Statistical Inference

Statistical Inference
• drawing conclusions about a population, based on a sample.
• uses properties of the sampling distribution and random sampling.
Example:
Population: GRE results for a new exam format on the quantitative section
Sample:
n=300 test scores
shape
Population
normal?
Sampling Dist. for X
approx. normal
mean
μ (unknown)
μ
SD
σ =100 (assume known)
σX =
σ
100
=
= 5.8
n
300
~95% of the Sampling Distribution is within ± 2iσ X of μ.
1) In ~95% of the samples of n=300, X is within +/- 11.6 pts of μ.
2) In ~95% of the samples of n=300, μ is within +/- 11.6 pts of X .
3) In ~95% of the samples of n=300, μ lies between X − 11.6 and X + 11.6 .
4) We are ~95% confident that we have one of the samples that gives an interval containing μ.
Hypothesis Testing
• Null Hypothesis (Ho ):
Ho: µ=54 cm
The population mean is 54 cm.
Ha: µ>54 (µa = 58)
σX =
4.5 cm
= 1.5 cm
9
α = .05
• Alternative Hypothesis (HA ): The population mean is greater than 54 cm.
p-value
• Measures the strength of the sample evidence against Ho
4 Steps for finding the Power in a test of hypotheses
1) Write the RR for Ho in terms of z-scores:
zs ≥ 1.645
2) Write the RR for Ho in terms of X :
X − 54
≥ 1.645 → X ≥ 56.47
1.5
• A small p-value gives strong evidence against Ho
• Definition:
The probability, computed assuming that Ho is true, of a sample result ( X ) as extreme
or more extreme than the one from our sample.
Rule of Thumb for the significance of p-values
3) Find the probability of a Type II error
if µ=58
4) Power = 1 – P(Type II Error) :
• If the p-value is less than .05, then our results are statistically significant at the .05 level
β ( µa ) = P(Z < zα − | µ0 − µa | / σ x )
β(58) = P(Accept Ho | Ha is true [µ=58] )
PWR(58) = 1 – β(58)
Ho: µ = 40 mpg
HA: µ < 40
Population Standard Deviation: σ = 6 mpg
Study Design
The number of observations needed to detect a true difference ∆ = µA - µo
at the α level of significance with power=1-β is
Significance Level:
α = .01
Sample Results:
A SRS of n = 16 gives X = 36.7
• 1-sided alternative:
n=
2
σ2
z + zβ )
2 ( α
∆
• 2-sided alternative:
n=
2
σ2
z + zβ )
2 ( α /2
∆
1) Write the rejection rule (RR) for Ho in terms of z-scores.
2) Write the rejection rule (RR) for Ho in terms of X .
3) Find the probability of a Type II error if µ=38
[i.e., β (38)]
a) Find the sample z-score (zs ).
b) State a conclusion for the test at the α = .01 level.
c) Find the p-value.
Ex: Find the sample size needed to detect a 2 mpg difference at the α=.01 level with 80% power.
Ho: µ = 40 mpg
HA: µ < 40
Population Standard Deviation: σ = 6 mpg
• So far we have assumed that σ was known for the population
100(1-α)% Confidence Interval for µ
⎛
σ ⎞
* CLT Æ Sampling Distribution for X is approximately N ⎜ µ , ⎟
n⎠
⎝
X −µ
* zs =
has a N(0,1) distribution
σ/ n
• In reality, we will rarely (if ever) know σ
* we estimate σ with s =
∑ (x − x )
where tα / 2 is the upper tail critical value with n-1 d.f.
Ex: A study on cholesterol levels in adult males eating fast food >3 times/week
A SRS of n = 30 gives X = 180.52 mg/dL
s = 41.23 mg/dL
i
n −1
1. Find a 95% CI for µ and state an interpretation of the interval
* Standard Deviation (of the mean) vs. Standard Error (of the mean)
σ
n
s
n
2
X −µ
has a Student’s T distribution with n-1 degrees of freedom
* ts =
s/ n
SDX =
x ± tα / 2
SE X =
Properties of the T-distribution
• Symmetric & bell shaped with mean = 0
• Larger spread than the N(0,1) distribution
• As the d.f. increase, the T-distribution
approaches the standard normal curve
• Table 3/C gives upper tail probabilities
• Developed by William S. Gosset
http://www.uvm.edu/~rsingle/stats/Gosset.html
2. Find t* such that there is 10% area to the left
Normal Distribution in blue
T Distribution df = 5 in green
T Distribution df = 10 in red
Ho: µ=170 vs. Ha: µ≠170 at the α=.05 l.o.s.
Suppose we want to find a p-value for the following set of hypotheses
Ho: µ=170 mg/dL
Ha: µ>170
s
n
Ex: (d.f. = 6)
1. Find t* such that there is 2% area to the right
2. Based on this CI, test …
ts =
H o : µ = µ0
X − µ0
s n
Rejection Region
Assumptions, Robustness, and Conditions for Valid CIs & T-Tests
p-value
1) H a : µ > µ0
ts ≥ tα
p − value = P(T ≥ ts )
2) H a : µ < µ0
ts ≤ −tα
p − value = P(T ≤ ts )
3) H a : µ ≠ µ0
| ts |≥ tα / 2
p − value = 2 ⋅ P(T ≥| ts |)
Ho: µ=165 mg/dL
• In reality, populations may be anywhere from slightly non-normal to very non-normal.
Robustness of the T-procedures
• The T-test and CI are called robust to the assumption of normality because p-values and
confidence levels are not greatly affected by violations of this assumption of normally
distributed populations, especially if sample sizes are large enough.
Conditions for a Valid 1-Sample CI and T-Test
Ha: µ>165
Sample Results:
• T-tests and CIs are based on the assumption that the population values being studied have a
Normal distribution.
A SRS of n = 30 gives X = 180.52
s = 41.23
a) Find the sample t-score (ts ).
b) Bracket the p-value.
c) State a conclusion for the test at the α = .05 level.
1) Write the rejection rule (RR) for Ho in terms of t-scores.
2) Write the rejection rule (RR) for Ho in terms of X .
• The data are a SRS from the population.
• The population must be large enough (at least 10 times larger than the sample size).
• Conditions for the sample:
o If n is small (n < 15), the data should not be grossly non-normal or contain outliers.
o If n is “medium” (15 ≤ n < 40), the data should not have strong skewness or outliers.
o If n is large (n ≥ 40), the T-procedures are robust to non-normality.
Checking if the conditions are met in your sample
• Always make a plot of the data to check for skewness and outliers before relying on
T-procedures in small samples.