Forecasting epidemic

:
Time series modelling
Dr Cho-Min-Naing
Medical Officer (Malaria/DHF)
The National Vector Borne Diseases Control
Project, Yangon, Insein PO, Myanmar

Email

Learning objectives:
At the end of this session,
1. the participant should
understand forecasting methods.
2. the participant should know
concepts behind forecasting models.

Performance objectives:
1. the participant should be able
to develop times series model for
forecasting epidemic.

. explicit approaches to forecasting. This does mean explicit systematic forecasting approaching can provide substantial benefits when used properly as all types and forms of forecasting techniques are made available within the existing data. Background: Unaided. This does not mean there is error free forecasts. subjective judgements to warn of forthcoming events and changes are not as accurate and effective as systematic.I.

1.II. Quantitative method 3. Technological method . Judgmental method 2. Forecasting methods: There are three major categories as stated below.

Judgmental method: Forecasts are made as individual judgements or by committee agreement or decisions. .1.

Quantitative method: To know what will happen. There are three subcategories of this method 2. but not why something happens. .1 Times series methods: Seek to identify historical patterns (using time as a reference) and then forecast using a time-based extrapolation of those patterns.2.

. 3. and desired outcomes as a basic for forecasting.2 Analogy-based methods: using historical and other analogies to make forecasts 3. political or economic nature.1 Extrapolative methods: using historical patterns and relationships as a basic for forecasts 3. thereby influencing future events}. societal.3. goals.4 Normative-based methods: {using objectives.3 Expert-based methods: 3. Technological method: Address long-term issues of a technological.

. we need to iterate the importance of selecting the appropriate forecasting techniques. In this context. medium term (3 months to 2 years). and long term (2 years or more). Selection points for the appropriate forecasting techniques. When we have to concern with application of forecasting in our decision making. time horizon can be divided into short term (1 to 3 months) immediate term (less than 1 month).III. there are six points that play an important role in determining the requirements for an appropriate technique.Time horizon: Generally. 1. The exact length of time used to classify these four categories is subject to vary by organization and situation.

Control versus planning: In control. Selection points for the appropriate forecasting techniques. 3. it is generally assumed that the existing patterns will continue in the future. Thus. (cont.) 2. On the planning side. the major emphasis is on identifying those patterns and extrapolating them into future. a forecasting method in such situation should be able to recognize changes in basic patterns or relationships at an early stage.III. and vice versa. 4. the greater the need for an automated forecasting procedure. the greater the level of detail (and frequency) that is required. the more accurate the forecasts. Level of aggregate detail: In general. . management by except is the general procedure. Number of items: The larger the number of items involved (all other things being equal).

the more difficult to forecast. In changing circumstances. Existing planning procedure: The greater the competition (all other things being equal). Constancy: Forecasting a situation that is constant over time is very different from forecasting one that is in a state of flux. 6. Based on the outcomes of forecasting models.) 5. (cont. rather than all at once. what is needed is a method that can adopt continually to reflect the most recent results and the latest information. In the stable situation a quantitative forecasting method can be adopted and checked periodically to reconfirm its appropriateness. The change can be made in a stepwise manner. .III. there is built – in resistance to change in any organization. Selection points for the appropriate forecasting techniques.

the conflict is what is desired and what should be desired. the conflict is what is and what is desired. . In the animal.IV. Concepts behind the times series analysis: In man.

It explains only what will happen.Time series forecasting treats the system as black box and makes no attempt to discover the factors affecting its behavior. . but not why something happens. The general formula for the time series model is Actual = pattern + randomness The common goal in the application of forecasting techniques is to minimize these deviations or errors in the forecast. The errors are defined as the differences between the actual value and what was predicted.

V. The minimal start-up time 4. Accuracy especially for short-term forecasting Disadvantage: Not having sound statistical theory behind the method Times series model can basically be classified into two types. assuming that the data can be broken down into the various components and a forecast obtained for each component. The decomposition method We will selectively present the decomposition method. additive model and . The simplicity of the procedures 2. Ease for computational procedures 3. Advantage: 1.

Clt. 2. ˜t ) Y^ = forecast y f = function Tr = trend Sn = seasonal variation Cl = cycle ˜ = error t = the time period being examined (t = 1. .The forecast for Y in the year t is generally written as Y^t = f (Trt.… i ). Snt.

. The seasonal factor of multiplicative model is a proportion (ratio) to the trends. We assume that the data is the sum of the time series components. Multiplicative model 1. We assume that the data is the product of the various components. If trend. If the data do not contain one of the components (e. • . Yt = Trt + Snt + Clt +  2. Suppose there is no cycle. or cycle is missing. Yt = Trt * Snt * Clt * ˜t 2. and thus magnitude of the seasonal swing is constant over time. The seasonal component is independent of trend.Additive model 1.g. seasonal variation. the the value is assumed to be 1. cycle) the value for that missing component is zero. then Yt = Trt + Snt +  t 3. then Yt = Trt * Snt * t 3. Suppose there is no cycle. and thus the magnitude of the seasonal swing increases or decreases according to the behaviour of trend.

Using the multiplicative decomposition method. . Case study: Quarterly malaria cases of a Township in Myanmar between 1984-1992 is shown in Table 1. a) calculate the centered moving average for the time series data.VI.

.Objectives of modelling : 1) to monitor the malaria situation in the study area and forecast with modelling.

we demonstrated a simple.Methods: 1. 3. This is a documentary study using time series data covering 1984 to 1992. 2. two-variable regression model using the . The dependent variable was the incidence of malaria occurring during a given time including both out-patient and in-patient malaria cases. For a starting point.

76732 Accuracy of Model MAPE: 494 MAD: 234 MSD: 101789 .309150 3 0. Times series (multiplicative decomposition method) Seasonal Indices Period Index 1 1.18483 2 0.738706 4 1.Results: The output for MINITAB program illustrating seasonal indices and centred moving average.

1 89899.Fig 1. The multiplicative decomposition method: Actual versus forecast values for malaria cases.4 206. 1984-92 1500 Actual Predicted Forecast Actual Predicted Forecast Cases 1000 500 MAPE: MAD: MSD: 0 0 10 20 30 quarterly time periods 40 249.9 .

Moving average model 1500 Actual Predicted Actual Predicted cases 1000 Moving Average 500 0 0 10 20 quarterly time periods 30 Length: 4 MAPE: 499 MAD: 256 MSD: 132728 .Fig 2.

043 R-Sq = 11.1 StDev 110.7 Coef 21.9 X Predictor Constant X S = 324.209 2.10 0.9% Dependent variable: quarterly malaria cases Independent variable: time .5 T 0.Linear regression model (simple.5% R-Sq(adj) = 8. two-variable model) in MINITAB The regression equation is Y = 21 + 10.850 10.932 5.19 P 0.

FIG 2 The linear regression model for malaria cases. 1984-92 Y = 21.9322X R-Sq = 0.115 1500 1000 se sa C 500 0 Regression 95% CI -500 95% PI 0 10 20 quarterlytime periods 30 40 .0603 + 10.

diagnostic tests taking account of statistical pathology are to be investigated. Graphing the actual values with the predicted values: closeness of the differences? [see Figure 1]Graphing the residual: .Evaluating the model: Before completing the analysis.

Discussion: 1. reaching levels above local average endemicity. 1.Epidemic of malaria: What? 1.1 Periodical rapid and great increase in malaria morbidity and perhaps mortality.2 A rapid increase in malaria morbidity and mortality in a given population (independent of .

Among diverse factors. It is worth emphasizing that the simple. 2.Points to ponder: 1. the selection of independent variables should be judiciously based on theoretical considerations. The preferred approach is to perform . twovariable regression model is limited in information. 3.

[Bradley.1982].A cautionary note: For real progress. . the mathematical modeller as well as the epidemiologist must have mud on his boots.

Berry G. 321-323. Armitage P. 1994. Oxford. 3rd ed. . In:Statistical Methods in Medical Research. Automatic selection procedures and colinearity. Blackwell Scientific Publications.References: 1.

In: Basic Econometric. 3rd ed. 227. Inc.References (cont): 4. In:Business Analysis Using Regression: A Case Book. Gujarati DN. Test of specification errors. Mcgraw-Hill. Inc. Foster DP. 1998. Summary regression case. Singapore. Springer-Verlag New York. Waterman RP. 5. . Stine RA.