A/B testing using chi-square to calculate the significance in an elegant way

The definition of ABtest

Goal:

For input data, each pair row is a group such as 0 and 1, 2 and 3. I want to test whether each group is significant using chi-square.

• If result is significant and any value in group is more than or equals 5 , it returns p-value, else it returns 'not significant'.
• If any value in group is less than 5, it returns 'sample size too small'.

My question is:

• My code seems so long and not so elegant and want to be more simple.

Notice: please don't use column grp as raw input to rewrite like groupby method because the raw data is not named a1 ca1, b1 cb1 . But I make sure each pair row (e.g. 0 and 1, 2 and 3) is a group.

Code:

import numpy as np
from scipy.stats import chi2_contingency
from itertools import chain, repeat

def abtest(df,tot,convert):
df = df.copy()
df['nonconvert'] = df[tot] - df[convert]
grp = np.split(df.index.values,df.index.size//2)
rst = []
for g in grp:
obs = df.loc[g][[convert,'nonconvert']].values
if (obs>=5).all():
_, p, _, _=chi2_contingency(obs)
if p<0.05:
rst.append(p)
else:
rst.append('not significant')
else:
rst.append('sample size too small')
rate = tot + '_' + convert + '_'+'test'
df[rate] = list(chain.from_iterable(zip(*repeat(rst, 2))))
del df['nonconvert']
return df

df = abtest(df=df,tot='tot',convert='lgn')


Input:

   grp    tot  lgn  read
0   a1   6300  360    30
1  ca1   2300   60     7
2   b1  26300  148     6
3  cb1  10501   15     3
4   c1  74600   36     2
5  cc1  26000    6     1


Output:

   grp    tot  lgn  read          tot_read_test     tot_lgn_test  \
0   a1   6300  360    30        not significant      4.68208e-09
1  ca1   2300   60     7        not significant      4.68208e-09
2   b1  26300  148     6  sample size too small      7.01275e-08
3  cb1  10501   15     3  sample size too small      7.01275e-08
4   c1  74600   36     2  sample size too small  not significant
5  cc1  26000    6     1  sample size too small  not significant

0        not significant
1        not significant
2  sample size too small
3  sample size too small
4  sample size too small
5  sample size too small


When you have a finite number of members in a group A and B. Instead of split into groups, hstack the DataFrame like this:

pd.concat(
[
], axis=1
)

    a_grp   a_tot   a_lgn   a_read  index   b_grp   b_tot   b_lgn   b_read
0   0   a1  6300    360 30  1   ca1 2300    60  7
1   2   b1  26300   148 6   3   cb1 10501   15  3
2   4   c1  74600   36  2   5   cc1 26000   6   1


Now you can replace the for-loop with 'a_' and 'b_' and df.apply() like

df.apply(apply_chi2)

def apply_chi2(df_ab):
if df_ab['a_'+convert] > df_ab['a_'+'nonconvert']:
return ...
obs = df_ab.a
return _, p, _, _=chi2_contingency(obs)


• read it carefully, the grp is not used Aug 29 '19 at 11:39
• sorry, but it doesn't seem more elegant than my code.
– Jack
Aug 29 '19 at 11:41

One minor thing that you can do is make your function reduce-friendly:

import numpy as np
from scipy.stats import chi2_contingency
from functools import reduce
from itertools import chain, repeat

def abtest(df, args):
tot, convert = args
df = df.copy()
df['nonconvert'] = df[tot] - df[convert]
grp = np.split(df.index.values,df.index.size//2)
rst = []

for g in grp:
obs = df.loc[g][[convert,'nonconvert']].values
if (obs>=5).all():
_, p, _, _=chi2_contingency(obs)
if p<0.05:
rst.append(p)
else:
rst.append('not significant')
else:
rst.append('sample size too small')

rate = tot + '_' + convert + '_'+'test'
df[rate] = list(chain.from_iterable(zip(*repeat(rst, 2))))
del df['nonconvert']
return df

df = reduce(abtest,
(