Sources of data The secondary data on the production and acreage of wheat were collected from the book “A Data Base on Agriculture and Food grains in Bangladesh (1947-48 to 1989-90)” (Hamid M. A. (1991) and “Statistical Year Book of Bangladesh” published by the Bangladesh Bureau of Statistics (BBS, 1992). The annual data on wheat production in metric ton and in acres were collected for 1947-48 to 2003-2004. The wheat production rate in percentage (wpr) was used in the analysis and it is calculated by a simple formula: Rate = (Annual wheat production / Annual cultivated area) 100. Data from 1949 to 1990 were taken from the Hamid M. A. (1991) and data from 1991 to 2001 are taken from Statistical Year Book of Bangladesh. The production 13,913 metric ton in 1982 sources from Hamid M. A. (1991) was detected as outlier and that was replaced from the data 1,28,845 metric ton recorded in the Statistical Year Book of Bangladesh (BBS,1992) which was more than 9 times bigger than the earlier value. Again, the highly significant structural change was detected Gujarati, D.N. (1995) in the data from 1976 and a dummy variable (dm) was used for structural change as: dm= 0 for wheat production rate before 1976 and dm-1 for wheat production rate from 1976. The secondary data on climatic factors of Dinajpur district during 1948-2004 for the wheat growing period (November–March) are collected from the Bangladesh Meteorological Department, Dhaka, Bangladesh. But all the climatic data of Dinajpur were not available for 1973-1980 and these missing data were estimated by applying univariate Box-Jenkin’s ARIMA (autoregressive integrated moving average) modeling techniques Pankraiz (1991). The residual’s stationary and normality were checked. The outliers were checked in these data and the detected outliers were replaced by the estimated value using the same techniques. In this research, the used climatic variables were the average minimum temperature in celcius (tmn), the average maximum temperature in celcius (tmx), the average dry bulb temperature in celcius (td), the average wet bulb temperature in celcius (tw), the frequency of average dry bulb temperature which is greater than 200C (ftd), the total rainfall in millimeter (ttr), the average maximum rainfall in millimeter (mxr), the average frequency of insignificant rainfall which is less than 5mm (rf), the average relative humidity in percentage (hu), the average sea level pressure in millibar (slp), the average cloud in octas (ac), the average maximum wind speed in knots (wmx), the average wind speed in knots (wv) and the average difference of morning and the afternoon relative humidity in percentage {hu(0-12)}. Method A multiple regression model was fitted to examine for the rate of wheat production data on climatic variables during November-March over the years 1948-2004. The three regression models were estimated using three predictor sets of historical climatic data. One dummy variable was included with each predictor set as the significant structural change viewed in the response variable from 1976 what was tested according to Gujarati (1995). The predictor set 1 included 15 variables namely dm, tmn, tw, ftd, ttr, td, hu, mxr, tmx, slp, ac, wmx, wv, hu(0-12) and rf. The predictor set 2 included 11 variables namely dm, tmn, tw, ttr, td, hu, tmx, slp, ac, wmx and wv. The predictor set 3 contained 10 variables namely dm, tmn, tw, ftd, ttr, td, hu, tmx, slp and ac. Multicollinearity was checked in the regression models for selecting the climatic variables through investigating the range of variance inflation factors (VIF). The variables were selected by using the backward elimination procedure started with full equation and dropped one variable at a time against the smallest insignificant t values. This process was stopped for the minimum absolute t- test became greater than 1. Draper N.R and Smith H. (1981). Normality and stationarity for residuals were also checked for selecting the variables. But the residuals followed normality and first order auto-correlated structure. So it was tried to refit and reexamine the new regression coefficients taking the autocorrelation into account. The new regression models were estimated according to Cochrane and Orcutt (1949) iterative procedure, which satisfied the assumption of uncorrelated errors with the same procedure. To obtain robust models, outliers and hi-leverage points were identified applying some modern diagnostics tools namely deleted Studentized residuals and other residual based techniques. Chatterjee, S. and Hadi, A. S. (1988). Finally the three appropriate models were obtained and among the three obtained models, one model was selected having the highest F and R2 values.