Change mean imputation in MICE package MICE Steps
The chained equation process can be broken down into six general steps:
Step 1: A simple imputation, such as imputing the mean, is performed for every missing
value in the dataset. These mean imputations can be thought of as “place holders.”
Step 2: The “place holder” mean imputations for one variable (“var”) are set back to missing.
Step 3: The observed values from the variable “var” in Step 2 are regressed on the other variables in the imputation model, which may or may not consist of all of the variables in the dataset. In other words, “var” is the dependent variable in a regression model and all the other variables are independent variables in the regression model. These regression models operate under the same assumptions that one would make when performing (e.g.,) linear, logistic, or Poison regression models outside of the context of imputing missing data.
Step 4: The missing values for “var” are then replaced with predictions (imputations) from the regression model. When “var” is subsequently used as an  independent variable in the regression models for other variables, both the observed and these imputed values will be used.
Step 5: Steps 2–4 are then repeated for each variable that has missing data. The cycling through each of the variables constitutes one iteration or “cycle.” At the end of one cycle all of the missing values have been replaced with predictions from regressions that reflect the relationships observed in the data.
Step 6: Steps 2 through 4 are repeated for a number of cycles, with the imputations being updated at each cycle. The number of cycles to be performed can be specified by the researcher. At the end of these cycles the final imputations are retained, resulting in one imputed dataset. Generally, ten cycles are performed (Raghunathan et al., 2002); however, research is needed to identify the optimal number of cycles when imputing data under different conditions. The idea is that by the end of the cycles the distribution
of the parameters governing the imputations (e.g., the coefficients in the regression models) should have converged in the sense of becoming stable. This will, for example, avoid dependence on the order in which the variables are imputed. In practice, researchers can check the convergence by, for example, comparing the regression models at subsequent cycles, as discussed in He et al. (2009). Different MICE software packages vary somewhat in their exact implementation of this algorithm (e.g., in the order in which the variables are imputed), but the general strategy is the same.
I am interesting to know why mice use mean imputation (step 1) as first imputation , If I change mean imputation from step(1) to another imputations like (classification, neural-network , ... or another method for imputation) is that affect the accuracy of MICE imputation to higher accuracy.
 A: The aregImpute function in the R Hmisc package implements a mice-like algorithm but relaxes linearity assumptions.  It starts things off with the median (mode for categorical variables).  I don't think the result is very sensitive to this but more research is needed about the affect of the order of presentation of variables and about what happens when many variables are missing on the same subject for many subjects.
There is some guidance in the literature about how many iterations should be done, with White et al suggesting $100f$ where $f$ is the fraction of incomplete observations.  See these references:


*

*http://www.citeulike.org/user/harrelfe/article/13265871

*http://www.citeulike.org/user/harrelfe/article/13265938

*http://www.citeulike.org/user/harrelfe/article/13265633
A: The values in step 1 are only a starting point.
One of them actually is never used!
So most likely, the method is not very sensitive to these starting points. Even less if you plug in different advanced methods that may actually yielf very similar results.
It may be worth trying with random imputation in step 1. Instead of the mean, any random value from the same column is used.
Why don't you implement this variant of MICE, and let it run 1000 times. Then compare these 1000 results to the "mean" based standard MICE result. I'd expect a few runs to be better, a few more to be worse, but 95% of the runs will yield very similar quality to the mean-based MICE. Actually, there must even be some research paper that already tried this...
