Evaluation applications of regression analysis with time-series data

Abstract

Time-series analysis can be a powerful evaluation tool whose use was suggested as early as 1963 by Campbell and Stanley. A number of authors have used time-series analysis for evaluation purposes. For example, Gillings, Makuc, and Siegel (1981) used time-series analysis to assess the effect of regionalized perinatal care. Veney and Luckey (1983) used time-series analysis to evaluate the effect of changes in highway speed limit laws on highway accident deaths. Deutsch and Alt (1977) and Zimring (1975) used time-series analysis to assess the effect of gun control. Guerin and MacKinnon (1985) used it to examine child seat belt legislation. Valuable methodological contributions to the evaluation uses of timeseries analysis have been made by Mandell (1987) and by Sawyer (1987). Despite this work, which spans a period of nearly twenty years, time series analysis still receives little attention in the evaluation practice literature. The purpose of this paper is to describe the application of time-series analysis, and particularly the use of regression analysis for analyzing time series in a way that may make it more readily accessible to an evaluation practice audience. In particular, this paper will provide practical guidelines for applying regression analysis to time-series data of a type that may be readily available to decision makers in government agencies, health and social service agencies, ministries of health or social welfare in developing countries, or to other organizations that wish to better understand and evaluate program actions. The specific example used in this paper is drawn from the health sector and is an example from Sri Lanka concerned with tetanus toxoid immunization levels for expectant mothers and neonatal tetanus deaths. However, the techniques examined are broadly applicable to a variety of time-series analysis situations.’ The analysis methods discussed include linear and non-linear analysis with time as the independent variable, linear analysis with the level of tetanus toxoid immunizations as the independent variable, and a piecewise regression example that examines the interaction between time and both the start-up of a government immunization program and immunization levels. The paper also deals with problems of serial correlation of errors in time series and concludes with a discussion of interpretation and appropriate use of regression analysis in time-series situations.