|
1 |
|
2 ============ |
|
3 Statistics |
|
4 ============ |
|
5 .. contents:: |
|
6 :local: |
|
7 |
|
8 .. role:: def |
|
9 :class: def |
|
10 |
|
11 Markov inequality |
|
12 ================= |
|
13 |
|
14 :def:`Markov inequality`: :math:`P(X ≥ a) ≤ E[P]/a` for all :math:`a > 0`. |
|
15 |
|
16 Chebyshev inequality |
|
17 ==================== |
|
18 |
|
19 :def:`Chebyshev inequality`: if :math:`X` is a random variable with mean |
|
20 :math:`μ` and variance :math:`σ²` then |
|
21 |
|
22 .. math:: |
|
23 |
|
24 P(|X-μ| ≥ c) ≤ σ²/c² |
|
25 |
|
26 for all :math:`c > 0`. |
|
27 |
|
28 Central limit theorem |
|
29 ===================== |
|
30 |
|
31 :def:`Central limit theorem`: let :math:`X_1, ..., X_n, ...` be a sequence of |
|
32 independent identically distributed random variables with common mean :math:`μ` |
|
33 and variance :math:`σ²` and let: |
|
34 |
|
35 .. math:: |
|
36 |
|
37 Z_n = ((∑_{1≤i≤n} X_i) - n·μ) / (σ·sqrt(n)) |
|
38 |
|
39 Then CDF of :math:`Z_n` converge to standard normal CDF: |
|
40 |
|
41 .. math:: |
|
42 |
|
43 Φ(z) = 1/(2·π)·∫_{(-∞;z]} exp(-x²/2) 𝑑x |
|
44 |
|
45 lim_{n → ∞} P(Z_n ≤ z) = Φ(z) |
|
46 |
|
47 Null hypothesis |
|
48 =============== |
|
49 |
|
50 :def:`Null hypothesis` a statement that the phenomenon being studied produces no |
|
51 effect or makes no difference, assumption that effect actually due to chance. |
|
52 |
|
53 p-value |
|
54 ======= |
|
55 |
|
56 :def:`p-value` is the probability of the apparent effect under the null |
|
57 hypothesis. |
|
58 |
|
59 https://en.wikipedia.org/wiki/P-value |
|
60 Wikipedia page |
|
61 |
|
62 Significance level |
|
63 ================== |
|
64 |
|
65 If the p-value is less than or equal to the chosen :def:`significance level` |
|
66 (:math:`α`), the test suggests that the observed data are inconsistent with the |
|
67 null hypothesis, so the null hypothesis should be rejected. |
|
68 |
|
69 Hypothesis testing |
|
70 ================== |
|
71 |
|
72 :def:`Hypothesis testing` is process of interpretation of statistical |
|
73 significance of given null hypothesis based on observed p-value from sample with |
|
74 choosen significance level. |
|
75 |
|
76 After finishing hypothesis testing we should reject null hypothesis or fail to |
|
77 reject due to lack of enough evidence or ... |
|
78 |
|
79 Hypothesis testing only take into account: |
|
80 |
|
81 * that effect might be due to chance; that is, the difference might appear in a |
|
82 random sample, but not in the general population |
|
83 |
|
84 But it doesn't cover cases: |
|
85 |
|
86 * The effect might be real; that is, a similar difference would be seen in the |
|
87 general population. |
|
88 * The apparent effect might be due to a biased sampling process, so it would not |
|
89 appear in the general population. |
|
90 * The apparent effect might be due to measurement errors. |
|
91 |
|
92 Asymptotic approximation |
|
93 ======================== |
|
94 |
|
95 CLT say that sample mean distribution is approximated by normal distribution. |
|
96 |
|
97 With fair enough number of samples approximation is quite good. |
|
98 |
|
99 So during hypothesis testing usually researcher makes assumption that is is safe |
|
100 to replace unknown distribution of means for independent and identicaly |
|
101 distributed individual samples with approximation. |
|
102 |
|
103 For really small number of samples Student distribution is used instead of |
|
104 normal distribution. But again it means that researcher made assumption and you |
|
105 may not agree with it, so it is your right to reject any subsequent decision |
|
106 based on "wrong" assumption. |
|
107 |
|
108 Type I error |
|
109 ============ |
|
110 |
|
111 :def:`Type I error` is the incorrect rejection of a true null hypothesis (a |
|
112 *false positive*). |
|
113 |
|
114 Type I error rate is at most :math:`α` (significant level). |
|
115 |
|
116 The p-value of a test is the maximum false positive risk you would take by |
|
117 rejecting the null hypothesis. |
|
118 |
|
119 Type II error |
|
120 ============= |
|
121 |
|
122 :def:`Type II error` is failing to reject a false null hypothesis (a *false |
|
123 negative*). |
|
124 |
|
125 Probability of type II error usually called :math:`β`. |
|
126 |
|
127 Power |
|
128 ===== |
|
129 |
|
130 :def:`Power` is a probability to reject null hypothesis when it's false. So |
|
131 power probability is :math:`1-β`. |
|
132 |
|
133 Confidence interval |
|
134 =================== |
|
135 |
|
136 :def:`Confidence interval` |
|
137 |
|
138 https://en.wikipedia.org/wiki/Confidence_interval |
|
139 Wikipedia page |
|
140 |
|
141 |
|
142 Question |
|
143 ======== |
|
144 |
|
145 What to do with null hypothesis in classical inference? |
|
146 |
|
147 I successfully shirked stat classes 10 years ago (last night reading help me |
|
148 actually to pass exam) and now when I take several Coursera stat classes I have |
|
149 difficulties with understanding **null hypothesis**. Somehow with unclear |
|
150 intuition I passed quizzes but want to understand subject. |
|
151 |
|
152 Suppose we have population and sample some data from population. Reasonable |
|
153 question: is some property of sample make evidence to be true on population? |
|
154 |
|
155 Statistic is a real number that can be derived from population or sample. |
|
156 Classical example is a mean value. |
|
157 |
|
158 We ask is it statistically significant that statistic of population is near to |
|
159 statistic of sample. |
|
160 |