view logistic_regression_vif.py @ 0:bd196d7c1ca9 draft default tip

Imported from capsule None

#!/usr/bin/env python

import sys
from rpy import *
import numpy

def stop_err(msg):
    sys.stderr.write(msg)
    sys.exit()

infile = sys.argv[1]
y_col = int(sys.argv[2])-1
x_cols = sys.argv[3].split(',')
outfile = sys.argv[4]


print "Predictor columns: %s; Response column: %d" % ( x_cols, y_col+1 )
fout = open(outfile,'w')
elems = []
for i, line in enumerate( file( infile ) ):
    line = line.rstrip('\r\n')
    if len( line )>0 and not line.startswith( '#' ):
        elems = line.split( '\t' )
        break
    if i == 30:
        break # Hopefully we'll never get here...

if len( elems )<1:
    stop_err( "The data in your input dataset is either missing or not formatted properly." )

y_vals = []
x_vals = []

for k, col in enumerate(x_cols):
    x_cols[k] = int(col)-1
    x_vals.append([])

NA = 'NA'
for ind, line in enumerate( file( infile )):
    if line and not line.startswith( '#' ):
        try:
            fields = line.split("\t")
            try:
                yval = float(fields[y_col])
            except:
                yval = r('NA')
            y_vals.append(yval)
            for k, col in enumerate(x_cols):
                try:
                    xval = float(fields[col])
                except:
                    xval = r('NA')
                x_vals[k].append(xval)
        except:
            pass

x_vals1 = numpy.asarray(x_vals).transpose()

check1 = 0
check0 = 0
for i in y_vals:
    if i == 1:
        check1 = 1
    if i == 0:
        check0 = 1
if check1 == 0 or check0 == 0:
    sys.exit("Warning: logistic regression must have at least two classes")

for i in y_vals:
    if i not in [1, 0, r('NA')]:
        print >> fout, str(i)
        sys.exit("Warning: the current version of this tool can run only with two classes and need to be labeled as 0 and 1.")

dat = r.list(x=array(x_vals1), y=y_vals)
novif = 0
set_default_mode(NO_CONVERSION)
try:
    linear_model = r.glm(r("y ~ x"), data=r.na_exclude(dat), family="binomial")
except RException, rex:
    stop_err("Error performing logistic regression on the input data.\nEither the response column or one of the predictor columns contain only non-numeric or invalid values.")
if len(x_cols)>1:
    try:
        r('suppressPackageStartupMessages(library(car))')
        r.assign('dat', dat)
        r.assign('ncols', len(x_cols))
        vif = r.vif(r('glm(dat$y ~ ., data = na.exclude(data.frame(as.matrix(dat$x,ncol=ncols))->datx), family="binomial")'))
    except RException, rex:
        print rex
else:
    novif = 1

set_default_mode(BASIC_CONVERSION)

coeffs = linear_model.as_py()['coefficients']
null_deviance = linear_model.as_py()['null.deviance']
residual_deviance = linear_model.as_py()['deviance']
yintercept = coeffs['(Intercept)']
summary = r.summary(linear_model)
co = summary.get('coefficients', 'NA')
"""
if len(co) != len(x_vals)+1:
    stop_err("Stopped performing logistic regression on the input data, since one of the predictor columns contains only non-numeric or invalid values.")
"""

try:
    yintercept = r.round(float(yintercept), digits=10)
    pvaly = r.round(float(co[0][3]), digits=10)
except:
    pass
print >> fout, "response column\tc%d" % (y_col+1)
tempP = []
for i in x_cols:
    tempP.append('c'+str(i+1))
tempP = ','.join(tempP)
print >> fout, "predictor column(s)\t%s" % (tempP)
print >> fout, "Y-intercept\t%s" % (yintercept)
print >> fout, "p-value (Y-intercept)\t%s" % (pvaly)

if len(x_vals) == 1:    #Simple linear  regression case with 1 predictor variable
    try:
        slope = r.round(float(coeffs['x']), digits=10)
    except:
        slope = 'NA'
    try:
        pval = r.round(float(co[1][3]), digits=10)
    except:
        pval = 'NA'
    print >> fout, "Slope (c%d)\t%s" % ( x_cols[0]+1, slope )
    print >> fout, "p-value (c%d)\t%s" % ( x_cols[0]+1, pval )
else:    #Multiple regression case with >1 predictors
    ind = 1
    while ind < len(coeffs.keys()):
        try:
            slope = r.round(float(coeffs['x'+str(ind)]), digits=10)
        except:
            slope = 'NA'
        print >> fout, "Slope (c%d)\t%s" % ( x_cols[ind-1]+1, slope )
        try:
            pval = r.round(float(co[ind][3]), digits=10)
        except:
            pval = 'NA'
        print >> fout, "p-value (c%d)\t%s" % ( x_cols[ind-1]+1, pval )
        ind += 1

rsq = summary.get('r.squared','NA')

try:
    rsq = r.round(float((null_deviance-residual_deviance)/null_deviance), digits=5)
    null_deviance = r.round(float(null_deviance), digits=5)
    residual_deviance = r.round(float(residual_deviance), digits=5)
except:
    pass

print >> fout, "Null deviance\t%s" % (null_deviance)
print >> fout, "Residual deviance\t%s" % (residual_deviance)
print >> fout, "pseudo R-squared\t%s" % (rsq)
print >> fout, "\n"
print >> fout, 'vif'

if novif == 0:
    py_vif = vif.as_py()
    count = 0
    for i in sorted(py_vif.keys()):
        print >> fout, 'c'+str(x_cols[count]+1), str(py_vif[i])
        count += 1
elif novif == 1:
    print >> fout, "vif can calculate only when model have more than 1 predictor"