From Surveys to Surveillance Time Series Analysis Eric Holowaty, Sr. Scientist, Informatics Unit, CCO Michael Spinks, Sr. Res. Assoc., CCO. Under Construction !
Background l 1980s and 1990s - occasional surveys l precise estimates of rates, proportions, means, nos. affected l detecting important differences in estimates within survey l 2000s - surveillance systems l continuous, or at least frequent sampling l monitoring and assessing temporal patterns, incl. change point detection and sub-group differences over time
Background l What is desired periodicity of sampling? Depends on: l how rapidly variables actually change l how important it is to detect changes quickly l desired precision in describing temporal patterns, changes and differences
Definitions l Time series l sequence of data points, measured at successive times, and spaced apart at uniform time intervals l Time series analysis l methods and models that describe and explain temporal patterns, and forecast future patterns l Trend l long-term movement in an ordered series; may be temporal or just ordered strata
Typical Time Series for One Participating PHU
l Uncoordinated l Fragmented l Lack of smaller area data l Poorly analysed l Poor dissemination l Not timely l Difficult to access Risk Factor Surveillance in Ontario pre-RRFSS
l Pilot tested in Durham Region in 1999 l Available for Individual PHUs in Jan 2001 l 22 PHUs participating as of Dec 2004 l ?Province-wide coverage in 2005/06
RRFSS Population Coverage RRFSS (2003) respondents : 25,600 CCHS (2003) resp. : 37,000 87% of pop’n 22/37 PHUs
Ê Monthly data more suitable for detecting temporal changes Ë More flexibility re. aggregation - before / after comparisons; geographic areas; demographic groups Ì Seasonal effects can be better described and analysed Í (Robust SPC procedures permit timely detection of stat. signif. changes) Î LARGE sample size permits more precise analysis Ï Standard CORE of questions helps ensure comparability over time and with other geo. areas. Ð Flexible MODULES permit targetted sampling and invest. of local concerns Benefits of RRFSS
Fundamental Statistical Issues in Time Series Analysis l Accuracy and precision of estimates l precision ~ sample size and survey design l bias l differential access and response l reporting/measurement bias l changes in the measurement tool, incl. wording importance of bias in time series analysis depends on size and consistency l Statistical power l probability of detecting an important change in time series - slope; seasonality; change points
Estimating the rate of change over time l Estimating that slope differs from the null i.e., zero change l assumption of monotonic relationship e.g., linear or log-linear model or logistic l assumption of no change points in time series
Statistical Power to Detect Slope > Null l Power influenced by: l length of time series (k) l size of each sample (n) l measurement of interest (p or x or x) and its variance l alpha (Type I error) l underlying rate of change/slope (b)
Statistical Power of Trend Tests Sample Size From: MacNeill and Umphrey, 1997.
Statistical Power of Trend Tests Sample Size From: MacNeill and Umphrey, 1997.
Statistical Power of Trend Tests Sample Size From: MacNeill and Umphrey, 1997.
Monthly Estimates of ETS Exposure Trends - RRFSS GTA Aug01-Dec03
Quarterly Estimates of ETS Exposure Trends - RRFSS GTA Aug01-Dec03
Estimates of ETS Exposure Trends - RRFSS GTA Aug01-Dec03
Detecting abrupt changes in lengthy time series l Change-point methods e.g. JoinPoint l Control Charts l conventional p-charts l CUSUM charts l EWMA charts, with residuals
Monthly Estimates of Support for Bylaws - RRFSS GTA Jan02-Dec04
15/35 point estimates in violation of Western Electric rules
Monthly Estimates of Support for Bylaws - RRFSS GTA Jan02-Dec04
Plan l Complete analysis of definitions, incl. temporal consistency and CCHS consistency l Assign final sample weights l Production of point estimates for 2003 and 2004 l Age-standardized comparisons l Time series analysis