The Equivalence of Backward Elimination and Multiple Comparisons
Authors:
- Jan F. BjornstadJBJan F. Bjornstad
- Ronald W. ButlerRBRonald W. Butler
- 1 March 1988
- journal article
- Published by JSTOR in Journal of the American Statistical Association
- Vol. 83 (401) , 136
Abstract
We study backward elimination (BE) and multiple comparisons (MC) in a general linear model. The usual BE method based on partial correlations is equivalent to the MC statistics for the corresponding regression coefficients. In addition, it is shown that, typically, the usual MC statistics for the general MC problem are equivalent to partial correlations for BE. To obtain P values and critical values in BE and MC one needs to compute probabilities for the minimum and maximum of absolute values of correlated t variables. We investigate algorithms that can efficiently compute upper and lower bounds based on second-order Bonferroni inequalities and are capable of handling a large number of variables and comparisons. To be specific, consider a linear normal model with full rank, expressed in the form y(n × 1) = X(n × p) · θ(p × 1) + Z(n × q) · δ(q × 1) + ε(n × 1). Two seemingly different problems are considered: (1) BE of the columns in Z, assuming that X has been forced into the model; specifically, one-step variable elimination where x1,...,xp, z1,...,zq are the variables remaining in the model at the current step. The problem is to test if one of the zi's is to be eliminated from the model or not. (2) MC of a finite number m of linear combinations of δ, c'iδ for i = 1,...,m. Here ci is a q × 1 known vector and the following multiple testing problem is considered: Hi: c'1δ = 0 versus Ai: c'1δ ≠ 0 for i = 1,...,m. The BE method is presented as a partial correlation method and related to the multiple t method for δ. Moreover, it is shown that if m ≤ q then the usual MC method is also a BE method. When $m > q$ the simultaneous MC procedure turns out to be (under a condition on c1,...,cm) equivalent to a certain type of restricted BE method, based on partial multiple correlations. One consequence of this is that many of the customary MC procedures are based on partial correlations. To obtain the significance level in BE it is necessary to compute $\Pr(\min_{1 \leqq i \leqq q} |t_i| \leqq c)$, where (t1,...,tq) has a multivariate t distribution with nonzero correlations. To achieve joint level in MC one needs to compute $\Pr(\max_{1 \leqq i \leqq m} |t_i| > c)$, where (t1,...,tm) also has a multivariate t distribution. As mentioned earlier, upper and lower bounds using second-order Bonferroni inequalities are found with an algorithm that can handle large values of m and q. Hence this algorithm enables us to calculate both liberal and conservative simultaneous confidence intervals in MC for a large number of linear functions in δ. The liberal intervals are then informative about how conservative the conservative intervals are and vice versa.