Benjamin Olken MIT Sampling and Sample Size. Course Overview 1.What is evaluation? 2.Measuring impacts (outcomes, indicators) 3.Why randomize? 4.How to

Benjamin OlkenMIT

Sampling and Sample Size

Course Overview

1. What is evaluation?2. Measuring impacts (outcomes,

indicators)3. Why randomize?4. How to randomize5. Threats and Analysis6. Sampling and sample size7. RCT: Start to Finish8. Cost Effectiveness Analysis and

Scaling Up

Course Overview

1. What is evaluation?2. Measuring impacts (outcomes,

indicators)3. Why randomize?4. How to randomize5. Threats and Analysis6. Sampling and sample size7. RCT: Start to Finish8. Cost Effectiveness Analysis and

Scaling Up

What’s the average result?

• If you were to roll a die once, what is the “expected result”? (i.e. the average)

Possible results & probability: 1 die

1 2 3 4 5 60

1/6

Rolling 1 die: possible results & average

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.00

1/6

Frequency Average

What’s the average result?

• If you were to roll two dice once, what is the expected average of the two dice?

Rolling 2 dice:Possible totals & likelihood

2 3 4 5 6 7 8 9 10 11 12

Fre-quency

0.027777777777777

9

0.055555555555555

5

0.083333333333333

3

0.111111111111111

0.138888888888889

0.166666666666667

0.138888888888889

0.111111111111111

0.083333333333333

3

0.055555555555555

5

0.027777777777777

9

0

1/8

1/5

Rolling 2 dice: possible totals 12 possible totals, 36 permutations

Die 1

Die 2

2 3 4 5 6 7

3 4 5 6 7 8

4 5 6 7 8 9

5 6 7 8 9 10

6 7 8 9 10 11

7 8 9 10 11 12

Rolling 2 dice:Average score of dice & likelihood

1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6

Fre-quency

0.02777777777777

79

0.05555555555555

55

0.08333333333333

33

0.11111111111111

1

0.13888888888888

9

0.16666666666666

7

0.13888888888888

9

0.11111111111111

1

0.08333333333333

33

0.05555555555555

55

0.02777777777777

79

0

1/8

1/5

Lik

elih

oo

d

Outcomes and Permutations

• Putting together permutations, you get:1. All possible outcomes2. The likelihood of each of those

outcomes– Each column represents one possible

outcome (average result)– Each block within a column represents one

possible permutation (to obtain that average)

2.5

Rolling 3 dice:16 results 318, 216 permutations

1 1 1/3 1 2/3 2 2 1/3 2 2/3 3 3 1/3 3 2/3 4 4 1/3 4 2/3 5 5 1/3 5 2/3 6

Fre-quency

0.00462962962962964

0.01388888888888

89

0.02777777777777

79

0.04629629629629

64

0.06944444444444

45

0.09722222222222

22

0.11574074074074

1

0.125 0.125 0.11574074074074

1

0.09722222222222

22

0.06944444444444

45

0.04629629629629

64

0.02777777777777

79

0.01388888888888

89

0.00462962962962964

0

1/8

Rolling 4 dice: 21 results, 1296 permutations

1 1.25 1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4 4.25 4.5 4.75 5 5.25 5.5 5.75 60%

2%

4%

6%

8%

10%

12%

14%

Rolling 5 dice:26 results, 7776 permutations

1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6 5.8 60%

2%

4%

6%

8%

10%

12%

Looks like a bell curve, or a normal distribution

1

1.16

6666

6666

6667

1.33

3333

3333

3333 1.

5

1.66

6666

6666

6667

1.83

3333

3333

3333 2

2.16

6666

6666

6667

2.33

3333

3333

3333 2.

5

2.66

6666

6666

6667

2.83

3333

3333

3333 3

3.16

6666

6666

6667

3.33

3333

3333

3334 3.

5

3.66

6666

6666

6667

3.83

3333

3333

3334 4

4.16

6666

6666

6667

4.33

3333

3333

3334 4.

5

4.66

6666

6666

6667

4.83

3333

3333

3334 5

5.16

6666

6666

6667

5.33

3333

3333

3334 5.

5

5.66

6666

6666

6666

5.83

3333

3333

3334

0.0%

1.0%

2.0%

3.0%

4.0%

5.0%

6.0%

Rolling 10 dice:50 results, >60 million permutations

>99% of all rolls will yield an average between 3 and 4

1 1.72222222222222 2.44444444444444 3.16666666666666 3.88888888888888 4.6111111111111 5.333333333333320.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

Rolling 30 dice:150 results, 2 x 10 23 permutations*

>99% of all rolls will yield an average between 3 and 4

1 1.73333333333333 2.46666666666666 3.19999999999999 3.93333333333332 4.66666666666669 5.400000000000060.0%

0.5%

1.0%

1.5%

2.0%

Rolling 100 dice: 500 results, 6 x 10 77 permutations

Rolling dice: 2 lessons

1. The more dice you roll, the closer most averages are to the true average(the distribution gets “tighter”)-THE LAW OF LARGE NUMBERS-

2. The more dice you roll, the more the distribution of possible averages (the sampling distribution) looks like a bell curve (a normal distribution)-THE CENTRAL LIMIT THEOREM-

Which of these is more accurate?

I. II.

A. B. C.

0% 0%0%

A. I.B. II.C. Don’t know

Accuracy versus Precision

Precision

(Sampl

e Size)

Accuracy (Randomization)

truth

estimates

Accuracy versus Precision

Precision

(Sampl

e Size)

Accuracy (Randomization)

truth truthestimates

estimates

truth truthestimates

estimates

THE basic questions in statistics

• How confident can you be in your results?

• How big does your sample need to be?

THAT WAS JUST THE INTRODUCTION

Outline

• Sampling distributions– population distribution– sampling distribution– law of large numbers/central limit

theorem– standard deviation and standard error

• Detecting impact

Outline




Baseline test scores

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

frequency

test scores

Mean = 26

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

26 frequency

mean

test scores

Standard Deviation = 20

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

0

100

200

300

400

500

600

26 frequency sd

mean

test scores

1 Standard Deviation

Let’s do an experiment

• Take 1 Random test score from the pile of 16,000 tests

• Write down the value• Put the test back• Do these three steps again• And again• 8,000 times• This is like a random sample of

8,000 (with replacement)

Good, the average of the sample is about 26…

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-0.5%

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

4.0%

26mean

frequency

freq (N=1)

test scores

What can we say about this sample?

But…

• … I remember that as my sample goes, up, isn’t the sampling distribution supposed to turn into a bell curve?

• (Central Limit Theorem)• Is it that my sample isn’t large

enough?

This is the distribution of my sample of 8,000 students!

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-0.5%

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

4.0%

26mean

frequency

freq (N=1)

test scores

Population v. sampling distribution

Outline




How do we get from here…

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

100200300400500

test scores

To here…

This is the distribution of the population(Population Distribution)

This is the distribution of Means from all Random Samples(Sampling distribution)

Draw 10 random students, take the average, plot it: Do this 5 & 10 times.

Inadequate sample size

No clear distribution around population mean

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 520123456789

10

Frequency of Means With 5 Samples

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 520123456789

10


More sample means around population mean

Still spread a good deal

Draw 10 random students: 50 and 100 times

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 520123456789

10


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 520123456789

10

Frequency of Means with 100 Samples

Distribution now significantly more normal

Starting to see peaks

Draws 10 random students: 500 and 1000 times

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 520

10

20

30

40

50

60

70

80


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 520

10

20

30

40

50

60

70

80


Draw 10 Random students

• This is like a sample size of 10• What happens if we take a sample

size of 50?

N = 10 N = 50

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

2

4

6

8

10


1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

2

4

6

8

10


1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

2

4

6

8

10


1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

2

4

6

8

10


Draws of 10 Draws of 50

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

2

4

6

8

10

12

14


1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

5

10

15

20

Frequency of Means with 100 Samples

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

2

4

6

8

10

12

14


1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

4

8

12

16


Draws of 10 Draws of 50

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

102030405060708090


1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

20406080

100120140160


1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

102030405060708090


1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 520

20406080

100120140160


Outline




Population & sampling distribution: Draw 1 random student (from 8,000)

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-0.5%

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

4.0%

26mean

frequency

freq (N=1)

test scores

Sampling Distribution: Draw 4 random students (N=4)

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-0.5%

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

4.0%

4.5%

26mean

frequency

freq (N=4)

test scores

Law of Large Numbers : N=9

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-1.0%

0.0%

1.0%

2.0%

3.0%

4.0%

5.0%

6.0%

7.0%

26mean

frequency

freq (N=9)

test scores

Law of Large Numbers: N =100

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

0.0%

5.0%

10.0%

15.0%

20.0%

25.0%

26mean

frequency

freq (N=100)

test scores

The white line is a theoretical distribution

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-0.5%

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

4.0%

26

mean

frequency

dist_1

freq (N=1)

test scores

Central Limit Theorem: N=1

Central Limit Theorem : N=4

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-0.5%

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

4.0%

4.5%

26

mean

fre-quency

dist_4

freq (N=4)

test scores

Central Limit Theorem : N=9

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-1.0%

0.0%

1.0%

2.0%

3.0%

4.0%

5.0%

6.0%

7.0%

26

mean

frequency

dist_9

freq (N=9)

test scores

Central Limit Theorem : N =100

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

0.0%

5.0%

10.0%

15.0%

20.0%

25.0%

26

mean

frequency

dist_100

freq (N=100)

test scores

Outline




Standard deviation/error

• What’s the difference between the standard deviation and the standard error?

• The standard error = the standard deviation of the sampling distributions

Variance and Standard Deviation

• Variance = 400

• Standard Deviation = 20

• Standard Error = SE =

Standard Deviation/ Standard Error

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-0.5%

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

4.0%

26

mean

frequency

sd

dist_1

test scores

Sample size ↑ x4, SE ↓ ½

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-0.5%

0.0%

0.5%

1.0%

1.5%

2.0%

2.5%

3.0%

3.5%

4.0%

4.5%

26

mean

frequency

sd

se4

dist_4

test scores

Sample size ↑ x9, SE ↓ ?

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

-1.0%

0.0%

1.0%

2.0%

3.0%

4.0%

5.0%

6.0%

7.0%

26mean

frequency

sd

se9

dist_9

test scores

Sample size ↑ x100, SE ↓?

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

0.0%

5.0%

10.0%

15.0%

20.0%

25.0%

26

mean

frequency

sd

se100

dist_100

test scores

Outline

• Sampling distributions• Detecting impact

– significance– effect size – power– baseline and covariates– clustering– stratification


01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

50

100

150

200

250

300

350

400

450

500

test scores

We implement the Balsakhi Program (lecture 3)

Case 2: Remedial Education in IndiaEvaluating the Balsakhi Program

Incorporating random assignment into the program

Case 2: Remedial Education in IndiaEvaluating the Balsakhi Program

Incorporating random assignment into the program

After the balsakhi programs, these are the endline test scores

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

20

40

60

80

100

120

140

160

test scores

Endline test scores

0123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100

0

20

40

60

80

100

120

140

160

Endline test scores

The impact appears to be?

A. PositiveB. NegativeC. No impactD. Don’t know

0123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100

0

50

100

150

200

250

300

350

400

450

500


Stop! That was the control group. The treatment group is red.

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

20

40

60

80

100

120

140

160

control

treatment

test scores

Post-test: control & treatment

Is this impact statistically significant?

A. B. C.

0% 0%0%A. YesB. NoC. Don’t know

01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991000

20

40

60

80

100

120

140

160

control

treatment

control μ

treatment μ

test scores

Average Difference = 6 points

One experiment: 6 points

One experiment

Two experiments

A few more…

A few more…

Many more…

A whole lot more…

…

Running the experiment thousands of times…

By the Central Limit Theorem, these are normally distributed

Hypothesis testing

• In criminal law, most institutions follow the rule: “innocent until proven guilty”

• The presumption is that the accused is innocent and the burden is on the prosecutor to show guilt– The jury or judge starts with the “null

hypothesis” that the accused person is innocent

– The prosecutor has a hypothesis that the accused person is guilty

74

Hypothesis testing

• In program evaluation, instead of “presumption of innocence,” the rule is: “presumption of insignificance”

• The “Null hypothesis” (H0) is that there was no (zero) impact of the program

• The burden of proof is on the evaluator to show a significant effect of the program

Hypothesis testing: conclusions

• If it is very unlikely (less than a 5% probability) that the difference is solely due to chance: – We “reject our null hypothesis”

• We may now say: – “our program has a statistically

significant impact”

What is the significance level?

• Type I error: rejecting the null hypothesis even though it is true (false positive)

• Significance level: The probability that we will reject the null hypothesis even though it is true

Hypothesis testing: 95% confidence

YOU CONCLUDE

Effective No Effect

THE TRUTH

Effective Type II Error

(low power)

No Effect

Type I Error(5% of the time)

78

What is Power?

• Type II Error: Failing to reject the null hypothesis (concluding there is no difference), when indeed the null hypothesis is false.

• Power: If there is a measureable effect of our intervention (the null hypothesis is false), the probability that we will detect an effect (reject the null hypothesis)

Assume two effects: no effect and treatment effect β

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

Before the experiment

H0Hβ

Anything between lines cannot be distinguished from 0

Impose significance level of 5%

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

significanceHβH0

Shaded area shows % of time we would find Hβ true if it was

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

Can we distinguish Hβ from H0 ?

HβH0

What influences power?

• What are the factors that change the proportion of the research hypothesis that is shaded—i.e. the proportion that falls to the right (or left) of the null hypothesis curve?

• Understanding this helps us design more powerful experiments

83

Power: main ingredients

1. Effect Size2. Sample Size 3. Variance4. Proportion of sample in T vs. C5. Clustering



Effect Size: 1*SE

• Hypothesized effect size determines distance between means

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment


HβH0

Effect Size = 1*SE

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

significanceH0

Hβ

The Null Hypothesis would be rejected only 26% of the time

Power: 26%If the true impact was 1*SE…

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

HβH0

Bigger hypothesized effect size distributions farther apart

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

Effect Size: 3*SE

3*SE

Bigger Effect size means more power

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

Effect size 3*SE: Power= 91%

H0

Hβ

What effect size should you use when designing your experiment?

A. B. C. D.

25% 25%25%25%A. Smallest effect size that is still cost effective

B. Largest effect size you estimate your program to produce

C. BothD. Neither

• Let’s say we believe the impact on our participants is “3”

• What happens if take up is 1/3?• Let’s show this graphically

Effect size and take-up

Let’s say we believe the impact on our participants is “3”

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

Effect Size: 3*SE

3*SE

Take up is 33%. Effect size is 1/3rd

• Hypothesized effect size determines distance between means

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment


HβH0

Take-up is reflected in the effect size

Back to: Power = 26%

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

HβH0



By increasing sample size you increase…

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

Power: 91%

A. B. C. D. E.

0% 0% 0%0%0%

A. AccuracyB. PrecisionC. BothD. NeitherE. Don’t know

Power: Effect size = 1SD, Sample size = N

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

significance

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

significance

Power: Sample size = 4N

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

Power: 64%

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

significance

Power: Sample size = 9

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

Power: 91%



What are typical ways to reduce the underlying variance

A. B. C. D. E. F.

0% 0% 0%0%0%0%

A.Include covariates

B.Increase the sample

C.Do a baseline survey

D.All of the aboveE.A and BF.A and C

Variance

• There is sometimes very little we can do to reduce the noise

• The underlying variance is what it is• We can try to “absorb” variance:

– using a baseline– controlling for other variables

• In practice, controlling for other variables (besides the baseline outcome) buys you very little



Sample split: 50% C, 50% T

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

significanceH0

Hβ

Equal split gives distributions that are the same “fatness”

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

Power: 91%

If it’s not 50-50 split?

• What happens to the relative fatness if the split is not 50-50.

• Say 25-75?

Sample split: 25% C, 75% T

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

significanceH0

Hβ

Uneven distributions, not efficient, i.e. less power

-4 -3 -2 -1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

control

treatment

power

Power: 83%

Allocation to T v C

2

2

1

2

21 )(nn

XXsd

12

2

2

1

2

1)( 21 XXsd

15.13

4

1

1

3

1)( 21 XXsd



Clustered design: intuition

• You want to know how close the upcoming national elections will be

• Method 1: Randomly select 50 people from entire Indian population

• Method 2: Randomly select 5 families, and ask ten members of each family their opinion

114

Low intra-cluster correlation (Rho)

HIGH intra-cluster correlation (rho)

All uneducated people live in one village. People with only primary education live in another. College grads live in a third, etc. Rho on education will be..

A. B. C. D.

25% 25%25%25%A. HighB. LowC. No effect on rhoD. Don’t know

If rho is high, what is a more efficient way of increasing power?

A. B. C. D.

0% 0%0%0%

A. Include more clusters in the sample

B. Include more people in clusters

C. BothD. Don’t know

Testing multiple treatments

Control Group Balsakhi

CAL program Balsakhi + CAL

←0.15 SD→

↖

0.25 SD↘

↑0.15 SD

↓

↑0.10 SD

↓

←0.10 SD→

↗

0.05 SD↙

50 50

50 100100

100200

200

END

Documents

Benjamin Olken MIT Sampling and Sample Size. Course Overview 1.What is evaluation? 2.Measuring impacts (outcomes, indicators) 3.Why randomize? 4.How to