def chisquare(matrix):
sum_all = float(sum([sum(x) for x in matrix]))
def sum_col(i):
return sum(matrix[i])
def sum_row(i):
return sum([x[i] for x in matrix])
dim_x = len(matrix)
dim_y = len(matrix[0])
matrix_expected = []
for i in range(0, dim_x):
col = []
for j in range(0, dim_y):
element = sum_col(i)*sum_row(j)/sum_all
col.append(element)
matrix_expected.append(col)
matrix_chi = []
for i in range(0, dim_x):
col = []
for j in range(0, dim_y):
element = matrix[i][j]-matrix_expected[i][j]
element *= element
divide_by = matrix_expected[i][j] if matrix_expected[i][j]!=0 else 1
element /= divide_by
col.append(element)
matrix_chi.append(col)
chi = sum([sum(x) for x in matrix_chi])
return chi, stats.chisqprob(chi, (dim_x-1)*(dim_y-1))
def chisquare(matrix):
sum_all = float(sum([sum(x) for x in matrix]))
def sum_col(i):
return sum(matrix[i])
def sum_row(i):
return sum([x[i] for x in matrix])
dim_x = len(matrix)
dim_y = len(matrix[0])
matrix_expected = []
for i in range(0, dim_x):
col = []
for j in range(0, dim_y):
element = sum_col(i) * sum_row(j) / sum_all
col.append(element)
matrix_expected.append(col)
matrix_chi = []
for i in range(0, dim_x):
col = []
for j in range(0, dim_y):
element = matrix[i][j] - matrix_expected[i][j]
element *= element
divide_by = matrix_expected[i][
j] if matrix_expected[i][j] != 0 else 1
element /= divide_by
col.append(element)
matrix_chi.append(col)
chi = sum([sum(x) for x in matrix_chi])
return chi, stats.chisqprob(chi, (dim_x - 1) * (dim_y - 1))
def likelihoodRatioTest(rawString_n, rawString_d, df):
test_stat = -2 * (ln(rawString_n) - ln(rawString_d))
print round(test_stat, 2)
return stats.chisqprob(test_stat, df)
def calcLRT(LRT_FILE, results):
file = open(LRT_FILE, 'w')
for i in range(1, len(results)):
df_im1 = countFreeParameters(results[i-1])
df_i = countFreeParameters(results[i])
likelihood_im1 = results[i-1]['likelihood']
likelihood_i = results[i]['likelihood']
test_stat = -2 * (ln(likelihood_i) - ln(likelihood_im1))
delta_df = df_im1 - df_i
test_prob = stats.chisqprob(test_stat, delta_df)
# print entries for latex table
file.write(results[i]['model_name'] + " " + "&" + " " + "some description" + " " + "&" + " " + str(df_i) + " " + "&" + " " +
str(delta_df) + " " + "&" + " " + likelihood_i + " " + "&" + " " + str(round(test_stat, 2)) + " " + "&" + " " +
str(round(test_prob, 4)) + " " + "\\\\" + "\n")
print "LTR saved to file '" + LRT_FILE + "'"
data = [float(d) for d in open(argv[1], 'r')]
## compute unimodal model
uni = Gaussian(mean(data), std(data))
uni_loglike = sum(log(uni.pdf(d)) for d in data)
print 'Best singleton: {0}'.format(uni)
print 'Null LL: {0:4.6}'.format(uni_loglike)
## find best one
# set defaults
best_gaus = None
best_loglike = float('-inf')
stderr.write('Computing best model with random restarts...\n')
for i in xrange(_rand_restarts):
mix = GaussianMixture(data, _mu_min, _mu_max, _sigma_min, _sigma_max)
# I catch division errors from bad starts, and just throw them out...
for i in xrange(_n_iterations):
try:
mix.iterate()
if mix.loglike > best_loglike:
best_loglike = mix.loglike
best_gaus = mix
except (ZeroDivisionError, ValueError):
pass
print 'Best {0}'.format(best_gaus)
print 'Alternative LL: {0:4.6}'.format(best_gaus.loglike)
test_stat = -2 * uni_loglike + 2 * best_gaus.loglike
print 'Test statistic for LLR (Chi-sq, df=3): {0:4.6}'.format(test_stat)
print 'P = {0:4.6}'.format(chisqprob(test_stat, 3))