to PowerPoint
2.3.3. Addressing Systematic Bias
 There are two known systematic bias in the CISDA (see Appendix A): (1) underestimation of low concentrations of ice occurring with all satellite imaging sensors and (2) overestimation of total and multiyear ice concentration in remote areas. The influence of systematic underestimation of low ice concentrations on AIC is explored by comparing AIC with AIC calculated by excluding regions with sea ice concentrations less than 2/10 (AIC > 2/10). Correlations between the two AIC parameters are all greater than 0.99 and trend analysis repeated using AIC > 2/10 revealed no detectable difference in the results. It is therefore assumed that the bias in estimates of low ice concentrations within the CISDA has no significant impact on interannual variability or trends in AIC. The bias in old ice associated with nowcasting in the early years in remote areas of the high Arctic is not formally addressed but it is noted and taken into consideration in the interpretation of the results.
 There are two likely sources of systematic bias in the CISDA (see Appendix A): (1) the switch to the egg code in 1983 and (2) the operational introduction of RADARSAT-1 in 1996. To test for potential shifts in CISIRR AIC/MYIC that may have been introduced by changing chart techniques (i.e., switch to egg code) and source data over time (i.e., RADARSAT), a comparison is made between CISDA and three other sea ice data sets: sea ice concentration estimates from Scanning Multichannel Microwave Radiometer (SMMR)-Special Sensor Microwave/Imager (SSM/I) passive microwave sensors using the NASA Team algorithm [Cavalieri et al., 2008], MYI concentration estimates from a neural network analysis of SSMR and SSMI/I data [Belchansky et al., 2004], and sea ice concentration from the Hadley Centre data set (HadISST2.1) [Walsh, 1978; Knight, 1984; Rayner et al., 2003]. The sea ice algorithm used to estimate ice concentration from the passive microwave data is sensor specific in order to homogenize data between sensors [Cavalieri et al., 2008]. The data is available as monthly averages on a 25 km degree grid from 1979 to present from the National Snow and Ice Data Center. Estimates of MYI concentration are also on a 25 km degree grid and are available as monthly averages for January, February, and March from the International Arctic Research Centre. The Hadley data for Arctic sea ice concentration is a compilation of the Walsh  data set, the National Ice Center (NIC) ice charts [Knight, 1984] and the passive microwave record [Cavalieri et al., 2008]; it is available as monthly averages on a 1° grid at the British Atmospheric Data Center. Although the Hadley record of sea ice concentration is a blended data set, considerable effort has been made to homogenize the data which includes a correction for surface melt effects on retrievals from satellite microwave-based estimates.
 For the comparison to be valid not only must the baseline data sets have minimal time-varying bias, but they must be independent of CISDA. The passive microwave data was rarely used in chart preparation due to the high error in ice concentration estimates (4/10; see Table A3); when it was used it was only used to delineate the ice edge in the absence of any other satellite or ground observations. Post-1979, the main data source in the Hadley data set is the passive microwave record [Rayner et al., 2003] and pre-1979 the main data sources are the NIC ice charts and the Walsh data set. Although the CIS ice charts are not a data source in either the NIC ice charts or the Walsh data set, they likely share common source data prior to 1979. For this reason the Hadley data set is only used to check for bias in the early 1980s when the passive microwave record is too short to be used.
 For comparison, the CISIRR subregions are grouped into four large regions: southern Beaufort Sea region (Figure 4, Beaufort Sea), Canadian Arctic Archipelago region (Figure 4, western Arctic waterway, western Parry Channel, M'Clintock Channel, Franklin, western high Arctic, eastern high Arctic, Baffin Inlets, Foxe Basin, and Kane Basin), Baffin Bay region (Figure 4, Baffin Bay), and Hudson Bay region (Figure 4, Hudson Bay, Hudson Strait, Davis Strait, and north Labrador Sea). AIC and MYIC is calculated from each data set for each of these four regions. For consistency, all data is first interpolated onto a common 1° grid and seasonally averaged. The summer shipping season in all data sets is chosen as the July-August-September (JAS) average for the three Arctic regions and the July-August-September-October (JASO) average for the Hudson Bay region.
 The results for AIC are shown in Figure 6. In general, there is good agreement in interannual variability between CISDA, the Hadley Centre data set (HAD) and the passive microwave record (SSMI) and for all regions correlation coefficients range between 0.9 and 0.95. The Hadley data set is only used in the Beaufort Sea because before 1979 observations were scarce in all other regions [Rayner et al., 2003]. The offset between CISDA and SSMI (Figure 6) is a result of the consistent underestimation of low ice concentrations derived from the passive microwave sensors and has been investigated by Agnew and Howell . The Rodionov  regime shift detector is used to test for significant shifts in the (CIS-SSMI)/SSMI and the (CIS-HAD)/HAD metrics. It is assumed that a significant shift in this metric would be due to a bias in the CISDA data that is not present in the passive microwave or Hadley data sets. The regime shift detector requires no a priori hypothesis about the timing of a shift and here it is run with a Hubert weight parameter of 2, no prewhitening and cutoffs ranging from 5 to 15 years. No significant shift was detected in the CAA, Baffin Bay or Hudson Bay regions. A significant shift was detected in 1998 in the Beaufort Sea region for (CIS-SSMI)/SSMI which comes shortly after the introduction of RADARSAT-1 in 1996; the mean percent difference before 1998 is ∼40% and after 1998 it increase to ∼70%. We believe that this shift is more likely due to an increase in the passive microwave bias in underestimating the presence of sea ice in regions with low ice concentrations or heavily decayed ice and less likely due to an increase in observable low ice concentration areas from RADARSAT-1. Of the four regions, the Beaufort Sea has experienced the greatest increase in the length of the summer melt season [Stroeve et al., 2006] while maintaining considerably high ice concentrations in summer due to the large fraction of MYI. As the length of the melt season increases, ice concentrations decrease and the fraction of ponded ice increases. In the Beaufort Sea, 1998 was the warmest year on record [Atkinson et al., 2006] and with the exception of 2008 it was also the lightest ice year on record. Further supporting this argument is the fact that no significant shift was detected in any other region. In particular, no shift was detected in Hudson Bay where low ice concentrations are prominent during breakup and where historically there was less direct observation of the ice compared to the Beaufort Sea. It is expected that if there is a bias in summer AIC due to the introduction of RADARSAT-1, it would be greatest in Hudson Bay. As a final check, the two most likely influential changes, the switch to the egg code in 1983 and the introduction of RADARSAT-1 in 1996, are tested using a simple difference of means test on the adjacent periods [e.g., Hare and Mantua, 2000]. The 1996 shift is tested using (CIS-HAD)/HAD and (CIS-SSMI)/SSMI; the 1983 shift is tested using (CIS-HAD)/HAD. The difference in mean was not significant to the 95% confidence interval for either period in the metrics tested. In conclusion, no evidence of a time-varying bias was detected in the AIC time series.
Figure 6. Summer all ice coverage, 1980–2005, for the Beaufort Sea, Canadian Arctic Archipelago, Baffin Bay, and Hudson Bay regions. The time series are generated from the CISDA (CIS), Hadley data set (HAD), and passive microwave record (SSMI).
Download figure to PowerPoint
 The results for MYI are shown in Figure 7. For MYI, the CISDA is compared against the Belchansky et al.  data set that contains monthly estimates of MYI ice concentration for January, February, and March. The comparison was made in all three months, but because the results were similar in each month only February is shown and discussed here. There is a strong agreement between the two MYIC time series (Figure 7) and the correlation is high (r = 0.75). The Rodionov regime shift detector algorithm was also applied to the (CIS-SSMI)/SSMI metric with no prewhitening and cutoff lengths ranging from 5 to 15. No significant shifts were detected at the 95% confidence level. A difference of means test was used to check for a significant shift in 1996 with the introduction of RADARAT-1 and again no significant change was detected at the 95% confidence level. As in the AIC time series, no evidence of a time-varying bias was detected in the MYIC time series.
Figure 7. February multiyear ice coverage, 1980–2005, for the Beaufort Sea, Canadian Arctic Archipelago, and Baffin Bay regions combined. The time series are generated from the CISDA (CIS) and the passive microwave record (SSMI).
Download figure to PowerPoint