Many computational statistical problems (including some estimation problems, such as fitting GLMs) cannot be solved in a single calculation step - their algorithms must be iterated repeatedly until convergence. Such algorithms need starting-values at the beginning of the algorithm.

Many computational statistical problems cannot be solved in a single calculation step - their algorithms must be iterated repeatedly until convergence.

Such algorithms need starting-values at the beginning of the algorithm (and some require good estimates, ones not too far from the converged values).

Examples of algorithms needing start-values include generalized linear models, nonlinear least squares algorithms, the EM algorithm and Markov-chain Monte Carlo algorithms.

Reference: Wikipedia - Guess value