Proper Data Analysis Techniques to Reduce Experimental Error in Longitudinal Data

Abstract

This study presents findings of research conducted to improve analysis techniques of experimental data from coconut research. It highlights the ways of handling unaccountable variability due to the inconsistent temporal behavior of the experimental units in perennial crop research to obtain a precise research output. Properly designed field experiments are essential to identify the influence of independent variable/s on the dependent/s at the various stages. This document highlights the ways how the improved methodologies can be successfully used to reduce the experimental error in most commonly used experimental designs and types of analysis. The first example, the study on long term coconut fertilizer experiment designed as a randomized complete block design in Badalgama, Sri Lanka, compares different types of analyses via evaluating the model residuals and calculating the coefficient of variability (CV) to reduce the error and thereby improve the output. The statistical methods used in the first case study includes Repeated Measure Analysis of Variance (RMANOVA) as the classical method and Repeated Measure ANOVA (With single palm per plots), Linear Mix Model, and MANOVA with two Principal Components that represent approximately 80% variation of the data as dependent variables as improved methods. The model adequacy of each approach was accepted after testing normality, homogeneity of variance and independence of residuals. CV resulted from classical RMANOVA was 39.95%, while it was 39.2% from Repeated Measure ANOVA (With single palm per plots) and 16.51% from the Linear Mixed model. The lowermost CV (10.04%) resulted from MANOVA with two principal components indicating that it can be more powerfully used to analyze long term experiments of coconuts. The second example, the study on long term coconut fertilizer experiment from Bandirippuwa, Sri Lanka that failed to have normality assumption of parametric methods, illustrates appropriate types of the Non-Parametric analysis(F2-LD-F1) for the longitudinal data. The regularity of the results should be studied further with few more comparable data sets.